{
 "S1sh5::signage": {
  "fp": "e6eb3710ae845ccd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::09a4639f85639451": {
  "subjects": [],
  "subject_text": "인천 앞바다 수중·해저 쓰레기 더미\n푸른 수중 아래 각종 폐기물과 금속 기계 부품이 해저에 뒤엉켜 쌓인 공간. 수면에서 내려오는 빛이 물속에서 굴절되고 깊은 곳은 어둡다.",
  "identity": "canonical",
  "scope_id": "L01",
  "scope_role": "location_exterior",
  "scope_sha": "0de0d04a2937b921"
 },
 "groupbg::submerged_scrap": {
  "input_fingerprint": "2edc4e9d5753a4c4",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "submerged_scrap",
    "tags": [
     "S1sh10",
     "S1sh5"
    ]
   },
   "context_sig": "5d217eb5416bbe59"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 앞바다 수중·해저 쓰레기 더미: 푸른 바닷속에 각종 쓰레기 더미와 낡은 기계 잔해가 가라앉아 있는 환경이다. 해저 바닥에 묻힌 폐기물과 물고기들이 시각적으로 대조된다. (특징: 푸른빛의 해저 수중 질감; 플라스틱, 고철, 부서진 헬기 및 선박 잔해가 엉킨 쓰레기 산; 쓰레기 사이로 튀어나온 기계식 로봇 손가락; 유유히 헤엄치는 이름 모를 물고기 떼)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 물고기는 바닥에 잠긴 온갖 쓰레기 더미에 튀어나온 로봇 손가락을 발견한다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 앞바다 수중·해저 쓰레기 더미: 푸른 바닷속에 각종 쓰레기 더미와 낡은 기계 잔해가 가라앉아 있는 환경이다. 해저 바닥에 묻힌 폐기물과 물고기들이 시각적으로 대조된다. (특징: 푸른빛의 해저 수중 질감; 플라스틱, 고철, 부서진 헬기 및 선박 잔해가 엉킨 쓰레기 산; 쓰레기 사이로 튀어나온 기계식 로봇 손가락; 유유히 헤엄치는 이름 모를 물고기 떼)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 물고기는 바닥에 잠긴 온갖 쓰레기 더미에 튀어나온 로봇 손가락을 발견한다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_submerged_scrap_dec8f4.png",
  "asset_id": "d7bf68dc-faef-41b6-b836-d259645a41f1",
  "input_asset_ids": [
   "0b08acd4-b6fa-4198-8a71-d164e9940104"
  ],
  "origin_tag": "S1sh5",
  "place_text": "At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.",
  "origin_inputs": {
   "place_text": "At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.",
   "time_of_day_en": "day",
   "conti_asset_id": "0b08acd4-b6fa-4198-8a71-d164e9940104"
  }
 },
 "S1sh5::bgfirst_bg": {
  "input_fingerprint": "2d54f1dc87ffd61e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 밖 찰리의 로봇 손가락에 물고기의 주둥이가 닿은 극근접 구도.\n\nLOCATION (lock): At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From beside the fish's approach axis, observe the contact directly just above 찰리's exposed fingertip, looking slightly downward as the forward dolly settles. Center the touching snout and fingertip, with the fish entering from the right and the finger rising from refuse at lower left; retain surrounding water and refuse rather than isolating either object against an empty field. The fish attends to the fingertip while 찰리's face remains outside the frame, and the final reduction in camera distance is the only emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Submerged refuse pile (Resting on the seabed with a robot finger protruding); used as A narrow lower-frame context establishes where the finger emerges without obscuring the contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued blue underwater illumination and controlled contrast preserve the delicate contact and the robot finger's precise contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 밖 찰리의 로봇 손가락에 물고기의 주둥이가 닿은 극근접 구도.\n\nLOCATION (lock): At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From beside the fish's approach axis, observe the contact directly just above 찰리's exposed fingertip, looking slightly downward as the forward dolly settles. Center the touching snout and fingertip, with the fish entering from the right and the finger rising from refuse at lower left; retain surrounding water and refuse rather than isolating either object against an empty field. The fish attends to the fingertip while 찰리's face remains outside the frame, and the final reduction in camera distance is the only emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Submerged refuse pile (Resting on the seabed with a robot finger protruding); used as A narrow lower-frame context establishes where the finger emerges without obscuring the contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued blue underwater illumination and controlled contrast preserve the delicate contact and the robot finger's precise contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S1sh5__bgfirst_bg.png",
  "asset_id": "d60046fd-e313-4a32-bee1-005f41a2b337",
  "input_asset_ids": [
   "0b08acd4-b6fa-4198-8a71-d164e9940104",
   "d7bf68dc-faef-41b6-b836-d259645a41f1"
  ]
 },
 "S1sh5": {
  "input_fingerprint": "978e1c6086a1976f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖 찰리의 로봇 손가락에 물고기의 주둥이가 닿은 극근접 구도.\n\nLOCATION (lock): At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From beside the fish's approach axis, observe the contact directly just above 찰리's exposed fingertip, looking slightly downward as the forward dolly settles. Center the touching snout and fingertip, with the fish entering from the right and the finger rising from refuse at lower left; retain surrounding water and refuse rather than isolating either object against an empty field. The fish attends to the fingertip while 찰리's face remains outside the frame, and the final reduction in camera distance is the only emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Submerged refuse pile (Resting on the seabed with a robot finger protruding); used as A narrow lower-frame context establishes where the finger emerges without obscuring the contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued blue underwater illumination and controlled contrast preserve the delicate contact and the robot finger's precise contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, Charlie remains buried in seabed rubbish with a robot finger protruding from the heap. The old scrap-metal body is entangled in netting, and its chest bears a worn UBIC logo, even though these details remain buried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖 찰리의 로봇 손가락에 물고기의 주둥이가 닿은 극근접 구도.\n\nLOCATION (lock): At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From beside the fish's approach axis, observe the contact directly just above 찰리's exposed fingertip, looking slightly downward as the forward dolly settles. Center the touching snout and fingertip, with the fish entering from the right and the finger rising from refuse at lower left; retain surrounding water and refuse rather than isolating either object against an empty field. The fish attends to the fingertip while 찰리's face remains outside the frame, and the final reduction in camera distance is the only emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Submerged refuse pile (Resting on the seabed with a robot finger protruding); used as A narrow lower-frame context establishes where the finger emerges without obscuring the contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued blue underwater illumination and controlled contrast preserve the delicate contact and the robot finger's precise contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, Charlie remains buried in seabed rubbish with a robot finger protruding from the heap. The old scrap-metal body is entangled in netting, and its chest bears a worn UBIC logo, even though these details remain buried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖 찰리의 로봇 손가락에 물고기의 주둥이가 닿은 극근접 구도.\n\nLOCATION (lock): At the exposed edge of a submerged rubbish heap on the coastal seabed, in open blue seawater. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From beside the fish's approach axis, observe the contact directly just above 찰리's exposed fingertip, looking slightly downward as the forward dolly settles. Center the touching snout and fingertip, with the fish entering from the right and the finger rising from refuse at lower left; retain surrounding water and refuse rather than isolating either object against an empty field. The fish attends to the fingertip while 찰리's face remains outside the frame, and the final reduction in camera distance is the only emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Submerged refuse pile (Resting on the seabed with a robot finger protruding); used as A narrow lower-frame context establishes where the finger emerges without obscuring the contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued blue underwater illumination and controlled contrast preserve the delicate contact and the robot finger's precise contours.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the blue underwater daylight, Charlie remains buried in seabed rubbish with a robot finger protruding from the heap. The old scrap-metal body is entangled in netting, and its chest bears a worn UBIC logo, even though these details remain buried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S1sh5__bgfirst_bg.png",
     "asset_id": "d60046fd-e313-4a32-bee1-005f41a2b337",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S1sh5.png",
     "asset_id": "0b08acd4-b6fa-4198-8a71-d164e9940104",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_submerged_scrap_dec8f4.png",
     "asset_id": "d7bf68dc-faef-41b6-b836-d259645a41f1",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "오른쪽에서 나타난 물고기의 주둥이가 왼쪽 아래에서 위를 향해 뻗은 로봇 손가락 끝을 정확히 향해 맞닿아 있음.",
    "built_space": "푸른 바닷속 해저면의 쓰레기 더미가 배경을 채우고 있으며, 지시된 위치의 환경적 특성을 잘 반영함.",
    "entities": "물고기 한 마리와 샌드 베이지색 기계 재질의 로봇 손가락 하나가 나타나며, 찰리의 나머지 신체는 보이지 않음.",
    "hard_violations": [],
    "physics": "물고기는 물속에서 유영하며 떠 있고, 로봇 손가락은 쓰레기 더미 아래에 단단히 고정되어 지탱됨."
   },
   {
    "label": "B",
    "direction": "화면 오른쪽의 물고기 주둥이가 왼쪽 아래에서 튀어나온 로봇 손가락 끝부분에 정확하게 닿아 있음.",
    "built_space": "해저의 쓰레기 더미 환경이며, 근접 촬영으로 인해 배경의 쓰레기들과 푸른 바닷물이 제한된 영역 내에서 묘사됨.",
    "entities": "물고기의 머리 부분과 샌드 베이지색 기계 재질의 로봇 손가락 마디 일부만 화면에 포함됨.",
    "hard_violations": [],
    "physics": "물고기는 자체 부력으로 물속에 떠 있으며, 로봇 손가락은 보이지 않는 아래쪽 구조물(쓰레기 더미)에 의해 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "물고기와 로봇 손가락의 접촉은 묘사되었으나, 프레임이 너무 넓게 설정되어 물고기 전체가 화면에 잡히면서 지시문이 요구한 '극근접 구도'에 부합하지 않습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "물고기의 주둥이와 로봇 손가락의 접촉면을 화면 중앙에 크게 배치하여 텍스트가 요구한 '극근접 구도'의 스케일과 피사체 배치를 가장 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽에서 나타난 물고기의 주둥이가 왼쪽 아래에서 위를 향해 뻗은 로봇 손가락 끝을 정확히 향해 맞닿아 있음.",
        "built_space": "푸른 바닷속 해저면의 쓰레기 더미가 배경을 채우고 있으며, 지시된 위치의 환경적 특성을 잘 반영함.",
        "entities": "물고기 한 마리와 샌드 베이지색 기계 재질의 로봇 손가락 하나가 나타나며, 찰리의 나머지 신체는 보이지 않음.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 유영하며 떠 있고, 로봇 손가락은 쓰레기 더미 아래에 단단히 고정되어 지탱됨."
       },
       {
        "label": "B",
        "direction": "화면 오른쪽의 물고기 주둥이가 왼쪽 아래에서 튀어나온 로봇 손가락 끝부분에 정확하게 닿아 있음.",
        "built_space": "해저의 쓰레기 더미 환경이며, 근접 촬영으로 인해 배경의 쓰레기들과 푸른 바닷물이 제한된 영역 내에서 묘사됨.",
        "entities": "물고기의 머리 부분과 샌드 베이지색 기계 재질의 로봇 손가락 마디 일부만 화면에 포함됨.",
        "hard_violations": [],
        "physics": "물고기는 자체 부력으로 물속에 떠 있으며, 로봇 손가락은 보이지 않는 아래쪽 구조물(쓰레기 더미)에 의해 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "물고기와 로봇 손가락의 접촉은 묘사되었으나, 프레임이 너무 넓게 설정되어 물고기 전체가 화면에 잡히면서 지시문이 요구한 '극근접 구도'에 부합하지 않습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "물고기의 주둥이와 로봇 손가락의 접촉면을 화면 중앙에 크게 배치하여 텍스트가 요구한 '극근접 구도'의 스케일과 피사체 배치를 가장 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽에서 나타난 물고기의 주둥이가 왼쪽 아래에서 위를 향해 뻗은 로봇 손가락 끝을 정확히 향해 맞닿아 있음.",
        "built_space": "푸른 바닷속 해저면의 쓰레기 더미가 배경을 채우고 있으며, 지시된 위치의 환경적 특성을 잘 반영함.",
        "entities": "물고기 한 마리와 샌드 베이지색 기계 재질의 로봇 손가락 하나가 나타나며, 찰리의 나머지 신체는 보이지 않음.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 유영하며 떠 있고, 로봇 손가락은 쓰레기 더미 아래에 단단히 고정되어 지탱됨."
       },
       {
        "label": "B",
        "direction": "화면 오른쪽의 물고기 주둥이가 왼쪽 아래에서 튀어나온 로봇 손가락 끝부분에 정확하게 닿아 있음.",
        "built_space": "해저의 쓰레기 더미 환경이며, 근접 촬영으로 인해 배경의 쓰레기들과 푸른 바닷물이 제한된 영역 내에서 묘사됨.",
        "entities": "물고기의 머리 부분과 샌드 베이지색 기계 재질의 로봇 손가락 마디 일부만 화면에 포함됨.",
        "hard_violations": [],
        "physics": "물고기는 자체 부력으로 물속에 떠 있으며, 로봇 손가락은 보이지 않는 아래쪽 구조물(쓰레기 더미)에 의해 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "오른쪽 물고기의 주둥이와 왼쪽 아래 로봇 손가락의 접촉을 크게 잡아, 장소 전경보다 접촉 순간을 우선하는 근접 구도를 가장 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "접촉 방향과 장소는 맞지만, 참고 사진처럼 쓰레기 더미를 넓게 보여주어 지시된 극근접 접촉 구도보다 전경 설명에 치우쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "물고기가 화면 오른쪽에서 왼쪽 아래를 향하며 입술을 로봇 손가락 끝에 직접 대고 있다. 손가락은 왼쪽 아래에서 오른쪽 위로 이어지며, 두 대상의 접촉점은 화면 중앙보다 조금 오른쪽 위에 있다. 물고기는 손끝을 향하고 찰리의 얼굴은 보이지 않는다.",
        "built_space": "건축물은 없는 해저 공간이다. 왼쪽에는 케이블과 고철이 쌓인 경사면, 아래에는 파편과 그물, 오른쪽 뒤에는 누운 대형 원통 하나와 화면 가장자리에 격자형 폐기물 하나가 보인다. 참고 장소의 배치와 재질이 유지되며, 확대된 구도 때문에 일부 주변 폐기물은 잘려 있다. 쓰레기 맥락은 충분하지만 왼쪽과 아래에서 차지하는 면적은 지시된 좁은 하단 띠보다 크다.",
        "entities": "주요 물고기 한 마리의 머리와 몸통 일부, 마모된 관절식 로봇 손가락, 배경의 작은 물고기 떼와 해저 쓰레기가 보인다. 손가락의 금속 외피와 원형 관절은 장소 참고의 노출 부위와 잘 맞는다. 외피는 캐릭터 참고의 샌드 베이지보다 희고 회색에 가깝다. 찰리의 머리·몸통·나머지 팔다리와 가슴 표식은 보이지 않아 은폐 조건을 지킨다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 지느러미를 펼친 채 유영하며 손끝에 접촉하고 있어 공중에 떠 있는 물체가 아니다. 손가락은 하단 쓰레기에 묻힌 기부부터 금속 관절로 연속 연결되어 있고, 기부 주변의 고철과 케이블이 받치고 있다. 노출된 끝부분은 기계 구조의 돌출부로 읽히며, 분리되어 떠 있거나 찰리가 손 전체를 들어 제시하는 모습은 아니다."
       },
       {
        "label": "B",
        "direction": "물고기는 오른쪽 위에서 왼쪽 아래의 손가락 끝을 향하고, 주둥이가 끝부분에 닿아 있다. 손가락은 왼쪽 아래 쓰레기에서 오른쪽 위로 뻗는다. 접촉점은 중앙보다 오른쪽 위이며, 방향과 접촉 대상은 지시와 일치한다.",
        "built_space": "왼쪽 쓰레기 경사면과 넓은 하단 해저가 드러난다. 왼쪽 아래 녹색 격자 상자 하나, 오른쪽 중경의 누운 대형 원통 하나, 오른쪽 가장자리의 금속 격자 하나, 하단 중앙의 찌그러진 용기와 다수의 케이블·병·금속 파편이 보인다. 참고 장소의 배치를 가깝게 유지하지만, 쓰레기가 화면 하단 대부분을 차지하여 좁은 배경 맥락이라는 요구에서 벗어난다.",
        "entities": "주요 물고기 한 마리와 배경 물고기 떼, 세 마디가 뚜렷한 낡은 로봇 손가락, 그물과 해저 폐기물이 보인다. 로봇 부위는 장소 참고의 금속 외피와 관절을 따른다. 캐릭터 참고의 베이지색보다는 밝은 회색 계열이다. 얼굴·몸통·다른 팔다리는 가려져 있으며 가슴 로고가 보이지 않는 것은 요구에 부합한다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "물고기의 수중 자세와 펼쳐진 지느러미는 손끝으로 접근하는 유영으로 성립한다. 손가락 기부는 쓰레기와 케이블 사이에 박혀 있고 노출 마디들이 관절로 연결되어 있어 지지 경로가 보인다. 쓰레기는 해저에 쌓여 있으며, 지지 없이 분리되어 떠 있는 신체나 고형물은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "오른쪽 물고기의 주둥이와 왼쪽 아래 로봇 손가락의 접촉을 크게 잡아, 장소 전경보다 접촉 순간을 우선하는 근접 구도를 가장 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "접촉 방향과 장소는 맞지만, 참고 사진처럼 쓰레기 더미를 넓게 보여주어 지시된 극근접 접촉 구도보다 전경 설명에 치우쳤다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "물고기가 화면 오른쪽에서 왼쪽 아래를 향하며 입술을 로봇 손가락 끝에 직접 대고 있다. 손가락은 왼쪽 아래에서 오른쪽 위로 이어지며, 두 대상의 접촉점은 화면 중앙보다 조금 오른쪽 위에 있다. 물고기는 손끝을 향하고 찰리의 얼굴은 보이지 않는다.",
        "built_space": "건축물은 없는 해저 공간이다. 왼쪽에는 케이블과 고철이 쌓인 경사면, 아래에는 파편과 그물, 오른쪽 뒤에는 누운 대형 원통 하나와 화면 가장자리에 격자형 폐기물 하나가 보인다. 참고 장소의 배치와 재질이 유지되며, 확대된 구도 때문에 일부 주변 폐기물은 잘려 있다. 쓰레기 맥락은 충분하지만 왼쪽과 아래에서 차지하는 면적은 지시된 좁은 하단 띠보다 크다.",
        "entities": "주요 물고기 한 마리의 머리와 몸통 일부, 마모된 관절식 로봇 손가락, 배경의 작은 물고기 떼와 해저 쓰레기가 보인다. 손가락의 금속 외피와 원형 관절은 장소 참고의 노출 부위와 잘 맞는다. 외피는 캐릭터 참고의 샌드 베이지보다 희고 회색에 가깝다. 찰리의 머리·몸통·나머지 팔다리와 가슴 표식은 보이지 않아 은폐 조건을 지킨다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "물고기는 물속에서 지느러미를 펼친 채 유영하며 손끝에 접촉하고 있어 공중에 떠 있는 물체가 아니다. 손가락은 하단 쓰레기에 묻힌 기부부터 금속 관절로 연속 연결되어 있고, 기부 주변의 고철과 케이블이 받치고 있다. 노출된 끝부분은 기계 구조의 돌출부로 읽히며, 분리되어 떠 있거나 찰리가 손 전체를 들어 제시하는 모습은 아니다."
       },
       {
        "label": "A",
        "direction": "물고기는 오른쪽 위에서 왼쪽 아래의 손가락 끝을 향하고, 주둥이가 끝부분에 닿아 있다. 손가락은 왼쪽 아래 쓰레기에서 오른쪽 위로 뻗는다. 접촉점은 중앙보다 오른쪽 위이며, 방향과 접촉 대상은 지시와 일치한다.",
        "built_space": "왼쪽 쓰레기 경사면과 넓은 하단 해저가 드러난다. 왼쪽 아래 녹색 격자 상자 하나, 오른쪽 중경의 누운 대형 원통 하나, 오른쪽 가장자리의 금속 격자 하나, 하단 중앙의 찌그러진 용기와 다수의 케이블·병·금속 파편이 보인다. 참고 장소의 배치를 가깝게 유지하지만, 쓰레기가 화면 하단 대부분을 차지하여 좁은 배경 맥락이라는 요구에서 벗어난다.",
        "entities": "주요 물고기 한 마리와 배경 물고기 떼, 세 마디가 뚜렷한 낡은 로봇 손가락, 그물과 해저 폐기물이 보인다. 로봇 부위는 장소 참고의 금속 외피와 관절을 따른다. 캐릭터 참고의 베이지색보다는 밝은 회색 계열이다. 얼굴·몸통·다른 팔다리는 가려져 있으며 가슴 로고가 보이지 않는 것은 요구에 부합한다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "물고기의 수중 자세와 펼쳐진 지느러미는 손끝으로 접근하는 유영으로 성립한다. 손가락 기부는 쓰레기와 케이블 사이에 박혀 있고 노출 마디들이 관절로 연결되어 있어 지지 경로가 보인다. 쓰레기는 해저에 쌓여 있으며, 지지 없이 분리되어 떠 있는 신체나 고형물은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.381,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.381,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1381,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1381,
    "verdict_ko": "물고기와 로봇 손가락의 접촉은 묘사되었으나, 프레임이 너무 넓게 설정되어 물고기 전체가 화면에 잡히면서 지시문이 요구한 '극근접 구도'에 부합하지 않습니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "물고기의 주둥이와 로봇 손가락의 접촉면을 화면 중앙에 크게 배치하여 텍스트가 요구한 '극근접 구도'의 스케일과 피사체 배치를 가장 정확하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_submerged_scrap_dec8f4.png",
    "asset_id": "d7bf68dc-faef-41b6-b836-d259645a41f1",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0797-50d6-7649-826d-09d5e91daf43",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S1sh5__bgfirst_bg.png",
   "bg_asset_id": "d60046fd-e313-4a32-bee1-005f41a2b337",
   "bg_record_key": "S1sh5::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "submerged_scrap",
   "groupbg_asset_id": "d7bf68dc-faef-41b6-b836-d259645a41f1"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh10::signage": {
  "fp": "2c09e4089e24241d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S1sh10": {
  "input_fingerprint": "00728546fc64579f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 손가락이 돌출된 쓰레기 더미를 향한 크레인이 화면 대부분을 채우며 덮쳐오는 중인 한 찰나의 수중 구도.\n\nLOCATION (lock): In the open water immediately above a rubbish-covered coastal seabed, along the descending crane's path. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the exposed finger and complete the continuous steep upward tilt, directly observing the crane from slightly outside its descending path. Its approaching underside enters across the upper center, cropped by the top edge but occupying no more than two-fifths of the image, while surrounding water and descending bubbles retain spatial scale and make the approach threatening. Keep only a sliver of 찰리's protruding finger and refuse at the lower edge, with his face unseen; the changed viewing direction, rather than a camera relocation, carries the reveal.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane in the upper-center of the frame, midground, moves toward refuse pile beside the camera; Refuse pile beside the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Descending crane (Approaching the submerged refuse from above) — The underside faces the low camera obliquely, with its upper extent cropped by the frame; used as An overhead threat whose advancing edge implies the imminent engulfing movement; Bubbles (Arriving from above around the approaching crane); used as Connect the upper frame to the camera's position beside the refuse; Submerged refuse pile (Below the approaching crane); used as A small lower-edge anchor preserves continuity with the fingertip contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained blue underwater light separates the descending crane and bubbles without exaggerating brightness or contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains buried in the seabed rubbish with the same finger protruding, netting around the old body, and a worn UBIC chest logo. A large crane descends through the blue water toward the heap. Bubbles descend into the water from above.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 손가락이 돌출된 쓰레기 더미를 향한 크레인이 화면 대부분을 채우며 덮쳐오는 중인 한 찰나의 수중 구도.\n\nLOCATION (lock): In the open water immediately above a rubbish-covered coastal seabed, along the descending crane's path. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the exposed finger and complete the continuous steep upward tilt, directly observing the crane from slightly outside its descending path. Its approaching underside enters across the upper center, cropped by the top edge but occupying no more than two-fifths of the image, while surrounding water and descending bubbles retain spatial scale and make the approach threatening. Keep only a sliver of 찰리's protruding finger and refuse at the lower edge, with his face unseen; the changed viewing direction, rather than a camera relocation, carries the reveal.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane in the upper-center of the frame, midground, moves toward refuse pile beside the camera; Refuse pile beside the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Descending crane (Approaching the submerged refuse from above) — The underside faces the low camera obliquely, with its upper extent cropped by the frame; used as An overhead threat whose advancing edge implies the imminent engulfing movement; Bubbles (Arriving from above around the approaching crane); used as Connect the upper frame to the camera's position beside the refuse; Submerged refuse pile (Below the approaching crane); used as A small lower-edge anchor preserves continuity with the fingertip contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained blue underwater light separates the descending crane and bubbles without exaggerating brightness or contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains buried in the seabed rubbish with the same finger protruding, netting around the old body, and a worn UBIC chest logo. A large crane descends through the blue water toward the heap. Bubbles descend into the water from above.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 손가락이 돌출된 쓰레기 더미를 향한 크레인이 화면 대부분을 채우며 덮쳐오는 중인 한 찰나의 수중 구도.\n\nLOCATION (lock): In the open water immediately above a rubbish-covered coastal seabed, along the descending crane's path. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the exposed finger and complete the continuous steep upward tilt, directly observing the crane from slightly outside its descending path. Its approaching underside enters across the upper center, cropped by the top edge but occupying no more than two-fifths of the image, while surrounding water and descending bubbles retain spatial scale and make the approach threatening. Keep only a sliver of 찰리's protruding finger and refuse at the lower edge, with his face unseen; the changed viewing direction, rather than a camera relocation, carries the reveal.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Descending crane in the upper-center of the frame, midground, moves toward refuse pile beside the camera; Refuse pile beside the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Descending crane (Approaching the submerged refuse from above) — The underside faces the low camera obliquely, with its upper extent cropped by the frame; used as An overhead threat whose advancing edge implies the imminent engulfing movement; Bubbles (Arriving from above around the approaching crane); used as Connect the upper frame to the camera's position beside the refuse; Submerged refuse pile (Below the approaching crane); used as A small lower-edge anchor preserves continuity with the fingertip contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained blue underwater light separates the descending crane and bubbles without exaggerating brightness or contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless, buried in the rubbish on the seabed, with a robot finger protruding beyond the debris. His head, torso, and remaining limbs are concealed; their arrangement and his body's orientation are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains buried in the seabed rubbish with the same finger protruding, netting around the old body, and a worn UBIC chest logo. A large crane descends through the blue water toward the heap. Bubbles descend into the water from above.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 해저면에서 수면 쪽을 향해 가파르게 올려다보고 있으며, 상단에서 크레인이 화면 하단의 쓰레기 더미(로봇 손가락)를 향해 하강하고 있습니다.",
    "built_space": "수중 해저면으로, 바닥에는 다양한 쓰레기가 쌓여 있습니다. 화면 우측 하단에는 이전 샷에서 확인되는 원통형 파이프와 철망 구조물이 동일하게 배치되어 있습니다.",
    "entities": "찰리의 로봇 손가락(흰색 장갑판, 관절, 부착된 해조류 등 레퍼런스 일치)이 하단 중앙에 위치하고, 화면 상단에는 거대한 기계식 크레인이 위치합니다. 기포들이 크레인 주변에서 위로 상승하고 있습니다.",
    "hard_violations": [],
    "physics": "크레인은 화면 밖 상단 구조물에 연결되어 하강 중이며, 로봇 손가락은 해저면의 쓰레기 더미에 묻혀 지지받고 있습니다. 기포는 부력에 의해 자연스럽게 상승합니다."
   },
   {
    "label": "B",
    "direction": "카메라는 바닥에서 위를 향해 올려다보고 있으며, 위에서 아래로 크레인이 하강하며 바닥의 손가락을 향하고 있습니다.",
    "built_space": "쓰레기로 뒤덮인 해저면이며, 우측에 원통형 파이프와 철망 등 이전 샷의 주요 지형지물이 존재합니다.",
    "entities": "하단에 찰리의 로봇 손가락(레퍼런스와 일치하는 외형 및 웨더링)이 튀어나와 있고, 상단에 노란색이 감도는 기계식 크레인이 하강하고 있습니다. 주변으로 기포가 발생하고 있습니다.",
    "hard_violations": [],
    "physics": "크레인은 상단에서 매달려 내려오고 있으며, 손가락은 바닥의 쓰레기 더미에 고정되어 있습니다. 기포의 상승 방향이 중력/부력과 일치합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 로우 앵글(steep upward tilt) 구도와 이전 샷의 배경 요소(우측 파이프 등)를 완벽하게 유지했으며, 기계 손가락과 크레인의 사실적인 질감 표현이 우수합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상승하는 시점과 배경의 연속성을 잘 구현했으나, 크레인의 질감이 A에 비해 다소 덜 자연스럽고 작위적인 느낌이 듭니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 해저면에서 수면 쪽을 향해 가파르게 올려다보고 있으며, 상단에서 크레인이 화면 하단의 쓰레기 더미(로봇 손가락)를 향해 하강하고 있습니다.",
        "built_space": "수중 해저면으로, 바닥에는 다양한 쓰레기가 쌓여 있습니다. 화면 우측 하단에는 이전 샷에서 확인되는 원통형 파이프와 철망 구조물이 동일하게 배치되어 있습니다.",
        "entities": "찰리의 로봇 손가락(흰색 장갑판, 관절, 부착된 해조류 등 레퍼런스 일치)이 하단 중앙에 위치하고, 화면 상단에는 거대한 기계식 크레인이 위치합니다. 기포들이 크레인 주변에서 위로 상승하고 있습니다.",
        "hard_violations": [],
        "physics": "크레인은 화면 밖 상단 구조물에 연결되어 하강 중이며, 로봇 손가락은 해저면의 쓰레기 더미에 묻혀 지지받고 있습니다. 기포는 부력에 의해 자연스럽게 상승합니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바닥에서 위를 향해 올려다보고 있으며, 위에서 아래로 크레인이 하강하며 바닥의 손가락을 향하고 있습니다.",
        "built_space": "쓰레기로 뒤덮인 해저면이며, 우측에 원통형 파이프와 철망 등 이전 샷의 주요 지형지물이 존재합니다.",
        "entities": "하단에 찰리의 로봇 손가락(레퍼런스와 일치하는 외형 및 웨더링)이 튀어나와 있고, 상단에 노란색이 감도는 기계식 크레인이 하강하고 있습니다. 주변으로 기포가 발생하고 있습니다.",
        "hard_violations": [],
        "physics": "크레인은 상단에서 매달려 내려오고 있으며, 손가락은 바닥의 쓰레기 더미에 고정되어 있습니다. 기포의 상승 방향이 중력/부력과 일치합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 로우 앵글(steep upward tilt) 구도와 이전 샷의 배경 요소(우측 파이프 등)를 완벽하게 유지했으며, 기계 손가락과 크레인의 사실적인 질감 표현이 우수합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상승하는 시점과 배경의 연속성을 잘 구현했으나, 크레인의 질감이 A에 비해 다소 덜 자연스럽고 작위적인 느낌이 듭니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 해저면에서 수면 쪽을 향해 가파르게 올려다보고 있으며, 상단에서 크레인이 화면 하단의 쓰레기 더미(로봇 손가락)를 향해 하강하고 있습니다.",
        "built_space": "수중 해저면으로, 바닥에는 다양한 쓰레기가 쌓여 있습니다. 화면 우측 하단에는 이전 샷에서 확인되는 원통형 파이프와 철망 구조물이 동일하게 배치되어 있습니다.",
        "entities": "찰리의 로봇 손가락(흰색 장갑판, 관절, 부착된 해조류 등 레퍼런스 일치)이 하단 중앙에 위치하고, 화면 상단에는 거대한 기계식 크레인이 위치합니다. 기포들이 크레인 주변에서 위로 상승하고 있습니다.",
        "hard_violations": [],
        "physics": "크레인은 화면 밖 상단 구조물에 연결되어 하강 중이며, 로봇 손가락은 해저면의 쓰레기 더미에 묻혀 지지받고 있습니다. 기포는 부력에 의해 자연스럽게 상승합니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바닥에서 위를 향해 올려다보고 있으며, 위에서 아래로 크레인이 하강하며 바닥의 손가락을 향하고 있습니다.",
        "built_space": "쓰레기로 뒤덮인 해저면이며, 우측에 원통형 파이프와 철망 등 이전 샷의 주요 지형지물이 존재합니다.",
        "entities": "하단에 찰리의 로봇 손가락(레퍼런스와 일치하는 외형 및 웨더링)이 튀어나와 있고, 상단에 노란색이 감도는 기계식 크레인이 하강하고 있습니다. 주변으로 기포가 발생하고 있습니다.",
        "hard_violations": [],
        "physics": "크레인은 상단에서 매달려 내려오고 있으며, 손가락은 바닥의 쓰레기 더미에 고정되어 있습니다. 기포의 상승 방향이 중력/부력과 일치합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "크레인의 비스듬한 밑면과 상단 중앙 배치가 더 정확하고 손가락 노출도 상대적으로 적지만, 두 후보 모두 하단에 손가락과 쓰레기를 가느다랗게만 남기라는 구도를 충족하지 못한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "크레인이 손가락 위로 접근하는 관계는 맞지만, 집게의 정면성이 강하고 손가락과 전경 쓰레기가 크게 드러나 지정된 급격한 상향 틸트의 종착 구도에서 더 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "크레인의 열린 집게 끝들이 하단 중앙의 기계 손가락과 쓰레기 더미 쪽을 향한다. 낮은 카메라에 원형 밑면이 비스듬히 보이고 상부 연결부는 위쪽 경계 밖으로 이어진다. 기포가 크레인 양옆에서 수면과 하부 공간을 연결하지만, 정지 화면만으로 기포의 하강 방향은 확인되지 않는다. 찰리의 얼굴이나 시선은 보이지 않는다.",
        "built_space": "개방된 수중 해저이며 크레인 집게 한 개가 상단 중앙에 있다. 오른쪽에는 옆으로 누운 큰 원통형 폐기물 한 개와 철망 상자 한 개가 보이며, 주변에는 금속 조각과 그물이 쌓여 있어 이전 장면의 재료 구성이 이어진다. 크레인의 실제 실루엣 면적은 화면의 5분의 2 이내로 보이지만, 전경 쓰레기는 화면 하부 상당 부분을 차지하고 손가락도 여러 마디가 드러나 하단의 가느다란 흔적이라는 요구에 맞지 않는다.",
        "entities": "찰리에게 해당하는 부분은 오염되고 마모된 밝은 베이지색 금속 손가락 하나로, 검은 관절과 표면 부착물이 이전 장면과 유사하다. 인간 피부나 다른 인물은 없다. 얼굴, 몸통, 가슴 로고와 원자로는 가려져 있어 평가 대상이 아니다. 대형 금속 집게, 푸른 물, 기포, 그물과 해저 폐기물이 보인다. 배경 물고기들은 이전 장소의 수중 생태와 일치한다.",
        "hard_violations": [],
        "physics": "집게는 상단 경계 밖으로 이어지는 기계 연결부와 구동 장치에 매달려 있어 지지 없는 부유물로 보이지 않는다. 손가락의 기부와 하부는 쓰레기 및 그물에 묻혀 받쳐지고, 끝부분은 연결된 단단한 기계 마디로 돌출되어 있다. 몸 전체가 떠 있거나 손으로 물건을 들어 올리는 모습은 없다. 폐기물은 해저 더미에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "벌어진 크레인 집게가 하단 중앙의 손가락과 그 주변 쓰레기를 향해 내려오는 배치다. 다만 카메라에 넓은 직사각형 전면과 비교적 대칭적인 집게가 강하게 보여, 비스듬한 밑면을 관찰하는 요구는 A보다 약하다. 기포는 위쪽에서 집게 둘레로 이어지지만 이동 방향 자체는 확인할 수 없다. 찰리의 얼굴과 시선은 노출되지 않는다.",
        "built_space": "크레인 집게 한 개가 상단 중앙에 있고 연결부는 상단에서 잘린다. 오른쪽 해저에는 큰 원통형 폐기물 한 개와 철망 상자 한 개가 있으며, 하단에는 그물과 금속 잔해가 넓게 쌓여 있다. 장소의 재료는 이전 장면과 이어진다. 집게 실루엣은 화면 면적의 5분의 2 이내로 보이지만 아래쪽으로 더 길게 내려오며, 손가락 끝은 화면 높이의 약 3분의 2 지점까지 올라와 하단 가장자리의 작은 기준점이라는 지시에서 벗어난다.",
        "entities": "밝은 베이지색 장갑판, 검은 기계 관절, 해양 부착물이 있는 손가락 하나가 보여 찰리의 기계적 재질과 이전 장면의 마모 상태를 따른다. 다른 인물이나 인간 신체는 없다. 얼굴과 가슴 장치는 매몰되어 보이지 않는다. 크레인 집게, 기포, 푸른 물, 그물, 원통형 폐기물과 철망 상자가 있으며 배경에는 물고기들이 보인다.",
        "hard_violations": [],
        "physics": "집게는 위쪽으로 이어지는 굵은 기계 연결부와 구동부에 지지되어 있다. 손가락 기부는 쓰레기 속에 묻혀 있고 하부 마디 주변에 잔해와 그물이 접촉한다. 돌출된 끝은 연결된 기계 구조여서 지지 없이 떠 있는 별도 물체로 보이지 않는다. 해저 폐기물은 서로 겹쳐 바닥에 지지되며, 명백한 무지지 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "크레인의 비스듬한 밑면과 상단 중앙 배치가 더 정확하고 손가락 노출도 상대적으로 적지만, 두 후보 모두 하단에 손가락과 쓰레기를 가느다랗게만 남기라는 구도를 충족하지 못한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "크레인이 손가락 위로 접근하는 관계는 맞지만, 집게의 정면성이 강하고 손가락과 전경 쓰레기가 크게 드러나 지정된 급격한 상향 틸트의 종착 구도에서 더 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "크레인의 열린 집게 끝들이 하단 중앙의 기계 손가락과 쓰레기 더미 쪽을 향한다. 낮은 카메라에 원형 밑면이 비스듬히 보이고 상부 연결부는 위쪽 경계 밖으로 이어진다. 기포가 크레인 양옆에서 수면과 하부 공간을 연결하지만, 정지 화면만으로 기포의 하강 방향은 확인되지 않는다. 찰리의 얼굴이나 시선은 보이지 않는다.",
        "built_space": "개방된 수중 해저이며 크레인 집게 한 개가 상단 중앙에 있다. 오른쪽에는 옆으로 누운 큰 원통형 폐기물 한 개와 철망 상자 한 개가 보이며, 주변에는 금속 조각과 그물이 쌓여 있어 이전 장면의 재료 구성이 이어진다. 크레인의 실제 실루엣 면적은 화면의 5분의 2 이내로 보이지만, 전경 쓰레기는 화면 하부 상당 부분을 차지하고 손가락도 여러 마디가 드러나 하단의 가느다란 흔적이라는 요구에 맞지 않는다.",
        "entities": "찰리에게 해당하는 부분은 오염되고 마모된 밝은 베이지색 금속 손가락 하나로, 검은 관절과 표면 부착물이 이전 장면과 유사하다. 인간 피부나 다른 인물은 없다. 얼굴, 몸통, 가슴 로고와 원자로는 가려져 있어 평가 대상이 아니다. 대형 금속 집게, 푸른 물, 기포, 그물과 해저 폐기물이 보인다. 배경 물고기들은 이전 장소의 수중 생태와 일치한다.",
        "hard_violations": [],
        "physics": "집게는 상단 경계 밖으로 이어지는 기계 연결부와 구동 장치에 매달려 있어 지지 없는 부유물로 보이지 않는다. 손가락의 기부와 하부는 쓰레기 및 그물에 묻혀 받쳐지고, 끝부분은 연결된 단단한 기계 마디로 돌출되어 있다. 몸 전체가 떠 있거나 손으로 물건을 들어 올리는 모습은 없다. 폐기물은 해저 더미에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "벌어진 크레인 집게가 하단 중앙의 손가락과 그 주변 쓰레기를 향해 내려오는 배치다. 다만 카메라에 넓은 직사각형 전면과 비교적 대칭적인 집게가 강하게 보여, 비스듬한 밑면을 관찰하는 요구는 A보다 약하다. 기포는 위쪽에서 집게 둘레로 이어지지만 이동 방향 자체는 확인할 수 없다. 찰리의 얼굴과 시선은 노출되지 않는다.",
        "built_space": "크레인 집게 한 개가 상단 중앙에 있고 연결부는 상단에서 잘린다. 오른쪽 해저에는 큰 원통형 폐기물 한 개와 철망 상자 한 개가 있으며, 하단에는 그물과 금속 잔해가 넓게 쌓여 있다. 장소의 재료는 이전 장면과 이어진다. 집게 실루엣은 화면 면적의 5분의 2 이내로 보이지만 아래쪽으로 더 길게 내려오며, 손가락 끝은 화면 높이의 약 3분의 2 지점까지 올라와 하단 가장자리의 작은 기준점이라는 지시에서 벗어난다.",
        "entities": "밝은 베이지색 장갑판, 검은 기계 관절, 해양 부착물이 있는 손가락 하나가 보여 찰리의 기계적 재질과 이전 장면의 마모 상태를 따른다. 다른 인물이나 인간 신체는 없다. 얼굴과 가슴 장치는 매몰되어 보이지 않는다. 크레인 집게, 기포, 푸른 물, 그물, 원통형 폐기물과 철망 상자가 있으며 배경에는 물고기들이 보인다.",
        "hard_violations": [],
        "physics": "집게는 위쪽으로 이어지는 굵은 기계 연결부와 구동부에 지지되어 있다. 손가락 기부는 쓰레기 속에 묻혀 있고 하부 마디 주변에 잔해와 그물이 접촉한다. 돌출된 끝은 연결된 기계 구조여서 지지 없이 떠 있는 별도 물체로 보이지 않는다. 해저 폐기물은 서로 겹쳐 바닥에 지지되며, 명백한 무지지 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "지시된 로우 앵글(steep upward tilt) 구도와 이전 샷의 배경 요소(우측 파이프 등)를 완벽하게 유지했으며, 기계 손가락과 크레인의 사실적인 질감 표현이 우수합니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 상승하는 시점과 배경의 연속성을 잘 구현했으나, 크레인의 질감이 A에 비해 다소 덜 자연스럽고 작위적인 느낌이 듭니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S1sh5_sel.png",
    "asset_id": "ba5a19eb-ab3f-475a-9bd7-a1aaabf387d5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab079f-4520-71cd-9d1a-70f1ace523c4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S1sh5"
  },
  "locked_char_refs_kept_body_identity": [
   "찰리(C01)"
  ]
 },
 "S2sh1::signage": {
  "fp": "116765b74017f62a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S2sh1": {
  "input_fingerprint": "c785049530e89e9b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낮의 바다 멀리 거대한 쓰레기 수거선이 부유한 전경, 화면 위 자막은 ‘2069년, 인천’.\n\nLOCATION (lock): On the open coastal sea in daylight, with a massive rubbish-collection ship seen at a distance. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Establish a direct, high oblique view from far off the vessel's forward quarter, looking downward toward its bow and side at the still opening of the approach. Place the complete garbage collection vessel below center at less than two-fifths of the frame, surrounded by sea and scattered floating refuse, with no individually readable people. Reserve uncluttered space above the vessel for the exact editorial caption ‘2069년, 인천’, clearly separate from any physical surface.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Garbage collection vessel (Afloat at sea) — The bow and one side are visible together from a high forward-quarter angle; used as Provides the principal scale anchor below the caption space; Sea (Surrounding the vessel); used as Broad negative space establishes the vessel's isolation and distant scale; Floating refuse (Scattered on the water around the vessel); used as Small, separated forms distinguish the working environment without competing with the vessel.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with cool-neutral grading and controlled contrast gives the distant vessel a restrained, tangible presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large rubbish-collection ship travels through daylight waters littered with shattered helicopter wreckage, a half-broken vessel, and miscellaneous floating rubbish. The caption reads “2069년, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낮의 바다 멀리 거대한 쓰레기 수거선이 부유한 전경, 화면 위 자막은 ‘2069년, 인천’.\n\nLOCATION (lock): On the open coastal sea in daylight, with a massive rubbish-collection ship seen at a distance. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Establish a direct, high oblique view from far off the vessel's forward quarter, looking downward toward its bow and side at the still opening of the approach. Place the complete garbage collection vessel below center at less than two-fifths of the frame, surrounded by sea and scattered floating refuse, with no individually readable people. Reserve uncluttered space above the vessel for the exact editorial caption ‘2069년, 인천’, clearly separate from any physical surface.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Garbage collection vessel (Afloat at sea) — The bow and one side are visible together from a high forward-quarter angle; used as Provides the principal scale anchor below the caption space; Sea (Surrounding the vessel); used as Broad negative space establishes the vessel's isolation and distant scale; Floating refuse (Scattered on the water around the vessel); used as Small, separated forms distinguish the working environment without competing with the vessel.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with cool-neutral grading and controlled contrast gives the distant vessel a restrained, tangible presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large rubbish-collection ship travels through daylight waters littered with shattered helicopter wreckage, a half-broken vessel, and miscellaneous floating rubbish. The caption reads “2069년, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낮의 바다 멀리 거대한 쓰레기 수거선이 부유한 전경, 화면 위 자막은 ‘2069년, 인천’.\n\nLOCATION (lock): On the open coastal sea in daylight, with a massive rubbish-collection ship seen at a distance. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Establish a direct, high oblique view from far off the vessel's forward quarter, looking downward toward its bow and side at the still opening of the approach. Place the complete garbage collection vessel below center at less than two-fifths of the frame, surrounded by sea and scattered floating refuse, with no individually readable people. Reserve uncluttered space above the vessel for the exact editorial caption ‘2069년, 인천’, clearly separate from any physical surface.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Garbage collection vessel (Afloat at sea) — The bow and one side are visible together from a high forward-quarter angle; used as Provides the principal scale anchor below the caption space; Sea (Surrounding the vessel); used as Broad negative space establishes the vessel's isolation and distant scale; Floating refuse (Scattered on the water around the vessel); used as Small, separated forms distinguish the working environment without competing with the vessel.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with cool-neutral grading and controlled contrast gives the distant vessel a restrained, tangible presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large rubbish-collection ship travels through daylight waters littered with shattered helicopter wreckage, a half-broken vessel, and miscellaneous floating rubbish. The caption reads “2069년, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
    "built_space": "넓은 바다 한가운데 쓰레기 수거선이 위치하며, 배경으로 섬과 도시의 실루엣이 보입니다. 수거선의 구조는 레퍼런스와 일치합니다.",
    "entities": "레퍼런스와 일치하는 쓰레기 수거선, 부서진 헬기 잔해, 반파된 선박, 부유하는 쓰레기들이 모두 존재하나, '2069년, 인천' 자막은 없습니다. 사람은 보이지 않습니다.",
    "hard_violations": [],
    "physics": "수거선과 헬기 잔해, 부서진 선박, 쓰레기 등 모든 물체는 바다 수면 위에 부력으로 떠 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
    "built_space": "넓은 바다를 배경으로 쓰레기 수거선이 중앙 하단에 위치하며, 먼 배경에 섬과 도시가 보입니다. 수거선의 디테일이 레퍼런스와 잘 맞습니다.",
    "entities": "쓰레기 수거선, 헬기 잔해, 반파된 선박, 쓰레기 더미가 모두 정확히 묘사되었습니다. 화면 상단에 '2069년, 인천' 자막이 텍스트로 나타나 있습니다. 사람은 없습니다.",
    "hard_violations": [],
    "physics": "선박과 헬기의 동체, 파편 등 모든 요소가 수면 위에 정상적으로 떠서 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "쓰레기 수거선과 헬기 잔해, 부서진 선박 등 지시된 요소들이 잘 묘사되었으며, 특히 요구된 '2069년, 인천' 자막이 화면 상단에 정확히 구현되어 프롬프트 충실도가 가장 높습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "해양 쓰레기 수거선과 잔해물들의 배치는 프롬프트와 잘 일치하나, 필수적으로 요구된 화면 상단의 '2069년, 인천' 자막이 완전히 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
        "built_space": "넓은 바다 한가운데 쓰레기 수거선이 위치하며, 배경으로 섬과 도시의 실루엣이 보입니다. 수거선의 구조는 레퍼런스와 일치합니다.",
        "entities": "레퍼런스와 일치하는 쓰레기 수거선, 부서진 헬기 잔해, 반파된 선박, 부유하는 쓰레기들이 모두 존재하나, '2069년, 인천' 자막은 없습니다. 사람은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "수거선과 헬기 잔해, 부서진 선박, 쓰레기 등 모든 물체는 바다 수면 위에 부력으로 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
        "built_space": "넓은 바다를 배경으로 쓰레기 수거선이 중앙 하단에 위치하며, 먼 배경에 섬과 도시가 보입니다. 수거선의 디테일이 레퍼런스와 잘 맞습니다.",
        "entities": "쓰레기 수거선, 헬기 잔해, 반파된 선박, 쓰레기 더미가 모두 정확히 묘사되었습니다. 화면 상단에 '2069년, 인천' 자막이 텍스트로 나타나 있습니다. 사람은 없습니다.",
        "hard_violations": [],
        "physics": "선박과 헬기의 동체, 파편 등 모든 요소가 수면 위에 정상적으로 떠서 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "쓰레기 수거선과 헬기 잔해, 부서진 선박 등 지시된 요소들이 잘 묘사되었으며, 특히 요구된 '2069년, 인천' 자막이 화면 상단에 정확히 구현되어 프롬프트 충실도가 가장 높습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "해양 쓰레기 수거선과 잔해물들의 배치는 프롬프트와 잘 일치하나, 필수적으로 요구된 화면 상단의 '2069년, 인천' 자막이 완전히 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
        "built_space": "넓은 바다 한가운데 쓰레기 수거선이 위치하며, 배경으로 섬과 도시의 실루엣이 보입니다. 수거선의 구조는 레퍼런스와 일치합니다.",
        "entities": "레퍼런스와 일치하는 쓰레기 수거선, 부서진 헬기 잔해, 반파된 선박, 부유하는 쓰레기들이 모두 존재하나, '2069년, 인천' 자막은 없습니다. 사람은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "수거선과 헬기 잔해, 부서진 선박, 쓰레기 등 모든 물체는 바다 수면 위에 부력으로 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 선박의 우현 선수 쪽을 향해 하향 사선으로 향하고 있습니다.",
        "built_space": "넓은 바다를 배경으로 쓰레기 수거선이 중앙 하단에 위치하며, 먼 배경에 섬과 도시가 보입니다. 수거선의 디테일이 레퍼런스와 잘 맞습니다.",
        "entities": "쓰레기 수거선, 헬기 잔해, 반파된 선박, 쓰레기 더미가 모두 정확히 묘사되었습니다. 화면 상단에 '2069년, 인천' 자막이 텍스트로 나타나 있습니다. 사람은 없습니다.",
        "hard_violations": [],
        "physics": "선박과 헬기의 동체, 파편 등 모든 요소가 수면 위에 정상적으로 떠서 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "선수와 측면을 함께 내려다보는 원경, 중앙 아래의 완전한 선체, 정확한 상단 자막을 구현해 우세하지만 전경 잔해가 크고 햇빛이 요구보다 강하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "참조 선박과 전방 사선 시점은 유지했지만 필수 자막 ‘2069년, 인천’이 없고 선박도 더 커서 지정된 먼 원경 구도에 덜 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "선수는 화면 오른쪽 아래를 향하고 선미는 왼쪽 위에 있으며, 뒤쪽의 옅은 항적도 이 진행 방향과 맞는다. 카메라는 선수 전방의 높은 사선 위치에서 선수와 한쪽 측면, 갑판을 함께 내려다본다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "청소선 한 척에 후방의 흰색 선실 구조물 하나, 주 마스트 하나, 넓은 전방 작업 갑판 하나가 보인다. 선실 오른쪽에는 굽은 배기관 하나가 있고, 외곽 난간·여러 구명환·검은 타이어 및 연속 방현재가 배치되어 참조의 주요 구조와 대응한다. 선체 전체가 중앙 아래에 놓이고 가로 폭은 화면의 약 37%다. 상단에는 선박과 분리된 자막 공간이 있으나, 왼쪽 아래 헬기 잔해와 오른쪽 파손 선박이 상당히 크게 보인다.",
        "entities": "참조처럼 청색 선체와 흰색 선실을 가진 쓰레기 수거선이며, 태극기와 선체의 ‘해양청소 1’ 표기가 보인다. 바다, 부서진 헬기 한 대의 잔해, 반쯤 잠긴 파손 선박 한 척, 다수의 부유 쓰레기가 있다. 사람이나 얼굴은 보이지 않는다. 상단 자막은 요구한 ‘2069년, 인천’이다. 낮은 맞지만 수면의 강한 반짝임은 절제된 확산광 요구와 차이가 있다.",
        "hard_violations": [],
        "physics": "청소선은 하부 선체가 물에 잠겨 부력으로 지지되고, 선미 쪽 물결은 이동과 양립한다. 헬기 잔해와 파손 선박은 일부가 수면 아래에 잠긴 상태이며 공중에 떠 있지 않다. 헬기 로터는 동체 위 허브에 연결되어 있고 마스트·선실·갑판 설비도 각각 선체에 붙어 있다. 작은 쓰레기들은 수면에 접해 떠 있다."
       },
       {
        "label": "B",
        "direction": "선수는 오른쪽 아래, 선미는 왼쪽 위를 향하며 항적은 선미 뒤로 이어진다. 높은 전방 사선 카메라에서 선수와 측면 및 갑판이 함께 보여 요구한 관찰 방향은 맞는다. 사람의 시선이나 무기의 조준은 없다.",
        "built_space": "청소선 한 척에 후방 흰색 선실 구조물 하나, 주 마스트 하나, 전방 개방 갑판 하나와 선실 오른쪽 굽은 배기관 하나가 보인다. 외곽 난간, 여러 구명환과 타이어 방현재도 참조의 배치를 대체로 따른다. 선체는 중앙 아래에 온전히 들어오지만 가로 폭이 약 41%로 A보다 크다. 상단에는 여백이 있으나 자막은 없다. 왼쪽 아래 헬기 잔해와 오른쪽 가장자리에 잘린 파손 선박이 주변 바다의 여백을 차지한다.",
        "entities": "청색 선체·흰색 선실·태극기·‘해양청소 1’ 표기는 참조와 대응한다. 바다, 헬기 잔해 하나, 반파된 선박 하나와 흩어진 쓰레기가 보이고 사람은 없다. 필수 편집 자막 ‘2069년, 인천’은 누락되었다. 주간이지만 선명한 일광과 강한 수면 반사는 요구한 차분하고 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "청소선의 수선과 선미 항적이 보여 물의 부력과 추진으로 설명되는 상태다. 헬기 동체와 파손 선박은 수면에 걸쳐 부분적으로 잠겨 있으며, 로터는 헬기 허브에 연결되어 있다. 갑판 설비는 갑판 위에 고정되어 있고 부유 쓰레기는 물에 닿아 있다. 지지 없이 공중에 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "선수와 측면을 함께 내려다보는 원경, 중앙 아래의 완전한 선체, 정확한 상단 자막을 구현해 우세하지만 전경 잔해가 크고 햇빛이 요구보다 강하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "참조 선박과 전방 사선 시점은 유지했지만 필수 자막 ‘2069년, 인천’이 없고 선박도 더 커서 지정된 먼 원경 구도에 덜 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "선수는 화면 오른쪽 아래를 향하고 선미는 왼쪽 위에 있으며, 뒤쪽의 옅은 항적도 이 진행 방향과 맞는다. 카메라는 선수 전방의 높은 사선 위치에서 선수와 한쪽 측면, 갑판을 함께 내려다본다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "청소선 한 척에 후방의 흰색 선실 구조물 하나, 주 마스트 하나, 넓은 전방 작업 갑판 하나가 보인다. 선실 오른쪽에는 굽은 배기관 하나가 있고, 외곽 난간·여러 구명환·검은 타이어 및 연속 방현재가 배치되어 참조의 주요 구조와 대응한다. 선체 전체가 중앙 아래에 놓이고 가로 폭은 화면의 약 37%다. 상단에는 선박과 분리된 자막 공간이 있으나, 왼쪽 아래 헬기 잔해와 오른쪽 파손 선박이 상당히 크게 보인다.",
        "entities": "참조처럼 청색 선체와 흰색 선실을 가진 쓰레기 수거선이며, 태극기와 선체의 ‘해양청소 1’ 표기가 보인다. 바다, 부서진 헬기 한 대의 잔해, 반쯤 잠긴 파손 선박 한 척, 다수의 부유 쓰레기가 있다. 사람이나 얼굴은 보이지 않는다. 상단 자막은 요구한 ‘2069년, 인천’이다. 낮은 맞지만 수면의 강한 반짝임은 절제된 확산광 요구와 차이가 있다.",
        "hard_violations": [],
        "physics": "청소선은 하부 선체가 물에 잠겨 부력으로 지지되고, 선미 쪽 물결은 이동과 양립한다. 헬기 잔해와 파손 선박은 일부가 수면 아래에 잠긴 상태이며 공중에 떠 있지 않다. 헬기 로터는 동체 위 허브에 연결되어 있고 마스트·선실·갑판 설비도 각각 선체에 붙어 있다. 작은 쓰레기들은 수면에 접해 떠 있다."
       },
       {
        "label": "A",
        "direction": "선수는 오른쪽 아래, 선미는 왼쪽 위를 향하며 항적은 선미 뒤로 이어진다. 높은 전방 사선 카메라에서 선수와 측면 및 갑판이 함께 보여 요구한 관찰 방향은 맞는다. 사람의 시선이나 무기의 조준은 없다.",
        "built_space": "청소선 한 척에 후방 흰색 선실 구조물 하나, 주 마스트 하나, 전방 개방 갑판 하나와 선실 오른쪽 굽은 배기관 하나가 보인다. 외곽 난간, 여러 구명환과 타이어 방현재도 참조의 배치를 대체로 따른다. 선체는 중앙 아래에 온전히 들어오지만 가로 폭이 약 41%로 A보다 크다. 상단에는 여백이 있으나 자막은 없다. 왼쪽 아래 헬기 잔해와 오른쪽 가장자리에 잘린 파손 선박이 주변 바다의 여백을 차지한다.",
        "entities": "청색 선체·흰색 선실·태극기·‘해양청소 1’ 표기는 참조와 대응한다. 바다, 헬기 잔해 하나, 반파된 선박 하나와 흩어진 쓰레기가 보이고 사람은 없다. 필수 편집 자막 ‘2069년, 인천’은 누락되었다. 주간이지만 선명한 일광과 강한 수면 반사는 요구한 차분하고 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "청소선의 수선과 선미 항적이 보여 물의 부력과 추진으로 설명되는 상태다. 헬기 동체와 파손 선박은 수면에 걸쳐 부분적으로 잠겨 있으며, 로터는 헬기 허브에 연결되어 있다. 갑판 설비는 갑판 위에 고정되어 있고 부유 쓰레기는 물에 닿아 있다. 지지 없이 공중에 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "쓰레기 수거선과 헬기 잔해, 부서진 선박 등 지시된 요소들이 잘 묘사되었으며, 특히 요구된 '2069년, 인천' 자막이 화면 상단에 정확히 구현되어 프롬프트 충실도가 가장 높습니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "해양 쓰레기 수거선과 잔해물들의 배치는 프롬프트와 잘 일치하나, 필수적으로 요구된 화면 상단의 '2069년, 인천' 자막이 완전히 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_collection_vessel_sel.png",
    "asset_id": "cd16d01d-6027-4184-a812-921eea63de02",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07a3-cd69-779e-9195-443c51ac776f",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S2sh6::signage": {
  "fp": "3b01e20c41efea47",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::0abe8608199bea6e": {
  "subjects": [],
  "subject_text": "쓰레기 수거선 갑판\n바다와 하늘에 노출된 넓은 금속 갑판. 인양용 크레인 아래 폐기물과 고철, 어망이 뒤섞여 쌓여 있다.",
  "identity": "canonical",
  "scope_id": "L03",
  "scope_role": "location_exterior",
  "scope_sha": "46045a0860c8c36a"
 },
 "groupbg::collection_deck": {
  "input_fingerprint": "d42c28776ac13a83",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "collection_deck",
    "tags": [
     "S2sh6"
    ]
   },
   "context_sig": "f8b69cda1a292f70"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n쓰레기 수거선 갑판: 거대한 금속제 수거선의 갑판으로, 인양된 잡다한 고철과 부서진 기계들이 쏟아지는 작업 공간이다. (특징: 그물망에 몸이 감긴 채 쌓여 있는 고철 로봇 찰리 (금속 하드서페이스, 낡은 마감재); 갑판 위로 우수수 떨어지는 물 젖은 폐기물과 헬기 파편)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쓰레기 수거선의 갑판으로 우수수 떨어지는 쓰레기들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n쓰레기 수거선 갑판: 거대한 금속제 수거선의 갑판으로, 인양된 잡다한 고철과 부서진 기계들이 쏟아지는 작업 공간이다. (특징: 그물망에 몸이 감긴 채 쌓여 있는 고철 로봇 찰리 (금속 하드서페이스, 낡은 마감재); 갑판 위로 우수수 떨어지는 물 젖은 폐기물과 헬기 파편)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쓰레기 수거선의 갑판으로 우수수 떨어지는 쓰레기들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_collection_deck_b9e7b9.png",
  "asset_id": "62155cdd-f3f2-45bf-802f-7e8d5aeec2bf",
  "input_asset_ids": [
   "f8f2b853-c659-47d9-b6de-06b7d2d04916"
  ],
  "origin_tag": "S2sh6",
  "place_text": "On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.",
  "origin_inputs": {
   "place_text": "On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.",
   "time_of_day_en": "day",
   "conti_asset_id": "f8f2b853-c659-47d9-b6de-06b7d2d04916"
  }
 },
 "S2sh6::bgfirst_bg": {
  "input_fingerprint": "987d30726e9861bb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갑판 위 쓰레기 사이에 낡은 고철 로봇 찰리의 몸이 그물에 감긴 구도.\n\nLOCATION (lock): On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the deck-level approach from an elevated oblique position beside 찰리's torso, looking downward rather than along his facial axis in a direct observational view. Place his net-wrapped body diagonally across the central third, retaining his full bodily outline and adjacent refuse, with a few small foreground fragments creating depth without hiding him. His head lies turned away toward the surrounding refuse with no readable eye line; the narrowing camera distance is the emphasized change as the movement settles into the reveal.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Net around 찰리 (Wrapped around the robot's body) — The strands cross the visible torso and limbs without completely concealing their shapes; used as Makes the robot's entanglement legible within the wider deck context; Deck refuse (Accumulated around 찰리 after being dropped aboard); used as Uneven foreground and background groupings reveal a body among discarded objects; Collection vessel deck (Supporting the refuse and net-wrapped robot) — The deck plane recedes diagonally beneath the body; used as Visible gaps between the refuse establish a shared supporting plane.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and cool-neutral grading distinguish the net from 찰리's aged hard surfaces without adding a dramatic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갑판 위 쓰레기 사이에 낡은 고철 로봇 찰리의 몸이 그물에 감긴 구도.\n\nLOCATION (lock): On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the deck-level approach from an elevated oblique position beside 찰리's torso, looking downward rather than along his facial axis in a direct observational view. Place his net-wrapped body diagonally across the central third, retaining his full bodily outline and adjacent refuse, with a few small foreground fragments creating depth without hiding him. His head lies turned away toward the surrounding refuse with no readable eye line; the narrowing camera distance is the emphasized change as the movement settles into the reveal.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Net around 찰리 (Wrapped around the robot's body) — The strands cross the visible torso and limbs without completely concealing their shapes; used as Makes the robot's entanglement legible within the wider deck context; Deck refuse (Accumulated around 찰리 after being dropped aboard); used as Uneven foreground and background groupings reveal a body among discarded objects; Collection vessel deck (Supporting the refuse and net-wrapped robot) — The deck plane recedes diagonally beneath the body; used as Visible gaps between the refuse establish a shared supporting plane.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and cool-neutral grading distinguish the net from 찰리's aged hard surfaces without adding a dramatic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S2sh6__bgfirst_bg.png",
  "asset_id": "f591caab-24b5-4e16-98eb-efeffc3022c5",
  "input_asset_ids": [
   "f8f2b853-c659-47d9-b6de-06b7d2d04916",
   "62155cdd-f3f2-45bf-802f-7e8d5aeec2bf"
  ]
 },
 "S2sh6": {
  "input_fingerprint": "feef994a1c6d1cbd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쓰레기 사이에 낡은 고철 로봇 찰리의 몸이 그물에 감긴 구도.\n\nLOCATION (lock): On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the deck-level approach from an elevated oblique position beside 찰리's torso, looking downward rather than along his facial axis in a direct observational view. Place his net-wrapped body diagonally across the central third, retaining his full bodily outline and adjacent refuse, with a few small foreground fragments creating depth without hiding him. His head lies turned away toward the surrounding refuse with no readable eye line; the narrowing camera distance is the emphasized change as the movement settles into the reveal.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Net around 찰리 (Wrapped around the robot's body) — The strands cross the visible torso and limbs without completely concealing their shapes; used as Makes the robot's entanglement legible within the wider deck context; Deck refuse (Accumulated around 찰리 after being dropped aboard); used as Uneven foreground and background groupings reveal a body among discarded objects; Collection vessel deck (Supporting the refuse and net-wrapped robot) — The deck plane recedes diagonally beneath the body; used as Visible gaps between the refuse establish a shared supporting plane.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and cool-neutral grading distinguish the net from 찰리's aged hard surfaces without adding a dramatic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless among the rubbish deposited on the collection ship's deck, with a net wrapped around his body. The positions of his torso, head, arms, and legs within the net, and his body's orientation against the deck, are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies on the collection ship's deck, with Charlie's old scrap-metal body still wrapped in netting among it. The chest retains its worn UBIC logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쓰레기 사이에 낡은 고철 로봇 찰리의 몸이 그물에 감긴 구도.\n\nLOCATION (lock): On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the deck-level approach from an elevated oblique position beside 찰리's torso, looking downward rather than along his facial axis in a direct observational view. Place his net-wrapped body diagonally across the central third, retaining his full bodily outline and adjacent refuse, with a few small foreground fragments creating depth without hiding him. His head lies turned away toward the surrounding refuse with no readable eye line; the narrowing camera distance is the emphasized change as the movement settles into the reveal.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Net around 찰리 (Wrapped around the robot's body) — The strands cross the visible torso and limbs without completely concealing their shapes; used as Makes the robot's entanglement legible within the wider deck context; Deck refuse (Accumulated around 찰리 after being dropped aboard); used as Uneven foreground and background groupings reveal a body among discarded objects; Collection vessel deck (Supporting the refuse and net-wrapped robot) — The deck plane recedes diagonally beneath the body; used as Visible gaps between the refuse establish a shared supporting plane.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and cool-neutral grading distinguish the net from 찰리's aged hard surfaces without adding a dramatic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless among the rubbish deposited on the collection ship's deck, with a net wrapped around his body. The positions of his torso, head, arms, and legs within the net, and his body's orientation against the deck, are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies on the collection ship's deck, with Charlie's old scrap-metal body still wrapped in netting among it. The chest retains its worn UBIC logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갑판 위 쓰레기 사이에 낡은 고철 로봇 찰리의 몸이 그물에 감긴 구도.\n\nLOCATION (lock): On the exposed working deck of a rubbish-collection ship, among freshly deposited waste and tangled netting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the deck-level approach from an elevated oblique position beside 찰리's torso, looking downward rather than along his facial axis in a direct observational view. Place his net-wrapped body diagonally across the central third, retaining his full bodily outline and adjacent refuse, with a few small foreground fragments creating depth without hiding him. His head lies turned away toward the surrounding refuse with no readable eye line; the narrowing camera distance is the emphasized change as the movement settles into the reveal.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Net around 찰리 (Wrapped around the robot's body) — The strands cross the visible torso and limbs without completely concealing their shapes; used as Makes the robot's entanglement legible within the wider deck context; Deck refuse (Accumulated around 찰리 after being dropped aboard); used as Uneven foreground and background groupings reveal a body among discarded objects; Collection vessel deck (Supporting the refuse and net-wrapped robot) — The deck plane recedes diagonally beneath the body; used as Visible gaps between the refuse establish a shared supporting plane.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and cool-neutral grading distinguish the net from 찰리's aged hard surfaces without adding a dramatic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless among the rubbish deposited on the collection ship's deck, with a net wrapped around his body. The positions of his torso, head, arms, and legs within the net, and his body's orientation against the deck, are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rubbish lies on the collection ship's deck, with Charlie's old scrap-metal body still wrapped in netting among it. The chest retains its worn UBIC logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S2sh6__bgfirst_bg.png",
     "asset_id": "f591caab-24b5-4e16-98eb-efeffc3022c5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S2sh6.png",
     "asset_id": "f8f2b853-c659-47d9-b6de-06b7d2d04916",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_collection_deck_b9e7b9.png",
     "asset_id": "62155cdd-f3f2-45bf-802f-7e8d5aeec2bf",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 고개는 화면 바깥쪽 쓰레기 더미를 향해 꺾여 있으며, 시선은 드러나지 않음.",
    "built_space": "바다가 보이는 쓰레기 수거선 갑판 위로, 지시된 위치 레퍼런스의 구조물과 쓰레기 배치가 일치함.",
    "entities": "샌드 베이지 장갑판, 가슴의 푸른 에너지 원자로, UBIC 로고를 갖춘 찰리가 그물에 감긴 채 쓰레기 더미에 있음.",
    "hard_violations": [],
    "physics": "찰리의 무거운 기계 몸체는 쓰레기 더미 위에 완전히 밀착되어 중력에 순응하며 누워 있음."
   },
   {
    "label": "B",
    "direction": "로봇의 머리가 쓰레기 쪽으로 돌아가 있어 시선을 확인할 수 없음.",
    "built_space": "지시된 갑판의 형태, 난간, 크레인 등 구조물과 쓰레기 더미 배치가 레퍼런스와 동일함.",
    "entities": "그물에 감긴 샌드 베이지 로봇과 UBIC 로고는 보이나, 가슴 중앙의 푸른 에너지 원자로가 빛이 꺼진 짙은 색으로 잘못 묘사됨.",
    "hard_violations": [],
    "physics": "바닥과 쓰레기 더미 위에 로봇이 안정적으로 안착되어 중력 법칙을 잘 따름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터 가슴의 푸른 원자로와 UBIC 로고, 그물에 감긴 상태를 지시문과 레퍼런스에 맞게 정확히 묘사했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "환경과 배치, 로고 표현은 우수하나, 캐릭터 가슴의 핵심인 푸른 에너지 원자로 묘사가 누락되어 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 고개는 화면 바깥쪽 쓰레기 더미를 향해 꺾여 있으며, 시선은 드러나지 않음.",
        "built_space": "바다가 보이는 쓰레기 수거선 갑판 위로, 지시된 위치 레퍼런스의 구조물과 쓰레기 배치가 일치함.",
        "entities": "샌드 베이지 장갑판, 가슴의 푸른 에너지 원자로, UBIC 로고를 갖춘 찰리가 그물에 감긴 채 쓰레기 더미에 있음.",
        "hard_violations": [],
        "physics": "찰리의 무거운 기계 몸체는 쓰레기 더미 위에 완전히 밀착되어 중력에 순응하며 누워 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 머리가 쓰레기 쪽으로 돌아가 있어 시선을 확인할 수 없음.",
        "built_space": "지시된 갑판의 형태, 난간, 크레인 등 구조물과 쓰레기 더미 배치가 레퍼런스와 동일함.",
        "entities": "그물에 감긴 샌드 베이지 로봇과 UBIC 로고는 보이나, 가슴 중앙의 푸른 에너지 원자로가 빛이 꺼진 짙은 색으로 잘못 묘사됨.",
        "hard_violations": [],
        "physics": "바닥과 쓰레기 더미 위에 로봇이 안정적으로 안착되어 중력 법칙을 잘 따름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터 가슴의 푸른 원자로와 UBIC 로고, 그물에 감긴 상태를 지시문과 레퍼런스에 맞게 정확히 묘사했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "환경과 배치, 로고 표현은 우수하나, 캐릭터 가슴의 핵심인 푸른 에너지 원자로 묘사가 누락되어 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 고개는 화면 바깥쪽 쓰레기 더미를 향해 꺾여 있으며, 시선은 드러나지 않음.",
        "built_space": "바다가 보이는 쓰레기 수거선 갑판 위로, 지시된 위치 레퍼런스의 구조물과 쓰레기 배치가 일치함.",
        "entities": "샌드 베이지 장갑판, 가슴의 푸른 에너지 원자로, UBIC 로고를 갖춘 찰리가 그물에 감긴 채 쓰레기 더미에 있음.",
        "hard_violations": [],
        "physics": "찰리의 무거운 기계 몸체는 쓰레기 더미 위에 완전히 밀착되어 중력에 순응하며 누워 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 머리가 쓰레기 쪽으로 돌아가 있어 시선을 확인할 수 없음.",
        "built_space": "지시된 갑판의 형태, 난간, 크레인 등 구조물과 쓰레기 더미 배치가 레퍼런스와 동일함.",
        "entities": "그물에 감긴 샌드 베이지 로봇과 UBIC 로고는 보이나, 가슴 중앙의 푸른 에너지 원자로가 빛이 꺼진 짙은 색으로 잘못 묘사됨.",
        "hard_violations": [],
        "physics": "바닥과 쓰레기 더미 위에 로봇이 안정적으로 안착되어 중력 법칙을 잘 따름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "몸통 옆에서 비스듬히 내려다보는 가까운 관찰 시점과 대각선 전신 배치가 더 충실하지만, 몸이 중앙 3분의 1보다 크게 차지하고 장갑색은 기준보다 회색에 가깝다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샌드 베이지 장갑과 푸른 원자로는 더 정확하지만, 수평선과 배경을 넓게 담은 낮고 먼 시점이 지정된 하향 접근 구도보다 장소 사진의 구도를 따른다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리는 화면 왼쪽 위 쓰레기 쪽으로 돌아가 있고 얼굴 전면과 눈의 방향은 읽히지 않는다. 몸은 왼쪽 위 머리에서 오른쪽 아래 발로 이어지는 대각선이다. 카메라는 몸통 옆의 높은 사선 위치에서 가슴과 팔다리 윗면을 내려다본다. 무기나 조작 중인 물건은 없다.",
        "built_space": "오른쪽에 녹슨 선실 벽과 수직 사다리 1개, 뒤쪽에 소형 크레인 1대와 케이블 윈치 1기가 보인다. 뒤편 선측과 바다, 왼쪽 폐기물 더미의 배치는 장소 기준과 일치한다. 몸 주변과 오른쪽 뒤편에 젖은 갑판이 드러나 공통 지지면을 확인할 수 있다. 전경의 큰 상자와 봉투는 작은 파편만 두라는 지시보다 비중이 크지만 몸을 심하게 가리지는 않는다.",
        "entities": "사람 없이 기계 몸체 찰리 1대만 있다. 긴 육중한 팔, 상대적으로 짧은 다리, 각진 장갑과 기계 손이 보인다. 장갑은 심하게 부식된 회백색으로 샌드 베이지 기준보다 차갑고 어둡다. 가슴에는 닳은 UBIC 표기와 원형 원자로가 있으나 푸른 발광은 약하다. 얼굴은 돌아가 있어 흰 마스크의 세부는 판독할 수 없다. 그물은 몸통과 양팔·양다리를 가로지르면서 윤곽을 남기며, 금속 폐품과 봉투가 주변을 채운다.",
        "hard_violations": [],
        "physics": "머리와 등은 뒤쪽 폐품 더미에 기대고, 화면 왼쪽 팔과 손은 낮은 고철 더미에 놓여 있다. 반대 손은 허벅지와 인접 폐품 위에 내려앉아 있으며 다리와 발도 아래 폐기물에 받쳐져 있다. 그물은 장갑 표면을 따라 걸리고 남은 부분은 쓰레기 위로 처진다. 지지 없이 떠 있는 신체나 물건은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "머리는 화면 왼쪽 위의 폐기물 쪽으로 돌아가 있고 눈이나 얼굴 전면은 읽히지 않는다. 몸은 머리에서 오른쪽 아래 발로 대각선을 이룬다. 카메라도 아래를 보지만 A보다 낮고 멀게 느껴지며, 넓은 수평선과 배경 갑판이 강조되어 몸통 옆에서 가까워지는 관찰 시점은 약하다. 겨누거나 조작하는 물건은 없다.",
        "built_space": "오른쪽 선실 벽의 사다리 1개, 뒤쪽 소형 크레인 1대, 케이블 윈치 1기가 보이며 장소 사진의 주요 설비 배치를 유지한다. 왼쪽 쓰레기 더미와 뒤쪽 선측 사이로 젖은 갑판이 넓게 드러난다. 전신과 주변 폐기물은 프레임 안에 있으나 큰 전경 그물과 폐품이 상당한 면적을 차지한다. 수면과 갑판의 반사는 보이는 광원 및 표면과 모순되지 않는다.",
        "entities": "찰리 1대만 보이며 추가 인물은 없다. 샌드 베이지 장갑, 큰 팔과 손, 짧은 다리, 푸른 원형 가슴 원자로가 인물 기준에 비교적 가깝다. 가슴의 작은 UBIC 표기도 보인다. 머리가 돌아가 있으므로 흰 마스크와 눈의 세부는 노출되지 않는다. 몸통과 팔다리에 그물이 감겨 있고 폐금속, 용기, 봉투와 타이어가 주변에 놓여 있다.",
        "hard_violations": [],
        "physics": "머리와 상체는 경사진 폐기물 더미가 받치고, 화면 왼쪽 손은 아래쪽 쓰레기에 내려앉아 있다. 반대 손은 허벅지 부근에 놓이며 두 다리와 발도 폐품과 갑판 위에 지지된다. 그물은 몸의 돌출부를 따라 걸쳐지고 주변 바닥으로 이어진다. 정지한 몸이 스스로 팔다리를 들고 있거나 물건이 공중에 뜬 모습은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "몸통 옆에서 비스듬히 내려다보는 가까운 관찰 시점과 대각선 전신 배치가 더 충실하지만, 몸이 중앙 3분의 1보다 크게 차지하고 장갑색은 기준보다 회색에 가깝다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "샌드 베이지 장갑과 푸른 원자로는 더 정확하지만, 수평선과 배경을 넓게 담은 낮고 먼 시점이 지정된 하향 접근 구도보다 장소 사진의 구도를 따른다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리는 화면 왼쪽 위 쓰레기 쪽으로 돌아가 있고 얼굴 전면과 눈의 방향은 읽히지 않는다. 몸은 왼쪽 위 머리에서 오른쪽 아래 발로 이어지는 대각선이다. 카메라는 몸통 옆의 높은 사선 위치에서 가슴과 팔다리 윗면을 내려다본다. 무기나 조작 중인 물건은 없다.",
        "built_space": "오른쪽에 녹슨 선실 벽과 수직 사다리 1개, 뒤쪽에 소형 크레인 1대와 케이블 윈치 1기가 보인다. 뒤편 선측과 바다, 왼쪽 폐기물 더미의 배치는 장소 기준과 일치한다. 몸 주변과 오른쪽 뒤편에 젖은 갑판이 드러나 공통 지지면을 확인할 수 있다. 전경의 큰 상자와 봉투는 작은 파편만 두라는 지시보다 비중이 크지만 몸을 심하게 가리지는 않는다.",
        "entities": "사람 없이 기계 몸체 찰리 1대만 있다. 긴 육중한 팔, 상대적으로 짧은 다리, 각진 장갑과 기계 손이 보인다. 장갑은 심하게 부식된 회백색으로 샌드 베이지 기준보다 차갑고 어둡다. 가슴에는 닳은 UBIC 표기와 원형 원자로가 있으나 푸른 발광은 약하다. 얼굴은 돌아가 있어 흰 마스크의 세부는 판독할 수 없다. 그물은 몸통과 양팔·양다리를 가로지르면서 윤곽을 남기며, 금속 폐품과 봉투가 주변을 채운다.",
        "hard_violations": [],
        "physics": "머리와 등은 뒤쪽 폐품 더미에 기대고, 화면 왼쪽 팔과 손은 낮은 고철 더미에 놓여 있다. 반대 손은 허벅지와 인접 폐품 위에 내려앉아 있으며 다리와 발도 아래 폐기물에 받쳐져 있다. 그물은 장갑 표면을 따라 걸리고 남은 부분은 쓰레기 위로 처진다. 지지 없이 떠 있는 신체나 물건은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "머리는 화면 왼쪽 위의 폐기물 쪽으로 돌아가 있고 눈이나 얼굴 전면은 읽히지 않는다. 몸은 머리에서 오른쪽 아래 발로 대각선을 이룬다. 카메라도 아래를 보지만 A보다 낮고 멀게 느껴지며, 넓은 수평선과 배경 갑판이 강조되어 몸통 옆에서 가까워지는 관찰 시점은 약하다. 겨누거나 조작하는 물건은 없다.",
        "built_space": "오른쪽 선실 벽의 사다리 1개, 뒤쪽 소형 크레인 1대, 케이블 윈치 1기가 보이며 장소 사진의 주요 설비 배치를 유지한다. 왼쪽 쓰레기 더미와 뒤쪽 선측 사이로 젖은 갑판이 넓게 드러난다. 전신과 주변 폐기물은 프레임 안에 있으나 큰 전경 그물과 폐품이 상당한 면적을 차지한다. 수면과 갑판의 반사는 보이는 광원 및 표면과 모순되지 않는다.",
        "entities": "찰리 1대만 보이며 추가 인물은 없다. 샌드 베이지 장갑, 큰 팔과 손, 짧은 다리, 푸른 원형 가슴 원자로가 인물 기준에 비교적 가깝다. 가슴의 작은 UBIC 표기도 보인다. 머리가 돌아가 있으므로 흰 마스크와 눈의 세부는 노출되지 않는다. 몸통과 팔다리에 그물이 감겨 있고 폐금속, 용기, 봉투와 타이어가 주변에 놓여 있다.",
        "hard_violations": [],
        "physics": "머리와 상체는 경사진 폐기물 더미가 받치고, 화면 왼쪽 손은 아래쪽 쓰레기에 내려앉아 있다. 반대 손은 허벅지 부근에 놓이며 두 다리와 발도 폐품과 갑판 위에 지지된다. 그물은 몸의 돌출부를 따라 걸쳐지고 주변 바닥으로 이어진다. 정지한 몸이 스스로 팔다리를 들고 있거나 물건이 공중에 뜬 모습은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "캐릭터 가슴의 푸른 원자로와 UBIC 로고, 그물에 감긴 상태를 지시문과 레퍼런스에 맞게 정확히 묘사했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "환경과 배치, 로고 표현은 우수하나, 캐릭터 가슴의 핵심인 푸른 에너지 원자로 묘사가 누락되어 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_collection_deck_b9e7b9.png",
    "asset_id": "62155cdd-f3f2-45bf-802f-7e8d5aeec2bf",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07a8-3dd2-76c7-b842-db7a69dba243",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S2sh6__bgfirst_bg.png",
   "bg_asset_id": "f591caab-24b5-4e16-98eb-efeffc3022c5",
   "bg_record_key": "S2sh6::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "collection_deck",
   "groupbg_asset_id": "62155cdd-f3f2-45bf-802f-7e8d5aeec2bf"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh2::signage": {
  "fp": "66941d5968009c67",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4b7a5fa00e1b78d7": {
  "subjects": [],
  "subject_text": "고철 운반 덤프트럭 운전실\n운전대와 계기판, 라디오가 배치된 좁은 운전실. 넓은 앞유리를 통해 낮빛이 들어오며 대시보드가 전면을 차지한다.",
  "identity": "canonical",
  "scope_id": "L06",
  "scope_role": "location_interior",
  "scope_sha": "9003693cfacc42c9"
 },
 "groupbg::dump_cargo": {
  "input_fingerprint": "88859a0b14f2d68f",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "dump_cargo",
    "tags": [
     "S3sh2"
    ]
   },
   "context_sig": "17a267357ccf21f9"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n고철 운반 덤프트럭 운전실: 노후한 덤프트럭 내부로, 구형 조작 패널과 아날로그 라디오가 장착된 공간이다. (특징: 운전대를 잡고 꾸벅꾸벅 조는 운전사; 자율주행 시스템이 가동 중인 운전석 계기판 표지; 소리가 흘러나오는 낡은 카 라디오 장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 찰리를 포함해 고철을 가득 실은 덤프트럭이 달리고 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n고철 운반 덤프트럭 운전실: 노후한 덤프트럭 내부로, 구형 조작 패널과 아날로그 라디오가 장착된 공간이다. (특징: 운전대를 잡고 꾸벅꾸벅 조는 운전사; 자율주행 시스템이 가동 중인 운전석 계기판 표지; 소리가 흘러나오는 낡은 카 라디오 장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 찰리를 포함해 고철을 가득 실은 덤프트럭이 달리고 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_dump_cargo_6eb338.png",
  "asset_id": "d8fda544-14ad-4755-881b-5e32ad14c29b",
  "input_asset_ids": [
   "dee5c945-81f8-4b25-8542-09104728252c"
  ],
  "origin_tag": "S3sh2",
  "place_text": "Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.",
  "origin_inputs": {
   "place_text": "Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.",
   "time_of_day_en": "day",
   "conti_asset_id": "dee5c945-81f8-4b25-8542-09104728252c"
  }
 },
 "S3sh2::bgfirst_bg": {
  "input_fingerprint": "5a891601776fab2a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 덤프트럭 적재함의 빽빽한 고철 사이에 찰리의 몸체가 끼어 있는 내려다보기 구도.\n\nLOCATION (lock): Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rise, maintain the truck's longitudinal pace above the cargo-bed side edge and look steeply downward in a direct observational view. Position 찰리's wedged body slightly right of center, partially interrupted by densely packed scrap, while a strip of the cargo-bed edge at left establishes containment and scale. His face angles into the load without a readable gaze; hold the composition before the explicit cut into the cab rather than suggesting passage through the truck's structure.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Densely packed scrap (Filling the moving dump truck's cargo bed and trapping 찰리 among it); used as Irregular overlaps reveal the robot in fragments without enlarging individual scrap pieces; Dump truck cargo bed (Loaded with scrap while the truck travels) — The top of the side edge and the inward-facing cargo space are visible from above; used as The side edge fixes the camera's position relative to the moving load.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled cool-neutral contrast keep 찰리 distinguishable from the surrounding scrap without introducing a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 덤프트럭 적재함의 빽빽한 고철 사이에 찰리의 몸체가 끼어 있는 내려다보기 구도.\n\nLOCATION (lock): Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rise, maintain the truck's longitudinal pace above the cargo-bed side edge and look steeply downward in a direct observational view. Position 찰리's wedged body slightly right of center, partially interrupted by densely packed scrap, while a strip of the cargo-bed edge at left establishes containment and scale. His face angles into the load without a readable gaze; hold the composition before the explicit cut into the cab rather than suggesting passage through the truck's structure.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Densely packed scrap (Filling the moving dump truck's cargo bed and trapping 찰리 among it); used as Irregular overlaps reveal the robot in fragments without enlarging individual scrap pieces; Dump truck cargo bed (Loaded with scrap while the truck travels) — The top of the side edge and the inward-facing cargo space are visible from above; used as The side edge fixes the camera's position relative to the moving load.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled cool-neutral contrast keep 찰리 distinguishable from the surrounding scrap without introducing a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh2__bgfirst_bg.png",
  "asset_id": "22af4401-fec0-4da8-b380-ec1df9866e0e",
  "input_asset_ids": [
   "dee5c945-81f8-4b25-8542-09104728252c",
   "d8fda544-14ad-4755-881b-5e32ad14c29b"
  ]
 },
 "S3sh2": {
  "input_fingerprint": "29412f54398033a0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덤프트럭 적재함의 빽빽한 고철 사이에 찰리의 몸체가 끼어 있는 내려다보기 구도.\n\nLOCATION (lock): Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rise, maintain the truck's longitudinal pace above the cargo-bed side edge and look steeply downward in a direct observational view. Position 찰리's wedged body slightly right of center, partially interrupted by densely packed scrap, while a strip of the cargo-bed edge at left establishes containment and scale. His face angles into the load without a readable gaze; hold the composition before the explicit cut into the cab rather than suggesting passage through the truck's structure.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Densely packed scrap (Filling the moving dump truck's cargo bed and trapping 찰리 among it); used as Irregular overlaps reveal the robot in fragments without enlarging individual scrap pieces; Dump truck cargo bed (Loaded with scrap while the truck travels) — The top of the side edge and the inward-facing cargo space are visible from above; used as The side edge fixes the camera's position relative to the moving load.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled cool-neutral contrast keep 찰리 distinguishable from the surrounding scrap without introducing a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless amid the scrap metal loaded in the dump truck's cargo bed. The scrap surrounds his body, but the arrangement of his torso, head, arms, and legs and his facing direction are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moving dump truck's bed is packed with scrap metal, including Charlie's old, net-entangled body with its worn UBIC chest logo. The truck is operating autonomously in daylight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덤프트럭 적재함의 빽빽한 고철 사이에 찰리의 몸체가 끼어 있는 내려다보기 구도.\n\nLOCATION (lock): Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rise, maintain the truck's longitudinal pace above the cargo-bed side edge and look steeply downward in a direct observational view. Position 찰리's wedged body slightly right of center, partially interrupted by densely packed scrap, while a strip of the cargo-bed edge at left establishes containment and scale. His face angles into the load without a readable gaze; hold the composition before the explicit cut into the cab rather than suggesting passage through the truck's structure.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Densely packed scrap (Filling the moving dump truck's cargo bed and trapping 찰리 among it); used as Irregular overlaps reveal the robot in fragments without enlarging individual scrap pieces; Dump truck cargo bed (Loaded with scrap while the truck travels) — The top of the side edge and the inward-facing cargo space are visible from above; used as The side edge fixes the camera's position relative to the moving load.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled cool-neutral contrast keep 찰리 distinguishable from the surrounding scrap without introducing a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless amid the scrap metal loaded in the dump truck's cargo bed. The scrap surrounds his body, but the arrangement of his torso, head, arms, and legs and his facing direction are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moving dump truck's bed is packed with scrap metal, including Charlie's old, net-entangled body with its worn UBIC chest logo. The truck is operating autonomously in daylight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덤프트럭 적재함의 빽빽한 고철 사이에 찰리의 몸체가 끼어 있는 내려다보기 구도.\n\nLOCATION (lock): Inside the open-topped load bed of a moving dump truck, tightly packed with scrap metal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rise, maintain the truck's longitudinal pace above the cargo-bed side edge and look steeply downward in a direct observational view. Position 찰리's wedged body slightly right of center, partially interrupted by densely packed scrap, while a strip of the cargo-bed edge at left establishes containment and scale. His face angles into the load without a readable gaze; hold the composition before the explicit cut into the cab rather than suggesting passage through the truck's structure.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Densely packed scrap (Filling the moving dump truck's cargo bed and trapping 찰리 among it); used as Irregular overlaps reveal the robot in fragments without enlarging individual scrap pieces; Dump truck cargo bed (Loaded with scrap while the truck travels) — The top of the side edge and the inward-facing cargo space are visible from above; used as The side edge fixes the camera's position relative to the moving load.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled cool-neutral contrast keep 찰리 distinguishable from the surrounding scrap without introducing a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is motionless amid the scrap metal loaded in the dump truck's cargo bed. The scrap surrounds his body, but the arrangement of his torso, head, arms, and legs and his facing direction are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The moving dump truck's bed is packed with scrap metal, including Charlie's old, net-entangled body with its worn UBIC chest logo. The truck is operating autonomously in daylight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh2__bgfirst_bg.png",
     "asset_id": "22af4401-fec0-4da8-b380-ec1df9866e0e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S3sh2.png",
     "asset_id": "dee5c945-81f8-4b25-8542-09104728252c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_dump_cargo_6eb338.png",
     "asset_id": "d8fda544-14ad-4755-881b-5e32ad14c29b",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 앞을 향해 약한 하향각을 취함. 찰리의 얼굴은 정면 좌측을 향함.",
    "built_space": "덤프트럭 적재함 내부. 왼쪽 적재함 측면이 프레임에 걸쳐 있으며, 뒷배경으로 다리와 도시 스카이라인이 넓게 보임.",
    "entities": "찰리(베이지색 장갑, 흰 마스크, 가슴의 빛나는 원자로)가 고철 더미 위에 놓여 있음. 참조 이미지와 일치함.",
    "hard_violations": [],
    "physics": "찰리는 고철 위에 얹혀져 지지되고 있으며, 도로의 모션 블러를 통해 트럭이 이동 중임을 알 수 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 적재함 내부를 향해 가파르게 내려다보는 뷰. 찰리의 얼굴은 고철 더미 안쪽을 향해 틀어져 있음.",
    "built_space": "덤프트럭 적재함 내부. 왼쪽 가장자리에 적재함 벽면 일부가 보이고, 바깥으로는 하늘 없이 지나가는 도로 바닥만 보임.",
    "entities": "찰리(참조 이미지의 기계 몸체와 완벽히 일치)가 빽빽한 고철 사이에 파묻혀 있음.",
    "hard_violations": [],
    "physics": "찰리의 몸은 고철 사이에 깊숙이 끼어 안정적으로 지지받고 있으며, 왼쪽 도로의 모션 블러로 주행 상태를 잘 보여줌."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 '가파르게 내려다보는 구도(steeply downward)'를 정확히 따랐으며, 왼쪽 적재함 가장자리와 고철 사이에 끼인 찰리의 위치를 프레이밍에 완벽히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "내려다보는 구도 지시를 무시하고 장소 레퍼런스 이미지의 수평적 카메라 구도(스카이라인 노출)를 그대로 모방하여 프레이밍 조건에서 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 앞을 향해 약한 하향각을 취함. 찰리의 얼굴은 정면 좌측을 향함.",
        "built_space": "덤프트럭 적재함 내부. 왼쪽 적재함 측면이 프레임에 걸쳐 있으며, 뒷배경으로 다리와 도시 스카이라인이 넓게 보임.",
        "entities": "찰리(베이지색 장갑, 흰 마스크, 가슴의 빛나는 원자로)가 고철 더미 위에 놓여 있음. 참조 이미지와 일치함.",
        "hard_violations": [],
        "physics": "찰리는 고철 위에 얹혀져 지지되고 있으며, 도로의 모션 블러를 통해 트럭이 이동 중임을 알 수 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 적재함 내부를 향해 가파르게 내려다보는 뷰. 찰리의 얼굴은 고철 더미 안쪽을 향해 틀어져 있음.",
        "built_space": "덤프트럭 적재함 내부. 왼쪽 가장자리에 적재함 벽면 일부가 보이고, 바깥으로는 하늘 없이 지나가는 도로 바닥만 보임.",
        "entities": "찰리(참조 이미지의 기계 몸체와 완벽히 일치)가 빽빽한 고철 사이에 파묻혀 있음.",
        "hard_violations": [],
        "physics": "찰리의 몸은 고철 사이에 깊숙이 끼어 안정적으로 지지받고 있으며, 왼쪽 도로의 모션 블러로 주행 상태를 잘 보여줌."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 '가파르게 내려다보는 구도(steeply downward)'를 정확히 따랐으며, 왼쪽 적재함 가장자리와 고철 사이에 끼인 찰리의 위치를 프레이밍에 완벽히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "내려다보는 구도 지시를 무시하고 장소 레퍼런스 이미지의 수평적 카메라 구도(스카이라인 노출)를 그대로 모방하여 프레이밍 조건에서 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 앞을 향해 약한 하향각을 취함. 찰리의 얼굴은 정면 좌측을 향함.",
        "built_space": "덤프트럭 적재함 내부. 왼쪽 적재함 측면이 프레임에 걸쳐 있으며, 뒷배경으로 다리와 도시 스카이라인이 넓게 보임.",
        "entities": "찰리(베이지색 장갑, 흰 마스크, 가슴의 빛나는 원자로)가 고철 더미 위에 놓여 있음. 참조 이미지와 일치함.",
        "hard_violations": [],
        "physics": "찰리는 고철 위에 얹혀져 지지되고 있으며, 도로의 모션 블러를 통해 트럭이 이동 중임을 알 수 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 적재함 내부를 향해 가파르게 내려다보는 뷰. 찰리의 얼굴은 고철 더미 안쪽을 향해 틀어져 있음.",
        "built_space": "덤프트럭 적재함 내부. 왼쪽 가장자리에 적재함 벽면 일부가 보이고, 바깥으로는 하늘 없이 지나가는 도로 바닥만 보임.",
        "entities": "찰리(참조 이미지의 기계 몸체와 완벽히 일치)가 빽빽한 고철 사이에 파묻혀 있음.",
        "hard_violations": [],
        "physics": "찰리의 몸은 고철 사이에 깊숙이 끼어 안정적으로 지지받고 있으며, 왼쪽 도로의 모션 블러로 주행 상태를 잘 보여줌."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "가파른 내려다보기와 중앙 오른쪽에 고철로 부분 가려진 몸체는 잘 맞지만, 얼굴이 적재물 안쪽이 아니라 위로 드러나며 그물과 UBIC 표식은 확인되지 않습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "장소와 기계 외형은 충실하지만, 지평선과 운전실까지 보이는 낮은 시점이 지정된 가파른 내려다보기를 벗어나고 얼굴과 상체도 지나치게 드러납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리는 화면 오른쪽으로 기울었으나 흰 얼굴 면과 두 발광 눈은 위쪽 카메라 방향으로 노출됩니다. 적재물 안쪽으로 얼굴을 묻어 시선을 읽을 수 없게 하라는 지시와 다릅니다. 왼쪽 도로의 흐림은 차량 진행을 암시하지만 전진 방향 자체는 확정하기 어렵습니다. 조준하거나 사용하는 도구는 없습니다.",
        "built_space": "왼쪽 측벽 상단 한 줄과 오른쪽 측벽 하나, 상단에 일부 걸린 가로 벽이 보이는 개방형 적재함입니다. 녹슨 철제 벽 안을 휠, 철판, 빔, 배선, 원통형 폐부품이 밀집해 채워 장소의 재질과 구조를 따릅니다. 카메라는 왼쪽 가장자리 위에서 가파르게 내려다보며, 찰리는 중앙 오른쪽에 놓입니다. 고철의 겹침으로 몸 일부가 가려지고 지평선은 보이지 않습니다. 중복 설비나 부적절한 반사는 없습니다.",
        "entities": "등장 인물은 없고 찰리의 기계 몸체 하나만 있습니다. 마모된 샌드 베이지 장갑, 흰 마스크, 두 주황색 눈과 선 모양 입, 파란 원형 가슴 장치는 참조와 부합합니다. 큰 팔도 보이지만 가려진 하체의 전체 비율은 판단하지 않습니다. 몸 위에 얽힌 것은 주로 전선과 케이블로 보이며 그물 구조나 닳은 UBIC 글자는 식별되지 않습니다. 덤프트럭 적재함과 빽빽한 고철은 명확합니다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 고철 더미에 비스듬히 기대어 끼어 있고, 머리 뒤쪽과 어깨 주변에도 받치는 폐금속이 있습니다. 양팔과 보이는 손은 아래쪽 고철에 내려앉아 있으며 하체는 적재물 속에 묻혀 있습니다. 몸 위 케이블은 장갑과 고철에 걸쳐 놓여 있습니다. 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 왼쪽 아래의 고철 쪽으로 기울어 A보다 적재물 방향에 가깝습니다. 다만 흰 얼굴과 두 발광 눈, 입이 모두 읽혀 시선을 감추라는 지시는 충족하지 않습니다. 도로는 운전실 앞쪽으로 뻗고 노면 흐림은 주행을 암시합니다. 조준하거나 사용하는 도구는 없습니다.",
        "built_space": "좌우 측벽 각 하나와 앞쪽 가로 벽 하나가 보이며, 그 너머로 운전실 뒤판과 왼쪽 사이드미러 하나가 드러납니다. 녹슨 벽, 휠과 빔, 배선, 파란 철판 등은 장소 참조와 매우 가깝습니다. 그러나 도시 지평선과 하늘까지 보이는 완만한 하향 시점으로, 지정된 가파른 내려다보기 대신 장소 사진의 구도를 거의 따릅니다. 찰리는 중앙 오른쪽에 있으나 상체가 고철 위로 많이 드러납니다. 중복 설비나 불가능한 반사는 없습니다.",
        "entities": "찰리 한 몸체만 있으며 추가 사람은 없습니다. 샌드 베이지 장갑, 흰 마스크와 주황색 눈 두 개, 선 모양 입, 파란 원형 가슴 장치와 육중한 팔은 참조를 따릅니다. 하체는 가려져 전체 비율을 확정할 수 없습니다. 몸을 덮는 전선은 있지만 그물의 망 조직과 UBIC 표식은 확인되지 않습니다. 적재함과 고철은 분명하며, 운전자는 보이지 않아 자율주행 여부는 영상만으로 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "찰리는 뒤쪽 고철에 몸통을 기대어 반쯤 앉은 듯 끼어 있습니다. 아래로 뻗은 팔과 손은 원통형 부품과 배선 더미에 받쳐져 있고, 반대쪽 팔과 하체도 고철 속에 놓입니다. 기울어진 머리는 목 연결부에 붙어 있으며 공중에 분리되어 있지 않습니다. 고철과 케이블도 서로 기대거나 적재물에 걸쳐 있어 명백한 무지지 부유는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "가파른 내려다보기와 중앙 오른쪽에 고철로 부분 가려진 몸체는 잘 맞지만, 얼굴이 적재물 안쪽이 아니라 위로 드러나며 그물과 UBIC 표식은 확인되지 않습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "장소와 기계 외형은 충실하지만, 지평선과 운전실까지 보이는 낮은 시점이 지정된 가파른 내려다보기를 벗어나고 얼굴과 상체도 지나치게 드러납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리는 화면 오른쪽으로 기울었으나 흰 얼굴 면과 두 발광 눈은 위쪽 카메라 방향으로 노출됩니다. 적재물 안쪽으로 얼굴을 묻어 시선을 읽을 수 없게 하라는 지시와 다릅니다. 왼쪽 도로의 흐림은 차량 진행을 암시하지만 전진 방향 자체는 확정하기 어렵습니다. 조준하거나 사용하는 도구는 없습니다.",
        "built_space": "왼쪽 측벽 상단 한 줄과 오른쪽 측벽 하나, 상단에 일부 걸린 가로 벽이 보이는 개방형 적재함입니다. 녹슨 철제 벽 안을 휠, 철판, 빔, 배선, 원통형 폐부품이 밀집해 채워 장소의 재질과 구조를 따릅니다. 카메라는 왼쪽 가장자리 위에서 가파르게 내려다보며, 찰리는 중앙 오른쪽에 놓입니다. 고철의 겹침으로 몸 일부가 가려지고 지평선은 보이지 않습니다. 중복 설비나 부적절한 반사는 없습니다.",
        "entities": "등장 인물은 없고 찰리의 기계 몸체 하나만 있습니다. 마모된 샌드 베이지 장갑, 흰 마스크, 두 주황색 눈과 선 모양 입, 파란 원형 가슴 장치는 참조와 부합합니다. 큰 팔도 보이지만 가려진 하체의 전체 비율은 판단하지 않습니다. 몸 위에 얽힌 것은 주로 전선과 케이블로 보이며 그물 구조나 닳은 UBIC 글자는 식별되지 않습니다. 덤프트럭 적재함과 빽빽한 고철은 명확합니다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 고철 더미에 비스듬히 기대어 끼어 있고, 머리 뒤쪽과 어깨 주변에도 받치는 폐금속이 있습니다. 양팔과 보이는 손은 아래쪽 고철에 내려앉아 있으며 하체는 적재물 속에 묻혀 있습니다. 몸 위 케이블은 장갑과 고철에 걸쳐 놓여 있습니다. 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 왼쪽 아래의 고철 쪽으로 기울어 A보다 적재물 방향에 가깝습니다. 다만 흰 얼굴과 두 발광 눈, 입이 모두 읽혀 시선을 감추라는 지시는 충족하지 않습니다. 도로는 운전실 앞쪽으로 뻗고 노면 흐림은 주행을 암시합니다. 조준하거나 사용하는 도구는 없습니다.",
        "built_space": "좌우 측벽 각 하나와 앞쪽 가로 벽 하나가 보이며, 그 너머로 운전실 뒤판과 왼쪽 사이드미러 하나가 드러납니다. 녹슨 벽, 휠과 빔, 배선, 파란 철판 등은 장소 참조와 매우 가깝습니다. 그러나 도시 지평선과 하늘까지 보이는 완만한 하향 시점으로, 지정된 가파른 내려다보기 대신 장소 사진의 구도를 거의 따릅니다. 찰리는 중앙 오른쪽에 있으나 상체가 고철 위로 많이 드러납니다. 중복 설비나 불가능한 반사는 없습니다.",
        "entities": "찰리 한 몸체만 있으며 추가 사람은 없습니다. 샌드 베이지 장갑, 흰 마스크와 주황색 눈 두 개, 선 모양 입, 파란 원형 가슴 장치와 육중한 팔은 참조를 따릅니다. 하체는 가려져 전체 비율을 확정할 수 없습니다. 몸을 덮는 전선은 있지만 그물의 망 조직과 UBIC 표식은 확인되지 않습니다. 적재함과 고철은 분명하며, 운전자는 보이지 않아 자율주행 여부는 영상만으로 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "찰리는 뒤쪽 고철에 몸통을 기대어 반쯤 앉은 듯 끼어 있습니다. 아래로 뻗은 팔과 손은 원통형 부품과 배선 더미에 받쳐져 있고, 반대쪽 팔과 하체도 고철 속에 놓입니다. 기울어진 머리는 목 연결부에 붙어 있으며 공중에 분리되어 있지 않습니다. 고철과 케이블도 서로 기대거나 적재물에 걸쳐 있어 명백한 무지지 부유는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.143,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.143,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1143
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 '가파르게 내려다보는 구도(steeply downward)'를 정확히 따랐으며, 왼쪽 적재함 가장자리와 고철 사이에 끼인 찰리의 위치를 프레이밍에 완벽히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1143,
    "verdict_ko": "내려다보는 구도 지시를 무시하고 장소 레퍼런스 이미지의 수평적 카메라 구도(스카이라인 노출)를 그대로 모방하여 프레이밍 조건에서 크게 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_dump_cargo_6eb338.png",
    "asset_id": "d8fda544-14ad-4755-881b-5e32ad14c29b",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07b0-b85e-7929-9009-944ec651cc2c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh2__bgfirst_bg.png",
   "bg_asset_id": "22af4401-fec0-4da8-b380-ec1df9866e0e",
   "bg_record_key": "S3sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "dump_cargo",
   "groupbg_asset_id": "d8fda544-14ad-4755-881b-5e32ad14c29b"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh3::confined_fp_apt": {
  "applies": true,
  "reason_ko": "덤프트럭 운전석이라는 통제 장치가 집중된 밀폐 공간을 배경으로 하며, 운전석, 운전대, 앞유리의 위치와 방향이 정확히 배치되어야 이야기의 흐름이 깨지지 않기 때문입니다.",
  "input_fingerprint": "7569d0775e27dc9f"
 },
 "S3sh3::signage": {
  "fp": "37b5b85249a92206",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::7afce4fe93a5": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_7afce4fe93a5.png",
  "place_text": "Inside the dump truck's driving cab, at the driver's seat behind the steering wheel. Daylight enters through the windshield overlooking the road.",
  "input_fingerprint": "65c21f29ca66ea8f"
 },
 "S3sh3::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located on the left side of the dashboard, directly in front of the left seat.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "The camera is positioned on the right (passenger) side of the cabin, pointing diagonally forward and to the left towards the driver and the windshield.",
   "occupants": "A Driver occupies the left seat."
  },
  "mismatches": [],
  "scene_description_en": "The camera observes from the right passenger side, looking diagonally forward-left across the truck cab. In the near left foreground, the dump truck driver occupies the left seat, seen in profile facing forward. Just below the center of the frame, the steering wheel sits directly in front of the driver. The broad windshield spans the center and right background, facing forward to reveal the road ahead. No mirrors or reflective surfaces are present in the scene.",
  "fixed": true,
  "input_fingerprint": "6076f4776b83c1cf"
 },
 "S3sh3": {
  "input_fingerprint": "0793e68d0be34fd6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대와 앞유리 너머 도로를 포함한 운전석 구도 속 덤프트럭 운전사가 졸음으로 눈꺼풀을 늘어뜨린 모습.\n\nLOCATION (lock): Inside the dump truck's driving cab, at the driver's seat behind the steering wheel. Daylight enters through the windshield overlooking the road. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut into the cab and hold a direct observational view from the passenger side, slightly behind the driver's shoulder at seated eye height. Keep 덤프트럭 운전사's profile and upper body at left, the steering wheel below center, and the road visible through the windshield at right. Catch his head beginning to nod and his eyelids hanging low, with his unfocused gaze dropping toward the wheel rather than toward the camera.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Steering wheel (In front of the drowsing driver during autonomous travel) — Seen obliquely from the passenger side, below the driver's profile; used as Links the driver's nodding posture to the vehicle controls without implying active steering; Windshield (The road ahead is visible through it) — Viewed diagonally from inside the cab, with the forward road beyond; used as Maintains the travelling context beside the driver's face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daytime ambient light preserves readable eyelids and the road beyond without introducing an unsupported cabin light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dump truck continues autonomously along the daytime road with its bed full of scrap. Charlie remains in the load, an old net-entangled robot with a worn UBIC chest logo.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera observes from the right passenger side, looking diagonally forward-left across the truck cab. In the near left foreground, the dump truck driver occupies the left seat, seen in profile facing forward. Just below the center of the frame, the steering wheel sits directly in front of the driver. The broad windshield spans the center and right background, facing forward to reveal the road ahead. No mirrors or reflective surfaces are present in the scene.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대와 앞유리 너머 도로를 포함한 운전석 구도 속 덤프트럭 운전사가 졸음으로 눈꺼풀을 늘어뜨린 모습.\n\nLOCATION (lock): Inside the dump truck's driving cab, at the driver's seat behind the steering wheel. Daylight enters through the windshield overlooking the road. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daytime ambient light preserves readable eyelids and the road beyond without introducing an unsupported cabin light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dump truck continues autonomously along the daytime road with its bed full of scrap. Charlie remains in the load, an old net-entangled robot with a worn UBIC chest logo.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera observes from the right passenger side, looking diagonally forward-left across the truck cab. In the near left foreground, the dump truck driver occupies the left seat, seen in profile facing forward. Just below the center of the frame, the steering wheel sits directly in front of the driver. The broad windshield spans the center and right background, facing forward to reveal the road ahead. No mirrors or reflective surfaces are present in the scene.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대와 앞유리 너머 도로를 포함한 운전석 구도 속 덤프트럭 운전사가 졸음으로 눈꺼풀을 늘어뜨린 모습.\n\nLOCATION (lock): Inside the dump truck's driving cab, at the driver's seat behind the steering wheel. Daylight enters through the windshield overlooking the road. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly controlled daytime ambient light preserves readable eyelids and the road beyond without introducing an unsupported cabin light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dump truck continues autonomously along the daytime road with its bed full of scrap. Charlie remains in the load, an old net-entangled robot with a worn UBIC chest logo.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "운전자의 시선이 아래가 아닌 앞유리 너머 전방 도로를 뚜렷하게 향하고 있음.",
    "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 측에서 촬영된 구도로 설계도의 카메라 위치 및 뷰와 일치함.",
    "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리와 전방 도로 풍경.",
    "hard_violations": [],
    "physics": "좌석에 앉아 양손으로 스티어링 휠을 단단히 쥐고 능동적으로 조작하는 자세임."
   },
   {
    "label": "B",
    "direction": "운전자의 고개가 아래로 향하고 눈꺼풀이 무겁게 감겨 스티어링 휠 쪽으로 시선(또는 얼굴)이 떨어져 있음.",
    "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 쪽에서 운전자와 앞유리를 바라보는 구도로 설계도와 정확히 일치함.",
    "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리 및 너머로 보이는 낮 시간대의 도로.",
    "hard_violations": [],
    "physics": "좌석에 체중을 싣고 앉아 있으며, 양손은 스티어링 휠 위에 가볍게 얹혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "조수석 구도에서 눈꺼풀이 처지고 고개가 숙여진 졸음운전의 모습을 지시사항에 맞게 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 운전자가 눈을 뜨고 전방을 주시하며 적극적으로 조향하는 모습이어서 '졸음으로 눈꺼풀을 늘어뜨린 모습'이라는 핵심 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "운전자의 고개가 아래로 향하고 눈꺼풀이 무겁게 감겨 스티어링 휠 쪽으로 시선(또는 얼굴)이 떨어져 있음.",
        "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 쪽에서 운전자와 앞유리를 바라보는 구도로 설계도와 정확히 일치함.",
        "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리 및 너머로 보이는 낮 시간대의 도로.",
        "hard_violations": [],
        "physics": "좌석에 체중을 싣고 앉아 있으며, 양손은 스티어링 휠 위에 가볍게 얹혀 있음."
       },
       {
        "label": "A",
        "direction": "운전자의 시선이 아래가 아닌 앞유리 너머 전방 도로를 뚜렷하게 향하고 있음.",
        "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 측에서 촬영된 구도로 설계도의 카메라 위치 및 뷰와 일치함.",
        "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리와 전방 도로 풍경.",
        "hard_violations": [],
        "physics": "좌석에 앉아 양손으로 스티어링 휠을 단단히 쥐고 능동적으로 조작하는 자세임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "조수석 구도에서 눈꺼풀이 처지고 고개가 숙여진 졸음운전의 모습을 지시사항에 맞게 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 운전자가 눈을 뜨고 전방을 주시하며 적극적으로 조향하는 모습이어서 '졸음으로 눈꺼풀을 늘어뜨린 모습'이라는 핵심 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "운전자의 고개가 아래로 향하고 눈꺼풀이 무겁게 감겨 스티어링 휠 쪽으로 시선(또는 얼굴)이 떨어져 있음.",
        "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 쪽에서 운전자와 앞유리를 바라보는 구도로 설계도와 정확히 일치함.",
        "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리 및 너머로 보이는 낮 시간대의 도로.",
        "hard_violations": [],
        "physics": "좌석에 체중을 싣고 앉아 있으며, 양손은 스티어링 휠 위에 가볍게 얹혀 있음."
       },
       {
        "label": "A",
        "direction": "운전자의 시선이 아래가 아닌 앞유리 너머 전방 도로를 뚜렷하게 향하고 있음.",
        "built_space": "트럭 운전석 내부, 좌핸들 구조이며 조수석 측에서 촬영된 구도로 설계도의 카메라 위치 및 뷰와 일치함.",
        "entities": "덤프트럭 운전사(한국인 남성), 스티어링 휠, 앞유리와 전방 도로 풍경.",
        "hard_violations": [],
        "physics": "좌석에 앉아 양손으로 스티어링 휠을 단단히 쥐고 능동적으로 조작하는 자세임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 운전석 구도는 맞지만, 눈이 거의 완전히 감기고 고개가 비교적 들려 있어 막 꾸벅이며 운전대로 시선이 떨어지는 순간과 다르며 양손으로 직접 조향하는 인상이 강하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 운전자·하단 운전대·오른쪽 전방 도로의 구도와 낮게 처진 눈꺼풀, 운전대 쪽으로 떨어지는 시선이 더 정확하지만 양손의 조향 자세는 자율주행 연출에 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "운전자의 얼굴은 오른쪽 전방을 향하며 눈은 거의 완전히 감겨 실제 시선의 목표를 확인하기 어렵다. 턱이 크게 내려가지 않아 운전대로 시선과 고개가 떨어지는 동작은 약하다. 앞유리 너머 차량들은 도로를 따라 전방으로 향한다.",
        "built_space": "조수석 쪽의 앉은 눈높이에서 운전자의 오른쪽 옆모습과 상체를 보는 구도다. 운전자는 왼쪽 좌석에 있고, 운전대 하나가 하단 중앙, 앞유리 하나와 전방 도로가 오른쪽에 있다. 계기판 한 조와 기둥 손잡이 하나, 왼쪽 측창 및 외부 거울이 보이며 운전대와 계기판은 운전석 쪽으로 배치되어 있다. 도면의 핵심 배치와 일치하고 중복된 운전 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "운전자 한 명은 중년 동아시아계 남성으로 보이고, 때 묻은 작업복과 녹색 안전조끼를 착용했다. 인물 외모나 의상을 고정하는 참조는 없다. 대형 트럭 운전실, 운전대와 낮의 도로는 확인되지만 이 실내만으로 덤프트럭 적재함 형태까지 확인할 수는 없다. 고철과 찰리, 그물 및 UBIC 표식은 지정된 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [],
        "physics": "몸은 운전석과 등받이에 지지되어 있고 두 손은 운전대 양쪽을 실제로 잡고 있다. 운전대는 조향축과 대시보드에 연결되며 도로의 차량도 노면 위에 있다. 부유하거나 지지 없는 대상은 없지만, 양팔과 손의 자세는 자율주행 중 졸기보다 직접 조향하는 행동으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "운전자의 얼굴은 오른쪽으로 향하고 낮게 열린 눈의 시선은 전방 도로보다 아래쪽 운전대 부근으로 떨어진다. 카메라를 응시하지 않는다. 고개 숙임은 미약하지만 처진 눈꺼풀과 아래로 향한 시선은 요구한 졸음 순간에 더 가깝다. 앞유리 밖 차량들은 전방 도로를 따라 향한다.",
        "built_space": "운전자는 왼쪽 좌석에, 운전대 하나는 하단 중앙에, 앞유리 하나와 도로는 오른쪽에 놓인다. 조수석 쪽에서 어깨 뒤편에 가까운 앉은 눈높이로 본 미디엄 구도다. 계기판 한 조, 기둥 손잡이 하나, 측창과 좌석 등받이가 보인다. 운전대와 계기판의 방향 및 운전자 위치는 서로 맞고, 중복 설비나 불가능한 반사는 없다.",
        "entities": "중년 동아시아계 남성 운전자 한 명이 보이며 어두운 작업복에 반사띠가 있다. 자연스러운 눈과 낮게 처진 눈꺼풀이 확인된다. 운전실과 운전대, 낮의 전방 도로가 포함된다. 외모·의상 참조가 없으므로 해당 세부는 고정 조건과 충돌하지 않는다. 적재함의 고철과 찰리, 그물 및 UBIC 표식은 이 실내 구도에서 보이지 않는 것이 자연스럽다.",
        "hard_violations": [],
        "physics": "운전자의 몸은 좌석에 지지되고 등 뒤로 등받이가 보인다. 양손은 운전대 테두리를 잡으며 팔과 손목의 연결도 가능하다. 운전대는 조향축에 고정되어 있고 외부 차량은 노면에 놓인다. 지지 없는 물체나 불가능한 자세는 없지만, 운전대를 확실히 움켜쥔 양손은 자율주행 중 비조향 상태보다 능동적인 운전을 암시한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 운전석 구도는 맞지만, 눈이 거의 완전히 감기고 고개가 비교적 들려 있어 막 꾸벅이며 운전대로 시선이 떨어지는 순간과 다르며 양손으로 직접 조향하는 인상이 강하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "왼쪽 운전자·하단 운전대·오른쪽 전방 도로의 구도와 낮게 처진 눈꺼풀, 운전대 쪽으로 떨어지는 시선이 더 정확하지만 양손의 조향 자세는 자율주행 연출에 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "운전자의 얼굴은 오른쪽 전방을 향하며 눈은 거의 완전히 감겨 실제 시선의 목표를 확인하기 어렵다. 턱이 크게 내려가지 않아 운전대로 시선과 고개가 떨어지는 동작은 약하다. 앞유리 너머 차량들은 도로를 따라 전방으로 향한다.",
        "built_space": "조수석 쪽의 앉은 눈높이에서 운전자의 오른쪽 옆모습과 상체를 보는 구도다. 운전자는 왼쪽 좌석에 있고, 운전대 하나가 하단 중앙, 앞유리 하나와 전방 도로가 오른쪽에 있다. 계기판 한 조와 기둥 손잡이 하나, 왼쪽 측창 및 외부 거울이 보이며 운전대와 계기판은 운전석 쪽으로 배치되어 있다. 도면의 핵심 배치와 일치하고 중복된 운전 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "운전자 한 명은 중년 동아시아계 남성으로 보이고, 때 묻은 작업복과 녹색 안전조끼를 착용했다. 인물 외모나 의상을 고정하는 참조는 없다. 대형 트럭 운전실, 운전대와 낮의 도로는 확인되지만 이 실내만으로 덤프트럭 적재함 형태까지 확인할 수는 없다. 고철과 찰리, 그물 및 UBIC 표식은 지정된 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [],
        "physics": "몸은 운전석과 등받이에 지지되어 있고 두 손은 운전대 양쪽을 실제로 잡고 있다. 운전대는 조향축과 대시보드에 연결되며 도로의 차량도 노면 위에 있다. 부유하거나 지지 없는 대상은 없지만, 양팔과 손의 자세는 자율주행 중 졸기보다 직접 조향하는 행동으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "운전자의 얼굴은 오른쪽으로 향하고 낮게 열린 눈의 시선은 전방 도로보다 아래쪽 운전대 부근으로 떨어진다. 카메라를 응시하지 않는다. 고개 숙임은 미약하지만 처진 눈꺼풀과 아래로 향한 시선은 요구한 졸음 순간에 더 가깝다. 앞유리 밖 차량들은 전방 도로를 따라 향한다.",
        "built_space": "운전자는 왼쪽 좌석에, 운전대 하나는 하단 중앙에, 앞유리 하나와 도로는 오른쪽에 놓인다. 조수석 쪽에서 어깨 뒤편에 가까운 앉은 눈높이로 본 미디엄 구도다. 계기판 한 조, 기둥 손잡이 하나, 측창과 좌석 등받이가 보인다. 운전대와 계기판의 방향 및 운전자 위치는 서로 맞고, 중복 설비나 불가능한 반사는 없다.",
        "entities": "중년 동아시아계 남성 운전자 한 명이 보이며 어두운 작업복에 반사띠가 있다. 자연스러운 눈과 낮게 처진 눈꺼풀이 확인된다. 운전실과 운전대, 낮의 전방 도로가 포함된다. 외모·의상 참조가 없으므로 해당 세부는 고정 조건과 충돌하지 않는다. 적재함의 고철과 찰리, 그물 및 UBIC 표식은 이 실내 구도에서 보이지 않는 것이 자연스럽다.",
        "hard_violations": [],
        "physics": "운전자의 몸은 좌석에 지지되고 등 뒤로 등받이가 보인다. 양손은 운전대 테두리를 잡으며 팔과 손목의 연결도 가능하다. 운전대는 조향축에 고정되어 있고 외부 차량은 노면에 놓인다. 지지 없는 물체나 불가능한 자세는 없지만, 운전대를 확실히 움켜쥔 양손은 자율주행 중 비조향 상태보다 능동적인 운전을 암시한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1857,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "조수석 구도에서 눈꺼풀이 처지고 고개가 숙여진 졸음운전의 모습을 지시사항에 맞게 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "구도와 배경은 적절하나, 운전자가 눈을 뜨고 전방을 주시하며 적극적으로 조향하는 모습이어서 '졸음으로 눈꺼풀을 늘어뜨린 모습'이라는 핵심 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S3sh3_confinedfp.png",
    "asset_id": null,
    "role": null
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07b9-80a4-7fb9-9653-989d69b7eca8",
  "confined_fp": {
   "base_key": "confinedfp::7afce4fe93a5",
   "apt_reason": "덤프트럭 운전석이라는 통제 장치가 집중된 밀폐 공간을 배경으로 하며, 운전석, 운전대, 앞유리의 위치와 방향이 정확히 배치되어야 이야기의 흐름이 깨지지 않기 때문입니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh2::confined_fp_apt": {
  "applies": true,
  "reason_ko": "샷은 자동차 내부 운전석과 조수석을 배경으로 하며, 페드로가 운전대에 앉고 이현우가 조수석에 앉는 정확한 위치 배치가 중요하므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "9b959cbb3f1ea70c"
 },
 "S4sh2::signage": {
  "fp": "e82c582afe528a25",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "구겨진 도면"
   }
  ],
  "dropped": []
 },
 "confinedfp::e8dbaebcb6ee": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_e8dbaebcb6ee.png",
  "place_text": "Inside the front cabin of a car parked in an affluent residential street, between the driver's position and the adjacent passenger seat. Daylight enters through the car windows.",
  "input_fingerprint": "c5ea670faf068e6e"
 },
 "S4sh2::confined_fp": {
  "reads": {
   "controls": "Steering wheel at the front-left seat. Center console located between the left and right seats.",
   "mirrors": "No mirrors are indicated in the diagram.",
   "camera": "Positioned behind the center console, slightly offset to the right, pointing diagonally forward and left toward the front seats.",
   "occupants": "페드로 (Pedro) in the front-left seat. 이현우 (Lee Hyun-woo) in the front-right seat."
  },
  "mismatches": [],
  "scene_description_en": "The camera views the car interior from behind the center console, slightly right of the center line, facing forward. On the left side of the screen, Pedro occupies the driver's seat, viewed from behind and facing forward. A steering wheel is positioned directly in front of him on the far left. On the right side of the screen, Lee Hyun-woo sits in the passenger seat, also viewed from behind and facing forward. In the center of the frame, resting between the two occupants above the center console, a crumpled drawing is visible.",
  "fixed": false,
  "input_fingerprint": "1c2f56385569d39a"
 },
 "era_assess::6413f88d36030a77": {
  "subjects": [],
  "subject_text": "페드로의 자동차 내부\n운전석과 조수석이 나란한 자동차 앞좌석 공간. 운전대와 대시보드가 있고 창문을 통해 주택가의 낮빛이 들어온다.",
  "identity": "canonical",
  "scope_id": "L08",
  "scope_role": "location_interior",
  "scope_sha": "e36d30f081dccc20"
 },
 "S4sh2": {
  "input_fingerprint": "7cccc3f536b69da0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대 앞의 페드로와 조수석의 이현우 사이에 구겨진 도면이 펼쳐진 차 안 구도.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, between the driver's position and the adjacent passenger seat. Daylight enters through the car windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from behind the front seats, slightly toward the passenger side and above seated eye level, tilting gently down toward the space between the two teenagers. Place 페드로 in rear three-quarter view at left with a hand on the wheel and 이현우 at right leaning over the crumpled drawing, which occupies a small area below center with its marked face visible. 이현우 studies the drawing while 페드로 remains oriented toward the street ahead, establishing a held entry composition before the pan toward his gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crumpled drawing (Open between the occupants, with a poorly drawn, difficult-to-read plan) — The marked face tilts upward toward the elevated camera, exposing the confused drawing rather than the blank reverse; used as A small central narrative anchor connects the two different seated postures; Steering wheel (Held by 페드로 in the parked car) — Seen obliquely beyond his arm on the left side of the composition; used as Identifies his position as the driver without competing with the drawing; Front seats (Occupied by 페드로 and 이현우) — Their rear edges frame the view into the space between the occupants; used as Provide restrained foreground framing and preserve the rear-seat camera position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and cool-neutral grading retain natural facial variation and readable drawing marks without stylized color accents.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car is parked in an affluent residential district in daylight. The house plan is crumpled and badly drawn. 페드로: Sits behind the steering wheel with an earphone fitted in his ear. 이현우: Sits in the passenger seat, holding and examining the crumpled plan.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera views the car interior from behind the center console, slightly right of the center line, facing forward. On the left side of the screen, Pedro occupies the driver's seat, viewed from behind and facing forward. A steering wheel is positioned directly in front of him on the far left. On the right side of the screen, Lee Hyun-woo sits in the passenger seat, also viewed from behind and facing forward. In the center of the frame, resting between the two occupants above the center console, a crumpled drawing is visible.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대 앞의 페드로와 조수석의 이현우 사이에 구겨진 도면이 펼쳐진 차 안 구도.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, between the driver's position and the adjacent passenger seat. Daylight enters through the car windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and cool-neutral grading retain natural facial variation and readable drawing marks without stylized color accents.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car is parked in an affluent residential district in daylight. The house plan is crumpled and badly drawn. 페드로: Sits behind the steering wheel with an earphone fitted in his ear. 이현우: Sits in the passenger seat, holding and examining the crumpled plan.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera views the car interior from behind the center console, slightly right of the center line, facing forward. On the left side of the screen, Pedro occupies the driver's seat, viewed from behind and facing forward. A steering wheel is positioned directly in front of him on the far left. On the right side of the screen, Lee Hyun-woo sits in the passenger seat, also viewed from behind and facing forward. In the center of the frame, resting between the two occupants above the center console, a crumpled drawing is visible.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대 앞의 페드로와 조수석의 이현우 사이에 구겨진 도면이 펼쳐진 차 안 구도.\n\nLOCATION (lock): Inside the front cabin of a car parked in an affluent residential street, between the driver's position and the adjacent passenger seat. Daylight enters through the car windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and cool-neutral grading retain natural facial variation and readable drawing marks without stylized color accents.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car is parked in an affluent residential district in daylight. The house plan is crumpled and badly drawn. 페드로: Sits behind the steering wheel with an earphone fitted in his ear. 이현우: Sits in the passenger seat, holding and examining the crumpled plan.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh2_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "페드로",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh2_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "페드로",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "페드로는 앞쪽 거리를 향하고 있고, 이현우는 시선을 아래로 향해 구겨진 도면을 집중해서 보고 있음.",
    "built_space": "자동차의 앞좌석 뒷공간. 스티어링 휠은 운전석(왼쪽), 센터 콘솔은 가운데 위치함. 카메라는 조수석 쪽으로 치우쳐 두 사람 사이를 촬영하고 있으며, 앞좌석의 양쪽 등받이가 화면 양옆 시야를 프레이밍하고 있음. 룸미러에는 차량 뒤편의 풍경이 올바르게 반사되어 보임.",
    "entities": "페드로(왼쪽)는 비니와 짙은 녹색 자켓을 입고 머리 왼편에 이어폰을 착용함. 이현우(오른쪽)는 헝클어진 검은 머리에 피와 흙먼지가 묻은 어두운 셔츠를 입고 있음. 이현우가 들고 있는 구겨진 종이에는 레퍼런스와 일치하는 구조의 평면도가 선명하게 그려져 있음.",
    "hard_violations": [
     "[gemini-pro] 광학적 불가능(Staging/Physics): 카메라는 조수석(오른쪽) 뒤편에 위치하고 페드로는 정면을 바라보고 있으나, 페드로의 머리 왼쪽 실루엣에 온전한 '왼쪽 귀(왼쪽 측면에서 본 형태)'가 노출됨. 우측 후방 카메라 시점에서는 오른쪽 귀와 우측 뒤통수가 보여야 하므로 이는 기하학적으로 불가능한 렌더링임."
    ],
    "physics": "이현우의 양손이 구겨진 도면의 양쪽 가장자리를 쥐고 무릎 위쪽 공간에서 물리적으로 자연스럽게 지탱하고 있음. 페드로는 좌석에 엉덩이를 붙이고 올바르게 착석해 있음."
   },
   {
    "label": "B",
    "direction": "페드로는 정면 거리를 향하고, 이현우는 시선을 아래로 내려 도면을 향하고 있음.",
    "built_space": "차량 실내 앞좌석 뒤편. 왼쪽에 스티어링 휠, 가운데에 센터 콘솔이 있음. 조수석 쪽으로 약간 치우친 카메라 시점으로 두 사람 사이의 공간을 비추고 있으며, 앞좌석 등받이가 프레임 양쪽을 형성함.",
    "entities": "페드로(왼쪽)는 비니와 녹색 자켓을 입고 왼쪽 귀에 이어폰을 낀 모습. 이현우(오른쪽)는 어두운 색 셔츠에 핏자국이 묻어 있음. 도면은 레퍼런스와 유사한 평면도임.",
    "hard_violations": [
     "[gemini-pro] 광학적 불가능(Staging/Physics): 카메라가 페드로의 우측 후방에 위치하고 페드로가 정면 거리를 향하고 있음에도, 인물의 머리 왼편 실루엣에 왼쪽 귀가 노출되는 광학적/해부학적으로 불가능한 시점 오류가 발생함."
    ],
    "physics": "이현우의 두 손이 도면의 좌우를 쥐고 지탱하고 있으며, 페드로는 운전석에 정상적으로 착석해 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "조수석 쪽에 위치한 카메라에서 페드로의 왼쪽 귀가 보이는 광학적/기하학적 오류(Hard Violation)가 치명적이나, 구겨진 도면의 재질감과 차량 내부의 영화적 조명 및 디테일은 매우 뛰어납니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일하게 카메라 위치상 보일 수 없는 왼쪽 귀가 묘사된 치명적 앵글 오류가 있으며, 스티어링 휠의 형태 왜곡과 전체적인 사실감이 A에 비해 다소 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 앞쪽 거리를 향하고 있고, 이현우는 시선을 아래로 향해 구겨진 도면을 집중해서 보고 있음.",
        "built_space": "자동차의 앞좌석 뒷공간. 스티어링 휠은 운전석(왼쪽), 센터 콘솔은 가운데 위치함. 카메라는 조수석 쪽으로 치우쳐 두 사람 사이를 촬영하고 있으며, 앞좌석의 양쪽 등받이가 화면 양옆 시야를 프레이밍하고 있음. 룸미러에는 차량 뒤편의 풍경이 올바르게 반사되어 보임.",
        "entities": "페드로(왼쪽)는 비니와 짙은 녹색 자켓을 입고 머리 왼편에 이어폰을 착용함. 이현우(오른쪽)는 헝클어진 검은 머리에 피와 흙먼지가 묻은 어두운 셔츠를 입고 있음. 이현우가 들고 있는 구겨진 종이에는 레퍼런스와 일치하는 구조의 평면도가 선명하게 그려져 있음.",
        "hard_violations": [
         "광학적 불가능(Staging/Physics): 카메라는 조수석(오른쪽) 뒤편에 위치하고 페드로는 정면을 바라보고 있으나, 페드로의 머리 왼쪽 실루엣에 온전한 '왼쪽 귀(왼쪽 측면에서 본 형태)'가 노출됨. 우측 후방 카메라 시점에서는 오른쪽 귀와 우측 뒤통수가 보여야 하므로 이는 기하학적으로 불가능한 렌더링임."
        ],
        "physics": "이현우의 양손이 구겨진 도면의 양쪽 가장자리를 쥐고 무릎 위쪽 공간에서 물리적으로 자연스럽게 지탱하고 있음. 페드로는 좌석에 엉덩이를 붙이고 올바르게 착석해 있음."
       },
       {
        "label": "B",
        "direction": "페드로는 정면 거리를 향하고, 이현우는 시선을 아래로 내려 도면을 향하고 있음.",
        "built_space": "차량 실내 앞좌석 뒤편. 왼쪽에 스티어링 휠, 가운데에 센터 콘솔이 있음. 조수석 쪽으로 약간 치우친 카메라 시점으로 두 사람 사이의 공간을 비추고 있으며, 앞좌석 등받이가 프레임 양쪽을 형성함.",
        "entities": "페드로(왼쪽)는 비니와 녹색 자켓을 입고 왼쪽 귀에 이어폰을 낀 모습. 이현우(오른쪽)는 어두운 색 셔츠에 핏자국이 묻어 있음. 도면은 레퍼런스와 유사한 평면도임.",
        "hard_violations": [
         "광학적 불가능(Staging/Physics): 카메라가 페드로의 우측 후방에 위치하고 페드로가 정면 거리를 향하고 있음에도, 인물의 머리 왼편 실루엣에 왼쪽 귀가 노출되는 광학적/해부학적으로 불가능한 시점 오류가 발생함."
        ],
        "physics": "이현우의 두 손이 도면의 좌우를 쥐고 지탱하고 있으며, 페드로는 운전석에 정상적으로 착석해 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "조수석 쪽에 위치한 카메라에서 페드로의 왼쪽 귀가 보이는 광학적/기하학적 오류(Hard Violation)가 치명적이나, 구겨진 도면의 재질감과 차량 내부의 영화적 조명 및 디테일은 매우 뛰어납니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일하게 카메라 위치상 보일 수 없는 왼쪽 귀가 묘사된 치명적 앵글 오류가 있으며, 스티어링 휠의 형태 왜곡과 전체적인 사실감이 A에 비해 다소 떨어집니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 앞쪽 거리를 향하고 있고, 이현우는 시선을 아래로 향해 구겨진 도면을 집중해서 보고 있음.",
        "built_space": "자동차의 앞좌석 뒷공간. 스티어링 휠은 운전석(왼쪽), 센터 콘솔은 가운데 위치함. 카메라는 조수석 쪽으로 치우쳐 두 사람 사이를 촬영하고 있으며, 앞좌석의 양쪽 등받이가 화면 양옆 시야를 프레이밍하고 있음. 룸미러에는 차량 뒤편의 풍경이 올바르게 반사되어 보임.",
        "entities": "페드로(왼쪽)는 비니와 짙은 녹색 자켓을 입고 머리 왼편에 이어폰을 착용함. 이현우(오른쪽)는 헝클어진 검은 머리에 피와 흙먼지가 묻은 어두운 셔츠를 입고 있음. 이현우가 들고 있는 구겨진 종이에는 레퍼런스와 일치하는 구조의 평면도가 선명하게 그려져 있음.",
        "hard_violations": [
         "광학적 불가능(Staging/Physics): 카메라는 조수석(오른쪽) 뒤편에 위치하고 페드로는 정면을 바라보고 있으나, 페드로의 머리 왼쪽 실루엣에 온전한 '왼쪽 귀(왼쪽 측면에서 본 형태)'가 노출됨. 우측 후방 카메라 시점에서는 오른쪽 귀와 우측 뒤통수가 보여야 하므로 이는 기하학적으로 불가능한 렌더링임."
        ],
        "physics": "이현우의 양손이 구겨진 도면의 양쪽 가장자리를 쥐고 무릎 위쪽 공간에서 물리적으로 자연스럽게 지탱하고 있음. 페드로는 좌석에 엉덩이를 붙이고 올바르게 착석해 있음."
       },
       {
        "label": "B",
        "direction": "페드로는 정면 거리를 향하고, 이현우는 시선을 아래로 내려 도면을 향하고 있음.",
        "built_space": "차량 실내 앞좌석 뒤편. 왼쪽에 스티어링 휠, 가운데에 센터 콘솔이 있음. 조수석 쪽으로 약간 치우친 카메라 시점으로 두 사람 사이의 공간을 비추고 있으며, 앞좌석 등받이가 프레임 양쪽을 형성함.",
        "entities": "페드로(왼쪽)는 비니와 녹색 자켓을 입고 왼쪽 귀에 이어폰을 낀 모습. 이현우(오른쪽)는 어두운 색 셔츠에 핏자국이 묻어 있음. 도면은 레퍼런스와 유사한 평면도임.",
        "hard_violations": [
         "광학적 불가능(Staging/Physics): 카메라가 페드로의 우측 후방에 위치하고 페드로가 정면 거리를 향하고 있음에도, 인물의 머리 왼편 실루엣에 왼쪽 귀가 노출되는 광학적/해부학적으로 불가능한 시점 오류가 발생함."
        ],
        "physics": "이현우의 두 손이 도면의 좌우를 쥐고 지탱하고 있으며, 페드로는 운전석에 정상적으로 착석해 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 사람의 시선 대비와 뒷좌석 구도는 맞지만, 도면이 중앙 오른쪽에 크게 들려 있고 페드로가 운전대를 잡는 필수 동작이 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "중앙 아래의 비교적 작은 도면과 두 좌석 사이를 내려다보는 구도가 더 충실하지만, 운전대를 잡은 손과 이현우의 인이어 무전기는 확인되지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 페드로의 머리는 앞유리 너머 도로를 향하고, 오른쪽 이현우는 고개를 숙여 손에 든 도면을 본다. 도면의 표시 면은 이현우와 뒤쪽 카메라 양쪽에서 볼 수 있는 방향이지만, 위로 눕혀 펼치기보다 가슴 앞에 세워 들었다.",
        "built_space": "앞좌석 두 개가 전경 양쪽에 있고, 왼쪽 운전대 하나, 중앙 화면 하나, 변속기 하나, 중앙 콘솔 하나와 실내 후사경 하나가 보인다. 두 사람은 각각 운전석과 조수석에 자리한다. 카메라는 좌석 뒤에서 약간 조수석 쪽에 있으며, 도면은 중앙보다 오른쪽에서 상당한 면적을 차지한다. 창밖에는 낮의 담장과 수목이 있는 주택가가 보인다. 별도의 장소 사진은 제공되지 않아 정확한 장소 일치는 확인할 수 없다. 후사경의 어두운 부분상만으로 불가능한 반사라고 단정할 근거는 없다.",
        "entities": "인물은 두 명뿐이다. 페드로의 검은 비니와 낡은 녹색 계열 재킷, 보이는 귀의 이어폰은 참조와 부합한다. 이현우는 짧고 헝클어진 검은 머리와 얼룩진 어두운 셔츠를 착용하지만, 드러난 귀에서 인이어 무전기는 보이지 않는다. 두 사람의 얼굴 대부분이 가려져 정확한 얼굴 동일성이나 민족적 외양은 판단하기 어렵다. 종이는 실제로 구겨져 있고 표면에 서투른 평면도 선이 있다.",
        "hard_violations": [],
        "physics": "두 사람의 몸은 각각 좌석에 지지되고, 도면은 이현우의 양손이 양쪽 가장자리를 잡아 지지한다. 손과 종이의 접촉 및 종이의 굴곡은 자연스럽다. 페드로의 손은 운전대에서 식별되지 않아 요구된 접촉 동작을 확인할 수 없다. 지지 없이 떠 있는 물체나 인체는 없다."
       },
       {
        "label": "B",
        "direction": "페드로는 도면 쪽으로 돌아보지 않고 차량 전방을 향한다. 이현우의 고개와 시선은 양손 사이 도면으로 내려간다. 도면의 표시 면은 이현우와 후방 카메라를 향하며, 카메라에서도 선을 읽을 수 있다. 다만 도면은 위를 향해 넓게 눕기보다 비스듬히 세워져 있다.",
        "built_space": "전경에 앞좌석 등받이 두 개, 왼쪽에 운전대 하나, 중앙에 화면 하나와 변속 조작부 하나, 두 사람 사이에 콘솔 하나, 상단에 실내 후사경 하나가 보인다. 운전자와 동승자의 좌우 배치가 도식과 맞는다. 좌석 뒤의 다소 높은 카메라가 두 사람 사이를 내려다보며, 도면은 중앙 아래에 A보다 작게 놓인다. 다만 전경 등받이가 다소 넓은 면적을 차지한다. 창밖은 낮의 고급 주택가로 읽힌다. 후사경에는 어두운 실내 형태가 보이며 명백한 광학적 모순은 없다.",
        "entities": "두 인물 외에 추가 인물은 없다. 페드로는 참조의 검은 비니와 낡은 녹색 계열 재킷을 착용하고 귀에 작은 이어폰이 보인다. 이현우의 헝클어진 검은 머리, 마른 체형과 피·먼지 얼룩이 있는 어두운 셔츠는 요구에 부합하지만 인이어 무전기는 식별되지 않는다. 후면 위주라 얼굴 동일성은 충분히 검증할 수 없다. 도면은 구겨진 종이에 혼란스러운 방과 통로 형태가 그려진 실물 소품이다.",
        "hard_violations": [],
        "physics": "두 사람은 각자의 좌석에 앉아 있고, 이현우는 양손으로 종이를 잡고 몸을 앞으로 기울인다. 팔과 종이에 분명한 지지가 있으며 공중에 뜬 요소는 없다. 보이는 운전대 테두리에는 페드로의 손이 없어, 운전대를 잡고 있는 순간이라는 필수 동작이 드러나지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 사람의 시선 대비와 뒷좌석 구도는 맞지만, 도면이 중앙 오른쪽에 크게 들려 있고 페드로가 운전대를 잡는 필수 동작이 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "중앙 아래의 비교적 작은 도면과 두 좌석 사이를 내려다보는 구도가 더 충실하지만, 운전대를 잡은 손과 이현우의 인이어 무전기는 확인되지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 페드로의 머리는 앞유리 너머 도로를 향하고, 오른쪽 이현우는 고개를 숙여 손에 든 도면을 본다. 도면의 표시 면은 이현우와 뒤쪽 카메라 양쪽에서 볼 수 있는 방향이지만, 위로 눕혀 펼치기보다 가슴 앞에 세워 들었다.",
        "built_space": "앞좌석 두 개가 전경 양쪽에 있고, 왼쪽 운전대 하나, 중앙 화면 하나, 변속기 하나, 중앙 콘솔 하나와 실내 후사경 하나가 보인다. 두 사람은 각각 운전석과 조수석에 자리한다. 카메라는 좌석 뒤에서 약간 조수석 쪽에 있으며, 도면은 중앙보다 오른쪽에서 상당한 면적을 차지한다. 창밖에는 낮의 담장과 수목이 있는 주택가가 보인다. 별도의 장소 사진은 제공되지 않아 정확한 장소 일치는 확인할 수 없다. 후사경의 어두운 부분상만으로 불가능한 반사라고 단정할 근거는 없다.",
        "entities": "인물은 두 명뿐이다. 페드로의 검은 비니와 낡은 녹색 계열 재킷, 보이는 귀의 이어폰은 참조와 부합한다. 이현우는 짧고 헝클어진 검은 머리와 얼룩진 어두운 셔츠를 착용하지만, 드러난 귀에서 인이어 무전기는 보이지 않는다. 두 사람의 얼굴 대부분이 가려져 정확한 얼굴 동일성이나 민족적 외양은 판단하기 어렵다. 종이는 실제로 구겨져 있고 표면에 서투른 평면도 선이 있다.",
        "hard_violations": [],
        "physics": "두 사람의 몸은 각각 좌석에 지지되고, 도면은 이현우의 양손이 양쪽 가장자리를 잡아 지지한다. 손과 종이의 접촉 및 종이의 굴곡은 자연스럽다. 페드로의 손은 운전대에서 식별되지 않아 요구된 접촉 동작을 확인할 수 없다. 지지 없이 떠 있는 물체나 인체는 없다."
       },
       {
        "label": "A",
        "direction": "페드로는 도면 쪽으로 돌아보지 않고 차량 전방을 향한다. 이현우의 고개와 시선은 양손 사이 도면으로 내려간다. 도면의 표시 면은 이현우와 후방 카메라를 향하며, 카메라에서도 선을 읽을 수 있다. 다만 도면은 위를 향해 넓게 눕기보다 비스듬히 세워져 있다.",
        "built_space": "전경에 앞좌석 등받이 두 개, 왼쪽에 운전대 하나, 중앙에 화면 하나와 변속 조작부 하나, 두 사람 사이에 콘솔 하나, 상단에 실내 후사경 하나가 보인다. 운전자와 동승자의 좌우 배치가 도식과 맞는다. 좌석 뒤의 다소 높은 카메라가 두 사람 사이를 내려다보며, 도면은 중앙 아래에 A보다 작게 놓인다. 다만 전경 등받이가 다소 넓은 면적을 차지한다. 창밖은 낮의 고급 주택가로 읽힌다. 후사경에는 어두운 실내 형태가 보이며 명백한 광학적 모순은 없다.",
        "entities": "두 인물 외에 추가 인물은 없다. 페드로는 참조의 검은 비니와 낡은 녹색 계열 재킷을 착용하고 귀에 작은 이어폰이 보인다. 이현우의 헝클어진 검은 머리, 마른 체형과 피·먼지 얼룩이 있는 어두운 셔츠는 요구에 부합하지만 인이어 무전기는 식별되지 않는다. 후면 위주라 얼굴 동일성은 충분히 검증할 수 없다. 도면은 구겨진 종이에 혼란스러운 방과 통로 형태가 그려진 실물 소품이다.",
        "hard_violations": [],
        "physics": "두 사람은 각자의 좌석에 앉아 있고, 이현우는 양손으로 종이를 잡고 몸을 앞으로 기울인다. 팔과 종이에 분명한 지지가 있으며 공중에 뜬 요소는 없다. 보이는 운전대 테두리에는 페드로의 손이 없어, 운전대를 잡고 있는 순간이라는 필수 동작이 드러나지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.607
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.357
   },
   "violations": {
    "A": [
     "[gemini-pro] 광학적 불가능(Staging/Physics): 카메라는 조수석(오른쪽) 뒤편에 위치하고 페드로는 정면을 바라보고 있으나, 페드로의 머리 왼쪽 실루엣에 온전한 '왼쪽 귀(왼쪽 측면에서 본 형태)'가 노출됨. 우측 후방 카메라 시점에서는 오른쪽 귀와 우측 뒤통수가 보여야 하므로 이는 기하학적으로 불가능한 렌더링임."
    ],
    "B": [
     "[gemini-pro] 광학적 불가능(Staging/Physics): 카메라가 페드로의 우측 후방에 위치하고 페드로가 정면 거리를 향하고 있음에도, 인물의 머리 왼편 실루엣에 왼쪽 귀가 노출되는 광학적/해부학적으로 불가능한 시점 오류가 발생함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1357
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "조수석 쪽에 위치한 카메라에서 페드로의 왼쪽 귀가 보이는 광학적/기하학적 오류(Hard Violation)가 치명적이나, 구겨진 도면의 재질감과 차량 내부의 영화적 조명 및 디테일은 매우 뛰어납니다.  ★위반: [gemini-pro] 광학적 불가능(Staging/Physics): 카메라는 조수석(오른쪽) 뒤편에 위치하고 페드로는 정면을 바라보고 있으나, 페드로의 머리 왼쪽 실루엣에 온전한 '왼쪽 귀(왼쪽 측면에서 본 형태)'가 노출됨. 우측 후방 카메라 시점에서는 오른쪽 귀와 우측 뒤통수가 보여야 하므로 이는 기하학적으로 불가능한 렌더링임."
   },
   {
    "label": "B",
    "score": 1357,
    "verdict_ko": "A와 동일하게 카메라 위치상 보일 수 없는 왼쪽 귀가 묘사된 치명적 앵글 오류가 있으며, 스티어링 휠의 형태 왜곡과 전체적인 사실감이 A에 비해 다소 떨어집니다.  ★위반: [gemini-pro] 광학적 불가능(Staging/Physics): 카메라가 페드로의 우측 후방에 위치하고 페드로가 정면 거리를 향하고 있음에도, 인물의 머리 왼편 실루엣에 왼쪽 귀가 노출되는 광학적/해부학적으로 불가능한 시점 오류가 발생함."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh2_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "페드로",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07c4-23dc-7ad1-9cc5-65bf4488b750",
  "confined_fp": {
   "base_key": "confinedfp::e8dbaebcb6ee",
   "apt_reason": "샷은 자동차 내부 운전석과 조수석을 배경으로 하며, 페드로가 운전대에 앉고 이현우가 조수석에 앉는 정확한 위치 배치가 중요하므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh5::confined_fp_apt": {
  "applies": true,
  "reason_ko": "차량 내부 운전석이라는 명확한 조작 공간을 배경으로 하며 인물이 차창 너머 특정 방향을 가리키고 있으므로 차량 내 좌석의 위치와 가리키는 방향이 시각적으로 정확히 일치해야 합니다.",
  "input_fingerprint": "d4e7edf1b3690c21"
 },
 "S4sh5::signage": {
  "fp": "fc956bce7e125fc8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::03bcfbed1a4a": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_03bcfbed1a4a.png",
  "place_text": "At the driver's position inside a parked car in an affluent residential neighborhood. Daylight and the house across the street are visible through the side window.",
  "input_fingerprint": "0b6fac30edcf3f79"
 },
 "S4sh5::confined_fp": {
  "reads": {
   "controls": "Steering wheel located at the front-left seat.",
   "mirrors": "No mirrors are shown in the diagram.",
   "camera": "Positioned in the rear center seat, pointing forward and diagonally to the left towards the driver's seat and front-left window.",
   "occupants": "Pedro (페드로) is in the front-left driver's seat."
  },
  "mismatches": [
   "The text states Pedro's face is at the left edge and the indicated house/window is in the upper-right background. However, the diagram shows the camera aiming from the rear center toward the front-left window, which places the window on the left side of the frame and Pedro on the right."
  ],
  "scene_description_en": "The camera is stationed in the rear center of the vehicle, angled forward and to the left. Pedro occupies the front-left driver's seat in the midground, positioned on the right side of the camera's frame. A steering wheel sits directly in front of him. He is turned to his left, extending his left arm to point out the adjacent side window. The front-left window occupies the left background of the frame, showing the exterior view. No mirrors or other occupants are visible.",
  "fixed": true,
  "input_fingerprint": "1a32f2560929ae9c"
 },
 "S4sh5": {
  "input_fingerprint": "42737b55e6fce5b6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석의 페드로가 차창 너머 길 건너 고급주택을 손으로 가리키는 구도.\n\nLOCATION (lock): At the driver's position inside a parked car in an affluent residential neighborhood. Daylight and the house across the street are visible through the side window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Keep the same elevated rear-seat position and pan obliquely past 페드로's shoulder, directly observing his pointing hand and the house through the side window without changing camera distance. Let his turned three-quarter face remain at the left edge, his extended hand cross the lower middle, and the house sit in the upper-right background along the gesture; 이현우 stays outside this reframing. 페드로 looks toward the indicated house across the street, and the redirected viewing axis is the sole emphasized change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: House indicated by 페드로's pointing hand in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Car side window (Providing a view of the house across the street) — Viewed obliquely from behind the front seats, with the indicated house visible beyond; used as Frames the destination without making reflections the subject; Upscale house across the street (The house 페드로 is indicating) — Its street-facing exterior is seen obliquely beyond his pointing hand; used as Provides the background endpoint of the gesture in the upper-right portion of the image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding shot's subdued daylight and neutral contrast so the gesture, not a lighting shift, redirects attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 페드로 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the car's steering wheel, dashboard, window fittings, and daylight appearance from the reference. Exclude furnishings from the house across the street; they do not belong inside the car.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked opposite the targeted luxury house. The crumpled, poorly drawn house plan remains part of the car-interior setup. 페드로: Remains in the driver's seat with his earphone in place and one hand raised to point through the window.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the rear center of the vehicle, angled forward and to the left. Pedro occupies the front-left driver's seat in the midground, positioned on the right side of the camera's frame. A steering wheel sits directly in front of him. He is turned to his left, extending his left arm to point out the adjacent side window. The front-left window occupies the left background of the frame, showing the exterior view. No mirrors or other occupants are visible.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석의 페드로가 차창 너머 길 건너 고급주택을 손으로 가리키는 구도.\n\nLOCATION (lock): At the driver's position inside a parked car in an affluent residential neighborhood. Daylight and the house across the street are visible through the side window. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding shot's subdued daylight and neutral contrast so the gesture, not a lighting shift, redirects attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked opposite the targeted luxury house. The crumpled, poorly drawn house plan remains part of the car-interior setup. 페드로: Remains in the driver's seat with his earphone in place and one hand raised to point through the window.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the rear center of the vehicle, angled forward and to the left. Pedro occupies the front-left driver's seat in the midground, positioned on the right side of the camera's frame. A steering wheel sits directly in front of him. He is turned to his left, extending his left arm to point out the adjacent side window. The front-left window occupies the left background of the frame, showing the exterior view. No mirrors or other occupants are visible.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석의 페드로가 차창 너머 길 건너 고급주택을 손으로 가리키는 구도.\n\nLOCATION (lock): At the driver's position inside a parked car in an affluent residential neighborhood. Daylight and the house across the street are visible through the side window. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding shot's subdued daylight and neutral contrast so the gesture, not a lighting shift, redirects attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked opposite the targeted luxury house. The crumpled, poorly drawn house plan remains part of the car-interior setup. 페드로: Remains in the driver's seat with his earphone in place and one hand raised to point through the window.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh5_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "페드로",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh5_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "페드로",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "페드로의 왼손은 좌측 창밖의 집을 가리키나, 시선은 집이 아닌 카메라 렌즈를 정면으로 응시함.",
    "built_space": "차량 내부 좌핸들 운전석에 앉아 있으며, 카메라는 뒷좌석에서 좌측 창문을 향해 다이어그램의 공간 배치와 일치함.",
    "entities": "페드로(비니, 녹색 자켓, 이어폰)의 외형은 일치함. 창밖의 고급 주택이 보이나 도면은 생략됨.",
    "hard_violations": [],
    "physics": "시트에 착석한 자세와 몸통에 지지된 팔의 형태가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "페드로의 시선과 가리키는 왼손 모두 우측 창문 밖의 주택을 향하고 있음.",
    "built_space": "페드로가 좌측 시트에 앉아 있으나 우측에 스티어링 휠이 배치되어, 참조와 다른 우핸들 차량으로 공간이 변형됨.",
    "entities": "페드로의 인상착의가 일치하며, 화면 좌측 하단에 구겨진 도면이 있고 창밖에 집이 보임.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (우측 운전대를 잡고 있는 정체불명의 오른손)",
     "[gemini-pro] 공간 구조 위반 (지시된 좌핸들 배치를 무시하고 우핸들로 변형)"
    ],
    "physics": "화면 우측 하단에서 스티어링 휠을 잡은 손은 좌측에 앉은 페드로의 신체와 연결될 수 없는 불가능한 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "목표물을 보지 않고 카메라를 응시하는 연출 오류가 있으나, 참조된 좌핸들 차량의 구조를 유지하였고 해부학적 위반이 없어 선정됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "시선과 가리키는 방향은 정확하나, 차량이 우핸들로 왜곡되고 운전대를 잡은 불가능한 손이 나타나는 치명적 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로의 왼손은 좌측 창밖의 집을 가리키나, 시선은 집이 아닌 카메라 렌즈를 정면으로 응시함.",
        "built_space": "차량 내부 좌핸들 운전석에 앉아 있으며, 카메라는 뒷좌석에서 좌측 창문을 향해 다이어그램의 공간 배치와 일치함.",
        "entities": "페드로(비니, 녹색 자켓, 이어폰)의 외형은 일치함. 창밖의 고급 주택이 보이나 도면은 생략됨.",
        "hard_violations": [],
        "physics": "시트에 착석한 자세와 몸통에 지지된 팔의 형태가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "페드로의 시선과 가리키는 왼손 모두 우측 창문 밖의 주택을 향하고 있음.",
        "built_space": "페드로가 좌측 시트에 앉아 있으나 우측에 스티어링 휠이 배치되어, 참조와 다른 우핸들 차량으로 공간이 변형됨.",
        "entities": "페드로의 인상착의가 일치하며, 화면 좌측 하단에 구겨진 도면이 있고 창밖에 집이 보임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (우측 운전대를 잡고 있는 정체불명의 오른손)",
         "공간 구조 위반 (지시된 좌핸들 배치를 무시하고 우핸들로 변형)"
        ],
        "physics": "화면 우측 하단에서 스티어링 휠을 잡은 손은 좌측에 앉은 페드로의 신체와 연결될 수 없는 불가능한 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "목표물을 보지 않고 카메라를 응시하는 연출 오류가 있으나, 참조된 좌핸들 차량의 구조를 유지하였고 해부학적 위반이 없어 선정됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "시선과 가리키는 방향은 정확하나, 차량이 우핸들로 왜곡되고 운전대를 잡은 불가능한 손이 나타나는 치명적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "페드로의 왼손은 좌측 창밖의 집을 가리키나, 시선은 집이 아닌 카메라 렌즈를 정면으로 응시함.",
        "built_space": "차량 내부 좌핸들 운전석에 앉아 있으며, 카메라는 뒷좌석에서 좌측 창문을 향해 다이어그램의 공간 배치와 일치함.",
        "entities": "페드로(비니, 녹색 자켓, 이어폰)의 외형은 일치함. 창밖의 고급 주택이 보이나 도면은 생략됨.",
        "hard_violations": [],
        "physics": "시트에 착석한 자세와 몸통에 지지된 팔의 형태가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "페드로의 시선과 가리키는 왼손 모두 우측 창문 밖의 주택을 향하고 있음.",
        "built_space": "페드로가 좌측 시트에 앉아 있으나 우측에 스티어링 휠이 배치되어, 참조와 다른 우핸들 차량으로 공간이 변형됨.",
        "entities": "페드로의 인상착의가 일치하며, 화면 좌측 하단에 구겨진 도면이 있고 창밖에 집이 보임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (우측 운전대를 잡고 있는 정체불명의 오른손)",
         "공간 구조 위반 (지시된 좌핸들 배치를 무시하고 우핸들로 변형)"
        ],
        "physics": "화면 우측 하단에서 스티어링 휠을 잡은 손은 좌측에 앉은 페드로의 신체와 연결될 수 없는 불가능한 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "시선과 손끝이 오른쪽 위 주택으로 이어져 핵심 동작과 좌우 배치는 더 충실하지만, 지정된 높은 뒷좌석 시점보다 앞쪽에서 본 구도이고 얼굴과 손의 위치도 정확하지 않다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "뒷좌석에서 어깨 너머로 보는 시점은 가깝지만, 인물과 주택의 좌우 배치가 반대이고 페드로가 집이 아닌 카메라를 바라봐 핵심 시선 지시를 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 창밖 오른쪽 위의 주택 쪽으로 고개를 돌리고 있다. 뻗은 검지는 그 주택의 담장 위 수목과 외벽 방향을 가리켜, 지시 대상이 해당 주택으로 읽힌다. 다른 손은 운전대를 잡고 있다.",
        "built_space": "운전대 하나, 운전석 등받이와 머리받침 하나, 측면 창 하나, 사이드미러 하나, 문손잡이 하나와 대시보드 일부가 보인다. 인물은 운전대 뒤 운전석에 있고 주택은 도로 너머 오른쪽 위에 있다. 다만 가슴 앞면과 운전대를 잡은 손이 크게 보이는 관찰각은 지정된 높은 뒷좌석보다 앞좌석 동승자 쪽에 가깝다. 얼굴은 왼쪽 가장자리보다 안쪽이며 지시 손도 하단 중앙보다 높다. 거울 속 가로수와 담장에는 명백한 광학적 모순이 없다. 이전 장소 실사 참조는 제공되지 않아 차량 마감의 정확한 연속성은 확인할 수 없다.",
        "entities": "인물은 한 명뿐이며 이현우는 없다. 검은 비니, 그 아래 짙은 머리, 낡은 올리브색 집업과 회색 상의가 페드로 참조와 부합한다. 보이는 얼굴 일부는 젊은 남성으로 읽히지만 정확한 나이와 혼혈 배경은 단정할 수 없다. 귀에는 검은 유선 이어폰이 있고 왼쪽 아래에는 구겨진 주택 도면 일부가 보인다. 현대적인 고급주택과 낮 풍경이 있으며 바지와 신발은 평가할 만큼 보이지 않는다. 자막이나 도해의 표식은 없다.",
        "hard_violations": [],
        "physics": "몸통은 운전석에 앉아 지지되고, 지시하는 팔은 어깨와 팔꿈치에서 자연스럽게 뻗어 있다. 다른 손은 운전대 테두리를 실제로 감싼다. 구겨진 종이는 화면 아래의 어두운 천 표면에 얹혀 있다. 지지 없이 뜬 신체나 물체, 명백히 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "검지는 왼쪽 창밖 주택의 대문과 담장 방향을 가리킨다. 그러나 얼굴을 뒤로 돌린 페드로의 눈은 주택이 아니라 카메라를 향한다. 따라서 손짓의 대상은 맞지만 시선의 대상은 틀리다.",
        "built_space": "운전대 하나, 오른쪽 전경의 운전석 등받이와 머리받침 하나, 왼쪽 아래 다른 좌석 일부, 측면 창 하나, 사이드미러 하나, 문손잡이 하나와 창문 조작부가 보인다. 운전자는 운전대 뒤에 앉아 있고 카메라는 좌석 뒤에서 어깨 너머를 본다. 하지만 얼굴은 오른쪽, 주택은 왼쪽 위라 요구된 좌우 배치와 반대이며 높은 관찰각도 두드러지지 않는다. 사이드미러의 도로와 가로수 반사는 가능한 범위다. 이전 장소 실사 참조가 없어 차량 고정 부품의 정확한 일치는 확인할 수 없다.",
        "entities": "젊은 남성 한 명만 등장하며 검은 비니, 짙은 머리, 올리브색의 낡은 집업과 어두운 바지가 페드로 참조에 대체로 맞는다. 얼굴에는 참조보다 수염이 도드라져 조금 더 성숙해 보인다. 혼혈 배경은 외양만으로 확정할 수 없다. 귀에는 흰 무선 이어폰이 있다. 도로 건너 고급주택과 낮 풍경이 보인다. 구겨진 도면은 이 프레임에서 확인되지 않지만, 화면 밖에 있을 가능성을 배제할 수 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "하체와 몸통은 운전석에 지지되고 등받이가 뒤에 있다. 몸통과 목을 뒤로 돌리면서 한 팔을 창 쪽으로 드는 자세는 가능하다. 손과 팔은 연결되어 있고 검지를 펴는 동작도 자연스럽다. 이어폰은 귀에 꽂혀 있으며 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "시선과 손끝이 오른쪽 위 주택으로 이어져 핵심 동작과 좌우 배치는 더 충실하지만, 지정된 높은 뒷좌석 시점보다 앞쪽에서 본 구도이고 얼굴과 손의 위치도 정확하지 않다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "뒷좌석에서 어깨 너머로 보는 시점은 가깝지만, 인물과 주택의 좌우 배치가 반대이고 페드로가 집이 아닌 카메라를 바라봐 핵심 시선 지시를 어긴다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "페드로는 창밖 오른쪽 위의 주택 쪽으로 고개를 돌리고 있다. 뻗은 검지는 그 주택의 담장 위 수목과 외벽 방향을 가리켜, 지시 대상이 해당 주택으로 읽힌다. 다른 손은 운전대를 잡고 있다.",
        "built_space": "운전대 하나, 운전석 등받이와 머리받침 하나, 측면 창 하나, 사이드미러 하나, 문손잡이 하나와 대시보드 일부가 보인다. 인물은 운전대 뒤 운전석에 있고 주택은 도로 너머 오른쪽 위에 있다. 다만 가슴 앞면과 운전대를 잡은 손이 크게 보이는 관찰각은 지정된 높은 뒷좌석보다 앞좌석 동승자 쪽에 가깝다. 얼굴은 왼쪽 가장자리보다 안쪽이며 지시 손도 하단 중앙보다 높다. 거울 속 가로수와 담장에는 명백한 광학적 모순이 없다. 이전 장소 실사 참조는 제공되지 않아 차량 마감의 정확한 연속성은 확인할 수 없다.",
        "entities": "인물은 한 명뿐이며 이현우는 없다. 검은 비니, 그 아래 짙은 머리, 낡은 올리브색 집업과 회색 상의가 페드로 참조와 부합한다. 보이는 얼굴 일부는 젊은 남성으로 읽히지만 정확한 나이와 혼혈 배경은 단정할 수 없다. 귀에는 검은 유선 이어폰이 있고 왼쪽 아래에는 구겨진 주택 도면 일부가 보인다. 현대적인 고급주택과 낮 풍경이 있으며 바지와 신발은 평가할 만큼 보이지 않는다. 자막이나 도해의 표식은 없다.",
        "hard_violations": [],
        "physics": "몸통은 운전석에 앉아 지지되고, 지시하는 팔은 어깨와 팔꿈치에서 자연스럽게 뻗어 있다. 다른 손은 운전대 테두리를 실제로 감싼다. 구겨진 종이는 화면 아래의 어두운 천 표면에 얹혀 있다. 지지 없이 뜬 신체나 물체, 명백히 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "검지는 왼쪽 창밖 주택의 대문과 담장 방향을 가리킨다. 그러나 얼굴을 뒤로 돌린 페드로의 눈은 주택이 아니라 카메라를 향한다. 따라서 손짓의 대상은 맞지만 시선의 대상은 틀리다.",
        "built_space": "운전대 하나, 오른쪽 전경의 운전석 등받이와 머리받침 하나, 왼쪽 아래 다른 좌석 일부, 측면 창 하나, 사이드미러 하나, 문손잡이 하나와 창문 조작부가 보인다. 운전자는 운전대 뒤에 앉아 있고 카메라는 좌석 뒤에서 어깨 너머를 본다. 하지만 얼굴은 오른쪽, 주택은 왼쪽 위라 요구된 좌우 배치와 반대이며 높은 관찰각도 두드러지지 않는다. 사이드미러의 도로와 가로수 반사는 가능한 범위다. 이전 장소 실사 참조가 없어 차량 고정 부품의 정확한 일치는 확인할 수 없다.",
        "entities": "젊은 남성 한 명만 등장하며 검은 비니, 짙은 머리, 올리브색의 낡은 집업과 어두운 바지가 페드로 참조에 대체로 맞는다. 얼굴에는 참조보다 수염이 도드라져 조금 더 성숙해 보인다. 혼혈 배경은 외양만으로 확정할 수 없다. 귀에는 흰 무선 이어폰이 있다. 도로 건너 고급주택과 낮 풍경이 보인다. 구겨진 도면은 이 프레임에서 확인되지 않지만, 화면 밖에 있을 가능성을 배제할 수 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "하체와 몸통은 운전석에 지지되고 등받이가 뒤에 있다. 몸통과 목을 뒤로 돌리면서 한 팔을 창 쪽으로 드는 자세는 가능하다. 손과 팔은 연결되어 있고 검지를 펴는 동작도 자연스럽다. 이어폰은 귀에 꽂혀 있으며 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.667,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (우측 운전대를 잡고 있는 정체불명의 오른손)",
     "[gemini-pro] 공간 구조 위반 (지시된 좌핸들 배치를 무시하고 우핸들로 변형)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1667,
   "B": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1667,
    "verdict_ko": "목표물을 보지 않고 카메라를 응시하는 연출 오류가 있으나, 참조된 좌핸들 차량의 구조를 유지하였고 해부학적 위반이 없어 선정됨."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "시선과 가리키는 방향은 정확하나, 차량이 우핸들로 왜곡되고 운전대를 잡은 불가능한 손이 나타나는 치명적 오류가 발생함.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (우측 운전대를 잡고 있는 정체불명의 오른손) / [gemini-pro] 공간 구조 위반 (지시된 좌핸들 배치를 무시하고 우핸들로 변형)"
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh5_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "페드로",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07d1-5a03-718b-8eeb-638145523663",
  "confined_fp": {
   "base_key": "confinedfp::03bcfbed1a4a",
   "apt_reason": "차량 내부 운전석이라는 명확한 조작 공간을 배경으로 하며 인물이 차창 너머 특정 방향을 가리키고 있으므로 차량 내 좌석의 위치와 가리키는 방향이 시각적으로 정확히 일치해야 합니다.",
   "fixed": true,
   "mismatches": [
    "The text states Pedro's face is at the left edge and the indicated house/window is in the upper-right background. However, the diagram shows the camera aiming from the rear center toward the front-left window, which places the window on the left side of the frame and Pedro on the right."
   ]
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh2"
  }
 },
 "S4sh11::signage": {
  "fp": "5b5f368598c0fe5e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S4sh11::bgfirst_bg": {
  "input_fingerprint": "1a142c7fa185bba0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고급주택 앞길의 이현우가 통신 확인을 위해 귀에 손가락을 댄 근접 구도.\n\nLOCATION (lock): On the street immediately outside an upscale house, across from the parked getaway car.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral approach beside 이현우 at shoulder height, looking slightly upward into a side three-quarter close view rather than moving onto his forward-facing axis. Place his face left of center with the fingers touching his ear beside his cheek, leaving open space to the right toward the house and only a soft fragment of its exterior behind him. He keeps his attention on the house beyond the right edge while checking communication; emphasize only the reduced camera distance at this stage's end.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Upscale house exterior (Beside the road along 이현우's approach) — Only an oblique fragment is visible behind him; the part receiving his attention remains beyond the right edge; used as A subdued background fragment preserves location while leaving directional space ahead of his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained cool-neutral contrast keep the ear-touching gesture and his frustrated expression intimate but unsentimental.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고급주택 앞길의 이현우가 통신 확인을 위해 귀에 손가락을 댄 근접 구도.\n\nLOCATION (lock): On the street immediately outside an upscale house, across from the parked getaway car.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral approach beside 이현우 at shoulder height, looking slightly upward into a side three-quarter close view rather than moving onto his forward-facing axis. Place his face left of center with the fingers touching his ear beside his cheek, leaving open space to the right toward the house and only a soft fragment of its exterior behind him. He keeps his attention on the house beyond the right edge while checking communication; emphasize only the reduced camera distance at this stage's end.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Upscale house exterior (Beside the road along 이현우's approach) — Only an oblique fragment is visible behind him; the part receiving his attention remains beyond the right edge; used as A subdued background fragment preserves location while leaving directional space ahead of his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained cool-neutral contrast keep the ear-touching gesture and his frustrated expression intimate but unsentimental.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh11__bgfirst_bg.png",
  "asset_id": "f2e8e3e6-3b28-4cf6-810b-5a67a4bf5836",
  "input_asset_ids": [
   "caa1cead-b028-4463-8487-148f516c6fce",
   "c0e7057d-f698-4665-af76-084600319198"
  ]
 },
 "S4sh11": {
  "input_fingerprint": "ffdf576946e0ac19",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 앞길의 이현우가 통신 확인을 위해 귀에 손가락을 댄 근접 구도.\n\nLOCATION (lock): On the street immediately outside an upscale house, across from the parked getaway car. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral approach beside 이현우 at shoulder height, looking slightly upward into a side three-quarter close view rather than moving onto his forward-facing axis. Place his face left of center with the fingers touching his ear beside his cheek, leaving open space to the right toward the house and only a soft fragment of its exterior behind him. He keeps his attention on the house beyond the right edge while checking communication; emphasize only the reduced camera distance at this stage's end.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Upscale house exterior (Beside the road along 이현우's approach) — Only an oblique fragment is visible behind him; the part receiving his attention remains beyond the right edge; used as A subdued background fragment preserves location while leaving directional space ahead of his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained cool-neutral contrast keep the ear-touching gesture and his frustrated expression intimate but unsentimental.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the luxury house in daylight. 이현우: Is outside the car, approaching the luxury house across the road with a hand at his ear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 앞길의 이현우가 통신 확인을 위해 귀에 손가락을 댄 근접 구도.\n\nLOCATION (lock): On the street immediately outside an upscale house, across from the parked getaway car. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral approach beside 이현우 at shoulder height, looking slightly upward into a side three-quarter close view rather than moving onto his forward-facing axis. Place his face left of center with the fingers touching his ear beside his cheek, leaving open space to the right toward the house and only a soft fragment of its exterior behind him. He keeps his attention on the house beyond the right edge while checking communication; emphasize only the reduced camera distance at this stage's end.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Upscale house exterior (Beside the road along 이현우's approach) — Only an oblique fragment is visible behind him; the part receiving his attention remains beyond the right edge; used as A subdued background fragment preserves location while leaving directional space ahead of his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained cool-neutral contrast keep the ear-touching gesture and his frustrated expression intimate but unsentimental.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the luxury house in daylight. 이현우: Is outside the car, approaching the luxury house across the road with a hand at his ear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고급주택 앞길의 이현우가 통신 확인을 위해 귀에 손가락을 댄 근접 구도.\n\nLOCATION (lock): On the street immediately outside an upscale house, across from the parked getaway car. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral approach beside 이현우 at shoulder height, looking slightly upward into a side three-quarter close view rather than moving onto his forward-facing axis. Place his face left of center with the fingers touching his ear beside his cheek, leaving open space to the right toward the house and only a soft fragment of its exterior behind him. He keeps his attention on the house beyond the right edge while checking communication; emphasize only the reduced camera distance at this stage's end.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Upscale house exterior (Beside the road along 이현우's approach) — Only an oblique fragment is visible behind him; the part receiving his attention remains beyond the right edge; used as A subdued background fragment preserves location while leaving directional space ahead of his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained cool-neutral contrast keep the ear-touching gesture and his frustrated expression intimate but unsentimental.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The car remains parked across the road from the luxury house in daylight. 이현우: Is outside the car, approaching the luxury house across the road with a hand at his ear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh11__bgfirst_bg.png",
     "asset_id": "f2e8e3e6-3b28-4cf6-810b-5a67a4bf5836",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S4sh11.png",
     "asset_id": "caa1cead-b028-4463-8487-148f516c6fce",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L08B02.png",
     "asset_id": "c0e7057d-f698-4665-af76-084600319198",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀에 꽂힌 무전기에 닿아 있음.",
    "built_space": "우측 배경에 콘크리트 담장과 주택의 일부가 보이며, 좌측 배경에 주차된 검은 차량이 있음.",
    "entities": "이현우의 얼굴, 헝클어진 머리, 피 묻은 셔츠가 참조와 일치하며 인이어 무전기를 착용함.",
    "hard_violations": [],
    "physics": "자연스럽게 서서 팔을 들어 올려 귀에 손을 댄 자세가 구조적으로 안정됨."
   },
   {
    "label": "B",
    "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀 쪽을 향해 있음.",
    "built_space": "배경에 참조 이미지의 검은 대문과 콘크리트 담장이 측면 구도로 배치됨.",
    "entities": "이현우의 외형은 일치하나, 귀에 댄 손가락들이 다소 뭉개져 어색함. 무전기 착용함.",
    "hard_violations": [],
    "physics": "서 있는 자세 자체는 문제없으나, 손의 관절 형태가 다소 부자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 측면 근접 구도와 우측 여백을 잘 살렸으며, 자연스러운 손동작과 배경의 차량 배치로 현실감을 높였습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 배경 구조물은 지시를 잘 따랐으나, 귀에 댄 손의 해부학적 묘사가 뭉개져 디테일이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀에 꽂힌 무전기에 닿아 있음.",
        "built_space": "우측 배경에 콘크리트 담장과 주택의 일부가 보이며, 좌측 배경에 주차된 검은 차량이 있음.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 피 묻은 셔츠가 참조와 일치하며 인이어 무전기를 착용함.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 팔을 들어 올려 귀에 손을 댄 자세가 구조적으로 안정됨."
       },
       {
        "label": "B",
        "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀 쪽을 향해 있음.",
        "built_space": "배경에 참조 이미지의 검은 대문과 콘크리트 담장이 측면 구도로 배치됨.",
        "entities": "이현우의 외형은 일치하나, 귀에 댄 손가락들이 다소 뭉개져 어색함. 무전기 착용함.",
        "hard_violations": [],
        "physics": "서 있는 자세 자체는 문제없으나, 손의 관절 형태가 다소 부자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 측면 근접 구도와 우측 여백을 잘 살렸으며, 자연스러운 손동작과 배경의 차량 배치로 현실감을 높였습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 배경 구조물은 지시를 잘 따랐으나, 귀에 댄 손의 해부학적 묘사가 뭉개져 디테일이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀에 꽂힌 무전기에 닿아 있음.",
        "built_space": "우측 배경에 콘크리트 담장과 주택의 일부가 보이며, 좌측 배경에 주차된 검은 차량이 있음.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 피 묻은 셔츠가 참조와 일치하며 인이어 무전기를 착용함.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 팔을 들어 올려 귀에 손을 댄 자세가 구조적으로 안정됨."
       },
       {
        "label": "B",
        "direction": "시선은 우측 프레임 밖을 향하고, 오른손 검지는 귀 쪽을 향해 있음.",
        "built_space": "배경에 참조 이미지의 검은 대문과 콘크리트 담장이 측면 구도로 배치됨.",
        "entities": "이현우의 외형은 일치하나, 귀에 댄 손가락들이 다소 뭉개져 어색함. 무전기 착용함.",
        "hard_violations": [],
        "physics": "서 있는 자세 자체는 문제없으나, 손의 관절 형태가 다소 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴과 귀를 누르는 손을 더 가깝게 담고 오른쪽 시선 공간과 부드러운 주택 배경을 확보해 우세하지만, 카메라의 약한 올려다보기는 뚜렷하지 않다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "측면 삼사분면 시점과 귀 접촉은 충실하지만, 상체·차량·선명한 외벽까지 넓게 보여 최종 접근 단계의 친밀한 클로즈업과 절제된 배경 지시에서 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 두 눈이 화면 오른쪽 위의 화면 밖을 향하며, 주택 쪽에 주의를 둔 방향과 부합한다. 검지는 보이는 귀의 검은 인이어 장치를 직접 누른다. 렌즈를 바라보지 않으며 무기나 별도의 지향성 소품은 없다.",
        "built_space": "인물 뒤로 검은 대문 한 곳, 돌출 차양이 있는 현관 한 곳, 큰 검은 창틀의 창 한 곳, 노출 콘크리트 담장과 식재가 보인다. 대문 옆 기둥에는 위아래로 배치된 작은 금속 설비 두 개가 보여 참고 장소와 대응한다. 인물은 담장 밖 도로 쪽에 있으며, 창의 수목 반사에 명백한 광학적 모순은 없다. 배경은 흐려졌지만 요청한 외관의 작은 조각보다는 넓게 드러난다.",
        "entities": "젊은 동아시아계 남성 한 명이며, 짧고 헝클어진 검은 머리와 얼굴 윤곽이 이현우 참고 이미지에 가깝다. 국적은 외관만으로 확인할 수 없다. 어두운 낡은 셔츠에 먼지와 적갈색 핏자국이 있고 귀에는 작은 검은 인이어 장치가 있다. 찌푸린 눈썹과 벌어진 입술이 통신 확인 중의 긴장과 답답함을 나타낸다. 바지와 차량은 근접 프레임 밖이며 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "올린 손은 손목과 소매 속 팔로 이어지고 검지 끝이 인이어 장치에 닿아 있다. 장치는 귀에 끼워져 지지된다. 머리와 몸통의 연결 및 손가락 굽힘이 자연스럽다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만 공중에 뜬 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴은 오른쪽 측면 삼사분면으로 돌아가 있고 시선도 오른쪽 화면 밖 주택 방향을 향한다. 검지가 귀의 인이어 장치를 누르며 렌즈를 보지 않는다. 왼쪽 뒤 차량은 정지해 있는 모습이고 이동 동작은 없다.",
        "built_space": "오른쪽에 콘크리트 담장, 그 위 회색 석재 외벽과 큰 검은 창틀의 창 한 곳, 식재가 보인다. 인물 뒤에는 어두운 대문과 현관 일부가 가려져 있다. 고정 시설이 중복되지는 않으며 재료는 참고 장소와 맞는다. 다만 외벽과 담장이 상당히 선명하고 넓게 노출된다. 왼쪽 뒤 차량과 주택 사이에서 도로 건너편이라는 관계는 명확하게 확인되지 않는다. 창의 수목 반사는 가능한 배치다.",
        "entities": "젊은 동아시아계 남성 한 명으로 검은 헝클어진 머리, 마른 상체, 참고 인물과 유사한 얼굴을 갖는다. 어두운 셔츠의 먼지와 핏자국, 작은 검은 인이어 장치가 보인다. 눈썹을 약하게 찌푸린 경계하는 표정이나 좌절감은 절제되어 있다. 왼쪽 배경에는 검은 승용차 한 대가 일부 보인다. 바지는 프레임 밖이며 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "손과 팔이 자연스럽게 연결되고 검지가 귀에 끼운 장치에 접촉한다. 팔을 들어 통신을 확인하는 동작으로 가능한 자세다. 하체는 화면 밖이며 몸통이 떠 있다는 징후는 없다. 배경 차량은 보이는 바퀴로 도로에 지지되어 있고 지지 없이 떠 있는 소품은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴과 귀를 누르는 손을 더 가깝게 담고 오른쪽 시선 공간과 부드러운 주택 배경을 확보해 우세하지만, 카메라의 약한 올려다보기는 뚜렷하지 않다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "측면 삼사분면 시점과 귀 접촉은 충실하지만, 상체·차량·선명한 외벽까지 넓게 보여 최종 접근 단계의 친밀한 클로즈업과 절제된 배경 지시에서 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 두 눈이 화면 오른쪽 위의 화면 밖을 향하며, 주택 쪽에 주의를 둔 방향과 부합한다. 검지는 보이는 귀의 검은 인이어 장치를 직접 누른다. 렌즈를 바라보지 않으며 무기나 별도의 지향성 소품은 없다.",
        "built_space": "인물 뒤로 검은 대문 한 곳, 돌출 차양이 있는 현관 한 곳, 큰 검은 창틀의 창 한 곳, 노출 콘크리트 담장과 식재가 보인다. 대문 옆 기둥에는 위아래로 배치된 작은 금속 설비 두 개가 보여 참고 장소와 대응한다. 인물은 담장 밖 도로 쪽에 있으며, 창의 수목 반사에 명백한 광학적 모순은 없다. 배경은 흐려졌지만 요청한 외관의 작은 조각보다는 넓게 드러난다.",
        "entities": "젊은 동아시아계 남성 한 명이며, 짧고 헝클어진 검은 머리와 얼굴 윤곽이 이현우 참고 이미지에 가깝다. 국적은 외관만으로 확인할 수 없다. 어두운 낡은 셔츠에 먼지와 적갈색 핏자국이 있고 귀에는 작은 검은 인이어 장치가 있다. 찌푸린 눈썹과 벌어진 입술이 통신 확인 중의 긴장과 답답함을 나타낸다. 바지와 차량은 근접 프레임 밖이며 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "올린 손은 손목과 소매 속 팔로 이어지고 검지 끝이 인이어 장치에 닿아 있다. 장치는 귀에 끼워져 지지된다. 머리와 몸통의 연결 및 손가락 굽힘이 자연스럽다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만 공중에 뜬 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴은 오른쪽 측면 삼사분면으로 돌아가 있고 시선도 오른쪽 화면 밖 주택 방향을 향한다. 검지가 귀의 인이어 장치를 누르며 렌즈를 보지 않는다. 왼쪽 뒤 차량은 정지해 있는 모습이고 이동 동작은 없다.",
        "built_space": "오른쪽에 콘크리트 담장, 그 위 회색 석재 외벽과 큰 검은 창틀의 창 한 곳, 식재가 보인다. 인물 뒤에는 어두운 대문과 현관 일부가 가려져 있다. 고정 시설이 중복되지는 않으며 재료는 참고 장소와 맞는다. 다만 외벽과 담장이 상당히 선명하고 넓게 노출된다. 왼쪽 뒤 차량과 주택 사이에서 도로 건너편이라는 관계는 명확하게 확인되지 않는다. 창의 수목 반사는 가능한 배치다.",
        "entities": "젊은 동아시아계 남성 한 명으로 검은 헝클어진 머리, 마른 상체, 참고 인물과 유사한 얼굴을 갖는다. 어두운 셔츠의 먼지와 핏자국, 작은 검은 인이어 장치가 보인다. 눈썹을 약하게 찌푸린 경계하는 표정이나 좌절감은 절제되어 있다. 왼쪽 배경에는 검은 승용차 한 대가 일부 보인다. 바지는 프레임 밖이며 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "손과 팔이 자연스럽게 연결되고 검지가 귀에 끼운 장치에 접촉한다. 팔을 들어 통신을 확인하는 동작으로 가능한 자세다. 하체는 화면 밖이며 몸통이 떠 있다는 징후는 없다. 배경 차량은 보이는 바퀴로 도로에 지지되어 있고 지지 없이 떠 있는 소품은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지정된 측면 근접 구도와 우측 여백을 잘 살렸으며, 자연스러운 손동작과 배경의 차량 배치로 현실감을 높였습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "구도와 배경 구조물은 지시를 잘 따랐으나, 귀에 댄 손의 해부학적 묘사가 뭉개져 디테일이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L08B02.png",
    "asset_id": "c0e7057d-f698-4665-af76-084600319198",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07df-4dbb-781b-bbc8-9a78347687e6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh11__bgfirst_bg.png",
   "bg_asset_id": "f2e8e3e6-3b28-4cf6-810b-5a67a4bf5836",
   "bg_record_key": "S4sh11::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S5sh42::signage": {
  "fp": "df1763c9aa46d4a4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::42be3bd427d9404b": {
  "subjects": [],
  "subject_text": "고급주택 담장 안쪽·현관 앞·외벽 창가\n담장 안쪽에 현관과 외벽 통유리창이 이어지는 주택 외부. 현관문에는 구식 열쇠식 잠금장치가 있고 창가 옆으로 건물 모퉁이가 꺾인다.",
  "identity": "canonical",
  "scope_id": "L09",
  "scope_role": "location_exterior",
  "scope_sha": "1095951a0e3401e9"
 },
 "groupbg::glass_window_yard": {
  "input_fingerprint": "2b2ff161b4188763",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "glass_window_yard",
    "tags": [
     "S5sh42"
    ]
   },
   "context_sig": "db918b11a022dec9"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the ground outside the open living-room window of an upscale house, within its boundary wall.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n고급주택 담장 안쪽·현관 앞·외벽 창가: 높은 담장을 넘어서 진입하는 외벽 공간으로 구형 보안 장치와 통유리창이 보인다. (특징: 주택을 둘러싼 석조 또는 콘크리트 담벼락; 구형 아날로그 도어락이 설치된 현관문; 유리창 앞에 슬쩍 놓아둔 스마트폰)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 거실 유리창 쪽으로 다가가는 현우.\n- 이때 유리창 쪽에, 셰퍼드 한 마리가 이빨을 드러내며 으르렁거린다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the ground outside the open living-room window of an upscale house, within its boundary wall.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n고급주택 담장 안쪽·현관 앞·외벽 창가: 높은 담장을 넘어서 진입하는 외벽 공간으로 구형 보안 장치와 통유리창이 보인다. (특징: 주택을 둘러싼 석조 또는 콘크리트 담벼락; 구형 아날로그 도어락이 설치된 현관문; 유리창 앞에 슬쩍 놓아둔 스마트폰)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 거실 유리창 쪽으로 다가가는 현우.\n- 이때 유리창 쪽에, 셰퍼드 한 마리가 이빨을 드러내며 으르렁거린다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_window_yard_8fd2a5.png",
  "asset_id": "f82b5f69-5650-4242-9f2d-699966bb6919",
  "input_asset_ids": [
   "df9f06fa-18a1-452d-b4d4-a99db9c4bc48"
  ],
  "origin_tag": "S5sh42",
  "place_text": "On the ground outside the open living-room window of an upscale house, within its boundary wall.",
  "origin_inputs": {
   "place_text": "On the ground outside the open living-room window of an upscale house, within its boundary wall.",
   "time_of_day_en": "day",
   "conti_asset_id": "df9f06fa-18a1-452d-b4d4-a99db9c4bc48"
  }
 },
 "S5sh42::bgfirst_bg": {
  "input_fingerprint": "bebb2fb9e33b8de1",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 셰퍼드의 입 쪽으로 당겨진 로봇 머리의 인공 외피 아래 금속 내부가 드러난 소품 근접 구도.\n\nLOCATION (lock): On the ground outside the open living-room window of an upscale house, within its boundary wall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the low dolly beside the fallen robot's head near the living-room window, directly observing from just above floor level with a slight downward angle across the head toward 셰퍼드's mouth. Place the head of 젊은 여자 모습의 로봇 at lower left, occupying less than two-fifths of the image, and 셰퍼드's muzzle at right as the dog pulls the torn artificial covering toward itself, exposing the metal beneath without implying decapitation. The dog concentrates downward on its bite while the collapsed robot's rolled eyes remain unresponsive; retain part of her shoulder and the floor as scale references before the explicit cut back to the pursuit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Floor beside the living-room window (Supporting the fallen robot after impact); used as Visible around the head and shoulder to establish the low camera position and realistic scale; Living-room window opening (Opened earlier by the robot) — Only a small oblique edge of the open window area is retained behind the floor-level action; used as Maintains the location of the attack without distracting from the exposed robot interior.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled daytime ambient illumination reveals the torn artificial covering and exposed metal precisely without adding a sensational color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 셰퍼드의 입 쪽으로 당겨진 로봇 머리의 인공 외피 아래 금속 내부가 드러난 소품 근접 구도.\n\nLOCATION (lock): On the ground outside the open living-room window of an upscale house, within its boundary wall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the low dolly beside the fallen robot's head near the living-room window, directly observing from just above floor level with a slight downward angle across the head toward 셰퍼드's mouth. Place the head of 젊은 여자 모습의 로봇 at lower left, occupying less than two-fifths of the image, and 셰퍼드's muzzle at right as the dog pulls the torn artificial covering toward itself, exposing the metal beneath without implying decapitation. The dog concentrates downward on its bite while the collapsed robot's rolled eyes remain unresponsive; retain part of her shoulder and the floor as scale references before the explicit cut back to the pursuit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Floor beside the living-room window (Supporting the fallen robot after impact); used as Visible around the head and shoulder to establish the low camera position and realistic scale; Living-room window opening (Opened earlier by the robot) — Only a small oblique edge of the open window area is retained behind the floor-level action; used as Maintains the location of the attack without distracting from the exposed robot interior.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled daytime ambient illumination reveals the torn artificial covering and exposed metal precisely without adding a sensational color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S5sh42__bgfirst_bg.png",
  "asset_id": "504a1ad8-f3ea-47e6-8874-c67e355d8d69",
  "input_asset_ids": [
   "df9f06fa-18a1-452d-b4d4-a99db9c4bc48",
   "f82b5f69-5650-4242-9f2d-699966bb6919"
  ]
 },
 "S5sh42": {
  "input_fingerprint": "5fc3f607352beb6e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 셰퍼드의 입 쪽으로 당겨진 로봇 머리의 인공 외피 아래 금속 내부가 드러난 소품 근접 구도.\n\nLOCATION (lock): On the ground outside the open living-room window of an upscale house, within its boundary wall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the low dolly beside the fallen robot's head near the living-room window, directly observing from just above floor level with a slight downward angle across the head toward 셰퍼드's mouth. Place the head of 젊은 여자 모습의 로봇 at lower left, occupying less than two-fifths of the image, and 셰퍼드's muzzle at right as the dog pulls the torn artificial covering toward itself, exposing the metal beneath without implying decapitation. The dog concentrates downward on its bite while the collapsed robot's rolled eyes remain unresponsive; retain part of her shoulder and the floor as scale references before the explicit cut back to the pursuit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Floor beside the living-room window (Supporting the fallen robot after impact); used as Visible around the head and shoulder to establish the low camera position and realistic scale; Living-room window opening (Opened earlier by the robot) — Only a small oblique edge of the open window area is retained behind the floor-level action; used as Maintains the location of the attack without distracting from the exposed robot interior.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled daytime ambient illumination reveals the torn artificial covering and exposed metal precisely without adding a sensational color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The female-presenting robot is collapsed on the floor after being dropped by Hyunwoo, with the German shepherd biting her head and peeling back its artificial skin to expose metal underneath. Her torso's orientation and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 젊은 여자 모습의 로봇: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the entrance has an old key-operated lock. The female-presenting robot lies on the floor with torn artificial skin exposing the metal inside its head, and the shepherd is biting at that head.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 젊은 여자 모습의 로봇 (한국인 여성형 얼굴, 젊은 성인의 외모, 자연스러운 인간 얼굴, 검은 머리카락); 셰퍼드 (셰퍼드 견종, 곧게 선 귀, 긴 주둥이, 검정과 황갈색 털, 풍성한 꼬리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 셰퍼드의 입 쪽으로 당겨진 로봇 머리의 인공 외피 아래 금속 내부가 드러난 소품 근접 구도.\n\nLOCATION (lock): On the ground outside the open living-room window of an upscale house, within its boundary wall. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the low dolly beside the fallen robot's head near the living-room window, directly observing from just above floor level with a slight downward angle across the head toward 셰퍼드's mouth. Place the head of 젊은 여자 모습의 로봇 at lower left, occupying less than two-fifths of the image, and 셰퍼드's muzzle at right as the dog pulls the torn artificial covering toward itself, exposing the metal beneath without implying decapitation. The dog concentrates downward on its bite while the collapsed robot's rolled eyes remain unresponsive; retain part of her shoulder and the floor as scale references before the explicit cut back to the pursuit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Floor beside the living-room window (Supporting the fallen robot after impact); used as Visible around the head and shoulder to establish the low camera position and realistic scale; Living-room window opening (Opened earlier by the robot) — Only a small oblique edge of the open window area is retained behind the floor-level action; used as Maintains the location of the attack without distracting from the exposed robot interior.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled daytime ambient illumination reveals the torn artificial covering and exposed metal precisely without adding a sensational color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The female-presenting robot is collapsed on the floor after being dropped by Hyunwoo, with the German shepherd biting her head and peeling back its artificial skin to expose metal underneath. Her torso's orientation and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 젊은 여자 모습의 로봇: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the entrance has an old key-operated lock. The female-presenting robot lies on the floor with torn artificial skin exposing the metal inside its head, and the shepherd is biting at that head.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 젊은 여자 모습의 로봇 (한국인 여성형 얼굴, 젊은 성인의 외모, 자연스러운 인간 얼굴, 검은 머리카락); 셰퍼드 (셰퍼드 견종, 곧게 선 귀, 긴 주둥이, 검정과 황갈색 털, 풍성한 꼬리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 셰퍼드의 입 쪽으로 당겨진 로봇 머리의 인공 외피 아래 금속 내부가 드러난 소품 근접 구도.\n\nLOCATION (lock): On the ground outside the open living-room window of an upscale house, within its boundary wall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the low dolly beside the fallen robot's head near the living-room window, directly observing from just above floor level with a slight downward angle across the head toward 셰퍼드's mouth. Place the head of 젊은 여자 모습의 로봇 at lower left, occupying less than two-fifths of the image, and 셰퍼드's muzzle at right as the dog pulls the torn artificial covering toward itself, exposing the metal beneath without implying decapitation. The dog concentrates downward on its bite while the collapsed robot's rolled eyes remain unresponsive; retain part of her shoulder and the floor as scale references before the explicit cut back to the pursuit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Floor beside the living-room window (Supporting the fallen robot after impact); used as Visible around the head and shoulder to establish the low camera position and realistic scale; Living-room window opening (Opened earlier by the robot) — Only a small oblique edge of the open window area is retained behind the floor-level action; used as Maintains the location of the attack without distracting from the exposed robot interior.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled daytime ambient illumination reveals the torn artificial covering and exposed metal precisely without adding a sensational color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The female-presenting robot is collapsed on the floor after being dropped by Hyunwoo, with the German shepherd biting her head and peeling back its artificial skin to expose metal underneath. Her torso's orientation and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 젊은 여자 모습의 로봇: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room glass window is open, while the entrance has an old key-operated lock. The female-presenting robot lies on the floor with torn artificial skin exposing the metal inside its head, and the shepherd is biting at that head.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 젊은 여자 모습의 로봇 (한국인 여성형 얼굴, 젊은 성인의 외모, 자연스러운 인간 얼굴, 검은 머리카락); 셰퍼드 (셰퍼드 견종, 곧게 선 귀, 긴 주둥이, 검정과 황갈색 털, 풍성한 꼬리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S5sh42__bgfirst_bg.png",
     "asset_id": "504a1ad8-f3ea-47e6-8874-c67e355d8d69",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S5sh42.png",
     "asset_id": "df9f06fa-18a1-452d-b4d4-a99db9c4bc48",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 젊은 여자 모습의 로봇: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:871677>",
     "asset_id": "a9010707-81ea-4ded-a459-7766600ddc90",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 셰퍼드: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1168808>",
     "asset_id": "f1244d02-38bf-4afd-944f-26fd9cd43466",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_window_yard_8fd2a5.png",
     "asset_id": "f82b5f69-5650-4242-9f2d-699966bb6919",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 젊은 여자 모습의 로봇: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:871677>",
     "asset_id": "a9010707-81ea-4ded-a459-7766600ddc90",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 셰퍼드: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1168808>",
     "asset_id": "f1244d02-38bf-4afd-944f-26fd9cd43466",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "셰퍼드의 주둥이가 로봇 머리를 향해 있으며, 입으로 인공 피부를 강하게 물어 당기고 있음.",
    "built_space": "참조된 타일 바닥에 낮은 각도로 위치하며 배경에 열린 거실 창문이 보임. 지정된 화면 비율과 구도를 잘 유지함.",
    "entities": "셰퍼드와 로봇의 외형이 참조 이미지와 일치함. 내부 금속 구조와 늘어나는 피부 질감이 우수하나, 로봇의 눈은 뒤집히지 않고 감겨 있음.",
    "hard_violations": [],
    "physics": "로봇은 바닥에 누워 중력의 영향을 받고 있으며, 셰퍼드는 바닥을 단단히 디디고 서서 피부를 당기는 힘을 자연스럽게 보여줌."
   },
   {
    "label": "B",
    "direction": "셰퍼드가 로봇의 머리를 향해 입을 대고 물고 있으나, 로봇의 시선은 정면/위를 응시함.",
    "built_space": "타일 바닥과 배경 창문이 올바르게 배치되었으나, 로봇의 머리가 화면의 2/5 이상을 차지하여 지시된 비율보다 크게 잡힘.",
    "entities": "캐릭터 외형은 일치하나, 로봇의 눈이 생기 있게 떠 있어 무반응 상태로 보이지 않고 피부 묘사가 다소 뻣뻣함.",
    "hard_violations": [],
    "physics": "셰퍼드의 발과 로봇의 몸체 모두 바닥에 안정적으로 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "피부가 당겨지는 물리적 묘사와 카메라 구도가 지시사항을 잘 따랐으나, 로봇의 눈이 뒤집히지 않고 감겨 있는 점이 아쉬움."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "로봇의 눈이 또렷하게 떠 있어 반응이 없다는 지시를 어겼으며, 인공 피부가 당겨지는 질감보다는 단단한 껍질이 깨진 것처럼 묘사됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "셰퍼드의 주둥이가 로봇 머리를 향해 있으며, 입으로 인공 피부를 강하게 물어 당기고 있음.",
        "built_space": "참조된 타일 바닥에 낮은 각도로 위치하며 배경에 열린 거실 창문이 보임. 지정된 화면 비율과 구도를 잘 유지함.",
        "entities": "셰퍼드와 로봇의 외형이 참조 이미지와 일치함. 내부 금속 구조와 늘어나는 피부 질감이 우수하나, 로봇의 눈은 뒤집히지 않고 감겨 있음.",
        "hard_violations": [],
        "physics": "로봇은 바닥에 누워 중력의 영향을 받고 있으며, 셰퍼드는 바닥을 단단히 디디고 서서 피부를 당기는 힘을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "셰퍼드가 로봇의 머리를 향해 입을 대고 물고 있으나, 로봇의 시선은 정면/위를 응시함.",
        "built_space": "타일 바닥과 배경 창문이 올바르게 배치되었으나, 로봇의 머리가 화면의 2/5 이상을 차지하여 지시된 비율보다 크게 잡힘.",
        "entities": "캐릭터 외형은 일치하나, 로봇의 눈이 생기 있게 떠 있어 무반응 상태로 보이지 않고 피부 묘사가 다소 뻣뻣함.",
        "hard_violations": [],
        "physics": "셰퍼드의 발과 로봇의 몸체 모두 바닥에 안정적으로 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "피부가 당겨지는 물리적 묘사와 카메라 구도가 지시사항을 잘 따랐으나, 로봇의 눈이 뒤집히지 않고 감겨 있는 점이 아쉬움."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "로봇의 눈이 또렷하게 떠 있어 반응이 없다는 지시를 어겼으며, 인공 피부가 당겨지는 질감보다는 단단한 껍질이 깨진 것처럼 묘사됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "셰퍼드의 주둥이가 로봇 머리를 향해 있으며, 입으로 인공 피부를 강하게 물어 당기고 있음.",
        "built_space": "참조된 타일 바닥에 낮은 각도로 위치하며 배경에 열린 거실 창문이 보임. 지정된 화면 비율과 구도를 잘 유지함.",
        "entities": "셰퍼드와 로봇의 외형이 참조 이미지와 일치함. 내부 금속 구조와 늘어나는 피부 질감이 우수하나, 로봇의 눈은 뒤집히지 않고 감겨 있음.",
        "hard_violations": [],
        "physics": "로봇은 바닥에 누워 중력의 영향을 받고 있으며, 셰퍼드는 바닥을 단단히 디디고 서서 피부를 당기는 힘을 자연스럽게 보여줌."
       },
       {
        "label": "B",
        "direction": "셰퍼드가 로봇의 머리를 향해 입을 대고 물고 있으나, 로봇의 시선은 정면/위를 응시함.",
        "built_space": "타일 바닥과 배경 창문이 올바르게 배치되었으나, 로봇의 머리가 화면의 2/5 이상을 차지하여 지시된 비율보다 크게 잡힘.",
        "entities": "캐릭터 외형은 일치하나, 로봇의 눈이 생기 있게 떠 있어 무반응 상태로 보이지 않고 피부 묘사가 다소 뻣뻣함.",
        "hard_violations": [],
        "physics": "셰퍼드의 발과 로봇의 몸체 모두 바닥에 안정적으로 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 근접 시점, 좌하단 머리와 오른쪽 주둥이, 연결된 목과 일부 어깨가 지시와 가장 가깝지만, 눈이 뒤집힌 상태와 작은 창 가장자리만 남기는 배경 처리는 미흡합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "외피를 오른쪽으로 당기는 동작은 명확하지만, 몸통과 담장까지 넓어진 구도 및 감긴 눈이 지정된 소품 근접 구도와 무반응의 뒤집힌 눈에서 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "셰퍼드의 주둥이가 오른쪽에서 왼쪽 아래의 로봇 이마를 향하고, 이빨은 금속이 드러난 머리의 외피를 물고 있습니다. 개의 시선도 물고 있는 지점으로 내려갑니다. 외피는 입 쪽으로 들려 있으나 당겨진 길이는 짧습니다. 로봇의 두 눈은 열려 있고 카메라 쪽으로 비스듬히 향하며, 명확하게 뒤집힌 눈은 아닙니다.",
        "built_space": "회색 석재 바닥 위에 로봇과 개가 있고, 뒤에는 검은 창틀의 거실 창 영역 하나와 그 아래 석재 창턱이 보입니다. 창 너머 소파 일부와 화분, 벽 장식이 보여 참조 장소의 재료와 구성이 대체로 맞습니다. 머리는 좌하단에 화면 면적의 5분의 2 미만으로 놓이고 어깨 일부가 왼쪽에 남습니다. 바닥 가까운 시점은 맞지만 창과 실내가 작은 비스듬한 가장자리 이상으로 노출되며, 창의 열린 상태는 명확하지 않습니다. 중복 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "검은 머리와 젊은 한국인 여성형 얼굴을 가진 로봇 한 개체, 검정·황갈색 털과 긴 주둥이의 셰퍼드 한 마리가 보입니다. 로봇의 남색 의복과 노출된 기계식 목은 참조와 부합하며, 찢긴 이마 외피 아래 금속 부품이 있습니다. 개의 귀 끝과 꼬리는 프레임 밖이므로 확인 대상이 아닙니다. 추가 인물이나 문자는 없습니다.",
        "hard_violations": [],
        "physics": "로봇 머리는 옆으로 누워 머리카락과 함께 석재 바닥에 닿고, 어깨도 바닥에 놓여 있습니다. 금속 목이 머리와 몸통을 이어 절단된 머리로 보이지 않습니다. 들린 외피는 개의 이빨이 잡고 있으며 나머지는 머리에 붙어 있습니다. 개는 보이는 두 앞발로 바닥을 딛고 고개를 숙이고 있어 지지와 물어뜯는 동작이 성립합니다."
       },
       {
        "label": "B",
        "direction": "오른쪽 셰퍼드는 왼쪽 아래 로봇의 관자놀이 쪽 외피를 물고 자기 입 쪽으로 길게 당깁니다. 눈은 물고 있는 외피 부근을 내려다보며 목표가 맞습니다. 로봇 얼굴은 위를 향하지만 두 눈이 감겨 있어 지시된 뒤집힌 눈은 보이지 않습니다.",
        "built_space": "회색 석재 바닥, 뒤쪽 검은 창틀의 창 영역 하나와 석재 창턱, 왼쪽 담장 및 대문 하나가 보입니다. 참조 장소의 외장과 배치는 잘 이어지지만 담장·정원·하늘까지 넓게 포함해 창 가장자리만 남기라는 구도보다 배경이 큽니다. 머리는 하단 왼쪽에서 중앙에 걸치고, 어깨뿐 아니라 몸통 상당 부분이 전경을 차지합니다. 낮은 카메라 위치는 맞지만 소품 중심의 밀착감은 약합니다. 창의 열린 상태는 확실하지 않으며 설비 중복이나 불가능한 반사는 보이지 않습니다.",
        "entities": "검은 머리의 젊은 한국인 여성형 로봇과 곧은 귀, 긴 주둥이, 검정·황갈색 털의 셰퍼드가 각각 하나씩 있습니다. 얼굴과 남색 의복은 참조의 기본 특징을 따르지만, 의복 사이로 큰 금속 어깨 관절과 분절된 흉부 구조가 드러나 참조의 단정한 원피스 상체와 차이가 있습니다. 관자놀이 외피 아래 금속 기구가 명확히 노출됩니다. 추가 인물이나 문자는 없습니다.",
        "hard_violations": [],
        "physics": "로봇은 등을 대고 누워 뒤통수와 머리카락이 바닥에 닿고 몸통도 바닥에 놓여 있습니다. 머리는 기계식 목으로 몸통에 연결됩니다. 길게 들린 외피는 한쪽이 머리에 붙고 다른 쪽이 개의 입에 물려 있어 당기는 장력이 설명됩니다. 개의 두 앞발이 바닥을 지지하며, 떠 있는 신체나 지지 없는 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 근접 시점, 좌하단 머리와 오른쪽 주둥이, 연결된 목과 일부 어깨가 지시와 가장 가깝지만, 눈이 뒤집힌 상태와 작은 창 가장자리만 남기는 배경 처리는 미흡합니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "외피를 오른쪽으로 당기는 동작은 명확하지만, 몸통과 담장까지 넓어진 구도 및 감긴 눈이 지정된 소품 근접 구도와 무반응의 뒤집힌 눈에서 벗어납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "셰퍼드의 주둥이가 오른쪽에서 왼쪽 아래의 로봇 이마를 향하고, 이빨은 금속이 드러난 머리의 외피를 물고 있습니다. 개의 시선도 물고 있는 지점으로 내려갑니다. 외피는 입 쪽으로 들려 있으나 당겨진 길이는 짧습니다. 로봇의 두 눈은 열려 있고 카메라 쪽으로 비스듬히 향하며, 명확하게 뒤집힌 눈은 아닙니다.",
        "built_space": "회색 석재 바닥 위에 로봇과 개가 있고, 뒤에는 검은 창틀의 거실 창 영역 하나와 그 아래 석재 창턱이 보입니다. 창 너머 소파 일부와 화분, 벽 장식이 보여 참조 장소의 재료와 구성이 대체로 맞습니다. 머리는 좌하단에 화면 면적의 5분의 2 미만으로 놓이고 어깨 일부가 왼쪽에 남습니다. 바닥 가까운 시점은 맞지만 창과 실내가 작은 비스듬한 가장자리 이상으로 노출되며, 창의 열린 상태는 명확하지 않습니다. 중복 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "검은 머리와 젊은 한국인 여성형 얼굴을 가진 로봇 한 개체, 검정·황갈색 털과 긴 주둥이의 셰퍼드 한 마리가 보입니다. 로봇의 남색 의복과 노출된 기계식 목은 참조와 부합하며, 찢긴 이마 외피 아래 금속 부품이 있습니다. 개의 귀 끝과 꼬리는 프레임 밖이므로 확인 대상이 아닙니다. 추가 인물이나 문자는 없습니다.",
        "hard_violations": [],
        "physics": "로봇 머리는 옆으로 누워 머리카락과 함께 석재 바닥에 닿고, 어깨도 바닥에 놓여 있습니다. 금속 목이 머리와 몸통을 이어 절단된 머리로 보이지 않습니다. 들린 외피는 개의 이빨이 잡고 있으며 나머지는 머리에 붙어 있습니다. 개는 보이는 두 앞발로 바닥을 딛고 고개를 숙이고 있어 지지와 물어뜯는 동작이 성립합니다."
       },
       {
        "label": "A",
        "direction": "오른쪽 셰퍼드는 왼쪽 아래 로봇의 관자놀이 쪽 외피를 물고 자기 입 쪽으로 길게 당깁니다. 눈은 물고 있는 외피 부근을 내려다보며 목표가 맞습니다. 로봇 얼굴은 위를 향하지만 두 눈이 감겨 있어 지시된 뒤집힌 눈은 보이지 않습니다.",
        "built_space": "회색 석재 바닥, 뒤쪽 검은 창틀의 창 영역 하나와 석재 창턱, 왼쪽 담장 및 대문 하나가 보입니다. 참조 장소의 외장과 배치는 잘 이어지지만 담장·정원·하늘까지 넓게 포함해 창 가장자리만 남기라는 구도보다 배경이 큽니다. 머리는 하단 왼쪽에서 중앙에 걸치고, 어깨뿐 아니라 몸통 상당 부분이 전경을 차지합니다. 낮은 카메라 위치는 맞지만 소품 중심의 밀착감은 약합니다. 창의 열린 상태는 확실하지 않으며 설비 중복이나 불가능한 반사는 보이지 않습니다.",
        "entities": "검은 머리의 젊은 한국인 여성형 로봇과 곧은 귀, 긴 주둥이, 검정·황갈색 털의 셰퍼드가 각각 하나씩 있습니다. 얼굴과 남색 의복은 참조의 기본 특징을 따르지만, 의복 사이로 큰 금속 어깨 관절과 분절된 흉부 구조가 드러나 참조의 단정한 원피스 상체와 차이가 있습니다. 관자놀이 외피 아래 금속 기구가 명확히 노출됩니다. 추가 인물이나 문자는 없습니다.",
        "hard_violations": [],
        "physics": "로봇은 등을 대고 누워 뒤통수와 머리카락이 바닥에 닿고 몸통도 바닥에 놓여 있습니다. 머리는 기계식 목으로 몸통에 연결됩니다. 길게 들린 외피는 한쪽이 머리에 붙고 다른 쪽이 개의 입에 물려 있어 당기는 장력이 설명됩니다. 개의 두 앞발이 바닥을 지지하며, 떠 있는 신체나 지지 없는 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "피부가 당겨지는 물리적 묘사와 카메라 구도가 지시사항을 잘 따랐으나, 로봇의 눈이 뒤집히지 않고 감겨 있는 점이 아쉬움."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "로봇의 눈이 또렷하게 떠 있어 반응이 없다는 지시를 어겼으며, 인공 피부가 당겨지는 질감보다는 단단한 껍질이 깨진 것처럼 묘사됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_window_yard_8fd2a5.png",
    "asset_id": "f82b5f69-5650-4242-9f2d-699966bb6919",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 젊은 여자 모습의 로봇: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:871677>",
    "asset_id": "a9010707-81ea-4ded-a459-7766600ddc90",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 셰퍼드: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1168808>",
    "asset_id": "f1244d02-38bf-4afd-944f-26fd9cd43466",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07e5-0259-7cbe-b9af-da7c4cec6f00",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S5sh42__bgfirst_bg.png",
   "bg_asset_id": "504a1ad8-f3ea-47e6-8874-c67e355d8d69",
   "bg_record_key": "S5sh42::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "glass_window_yard",
   "groupbg_asset_id": "f82b5f69-5650-4242-9f2d-699966bb6919"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S5sh62::signage": {
  "fp": "5aebdf3bae7ebcae",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh62": {
  "input_fingerprint": "382f7ad025d56e86",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 여러 방향의 경찰과 총구 사이에 이현우가 포위된 주택 앞 넓은 구도.\n\nLOCATION (lock): In the outdoor space immediately in front of an upscale house, below the upstairs escape window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the house frontage at the elevated, oblique endpoint of the crane retreat, looking down with 이현우 off-center and his full body visible. Catch him checking his arrested momentum after rising from the landing, looking toward an officer beyond the left frame edge; distribute the visible police across foreground edges and deeper space, their guns converging inward without exaggerated foreground enlargement. Keep the officers at distinct phases of bracing and weight transfer, with differently angled shoulders and attention directed inward toward the surrounded man; emphasize the widened camera distance rather than a lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: House frontage (Behind the encirclement) — The exterior is seen obliquely behind 이현우; used as Establishes the landing location and the depth occupied by the police; Police guns (Raised and aimed at 이현우) — Barrels enter along different inward diagonals rather than pointing toward the lens; used as Builds converging lines around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast keep the encirclement legible without theatrical lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window and the second-floor escape window remain open. The dropped female-presenting robot remains on the floor with the metal inside its head exposed beneath torn artificial skin. 이현우: Is back on the ground outside the house after the jump, with a fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 경찰 right now, so 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 여러 방향의 경찰과 총구 사이에 이현우가 포위된 주택 앞 넓은 구도.\n\nLOCATION (lock): In the outdoor space immediately in front of an upscale house, below the upstairs escape window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the house frontage at the elevated, oblique endpoint of the crane retreat, looking down with 이현우 off-center and his full body visible. Catch him checking his arrested momentum after rising from the landing, looking toward an officer beyond the left frame edge; distribute the visible police across foreground edges and deeper space, their guns converging inward without exaggerated foreground enlargement. Keep the officers at distinct phases of bracing and weight transfer, with differently angled shoulders and attention directed inward toward the surrounded man; emphasize the widened camera distance rather than a lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: House frontage (Behind the encirclement) — The exterior is seen obliquely behind 이현우; used as Establishes the landing location and the depth occupied by the police; Police guns (Raised and aimed at 이현우) — Barrels enter along different inward diagonals rather than pointing toward the lens; used as Builds converging lines around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast keep the encirclement legible without theatrical lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window and the second-floor escape window remain open. The dropped female-presenting robot remains on the floor with the metal inside its head exposed beneath torn artificial skin. 이현우: Is back on the ground outside the house after the jump, with a fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 경찰 right now, so 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 여러 방향의 경찰과 총구 사이에 이현우가 포위된 주택 앞 넓은 구도.\n\nLOCATION (lock): In the outdoor space immediately in front of an upscale house, below the upstairs escape window. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the house frontage at the elevated, oblique endpoint of the crane retreat, looking down with 이현우 off-center and his full body visible. Catch him checking his arrested momentum after rising from the landing, looking toward an officer beyond the left frame edge; distribute the visible police across foreground edges and deeper space, their guns converging inward without exaggerated foreground enlargement. Keep the officers at distinct phases of bracing and weight transfer, with differently angled shoulders and attention directed inward toward the surrounded man; emphasize the widened camera distance rather than a lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: House frontage (Behind the encirclement) — The exterior is seen obliquely behind 이현우; used as Establishes the landing location and the depth occupied by the police; Police guns (Raised and aimed at 이현우) — Barrels enter along different inward diagonals rather than pointing toward the lens; used as Builds converging lines around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast keep the encirclement legible without theatrical lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The living-room window and the second-floor escape window remain open. The dropped female-presenting robot remains on the floor with the metal inside its head exposed beneath torn artificial skin. 이현우: Is back on the ground outside the house after the jump, with a fresh dog-bite injury to his leg.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 경찰 right now, so 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 프레임 왼쪽 밖을 바라보며, 경찰들의 총구는 모두 중앙의 이현우를 향해 조준되어 있음.",
    "built_space": "레퍼런스와 동일한 구조(왼쪽 검은 문, 오른쪽 콘크리트 벽)의 주택 앞 넓은 도로이며 하이앵글 구도를 정확히 구현함.",
    "entities": "이현우는 오른쪽 귀에 인이어 무전기를 착용하고 왼쪽 다리에 명시된 새로운 상처(핏자국)가 묘사됨.",
    "hard_violations": [
     "[gpt-high] 참고에 없는 다수의 경광등 장착 차량과 전경 차량을 추가하여, 허용되지 않은 소품이 포위 공간을 구성한다.",
     "[gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 장비에 'POLICE' 등의 문구를 새로 넣었다."
    ],
    "physics": "모든 인물이 지면에 안정적으로 서 있고 총기를 제대로 파지함. 단, 조명과 그림자 방향이 레퍼런스와 반대로 형성됨."
   },
   {
    "label": "B",
    "direction": "이현우는 왼쪽을 향해 시선을 두고 있으며, 둘러싼 경찰들의 총구는 이현우를 향함.",
    "built_space": "주택 앞 도로이나, 검은 대문과 계단이 오른쪽에 배치되어 레퍼런스의 공간 구조를 임의로 좌우 반전시킴.",
    "entities": "이현우의 외형은 레퍼런스와 비슷하나, 프롬프트가 요구한 인이어 무전기와 다리의 새로운 상처가 완전히 누락됨.",
    "hard_violations": [
     "[gpt-high] 참고에 있는 승용차 외에 여러 차량과 경광등 장착 차량을 추가하여, 허용되지 않은 소품이 전경과 포위 배치의 큰 부분을 차지한다.",
     "[gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 조끼에 '경찰' 등의 문구를 새로 넣었다."
    ],
    "physics": "인물들이 지면에 잘 고정되어 있고 장비를 자연스럽게 들고 있음. 조명 방향은 레퍼런스와 일치함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "명시된 인이어 무전기와 다리의 상처를 정확히 묘사하고 레퍼런스의 공간 구조(왼쪽 대문)를 잘 유지하여 지시를 충실히 이행함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "필수 요소인 인이어 무전기와 다리 상처가 누락되었고, 공간 구조(오른쪽 대문)를 반전시켜 위치 일관성을 어김."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프레임 왼쪽 밖을 바라보며, 경찰들의 총구는 모두 중앙의 이현우를 향해 조준되어 있음.",
        "built_space": "레퍼런스와 동일한 구조(왼쪽 검은 문, 오른쪽 콘크리트 벽)의 주택 앞 넓은 도로이며 하이앵글 구도를 정확히 구현함.",
        "entities": "이현우는 오른쪽 귀에 인이어 무전기를 착용하고 왼쪽 다리에 명시된 새로운 상처(핏자국)가 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 안정적으로 서 있고 총기를 제대로 파지함. 단, 조명과 그림자 방향이 레퍼런스와 반대로 형성됨."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽을 향해 시선을 두고 있으며, 둘러싼 경찰들의 총구는 이현우를 향함.",
        "built_space": "주택 앞 도로이나, 검은 대문과 계단이 오른쪽에 배치되어 레퍼런스의 공간 구조를 임의로 좌우 반전시킴.",
        "entities": "이현우의 외형은 레퍼런스와 비슷하나, 프롬프트가 요구한 인이어 무전기와 다리의 새로운 상처가 완전히 누락됨.",
        "hard_violations": [],
        "physics": "인물들이 지면에 잘 고정되어 있고 장비를 자연스럽게 들고 있음. 조명 방향은 레퍼런스와 일치함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "명시된 인이어 무전기와 다리의 상처를 정확히 묘사하고 레퍼런스의 공간 구조(왼쪽 대문)를 잘 유지하여 지시를 충실히 이행함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "필수 요소인 인이어 무전기와 다리 상처가 누락되었고, 공간 구조(오른쪽 대문)를 반전시켜 위치 일관성을 어김."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프레임 왼쪽 밖을 바라보며, 경찰들의 총구는 모두 중앙의 이현우를 향해 조준되어 있음.",
        "built_space": "레퍼런스와 동일한 구조(왼쪽 검은 문, 오른쪽 콘크리트 벽)의 주택 앞 넓은 도로이며 하이앵글 구도를 정확히 구현함.",
        "entities": "이현우는 오른쪽 귀에 인이어 무전기를 착용하고 왼쪽 다리에 명시된 새로운 상처(핏자국)가 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 안정적으로 서 있고 총기를 제대로 파지함. 단, 조명과 그림자 방향이 레퍼런스와 반대로 형성됨."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽을 향해 시선을 두고 있으며, 둘러싼 경찰들의 총구는 이현우를 향함.",
        "built_space": "주택 앞 도로이나, 검은 대문과 계단이 오른쪽에 배치되어 레퍼런스의 공간 구조를 임의로 좌우 반전시킴.",
        "entities": "이현우의 외형은 레퍼런스와 비슷하나, 프롬프트가 요구한 인이어 무전기와 다리의 새로운 상처가 완전히 누락됨.",
        "hard_violations": [],
        "physics": "인물들이 지면에 잘 고정되어 있고 장비를 자연스럽게 들고 있음. 조명 방향은 레퍼런스와 일치함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "더 멀고 높은 사선 시점과 작은 전신 크기가 지정된 와이드 구도에 가깝지만, 오른쪽을 보는 시선과 정적인 자세, 추가 차량·문구는 지시와 어긋난다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "경찰의 자세 차이와 다리 혈흔은 더 분명하지만, 전경 경찰이 과대하게 차지하고 이현우도 더 크게 잡혀 크레인 후퇴 종점의 구도에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 오른쪽 경찰 쪽으로 고개와 눈을 돌려, 요구된 왼쪽 프레임 밖 경찰을 보지 않는다. 좌우 가장자리와 아래쪽 경찰의 총은 대체로 이현우의 상체 방향으로 안쪽 대각선을 만든다. 뒤쪽 두 경찰은 카메라 쪽 정면성이 강해 이현우에게 정확히 조준선이 모이는지는 덜 분명하다.",
        "built_space": "회색 콘크리트 담, 어두운 금속 대문, 검은 창틀과 담 위 관목이 주택 뒤편을 사선으로 채워 참고 장소의 주요 재료를 유지한다. 중앙의 큰 대문 하나와 오른쪽 보행 출입구 하나, 그 안 계단이 보인다. 경찰은 후방 두 명, 좌우 가장자리 각 한 명, 하단 전경 두 명이 명확하게 구분된다. 이현우는 중앙보다 약간 왼쪽 도로에 전신으로 서 있다. 두 창문의 개방 상태와 실내 로봇은 이 구도에서 확인할 수 없다. 카메라는 높은 사선 시점이며 B보다 후퇴한 거리감이 강하다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 오염된 어두운 셔츠·갈색 계열 바지가 참고와 대체로 맞는다. 작은 얼굴 때문에 정확한 동일인 여부와 인이어는 확정하기 어렵다. 보이는 다리에는 신선한 개 물림 상처가 뚜렷하지 않다. 경찰은 검은 전술복과 총기를 갖춘 사람으로 표현된다. 참고의 검은 승용차 외에 전경 차량과 경광등 차량 등이 추가되어 있고, 경찰 조끼에는 참고가 정하지 않은 문구가 있다.",
        "hard_violations": [
         "참고에 있는 승용차 외에 여러 차량과 경광등 장착 차량을 추가하여, 허용되지 않은 소품이 전경과 포위 배치의 큰 부분을 차지한다.",
         "참고나 지시가 표기를 확정하지 않은 경찰 조끼에 '경찰' 등의 문구를 새로 넣었다."
        ],
        "physics": "이현우의 양발은 아스팔트에 닿아 체중을 지지한다. 다만 양팔을 내리고 비교적 곧게 서 있어 착지 후 일어서다 관성을 억제하는 순간보다는 정지한 자세로 읽힌다. 전신이 보이는 경찰은 벌린 두 발로 서 있고, 잘린 경찰들의 상체에도 공중에 떠 있다는 증거는 없다. 총기는 손과 어깨로 지지되며 차량도 지면 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 오른쪽 가장자리 경찰을 향하며, 지정된 왼쪽 프레임 밖 목표와 반대다. 왼쪽과 오른쪽 전경 경찰의 총구는 이현우의 몸통 쪽으로 향하고, 오른쪽 후방 경찰도 왼쪽 안쪽을 겨눈다. 왼쪽 후방 경찰들은 오른쪽 안쪽으로 총과 어깨를 돌려 포위 방향을 만든다. 총구가 렌즈를 직접 겨누는 구성은 아니다.",
        "built_space": "긴 회색 콘크리트 담, 검은 금속 대문 하나, 출입 제어 장치가 있는 기둥, 검은 창틀과 관목이 보인다. 참고의 재료와 주택가 성격은 유지된다. 이현우 주변에는 경찰 일곱 명이 전경 좌우, 왼쪽 차량 옆, 후방, 오른쪽 담 앞에 나뉘어 있다. 그러나 하단 좌우 경찰과 차량이 매우 크게 잘려 들어오고 이현우도 화면 높이의 상당 부분을 차지해, 요구된 먼 크레인 와이드보다 가까운 포위 장면이다. 실내 로봇과 위층 탈출 창문은 보이지 않아 상태를 판단할 수 없다.",
        "entities": "이현우의 젊은 동아시아계 얼굴, 헝클어진 검은 머리, 마른 체격, 오염된 셔츠와 바지는 참고와 대체로 일치하고 귀의 검은 인이어도 보인다. 화면 오른쪽 바짓단 부근에는 혈흔이 있지만 개에게 물린 상처 형태까지 확인되지는 않는다. 경찰들은 헬멧과 전술복을 착용하고 소총을 들고 있다. 참고의 승용차 외에 여러 경광등 차량이 추가되었으며 전경 경찰 장비에는 영문 표기가 들어가 있다.",
        "hard_violations": [
         "참고에 없는 다수의 경광등 장착 차량과 전경 차량을 추가하여, 허용되지 않은 소품이 포위 공간을 구성한다.",
         "참고나 지시가 표기를 확정하지 않은 경찰 장비에 'POLICE' 등의 문구를 새로 넣었다."
        ],
        "physics": "이현우는 양발을 지면에 붙이고 서 있으며 한쪽 다리에 약간 더 체중을 싣는다. 그래도 착지 후 관성을 멈추는 동작은 약하고 팔을 내린 정지 자세에 가깝다. 후방 경찰은 무릎 굽힘과 발 간격이 서로 달라 버티기와 체중 이동이 A보다 다양하다. 총은 손으로 쥐고 개머리판을 어깨에 대고 있으며, 보이는 인물과 물체에 명백히 지지 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "더 멀고 높은 사선 시점과 작은 전신 크기가 지정된 와이드 구도에 가깝지만, 오른쪽을 보는 시선과 정적인 자세, 추가 차량·문구는 지시와 어긋난다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "경찰의 자세 차이와 다리 혈흔은 더 분명하지만, 전경 경찰이 과대하게 차지하고 이현우도 더 크게 잡혀 크레인 후퇴 종점의 구도에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 화면 오른쪽 경찰 쪽으로 고개와 눈을 돌려, 요구된 왼쪽 프레임 밖 경찰을 보지 않는다. 좌우 가장자리와 아래쪽 경찰의 총은 대체로 이현우의 상체 방향으로 안쪽 대각선을 만든다. 뒤쪽 두 경찰은 카메라 쪽 정면성이 강해 이현우에게 정확히 조준선이 모이는지는 덜 분명하다.",
        "built_space": "회색 콘크리트 담, 어두운 금속 대문, 검은 창틀과 담 위 관목이 주택 뒤편을 사선으로 채워 참고 장소의 주요 재료를 유지한다. 중앙의 큰 대문 하나와 오른쪽 보행 출입구 하나, 그 안 계단이 보인다. 경찰은 후방 두 명, 좌우 가장자리 각 한 명, 하단 전경 두 명이 명확하게 구분된다. 이현우는 중앙보다 약간 왼쪽 도로에 전신으로 서 있다. 두 창문의 개방 상태와 실내 로봇은 이 구도에서 확인할 수 없다. 카메라는 높은 사선 시점이며 B보다 후퇴한 거리감이 강하다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 오염된 어두운 셔츠·갈색 계열 바지가 참고와 대체로 맞는다. 작은 얼굴 때문에 정확한 동일인 여부와 인이어는 확정하기 어렵다. 보이는 다리에는 신선한 개 물림 상처가 뚜렷하지 않다. 경찰은 검은 전술복과 총기를 갖춘 사람으로 표현된다. 참고의 검은 승용차 외에 전경 차량과 경광등 차량 등이 추가되어 있고, 경찰 조끼에는 참고가 정하지 않은 문구가 있다.",
        "hard_violations": [
         "참고에 있는 승용차 외에 여러 차량과 경광등 장착 차량을 추가하여, 허용되지 않은 소품이 전경과 포위 배치의 큰 부분을 차지한다.",
         "참고나 지시가 표기를 확정하지 않은 경찰 조끼에 '경찰' 등의 문구를 새로 넣었다."
        ],
        "physics": "이현우의 양발은 아스팔트에 닿아 체중을 지지한다. 다만 양팔을 내리고 비교적 곧게 서 있어 착지 후 일어서다 관성을 억제하는 순간보다는 정지한 자세로 읽힌다. 전신이 보이는 경찰은 벌린 두 발로 서 있고, 잘린 경찰들의 상체에도 공중에 떠 있다는 증거는 없다. 총기는 손과 어깨로 지지되며 차량도 지면 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우의 시선은 오른쪽 가장자리 경찰을 향하며, 지정된 왼쪽 프레임 밖 목표와 반대다. 왼쪽과 오른쪽 전경 경찰의 총구는 이현우의 몸통 쪽으로 향하고, 오른쪽 후방 경찰도 왼쪽 안쪽을 겨눈다. 왼쪽 후방 경찰들은 오른쪽 안쪽으로 총과 어깨를 돌려 포위 방향을 만든다. 총구가 렌즈를 직접 겨누는 구성은 아니다.",
        "built_space": "긴 회색 콘크리트 담, 검은 금속 대문 하나, 출입 제어 장치가 있는 기둥, 검은 창틀과 관목이 보인다. 참고의 재료와 주택가 성격은 유지된다. 이현우 주변에는 경찰 일곱 명이 전경 좌우, 왼쪽 차량 옆, 후방, 오른쪽 담 앞에 나뉘어 있다. 그러나 하단 좌우 경찰과 차량이 매우 크게 잘려 들어오고 이현우도 화면 높이의 상당 부분을 차지해, 요구된 먼 크레인 와이드보다 가까운 포위 장면이다. 실내 로봇과 위층 탈출 창문은 보이지 않아 상태를 판단할 수 없다.",
        "entities": "이현우의 젊은 동아시아계 얼굴, 헝클어진 검은 머리, 마른 체격, 오염된 셔츠와 바지는 참고와 대체로 일치하고 귀의 검은 인이어도 보인다. 화면 오른쪽 바짓단 부근에는 혈흔이 있지만 개에게 물린 상처 형태까지 확인되지는 않는다. 경찰들은 헬멧과 전술복을 착용하고 소총을 들고 있다. 참고의 승용차 외에 여러 경광등 차량이 추가되었으며 전경 경찰 장비에는 영문 표기가 들어가 있다.",
        "hard_violations": [
         "참고에 없는 다수의 경광등 장착 차량과 전경 차량을 추가하여, 허용되지 않은 소품이 포위 공간을 구성한다.",
         "참고나 지시가 표기를 확정하지 않은 경찰 장비에 'POLICE' 등의 문구를 새로 넣었다."
        ],
        "physics": "이현우는 양발을 지면에 붙이고 서 있으며 한쪽 다리에 약간 더 체중을 싣는다. 그래도 착지 후 관성을 멈추는 동작은 약하고 팔을 내린 정지 자세에 가깝다. 후방 경찰은 무릎 굽힘과 발 간격이 서로 달라 버티기와 체중 이동이 A보다 다양하다. 총은 손으로 쥐고 개머리판을 어깨에 대고 있으며, 보이는 인물과 물체에 명백히 지지 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gpt-high] 참고에 있는 승용차 외에 여러 차량과 경광등 장착 차량을 추가하여, 허용되지 않은 소품이 전경과 포위 배치의 큰 부분을 차지한다.",
     "[gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 조끼에 '경찰' 등의 문구를 새로 넣었다."
    ],
    "A": [
     "[gpt-high] 참고에 없는 다수의 경광등 장착 차량과 전경 차량을 추가하여, 허용되지 않은 소품이 포위 공간을 구성한다.",
     "[gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 장비에 'POLICE' 등의 문구를 새로 넣었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "명시된 인이어 무전기와 다리의 상처를 정확히 묘사하고 레퍼런스의 공간 구조(왼쪽 대문)를 잘 유지하여 지시를 충실히 이행함.  ★위반: [gpt-high] 참고에 없는 다수의 경광등 장착 차량과 전경 차량을 추가하여, 허용되지 않은 소품이 포위 공간을 구성한다. / [gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 장비에 'POLICE' 등의 문구를 새로 넣었다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "필수 요소인 인이어 무전기와 다리 상처가 누락되었고, 공간 구조(오른쪽 대문)를 반전시켜 위치 일관성을 어김.  ★위반: [gpt-high] 참고에 있는 승용차 외에 여러 차량과 경광등 장착 차량을 추가하여, 허용되지 않은 소품이 전경과 포위 배치의 큰 부분을 차지한다. / [gpt-high] 참고나 지시가 표기를 확정하지 않은 경찰 조끼에 '경찰' 등의 문구를 새로 넣었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S4sh11_sel.png",
    "asset_id": "c71dadd2-c3e3-4d7f-82d5-d7131be63dda",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07ee-0875-7158-b6ae-d1627c727c89",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh11"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S5sh68::signage": {
  "fp": "1a12d33d3bb225c6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh68": {
  "input_fingerprint": "25d7beffdcfc4a55",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경찰에게 붙잡힌 이현우가 빈 정차 자리를 향해 경멸 어린 표정으로 욕설을 내뱉는 측면 구도.\n\nLOCATION (lock): Outside the upscale house beside the residential street, facing the now-empty parking position across the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the prescribed descent and diagonal approach into a shoulder-height lateral track, observing 이현우 directly in left profile with his upper body and both officers' grips retained. Place him in the right half, his restrained torso continuing with the escort while his face turns toward the vacant stopping place beyond the left edge, mouth caught mid-curse; the officers look down along their holds rather than toward the lens. Emphasize the reduced camera distance, keeping the established daylight treatment and leaving open road space to the left without inserting 페드로 or a car.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Road beside the house (No departing car or 페드로 visible) — An unoccupied stretch extends toward the left frame edge; used as Provides negative space in the direction of 이현우's accusation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the previous shot's subdued daylight and controlled contrast, allowing the contemptuous expression to carry the emotional change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the house exterior, adjoining ground, and daylight appearance from the reference. Exclude the departed getaway car and its driver; the former parking spot must remain empty.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The former parking space is empty; the getaway car is gone. The two opened house windows and the fallen robot's torn head covering remain unchanged. 이현우: Is being led away under arrest and still has the fresh dog-bite injury to his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경찰에게 붙잡힌 이현우가 빈 정차 자리를 향해 경멸 어린 표정으로 욕설을 내뱉는 측면 구도.\n\nLOCATION (lock): Outside the upscale house beside the residential street, facing the now-empty parking position across the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the prescribed descent and diagonal approach into a shoulder-height lateral track, observing 이현우 directly in left profile with his upper body and both officers' grips retained. Place him in the right half, his restrained torso continuing with the escort while his face turns toward the vacant stopping place beyond the left edge, mouth caught mid-curse; the officers look down along their holds rather than toward the lens. Emphasize the reduced camera distance, keeping the established daylight treatment and leaving open road space to the left without inserting 페드로 or a car.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Road beside the house (No departing car or 페드로 visible) — An unoccupied stretch extends toward the left frame edge; used as Provides negative space in the direction of 이현우's accusation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the previous shot's subdued daylight and controlled contrast, allowing the contemptuous expression to carry the emotional change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the house exterior, adjoining ground, and daylight appearance from the reference. Exclude the departed getaway car and its driver; the former parking spot must remain empty.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The former parking space is empty; the getaway car is gone. The two opened house windows and the fallen robot's torn head covering remain unchanged. 이현우: Is being led away under arrest and still has the fresh dog-bite injury to his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 경찰에게 붙잡힌 이현우가 빈 정차 자리를 향해 경멸 어린 표정으로 욕설을 내뱉는 측면 구도.\n\nLOCATION (lock): Outside the upscale house beside the residential street, facing the now-empty parking position across the road. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the prescribed descent and diagonal approach into a shoulder-height lateral track, observing 이현우 directly in left profile with his upper body and both officers' grips retained. Place him in the right half, his restrained torso continuing with the escort while his face turns toward the vacant stopping place beyond the left edge, mouth caught mid-curse; the officers look down along their holds rather than toward the lens. Emphasize the reduced camera distance, keeping the established daylight treatment and leaving open road space to the left without inserting 페드로 or a car.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Road beside the house (No departing car or 페드로 visible) — An unoccupied stretch extends toward the left frame edge; used as Provides negative space in the direction of 이현우's accusation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the previous shot's subdued daylight and controlled contrast, allowing the contemptuous expression to carry the emotional change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the house exterior, adjoining ground, and daylight appearance from the reference. Exclude the departed getaway car and its driver; the former parking spot must remain empty.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The former parking space is empty; the getaway car is gone. The two opened house windows and the fallen robot's torn head covering remain unchanged. 이현우: Is being led away under arrest and still has the fresh dog-bite injury to his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 왼쪽 빈 도로를 향해 시선을 두고 욕설을 내뱉고 있으며, 두 경찰은 자신들의 결박 부위를 내려다봄.",
    "built_space": "레퍼런스와 동일한 주택가 도로변이며, 프레임 왼쪽에 요구된 빈 도로 공간이 확보됨.",
    "entities": "이현우의 인상착의(의상, 피자국, 인이어)가 레퍼런스와 일치하며, 두 명의 경찰이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "두 경찰이 이현우의 양팔을 붙잡고 이동하는 물리적 제압과 체중 지지가 자연스러움."
   },
   {
    "label": "B",
    "direction": "이현우는 왼쪽을 향해 입을 벌리고 있으나, 뒤쪽 경찰은 결박 부위가 아닌 앞(이현우의 등 부위)을 향해 시선을 둠.",
    "built_space": "레퍼런스와 동일한 주택가 배경이며, 왼쪽에 빈 도로가 올바르게 배치됨.",
    "entities": "이현우의 인상착의는 참조와 유사하게 묘사되었으나, 뒤쪽 경찰의 팔과 손이 전혀 보이지 않음.",
    "hard_violations": [],
    "physics": "앞쪽 경찰만 두 손으로 이현우의 왼팔을 잡고 있으며, 뒤쪽 경찰의 물리적 제압 동작이 누락되어 불안정함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 두 경찰의 결박 상태, 인물의 측면 구도, 빈 도로 여백, 그리고 시선 처리를 모두 매우 정확하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "측면 구도와 배경은 적절하나, 두 경찰의 결박 상태가 모두 보여야 한다는 명시적 지시를 누락함(뒤쪽 경찰의 손이 없음)."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽 빈 도로를 향해 시선을 두고 욕설을 내뱉고 있으며, 두 경찰은 자신들의 결박 부위를 내려다봄.",
        "built_space": "레퍼런스와 동일한 주택가 도로변이며, 프레임 왼쪽에 요구된 빈 도로 공간이 확보됨.",
        "entities": "이현우의 인상착의(의상, 피자국, 인이어)가 레퍼런스와 일치하며, 두 명의 경찰이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "두 경찰이 이현우의 양팔을 붙잡고 이동하는 물리적 제압과 체중 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽을 향해 입을 벌리고 있으나, 뒤쪽 경찰은 결박 부위가 아닌 앞(이현우의 등 부위)을 향해 시선을 둠.",
        "built_space": "레퍼런스와 동일한 주택가 배경이며, 왼쪽에 빈 도로가 올바르게 배치됨.",
        "entities": "이현우의 인상착의는 참조와 유사하게 묘사되었으나, 뒤쪽 경찰의 팔과 손이 전혀 보이지 않음.",
        "hard_violations": [],
        "physics": "앞쪽 경찰만 두 손으로 이현우의 왼팔을 잡고 있으며, 뒤쪽 경찰의 물리적 제압 동작이 누락되어 불안정함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 두 경찰의 결박 상태, 인물의 측면 구도, 빈 도로 여백, 그리고 시선 처리를 모두 매우 정확하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "측면 구도와 배경은 적절하나, 두 경찰의 결박 상태가 모두 보여야 한다는 명시적 지시를 누락함(뒤쪽 경찰의 손이 없음)."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽 빈 도로를 향해 시선을 두고 욕설을 내뱉고 있으며, 두 경찰은 자신들의 결박 부위를 내려다봄.",
        "built_space": "레퍼런스와 동일한 주택가 도로변이며, 프레임 왼쪽에 요구된 빈 도로 공간이 확보됨.",
        "entities": "이현우의 인상착의(의상, 피자국, 인이어)가 레퍼런스와 일치하며, 두 명의 경찰이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "두 경찰이 이현우의 양팔을 붙잡고 이동하는 물리적 제압과 체중 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽을 향해 입을 벌리고 있으나, 뒤쪽 경찰은 결박 부위가 아닌 앞(이현우의 등 부위)을 향해 시선을 둠.",
        "built_space": "레퍼런스와 동일한 주택가 배경이며, 왼쪽에 빈 도로가 올바르게 배치됨.",
        "entities": "이현우의 인상착의는 참조와 유사하게 묘사되었으나, 뒤쪽 경찰의 팔과 손이 전혀 보이지 않음.",
        "hard_violations": [],
        "physics": "앞쪽 경찰만 두 손으로 이현우의 왼팔을 잡고 있으며, 뒤쪽 경찰의 물리적 제압 동작이 누락되어 불안정함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 인물 배치와 왼쪽 빈 도로, 욕설하는 표정은 충실하지만 두 경찰이 뒤쪽에 겹쳐 먼 쪽 경찰의 제압 접촉이 가려진다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "가까운 어깨 높이의 측면 미디엄 구도에서 왼쪽을 향한 욕설과 양쪽 경찰의 팔 제압, 아래로 향한 경찰 시선을 가장 명확하게 실현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽 바깥을 바라보며 입을 벌리고 있어 빈 정차 자리를 향한 비난으로 읽힌다. 왼쪽 도로에는 차나 사람이 없다. 두 경찰은 렌즈가 아니라 붙잡은 팔 쪽으로 고개를 숙인다. 몸통은 앞으로 기울고 팔은 뒤로 제압되어 있으나 실제 호송 진행 방향은 정지 화면만으로 확정하기 어렵다.",
        "built_space": "오른쪽 절반에 이현우와 경찰 두 명이 모여 있고 왼쪽에는 넓은 도로 여백이 남는다. 배경에는 큰 짙은색 대문 한 곳, 그 왼쪽의 좁은 출입문 한 곳, 기둥의 작은 설비 두 개, 밝은 석재 담장과 화단, 보도 및 노란 도로 가장자리 선이 보인다. 참고 장소의 주요 재료와 배치가 이어진다. 상부 창은 일부만 보여 열린 창 두 개의 상태는 확인할 수 없다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "이현우 한 명과 검은 전술복·헬멧을 착용한 경찰 두 명이 보인다. 이현우는 앳된 동아시아계 남성 외모, 마른 체격, 헝클어진 짧은 검은 머리, 귀의 검은 인이어, 피와 먼지가 묻은 어두운 셔츠 및 갈색 계열 바지와 벨트를 갖춰 참고 인물과 대체로 일치한다. 국적은 외형만으로 확인할 수 없다. 페드로와 차량은 없다. 다리의 개 물림 상처와 쓰러진 로봇의 머리 덮개는 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "가까운 경찰의 장갑 낀 손이 이현우의 위팔을 확실히 붙잡고 있다. 다른 경찰의 제압 손과 접촉점은 겹친 몸과 팔 뒤에 대부분 가려져 두 사람의 그립을 각각 명료하게 확인하기 어렵다. 이현우의 앞으로 기운 상체와 뒤로 잡힌 팔은 물리적으로 가능한 제압 자세다. 하체는 아래로 이어지며 발은 프레임 밖이고, 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 얼굴은 선명한 왼쪽 측면이며 눈과 열린 입이 왼쪽 프레임 밖의 빈 도로 방향을 향한다. 경찰 두 명은 각자 붙잡고 있는 팔 쪽을 내려다보고 렌즈를 보지 않는다. 이현우의 몸통은 경찰 사이에서 제압된 채 얼굴만 왼쪽으로 돌린 관계가 잘 드러난다. 화면 안에 조준하는 무기는 없다.",
        "built_space": "어깨 높이의 가까운 미디엄 구도로 이현우가 오른쪽 절반을 차지하고 왼쪽 도로가 비어 있다. 경찰은 그의 양옆에서 팔에 접근한다. 뒤에는 큰 짙은색 대문 한 곳, 그 왼쪽의 좁은 출입문 한 곳, 작은 기둥 설비 두 개, 석재 담장과 화단, 보도 및 노란 가장자리 선이 이어져 참고 장소와 부합한다. 창 일부는 보이지만 두 창의 개방 상태까지 판독되지는 않는다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "이현우와 경찰 두 명만 보이며 추가 인물이나 차량은 없다. 이현우의 앳된 동아시아계 남성 얼굴, 마른 체격, 헝클어진 검은 머리, 검은 인이어, 피와 먼지가 묻은 어두운 셔츠, 갈색 벨트와 바지가 참고와 대체로 일치한다. 경찰 두 명은 검은 헬멧과 전술복을 착용한다. 프레임 밖의 다리 상처 및 로봇 머리 덮개 상태는 판단할 수 없으며, 페드로는 등장하지 않는다.",
        "hard_violations": [],
        "physics": "화면 중앙 경찰의 장갑 낀 손은 이현우의 먼 쪽 위팔을 잡고, 오른쪽 경찰의 손은 가까운 쪽 위팔을 감싸 양쪽 제압 접촉이 모두 보인다. 팔과 어깨의 굽힘은 해당 제압으로 가능한 자세이며 몸통도 자연스럽게 연결된다. 발은 프레임 밖이지만 보이는 하체와 경찰의 벌어진 다리는 서서 호송하는 자세에 부합한다. 지지 없이 떠 있는 사람이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 인물 배치와 왼쪽 빈 도로, 욕설하는 표정은 충실하지만 두 경찰이 뒤쪽에 겹쳐 먼 쪽 경찰의 제압 접촉이 가려진다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "가까운 어깨 높이의 측면 미디엄 구도에서 왼쪽을 향한 욕설과 양쪽 경찰의 팔 제압, 아래로 향한 경찰 시선을 가장 명확하게 실현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 화면 왼쪽 바깥을 바라보며 입을 벌리고 있어 빈 정차 자리를 향한 비난으로 읽힌다. 왼쪽 도로에는 차나 사람이 없다. 두 경찰은 렌즈가 아니라 붙잡은 팔 쪽으로 고개를 숙인다. 몸통은 앞으로 기울고 팔은 뒤로 제압되어 있으나 실제 호송 진행 방향은 정지 화면만으로 확정하기 어렵다.",
        "built_space": "오른쪽 절반에 이현우와 경찰 두 명이 모여 있고 왼쪽에는 넓은 도로 여백이 남는다. 배경에는 큰 짙은색 대문 한 곳, 그 왼쪽의 좁은 출입문 한 곳, 기둥의 작은 설비 두 개, 밝은 석재 담장과 화단, 보도 및 노란 도로 가장자리 선이 보인다. 참고 장소의 주요 재료와 배치가 이어진다. 상부 창은 일부만 보여 열린 창 두 개의 상태는 확인할 수 없다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "이현우 한 명과 검은 전술복·헬멧을 착용한 경찰 두 명이 보인다. 이현우는 앳된 동아시아계 남성 외모, 마른 체격, 헝클어진 짧은 검은 머리, 귀의 검은 인이어, 피와 먼지가 묻은 어두운 셔츠 및 갈색 계열 바지와 벨트를 갖춰 참고 인물과 대체로 일치한다. 국적은 외형만으로 확인할 수 없다. 페드로와 차량은 없다. 다리의 개 물림 상처와 쓰러진 로봇의 머리 덮개는 프레임에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "가까운 경찰의 장갑 낀 손이 이현우의 위팔을 확실히 붙잡고 있다. 다른 경찰의 제압 손과 접촉점은 겹친 몸과 팔 뒤에 대부분 가려져 두 사람의 그립을 각각 명료하게 확인하기 어렵다. 이현우의 앞으로 기운 상체와 뒤로 잡힌 팔은 물리적으로 가능한 제압 자세다. 하체는 아래로 이어지며 발은 프레임 밖이고, 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 얼굴은 선명한 왼쪽 측면이며 눈과 열린 입이 왼쪽 프레임 밖의 빈 도로 방향을 향한다. 경찰 두 명은 각자 붙잡고 있는 팔 쪽을 내려다보고 렌즈를 보지 않는다. 이현우의 몸통은 경찰 사이에서 제압된 채 얼굴만 왼쪽으로 돌린 관계가 잘 드러난다. 화면 안에 조준하는 무기는 없다.",
        "built_space": "어깨 높이의 가까운 미디엄 구도로 이현우가 오른쪽 절반을 차지하고 왼쪽 도로가 비어 있다. 경찰은 그의 양옆에서 팔에 접근한다. 뒤에는 큰 짙은색 대문 한 곳, 그 왼쪽의 좁은 출입문 한 곳, 작은 기둥 설비 두 개, 석재 담장과 화단, 보도 및 노란 가장자리 선이 이어져 참고 장소와 부합한다. 창 일부는 보이지만 두 창의 개방 상태까지 판독되지는 않는다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "이현우와 경찰 두 명만 보이며 추가 인물이나 차량은 없다. 이현우의 앳된 동아시아계 남성 얼굴, 마른 체격, 헝클어진 검은 머리, 검은 인이어, 피와 먼지가 묻은 어두운 셔츠, 갈색 벨트와 바지가 참고와 대체로 일치한다. 경찰 두 명은 검은 헬멧과 전술복을 착용한다. 프레임 밖의 다리 상처 및 로봇 머리 덮개 상태는 판단할 수 없으며, 페드로는 등장하지 않는다.",
        "hard_violations": [],
        "physics": "화면 중앙 경찰의 장갑 낀 손은 이현우의 먼 쪽 위팔을 잡고, 오른쪽 경찰의 손은 가까운 쪽 위팔을 감싸 양쪽 제압 접촉이 모두 보인다. 팔과 어깨의 굽힘은 해당 제압으로 가능한 자세이며 몸통도 자연스럽게 연결된다. 발은 프레임 밖이지만 보이는 하체와 경찰의 벌어진 다리는 서서 호송하는 자세에 부합한다. 지지 없이 떠 있는 사람이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.46
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.46
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1460
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 두 경찰의 결박 상태, 인물의 측면 구도, 빈 도로 여백, 그리고 시선 처리를 모두 매우 정확하게 구현함."
   },
   {
    "label": "B",
    "score": 1460,
    "verdict_ko": "측면 구도와 배경은 적절하나, 두 경찰의 결박 상태가 모두 보여야 한다는 명시적 지시를 누락함(뒤쪽 경찰의 손이 없음)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S5sh62_sel.png",
    "asset_id": "4a27f2dc-f708-411c-af22-eea9ecda0442",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07f5-2b52-70d1-bbf1-408c6197b973",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh62"
  }
 },
 "S6sh15::signage": {
  "fp": "13cba21b6283bc28",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::973d94f3b2b9b284": {
  "subjects": [],
  "subject_text": "인천 난민촌 쓰레기장\n폐가전과 고철, 로봇 부품이 뒤섞여 거대한 산처럼 쌓인 야외 쓰레기장. 둔덕과 폐기물 사이로 좁은 틈과 작업할 바닥이 드러난다.",
  "identity": "canonical",
  "scope_id": "L13",
  "scope_role": "location_exterior",
  "scope_sha": "aefe3b3a7f9672a8"
 },
 "groupbg::scrap_heap_encounter": {
  "input_fingerprint": "d1a69d72038bbb3a",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "scrap_heap_encounter",
    "tags": [
     "S6sh15",
     "S6sh20"
    ]
   },
   "context_sig": "a6da3a661cd7267c"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the surface of a towering rubbish heap in an open-air refugee-settlement dump.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 쓰레기장: 난민 구역 외곽에 위치한 거대한 쓰레기 산으로, 온갖 고철, 모포, 낡은 가전이 널려 있다. (특징: 녹슬고 부서진 헬리콥터 잔해 및 거대한 폐품 산; 먼지를 막는 마스크와 수선된 헌 옷을 입은 앰버(한국계 백인 혼혈 여자아이)와 라울(라틴계 흑인 혼혈 남자아이); 공구 주머니와 렌치를 들고 로봇 팔을 해체하는 앰버; 부서진 구형 주크박스와 널브러진 LP판들; 쓰레기 더미를 뚫고 일어나는 고릴라형 외형의 기계 관절 로봇 찰리 (푸른 발광 눈 장착); 외투와 모자를 걸쳐 사람처럼 위장한 찰리의 뚱뚱한 실루엣)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쓰레기 더미에서 몸을 서서히 드러내는 건.... 고릴라 모양의 로봇이다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the surface of a towering rubbish heap in an open-air refugee-settlement dump.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 쓰레기장: 난민 구역 외곽에 위치한 거대한 쓰레기 산으로, 온갖 고철, 모포, 낡은 가전이 널려 있다. (특징: 녹슬고 부서진 헬리콥터 잔해 및 거대한 폐품 산; 먼지를 막는 마스크와 수선된 헌 옷을 입은 앰버(한국계 백인 혼혈 여자아이)와 라울(라틴계 흑인 혼혈 남자아이); 공구 주머니와 렌치를 들고 로봇 팔을 해체하는 앰버; 부서진 구형 주크박스와 널브러진 LP판들; 쓰레기 더미를 뚫고 일어나는 고릴라형 외형의 기계 관절 로봇 찰리 (푸른 발광 눈 장착); 외투와 모자를 걸쳐 사람처럼 위장한 찰리의 뚱뚱한 실루엣)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쓰레기 더미에서 몸을 서서히 드러내는 건.... 고릴라 모양의 로봇이다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrap_heap_encounter_ab6370.png",
  "asset_id": "c12f180b-34fd-4d2d-a22f-0493ef01b519",
  "input_asset_ids": [
   "70df02ad-c681-4737-b149-35ce29562987"
  ],
  "origin_tag": "S6sh15",
  "place_text": "At the surface of a towering rubbish heap in an open-air refugee-settlement dump.",
  "origin_inputs": {
   "place_text": "At the surface of a towering rubbish heap in an open-air refugee-settlement dump.",
   "time_of_day_en": "day",
   "conti_asset_id": "70df02ad-c681-4737-b149-35ce29562987"
  }
 },
 "S6sh15::bgfirst_bg": {
  "input_fingerprint": "8bfcfaf920c6cffe",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 밖으로 육중한 고릴라 형태의 상체를 완전히 드러낸 찰리 주위로 흩뿌려진 파편들이 공중에 떠 있는 정면 구도.\n\nLOCATION (lock): At the surface of a towering rubbish heap in an open-air refugee-settlement dump.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from close to ground level beside the fallen children, looking upward along the inherited diagonal rather than Charlie's frontal axis as the dolly retreat briefly steadies. Place 찰리's fully exposed gorilla-shaped upper body in the upper middle, occupying less than half the image, with masked 앰버 and 라울 partially visible at the lower corners, leaning back on the rubbish and looking up while 찰리 looks down toward them. Preserve 앰버 as a ten-year-old Korean–white mixed-heritage girl and 라울 as a ten-year-old Latino–Black mixed-heritage boy, emphasizing 찰리's rising position without suspended debris or an added particle effect.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heap (Parted around 찰리's emerging body); used as Frames the exposed torso from below and supports the children's lower-edge placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight preserves precise robot contours, with the explicitly blue illumination confined to 찰리's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 밖으로 육중한 고릴라 형태의 상체를 완전히 드러낸 찰리 주위로 흩뿌려진 파편들이 공중에 떠 있는 정면 구도.\n\nLOCATION (lock): At the surface of a towering rubbish heap in an open-air refugee-settlement dump.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from close to ground level beside the fallen children, looking upward along the inherited diagonal rather than Charlie's frontal axis as the dolly retreat briefly steadies. Place 찰리's fully exposed gorilla-shaped upper body in the upper middle, occupying less than half the image, with masked 앰버 and 라울 partially visible at the lower corners, leaning back on the rubbish and looking up while 찰리 looks down toward them. Preserve 앰버 as a ten-year-old Korean–white mixed-heritage girl and 라울 as a ten-year-old Latino–Black mixed-heritage boy, emphasizing 찰리's rising position without suspended debris or an added particle effect.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heap (Parted around 찰리's emerging body); used as Frames the exposed torso from below and supports the children's lower-edge placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight preserves precise robot contours, with the explicitly blue illumination confined to 찰리's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh15__bgfirst_bg.png",
  "asset_id": "7ff350d0-f4f3-4922-b759-f2a2ca7fc8a6",
  "input_asset_ids": [
   "70df02ad-c681-4737-b149-35ce29562987",
   "c12f180b-34fd-4d2d-a22f-0493ef01b519"
  ]
 },
 "S6sh15": {
  "input_fingerprint": "98ea3e59510caa84",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖으로 육중한 고릴라 형태의 상체를 완전히 드러낸 찰리 주위로 흩뿌려진 파편들이 공중에 떠 있는 정면 구도.\n\nLOCATION (lock): At the surface of a towering rubbish heap in an open-air refugee-settlement dump. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from close to ground level beside the fallen children, looking upward along the inherited diagonal rather than Charlie's frontal axis as the dolly retreat briefly steadies. Place 찰리's fully exposed gorilla-shaped upper body in the upper middle, occupying less than half the image, with masked 앰버 and 라울 partially visible at the lower corners, leaning back on the rubbish and looking up while 찰리 looks down toward them. Preserve 앰버 as a ten-year-old Korean–white mixed-heritage girl and 라울 as a ten-year-old Latino–Black mixed-heritage boy, emphasizing 찰리's rising position without suspended debris or an added particle effect.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heap (Parted around 찰리's emerging body); used as Frames the exposed torso from below and supports the children's lower-edge placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight preserves precise robot contours, with the explicitly blue illumination confined to 찰리's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish dump contains a blanket with jukebox parts, tools, and LP records, and the reassembled jukebox has its horn-shaped speaker attached. Charlie's dirty, old gorilla-shaped body is emerging from the rubbish, retaining its netting and worn UBIC chest logo. 앰버: Wears a mask and a tool pouch at her waist; she has fallen backward in fright. 라울: Has fallen backward beside the rubbish heap and remains frightened.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖으로 육중한 고릴라 형태의 상체를 완전히 드러낸 찰리 주위로 흩뿌려진 파편들이 공중에 떠 있는 정면 구도.\n\nLOCATION (lock): At the surface of a towering rubbish heap in an open-air refugee-settlement dump. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from close to ground level beside the fallen children, looking upward along the inherited diagonal rather than Charlie's frontal axis as the dolly retreat briefly steadies. Place 찰리's fully exposed gorilla-shaped upper body in the upper middle, occupying less than half the image, with masked 앰버 and 라울 partially visible at the lower corners, leaning back on the rubbish and looking up while 찰리 looks down toward them. Preserve 앰버 as a ten-year-old Korean–white mixed-heritage girl and 라울 as a ten-year-old Latino–Black mixed-heritage boy, emphasizing 찰리's rising position without suspended debris or an added particle effect.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heap (Parted around 찰리's emerging body); used as Frames the exposed torso from below and supports the children's lower-edge placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight preserves precise robot contours, with the explicitly blue illumination confined to 찰리's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish dump contains a blanket with jukebox parts, tools, and LP records, and the reassembled jukebox has its horn-shaped speaker attached. Charlie's dirty, old gorilla-shaped body is emerging from the rubbish, retaining its netting and worn UBIC chest logo. 앰버: Wears a mask and a tool pouch at her waist; she has fallen backward in fright. 라울: Has fallen backward beside the rubbish heap and remains frightened.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 밖으로 육중한 고릴라 형태의 상체를 완전히 드러낸 찰리 주위로 흩뿌려진 파편들이 공중에 떠 있는 정면 구도.\n\nLOCATION (lock): At the surface of a towering rubbish heap in an open-air refugee-settlement dump. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from close to ground level beside the fallen children, looking upward along the inherited diagonal rather than Charlie's frontal axis as the dolly retreat briefly steadies. Place 찰리's fully exposed gorilla-shaped upper body in the upper middle, occupying less than half the image, with masked 앰버 and 라울 partially visible at the lower corners, leaning back on the rubbish and looking up while 찰리 looks down toward them. Preserve 앰버 as a ten-year-old Korean–white mixed-heritage girl and 라울 as a ten-year-old Latino–Black mixed-heritage boy, emphasizing 찰리's rising position without suspended debris or an added particle effect.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heap (Parted around 찰리's emerging body); used as Frames the exposed torso from below and supports the children's lower-edge placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight preserves precise robot contours, with the explicitly blue illumination confined to 찰리's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish dump contains a blanket with jukebox parts, tools, and LP records, and the reassembled jukebox has its horn-shaped speaker attached. Charlie's dirty, old gorilla-shaped body is emerging from the rubbish, retaining its netting and worn UBIC chest logo. 앰버: Wears a mask and a tool pouch at her waist; she has fallen backward in fright. 라울: Has fallen backward beside the rubbish heap and remains frightened.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh15__bgfirst_bg.png",
     "asset_id": "7ff350d0-f4f3-4922-b759-f2a2ca7fc8a6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S6sh15.png",
     "asset_id": "70df02ad-c681-4737-b149-35ce29562987",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrap_heap_encounter_ab6370.png",
     "asset_id": "c12f180b-34fd-4d2d-a22f-0493ef01b519",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "왼쪽 앰버와 오른쪽 라울은 고개를 위로 들어 중앙의 찰리를 바라본다. 찰리의 얼굴과 눈은 거의 수평으로 카메라 쪽을 향해, 아래에 누운 아이들에게 시선이 내려가는 관계가 뚜렷하지 않다. 무기나 손에 든 지향성 물건은 없다. 오른쪽 주크박스의 혼 입구는 오른쪽 위를 향한다.",
    "built_space": "야외 쓰레기 경사면 왼쪽 위에 폐헬기 한 대, 오른쪽에 혼 하나가 붙은 주크박스 한 대, 아래 중앙 왼쪽에 열린 공구함 하나가 보인다. 중앙 아래에는 푸른 천과 여러 장의 LP가 놓여 있어 장소 참조의 주요 배치와 재료를 유지한다. 찰리의 허리 주변을 쓰레기가 둘러싸고, 아이들은 양쪽 하단 경사면에 기대어 있다. 카메라는 아이들 가까이에서 올려다보지만 찰리의 정면 축에 가까워 요구된 비스듬한 접근은 약하다. 찰리의 노출 상체는 상단 중앙에 있고 화면 면적의 절반 미만이다.",
    "entities": "찰리 한 기와 아이 둘만 보인다. 찰리는 낡은 베이지 장갑, 긴 육중한 팔, 푸른 눈, 걸친 그물과 읽을 수 있는 UBIC 가슴 표시를 갖췄다. 그러나 얼굴은 콧구멍과 턱이 조형된 사실적인 고릴라형으로, 인물 참조의 매끈한 흰 마스크와 점 두 개·입 선 형태가 아니다. 원형 가슴 원자로도 보이지 않는다. 앰버는 금발의 어린 여자아이로 보이며 마스크와 허리 공구 주머니가 있다. 라울은 짙은 피부와 뒤로 모은 곱슬머리의 어린 남자아이로 보이지만 마스크는 확인되지 않는다. 두 아이의 얼굴이 측후면으로 가려져 정확한 혼혈 외모와 참조 얼굴 일치까지는 확인하기 어렵다. 의복은 참조의 단색 티셔츠 대신 낡은 겉옷이다.",
    "hard_violations": [],
    "physics": "아이들은 쓰레기 위에 엉덩이와 몸통을 기대고 무릎을 굽힌 자세로, 뒤로 넘어진 뒤 버티는 상태가 가능하다. 찰리의 아래 몸통은 더미 속에 이어지고 왼쪽 화면의 손은 잔해에 닿아 있어 떠 있는 몸으로 보이지 않는다. 주크박스와 공구함, LP는 잔해나 천 위에 놓여 있고 그물은 장갑에 걸려 늘어진다. 지지 없이 공중에 멈춘 파편이나 추가 입자 효과는 보이지 않는다."
   },
   {
    "label": "A",
    "direction": "양쪽 아래의 아이들은 중앙 위 찰리를 올려다본다. 찰리는 목과 얼굴을 아래로 기울여 두 아이가 있는 전경 쪽을 향하므로, 요구된 내려다보는 관계가 A보다 명확하다. 개별 아이 중 누구를 응시하는지는 점 형태의 눈만으로 확정하기 어렵다. 무기나 손에 든 지향성 물건은 없고, 주크박스의 혼 입구는 오른쪽 위를 향한다.",
    "built_space": "왼쪽 위 폐헬기 한 대, 오른쪽의 주크박스 한 대와 부착된 혼 하나, 아래 왼쪽 중앙의 열린 공구함 하나가 보인다. 중앙 전경의 푸른 천과 LP 여러 장, 겹겹의 녹슨 금속 잔해가 장소 참조와 잘 대응한다. 찰리는 갈라진 쓰레기 더미 사이에서 상체를 드러내고 상단 중앙에서 화면 절반 미만을 차지한다. 아이들은 양쪽 하단에 부분적으로 잘려 있으며 쓰레기에 뒤로 기대어 있다. 지면 가까운 상향 시점은 맞지만 중앙 정면에 가까운 축이고, 주변 잔해가 많이 보여 요구된 미디엄 숏보다 다소 넓게 읽힌다.",
    "entities": "찰리 한 기, 앰버와 라울 두 아이만 있다. 찰리의 흰 마스크형 얼굴, 두 원형 눈과 짧은 입 선, 각진 베이지 장갑과 굵고 긴 팔은 인물 참조에 더 가깝다. 낡은 표면과 그물도 보인다. 원형 가슴 원자로는 있지만 강하게 푸르게 발광하여 청색 조명을 눈에만 한정하라는 지시와 충돌한다. UBIC 문구는 확인되지 않는다. 앰버는 금발을 묶은 어린 여자아이로 마스크를 착용했고, 라울은 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보인다. 라울의 마스크는 확인되지 않으며, 측후면이라 두 아이의 정확한 참조 얼굴과 혼혈 외모는 판별에 제한이 있다. 두 아이 모두 참조의 티셔츠 대신 낡은 겉옷을 입었다. 앰버의 허리 공구 주머니는 가려져 확인하기 어렵다.",
    "hard_violations": [],
    "physics": "아이들의 엉덩이와 등은 쓰레기 경사면에 받쳐져 있고 다리는 앞쪽으로 굽혀져 있어 뒤로 넘어진 자세가 성립한다. 찰리의 아래 몸통은 잔해 속에 묻혀 있고 양팔이 주변 잔해 쪽으로 내려와 있어 상승 중 몸을 지탱하는 구성이 가능하다. 그물은 어깨와 가슴에 걸려 중력 방향으로 늘어진다. 주크박스는 잔해에 기대어 있고 혼은 상단에 부착되어 있으며 공구함과 음반도 바닥 재료 위에 놓여 있다. 떠 있는 몸이나 지지 없는 물체, 공중 파편 효과는 보이지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낮은 시점과 하단의 넘어진 아이들, 그물·UBIC 표시는 맞지만, 찰리가 아이들을 내려다보기보다 정면을 보며 흰 마스크형 얼굴도 인물 참조와 다르다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "아이들을 향해 숙인 찰리의 머리와 참조에 가까운 기계 얼굴, 상단 중앙 배치가 더 충실하지만, 가슴의 청색 발광과 정면에 가까운 카메라 축은 지시에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 앰버와 오른쪽 라울은 고개를 위로 들어 중앙의 찰리를 바라본다. 찰리의 얼굴과 눈은 거의 수평으로 카메라 쪽을 향해, 아래에 누운 아이들에게 시선이 내려가는 관계가 뚜렷하지 않다. 무기나 손에 든 지향성 물건은 없다. 오른쪽 주크박스의 혼 입구는 오른쪽 위를 향한다.",
        "built_space": "야외 쓰레기 경사면 왼쪽 위에 폐헬기 한 대, 오른쪽에 혼 하나가 붙은 주크박스 한 대, 아래 중앙 왼쪽에 열린 공구함 하나가 보인다. 중앙 아래에는 푸른 천과 여러 장의 LP가 놓여 있어 장소 참조의 주요 배치와 재료를 유지한다. 찰리의 허리 주변을 쓰레기가 둘러싸고, 아이들은 양쪽 하단 경사면에 기대어 있다. 카메라는 아이들 가까이에서 올려다보지만 찰리의 정면 축에 가까워 요구된 비스듬한 접근은 약하다. 찰리의 노출 상체는 상단 중앙에 있고 화면 면적의 절반 미만이다.",
        "entities": "찰리 한 기와 아이 둘만 보인다. 찰리는 낡은 베이지 장갑, 긴 육중한 팔, 푸른 눈, 걸친 그물과 읽을 수 있는 UBIC 가슴 표시를 갖췄다. 그러나 얼굴은 콧구멍과 턱이 조형된 사실적인 고릴라형으로, 인물 참조의 매끈한 흰 마스크와 점 두 개·입 선 형태가 아니다. 원형 가슴 원자로도 보이지 않는다. 앰버는 금발의 어린 여자아이로 보이며 마스크와 허리 공구 주머니가 있다. 라울은 짙은 피부와 뒤로 모은 곱슬머리의 어린 남자아이로 보이지만 마스크는 확인되지 않는다. 두 아이의 얼굴이 측후면으로 가려져 정확한 혼혈 외모와 참조 얼굴 일치까지는 확인하기 어렵다. 의복은 참조의 단색 티셔츠 대신 낡은 겉옷이다.",
        "hard_violations": [],
        "physics": "아이들은 쓰레기 위에 엉덩이와 몸통을 기대고 무릎을 굽힌 자세로, 뒤로 넘어진 뒤 버티는 상태가 가능하다. 찰리의 아래 몸통은 더미 속에 이어지고 왼쪽 화면의 손은 잔해에 닿아 있어 떠 있는 몸으로 보이지 않는다. 주크박스와 공구함, LP는 잔해나 천 위에 놓여 있고 그물은 장갑에 걸려 늘어진다. 지지 없이 공중에 멈춘 파편이나 추가 입자 효과는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "양쪽 아래의 아이들은 중앙 위 찰리를 올려다본다. 찰리는 목과 얼굴을 아래로 기울여 두 아이가 있는 전경 쪽을 향하므로, 요구된 내려다보는 관계가 A보다 명확하다. 개별 아이 중 누구를 응시하는지는 점 형태의 눈만으로 확정하기 어렵다. 무기나 손에 든 지향성 물건은 없고, 주크박스의 혼 입구는 오른쪽 위를 향한다.",
        "built_space": "왼쪽 위 폐헬기 한 대, 오른쪽의 주크박스 한 대와 부착된 혼 하나, 아래 왼쪽 중앙의 열린 공구함 하나가 보인다. 중앙 전경의 푸른 천과 LP 여러 장, 겹겹의 녹슨 금속 잔해가 장소 참조와 잘 대응한다. 찰리는 갈라진 쓰레기 더미 사이에서 상체를 드러내고 상단 중앙에서 화면 절반 미만을 차지한다. 아이들은 양쪽 하단에 부분적으로 잘려 있으며 쓰레기에 뒤로 기대어 있다. 지면 가까운 상향 시점은 맞지만 중앙 정면에 가까운 축이고, 주변 잔해가 많이 보여 요구된 미디엄 숏보다 다소 넓게 읽힌다.",
        "entities": "찰리 한 기, 앰버와 라울 두 아이만 있다. 찰리의 흰 마스크형 얼굴, 두 원형 눈과 짧은 입 선, 각진 베이지 장갑과 굵고 긴 팔은 인물 참조에 더 가깝다. 낡은 표면과 그물도 보인다. 원형 가슴 원자로는 있지만 강하게 푸르게 발광하여 청색 조명을 눈에만 한정하라는 지시와 충돌한다. UBIC 문구는 확인되지 않는다. 앰버는 금발을 묶은 어린 여자아이로 마스크를 착용했고, 라울은 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보인다. 라울의 마스크는 확인되지 않으며, 측후면이라 두 아이의 정확한 참조 얼굴과 혼혈 외모는 판별에 제한이 있다. 두 아이 모두 참조의 티셔츠 대신 낡은 겉옷을 입었다. 앰버의 허리 공구 주머니는 가려져 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "아이들의 엉덩이와 등은 쓰레기 경사면에 받쳐져 있고 다리는 앞쪽으로 굽혀져 있어 뒤로 넘어진 자세가 성립한다. 찰리의 아래 몸통은 잔해 속에 묻혀 있고 양팔이 주변 잔해 쪽으로 내려와 있어 상승 중 몸을 지탱하는 구성이 가능하다. 그물은 어깨와 가슴에 걸려 중력 방향으로 늘어진다. 주크박스는 잔해에 기대어 있고 혼은 상단에 부착되어 있으며 공구함과 음반도 바닥 재료 위에 놓여 있다. 떠 있는 몸이나 지지 없는 물체, 공중 파편 효과는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낮은 시점과 하단의 넘어진 아이들, 그물·UBIC 표시는 맞지만, 찰리가 아이들을 내려다보기보다 정면을 보며 흰 마스크형 얼굴도 인물 참조와 다르다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "아이들을 향해 숙인 찰리의 머리와 참조에 가까운 기계 얼굴, 상단 중앙 배치가 더 충실하지만, 가슴의 청색 발광과 정면에 가까운 카메라 축은 지시에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 앰버와 오른쪽 라울은 고개를 위로 들어 중앙의 찰리를 바라본다. 찰리의 얼굴과 눈은 거의 수평으로 카메라 쪽을 향해, 아래에 누운 아이들에게 시선이 내려가는 관계가 뚜렷하지 않다. 무기나 손에 든 지향성 물건은 없다. 오른쪽 주크박스의 혼 입구는 오른쪽 위를 향한다.",
        "built_space": "야외 쓰레기 경사면 왼쪽 위에 폐헬기 한 대, 오른쪽에 혼 하나가 붙은 주크박스 한 대, 아래 중앙 왼쪽에 열린 공구함 하나가 보인다. 중앙 아래에는 푸른 천과 여러 장의 LP가 놓여 있어 장소 참조의 주요 배치와 재료를 유지한다. 찰리의 허리 주변을 쓰레기가 둘러싸고, 아이들은 양쪽 하단 경사면에 기대어 있다. 카메라는 아이들 가까이에서 올려다보지만 찰리의 정면 축에 가까워 요구된 비스듬한 접근은 약하다. 찰리의 노출 상체는 상단 중앙에 있고 화면 면적의 절반 미만이다.",
        "entities": "찰리 한 기와 아이 둘만 보인다. 찰리는 낡은 베이지 장갑, 긴 육중한 팔, 푸른 눈, 걸친 그물과 읽을 수 있는 UBIC 가슴 표시를 갖췄다. 그러나 얼굴은 콧구멍과 턱이 조형된 사실적인 고릴라형으로, 인물 참조의 매끈한 흰 마스크와 점 두 개·입 선 형태가 아니다. 원형 가슴 원자로도 보이지 않는다. 앰버는 금발의 어린 여자아이로 보이며 마스크와 허리 공구 주머니가 있다. 라울은 짙은 피부와 뒤로 모은 곱슬머리의 어린 남자아이로 보이지만 마스크는 확인되지 않는다. 두 아이의 얼굴이 측후면으로 가려져 정확한 혼혈 외모와 참조 얼굴 일치까지는 확인하기 어렵다. 의복은 참조의 단색 티셔츠 대신 낡은 겉옷이다.",
        "hard_violations": [],
        "physics": "아이들은 쓰레기 위에 엉덩이와 몸통을 기대고 무릎을 굽힌 자세로, 뒤로 넘어진 뒤 버티는 상태가 가능하다. 찰리의 아래 몸통은 더미 속에 이어지고 왼쪽 화면의 손은 잔해에 닿아 있어 떠 있는 몸으로 보이지 않는다. 주크박스와 공구함, LP는 잔해나 천 위에 놓여 있고 그물은 장갑에 걸려 늘어진다. 지지 없이 공중에 멈춘 파편이나 추가 입자 효과는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "양쪽 아래의 아이들은 중앙 위 찰리를 올려다본다. 찰리는 목과 얼굴을 아래로 기울여 두 아이가 있는 전경 쪽을 향하므로, 요구된 내려다보는 관계가 A보다 명확하다. 개별 아이 중 누구를 응시하는지는 점 형태의 눈만으로 확정하기 어렵다. 무기나 손에 든 지향성 물건은 없고, 주크박스의 혼 입구는 오른쪽 위를 향한다.",
        "built_space": "왼쪽 위 폐헬기 한 대, 오른쪽의 주크박스 한 대와 부착된 혼 하나, 아래 왼쪽 중앙의 열린 공구함 하나가 보인다. 중앙 전경의 푸른 천과 LP 여러 장, 겹겹의 녹슨 금속 잔해가 장소 참조와 잘 대응한다. 찰리는 갈라진 쓰레기 더미 사이에서 상체를 드러내고 상단 중앙에서 화면 절반 미만을 차지한다. 아이들은 양쪽 하단에 부분적으로 잘려 있으며 쓰레기에 뒤로 기대어 있다. 지면 가까운 상향 시점은 맞지만 중앙 정면에 가까운 축이고, 주변 잔해가 많이 보여 요구된 미디엄 숏보다 다소 넓게 읽힌다.",
        "entities": "찰리 한 기, 앰버와 라울 두 아이만 있다. 찰리의 흰 마스크형 얼굴, 두 원형 눈과 짧은 입 선, 각진 베이지 장갑과 굵고 긴 팔은 인물 참조에 더 가깝다. 낡은 표면과 그물도 보인다. 원형 가슴 원자로는 있지만 강하게 푸르게 발광하여 청색 조명을 눈에만 한정하라는 지시와 충돌한다. UBIC 문구는 확인되지 않는다. 앰버는 금발을 묶은 어린 여자아이로 마스크를 착용했고, 라울은 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보인다. 라울의 마스크는 확인되지 않으며, 측후면이라 두 아이의 정확한 참조 얼굴과 혼혈 외모는 판별에 제한이 있다. 두 아이 모두 참조의 티셔츠 대신 낡은 겉옷을 입었다. 앰버의 허리 공구 주머니는 가려져 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "아이들의 엉덩이와 등은 쓰레기 경사면에 받쳐져 있고 다리는 앞쪽으로 굽혀져 있어 뒤로 넘어진 자세가 성립한다. 찰리의 아래 몸통은 잔해 속에 묻혀 있고 양팔이 주변 잔해 쪽으로 내려와 있어 상승 중 몸을 지탱하는 구성이 가능하다. 그물은 어깨와 가슴에 걸려 중력 방향으로 늘어진다. 주크박스는 잔해에 기대어 있고 혼은 상단에 부착되어 있으며 공구함과 음반도 바닥 재료 위에 놓여 있다. 떠 있는 몸이나 지지 없는 물체, 공중 파편 효과는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 6,
   "A": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "낮은 시점과 하단의 넘어진 아이들, 그물·UBIC 표시는 맞지만, 찰리가 아이들을 내려다보기보다 정면을 보며 흰 마스크형 얼굴도 인물 참조와 다르다."
   },
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "아이들을 향해 숙인 찰리의 머리와 참조에 가까운 기계 얼굴, 상단 중앙 배치가 더 충실하지만, 가슴의 청색 발광과 정면에 가까운 카메라 축은 지시에서 벗어난다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrap_heap_encounter_ab6370.png",
    "asset_id": "c12f180b-34fd-4d2d-a22f-0493ef01b519",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1334467>",
    "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab07f9-94b8-778b-83e0-6ae4f1f46976",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh15__bgfirst_bg.png",
   "bg_asset_id": "7ff350d0-f4f3-4922-b759-f2a2ca7fc8a6",
   "bg_record_key": "S6sh15::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "scrap_heap_encounter",
   "groupbg_asset_id": "c12f180b-34fd-4d2d-a22f-0493ef01b519"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C05",
   "C06"
  ]
 },
 "S6sh20::signage": {
  "fp": "e6255a3274fb2613",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S6sh20": {
  "input_fingerprint": "da26c1ce9728325a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 다가온 앰버의 작은 몸을 두꺼운 기계 팔로 단숨에 감싸 와락 끌어안은 찰리의 모습.\n\nLOCATION (lock): On the rubbish-strewn ground beside the disturbed heap in a refugee-settlement dump. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward track at 앰버's shoulder height on the established side of the pair, using a slightly upward three-quarter view to observe the embrace directly. Place 앰버 left of center inside the span of 찰리's arms, with his face above and to the right, both forearms readable without enlarging either hand through foreground perspective; she stiffens in bewilderment and looks up toward him as he bends his attention down to her. Emphasize their changed relative positions as the gap closes, preserving the exposure and allowing a narrow border of rubbish around the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding rubbish (Remaining around the pair after 찰리's emergence); used as Provides a restrained environmental border without competing with the enclosing arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and localized blue eye illumination, expressing warmth through the embrace rather than a new lighting color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding scrap heaps, exposed rubbish, and daylight colors from the reference. Exclude the airborne fragments from the robot's emergence; do not repeat that burst of debris.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The assembled jukebox retains its attached horn speaker, with tools and LPs remaining on the blanket. Charlie is upright with blue-lit eyes and stiff joints, retaining the dirty old body, entangling netting, and worn chest logo. 앰버: Has approached the robot closely and looks bewildered, still wearing her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 다가온 앰버의 작은 몸을 두꺼운 기계 팔로 단숨에 감싸 와락 끌어안은 찰리의 모습.\n\nLOCATION (lock): On the rubbish-strewn ground beside the disturbed heap in a refugee-settlement dump. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward track at 앰버's shoulder height on the established side of the pair, using a slightly upward three-quarter view to observe the embrace directly. Place 앰버 left of center inside the span of 찰리's arms, with his face above and to the right, both forearms readable without enlarging either hand through foreground perspective; she stiffens in bewilderment and looks up toward him as he bends his attention down to her. Emphasize their changed relative positions as the gap closes, preserving the exposure and allowing a narrow border of rubbish around the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding rubbish (Remaining around the pair after 찰리's emergence); used as Provides a restrained environmental border without competing with the enclosing arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and localized blue eye illumination, expressing warmth through the embrace rather than a new lighting color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding scrap heaps, exposed rubbish, and daylight colors from the reference. Exclude the airborne fragments from the robot's emergence; do not repeat that burst of debris.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The assembled jukebox retains its attached horn speaker, with tools and LPs remaining on the blanket. Charlie is upright with blue-lit eyes and stiff joints, retaining the dirty old body, entangling netting, and worn chest logo. 앰버: Has approached the robot closely and looks bewildered, still wearing her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 다가온 앰버의 작은 몸을 두꺼운 기계 팔로 단숨에 감싸 와락 끌어안은 찰리의 모습.\n\nLOCATION (lock): On the rubbish-strewn ground beside the disturbed heap in a refugee-settlement dump. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward track at 앰버's shoulder height on the established side of the pair, using a slightly upward three-quarter view to observe the embrace directly. Place 앰버 left of center inside the span of 찰리's arms, with his face above and to the right, both forearms readable without enlarging either hand through foreground perspective; she stiffens in bewilderment and looks up toward him as he bends his attention down to her. Emphasize their changed relative positions as the gap closes, preserving the exposure and allowing a narrow border of rubbish around the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding rubbish (Remaining around the pair after 찰리's emergence); used as Provides a restrained environmental border without competing with the enclosing arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and localized blue eye illumination, expressing warmth through the embrace rather than a new lighting color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding scrap heaps, exposed rubbish, and daylight colors from the reference. Exclude the airborne fragments from the robot's emergence; do not repeat that burst of debris.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The assembled jukebox retains its attached horn speaker, with tools and LPs remaining on the blanket. Charlie is upright with blue-lit eyes and stiff joints, retaining the dirty old body, entangling netting, and worn chest logo. 앰버: Has approached the robot closely and looks bewildered, still wearing her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "B",
    "direction": "중앙 왼쪽의 앰버는 눈을 위쪽 오른편으로 들어 찰리의 얼굴을 바라본다. 찰리는 얼굴을 아래쪽 왼편의 앰버에게 기울인다. 두꺼운 팔과 손은 앰버의 등과 몸통을 향해 안쪽으로 모여 있어, 다가온 아이를 자신에게 끌어안는 방향이 분명하다.",
    "built_space": "야외 고철 더미 사이를 인물들이 채운다. 왼쪽 위에는 헬리콥터 잔해 한 대, 오른쪽 뒤에는 혼 스피커 하나가 붙은 주크박스 한 대, 왼쪽 아래에는 열린 공구함 하나가 보인다. 오른쪽 아래에는 음반이 놓여 있다. 이전 장면의 주요 좌우 배치와 금속 쓰레기 환경이 유지되며, 고정 시설의 중복은 없다. 인물들은 허리 부근에서 잘리는 미디엄 구도로 보이고, 카메라는 약간 낮은 사선에서 포옹을 관찰한다.",
    "entities": "등장인물은 앰버와 찰리뿐이다. 앰버는 어린 여자아이의 얼굴과 체격, 밝은 피부, 금발, 얼굴에 착용한 방진 마스크, 때 묻은 카키 작업복과 가죽 공구 주머니를 갖췄다. 마스크 때문에 얼굴 전체와 혼혈 정체성까지 확정하기는 어렵지만 보이는 특징은 참조와 부합한다. 찰리는 낡은 샌드 베이지 장갑, 육중한 기계 팔, 흰 얼굴판, 파란 두 눈과 선형 입, 청색 원형 가슴 장치 및 얽힌 그물을 유지한다. 먼 쪽 전완은 앰버와 손에 상당 부분 가려져 가까운 전완만큼 명료하지 않다. 낮의 자연광과 금속 표면의 마모가 유지되며, 공중으로 튀는 파편이나 추가 문구는 없다.",
    "hard_violations": [],
    "physics": "찰리의 가까운 팔은 앰버의 몸통 앞을 가로질러 감싸고, 반대쪽 손은 등과 어깨 부근에 접촉한다. 손과 전완은 손목 및 팔꿈치 관절로 이어지며, 아이의 몸은 찰리의 가슴과 팔에 밀착되어 있다. 두 인물의 하체와 발은 화면 밖이므로 접지는 직접 확인할 수 없지만 공중에 뜬 자세는 보이지 않는다. 그물은 어깨와 몸체에 걸려 아래로 처지고, 주변 공구함과 주크박스는 쓰레기 더미에 받쳐져 있다."
   },
   {
    "label": "A",
    "direction": "앰버는 오른쪽 위의 찰리 얼굴을 올려다본다. 찰리의 머리도 앰버 쪽인 아래쪽 왼편으로 기울어 있지만 얼굴판은 A보다 카메라에 정면으로 열려 있다. 양팔은 앰버의 양옆에서 안으로 굽어 몸통을 둘러싸고, 두 손은 아이의 등과 허리 쪽으로 모인다.",
    "built_space": "왼쪽 위의 헬리콥터 잔해 한 대, 오른쪽 뒤의 혼 스피커가 달린 주크박스 한 대, 왼쪽 아래의 열린 공구함 하나가 보인다. 양옆과 아래의 고철 더미가 이전 장면과 같은 장소를 형성한다. 시설 중복이나 불가능한 반사는 없다. 앰버는 중앙 왼쪽, 찰리의 얼굴은 그 위쪽 오른편에 놓이지만 찰리의 양 어깨와 가슴을 거의 정면으로 보는 구도라 지정된 사선 관찰 시점은 A보다 약하다.",
    "entities": "앰버와 찰리 외의 인물은 없다. 앰버의 어린 외형, 금발과 밝은 피부, 착용 중인 방진 마스크, 오염된 카키 작업복과 허리 공구 주머니가 유지된다. 가려진 얼굴만으로 정확한 혈통을 판별할 수는 없다. 찰리의 흰 기계 얼굴, 파란 눈 두 개, 선형 입, 낡은 베이지 장갑판, 푸른 원형 가슴 장치와 그물이 참조에 부합한다. 두 전완과 두 손이 모두 비교적 명확하게 보인다. 주크박스에는 혼이 부착되어 있고 공구도 보이나 음반과 담요의 상태는 이 구도에서 분명하지 않다. 추가 인물, 비산 파편, 삽입 문구는 없다.",
    "hard_violations": [],
    "physics": "양쪽 팔꿈치가 굽혀져 두 전완이 아이 주위를 감싸며, 손은 아이의 등과 몸통 부근에 실제로 닿아 있다. 손목과 기계 관절의 연결에 명백한 단절은 없다. 앰버의 몸은 찰리의 가슴과 팔에 기대어 있고, 발이 잘린 구도일 뿐 부유를 나타내는 자세는 아니다. 그물은 장갑판에 걸쳐 처지고 주크박스와 공구함은 고철 위에 놓여 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "약간 올려다보는 사선 미디엄 구도와 서로를 향한 시선, 몸을 밀착시킨 포옹이 더 정확하지만 먼 쪽 전완은 일부 가려진다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 전완이 만드는 포옹은 명확하나 구도가 더 정면에 가깝고 찰리가 앰버에게 고개를 숙여 집중하는 정도가 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 왼쪽의 앰버는 눈을 위쪽 오른편으로 들어 찰리의 얼굴을 바라본다. 찰리는 얼굴을 아래쪽 왼편의 앰버에게 기울인다. 두꺼운 팔과 손은 앰버의 등과 몸통을 향해 안쪽으로 모여 있어, 다가온 아이를 자신에게 끌어안는 방향이 분명하다.",
        "built_space": "야외 고철 더미 사이를 인물들이 채운다. 왼쪽 위에는 헬리콥터 잔해 한 대, 오른쪽 뒤에는 혼 스피커 하나가 붙은 주크박스 한 대, 왼쪽 아래에는 열린 공구함 하나가 보인다. 오른쪽 아래에는 음반이 놓여 있다. 이전 장면의 주요 좌우 배치와 금속 쓰레기 환경이 유지되며, 고정 시설의 중복은 없다. 인물들은 허리 부근에서 잘리는 미디엄 구도로 보이고, 카메라는 약간 낮은 사선에서 포옹을 관찰한다.",
        "entities": "등장인물은 앰버와 찰리뿐이다. 앰버는 어린 여자아이의 얼굴과 체격, 밝은 피부, 금발, 얼굴에 착용한 방진 마스크, 때 묻은 카키 작업복과 가죽 공구 주머니를 갖췄다. 마스크 때문에 얼굴 전체와 혼혈 정체성까지 확정하기는 어렵지만 보이는 특징은 참조와 부합한다. 찰리는 낡은 샌드 베이지 장갑, 육중한 기계 팔, 흰 얼굴판, 파란 두 눈과 선형 입, 청색 원형 가슴 장치 및 얽힌 그물을 유지한다. 먼 쪽 전완은 앰버와 손에 상당 부분 가려져 가까운 전완만큼 명료하지 않다. 낮의 자연광과 금속 표면의 마모가 유지되며, 공중으로 튀는 파편이나 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "찰리의 가까운 팔은 앰버의 몸통 앞을 가로질러 감싸고, 반대쪽 손은 등과 어깨 부근에 접촉한다. 손과 전완은 손목 및 팔꿈치 관절로 이어지며, 아이의 몸은 찰리의 가슴과 팔에 밀착되어 있다. 두 인물의 하체와 발은 화면 밖이므로 접지는 직접 확인할 수 없지만 공중에 뜬 자세는 보이지 않는다. 그물은 어깨와 몸체에 걸려 아래로 처지고, 주변 공구함과 주크박스는 쓰레기 더미에 받쳐져 있다."
       },
       {
        "label": "B",
        "direction": "앰버는 오른쪽 위의 찰리 얼굴을 올려다본다. 찰리의 머리도 앰버 쪽인 아래쪽 왼편으로 기울어 있지만 얼굴판은 A보다 카메라에 정면으로 열려 있다. 양팔은 앰버의 양옆에서 안으로 굽어 몸통을 둘러싸고, 두 손은 아이의 등과 허리 쪽으로 모인다.",
        "built_space": "왼쪽 위의 헬리콥터 잔해 한 대, 오른쪽 뒤의 혼 스피커가 달린 주크박스 한 대, 왼쪽 아래의 열린 공구함 하나가 보인다. 양옆과 아래의 고철 더미가 이전 장면과 같은 장소를 형성한다. 시설 중복이나 불가능한 반사는 없다. 앰버는 중앙 왼쪽, 찰리의 얼굴은 그 위쪽 오른편에 놓이지만 찰리의 양 어깨와 가슴을 거의 정면으로 보는 구도라 지정된 사선 관찰 시점은 A보다 약하다.",
        "entities": "앰버와 찰리 외의 인물은 없다. 앰버의 어린 외형, 금발과 밝은 피부, 착용 중인 방진 마스크, 오염된 카키 작업복과 허리 공구 주머니가 유지된다. 가려진 얼굴만으로 정확한 혈통을 판별할 수는 없다. 찰리의 흰 기계 얼굴, 파란 눈 두 개, 선형 입, 낡은 베이지 장갑판, 푸른 원형 가슴 장치와 그물이 참조에 부합한다. 두 전완과 두 손이 모두 비교적 명확하게 보인다. 주크박스에는 혼이 부착되어 있고 공구도 보이나 음반과 담요의 상태는 이 구도에서 분명하지 않다. 추가 인물, 비산 파편, 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "양쪽 팔꿈치가 굽혀져 두 전완이 아이 주위를 감싸며, 손은 아이의 등과 몸통 부근에 실제로 닿아 있다. 손목과 기계 관절의 연결에 명백한 단절은 없다. 앰버의 몸은 찰리의 가슴과 팔에 기대어 있고, 발이 잘린 구도일 뿐 부유를 나타내는 자세는 아니다. 그물은 장갑판에 걸쳐 처지고 주크박스와 공구함은 고철 위에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "약간 올려다보는 사선 미디엄 구도와 서로를 향한 시선, 몸을 밀착시킨 포옹이 더 정확하지만 먼 쪽 전완은 일부 가려진다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 전완이 만드는 포옹은 명확하나 구도가 더 정면에 가깝고 찰리가 앰버에게 고개를 숙여 집중하는 정도가 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 왼쪽의 앰버는 눈을 위쪽 오른편으로 들어 찰리의 얼굴을 바라본다. 찰리는 얼굴을 아래쪽 왼편의 앰버에게 기울인다. 두꺼운 팔과 손은 앰버의 등과 몸통을 향해 안쪽으로 모여 있어, 다가온 아이를 자신에게 끌어안는 방향이 분명하다.",
        "built_space": "야외 고철 더미 사이를 인물들이 채운다. 왼쪽 위에는 헬리콥터 잔해 한 대, 오른쪽 뒤에는 혼 스피커 하나가 붙은 주크박스 한 대, 왼쪽 아래에는 열린 공구함 하나가 보인다. 오른쪽 아래에는 음반이 놓여 있다. 이전 장면의 주요 좌우 배치와 금속 쓰레기 환경이 유지되며, 고정 시설의 중복은 없다. 인물들은 허리 부근에서 잘리는 미디엄 구도로 보이고, 카메라는 약간 낮은 사선에서 포옹을 관찰한다.",
        "entities": "등장인물은 앰버와 찰리뿐이다. 앰버는 어린 여자아이의 얼굴과 체격, 밝은 피부, 금발, 얼굴에 착용한 방진 마스크, 때 묻은 카키 작업복과 가죽 공구 주머니를 갖췄다. 마스크 때문에 얼굴 전체와 혼혈 정체성까지 확정하기는 어렵지만 보이는 특징은 참조와 부합한다. 찰리는 낡은 샌드 베이지 장갑, 육중한 기계 팔, 흰 얼굴판, 파란 두 눈과 선형 입, 청색 원형 가슴 장치 및 얽힌 그물을 유지한다. 먼 쪽 전완은 앰버와 손에 상당 부분 가려져 가까운 전완만큼 명료하지 않다. 낮의 자연광과 금속 표면의 마모가 유지되며, 공중으로 튀는 파편이나 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "찰리의 가까운 팔은 앰버의 몸통 앞을 가로질러 감싸고, 반대쪽 손은 등과 어깨 부근에 접촉한다. 손과 전완은 손목 및 팔꿈치 관절로 이어지며, 아이의 몸은 찰리의 가슴과 팔에 밀착되어 있다. 두 인물의 하체와 발은 화면 밖이므로 접지는 직접 확인할 수 없지만 공중에 뜬 자세는 보이지 않는다. 그물은 어깨와 몸체에 걸려 아래로 처지고, 주변 공구함과 주크박스는 쓰레기 더미에 받쳐져 있다."
       },
       {
        "label": "A",
        "direction": "앰버는 오른쪽 위의 찰리 얼굴을 올려다본다. 찰리의 머리도 앰버 쪽인 아래쪽 왼편으로 기울어 있지만 얼굴판은 A보다 카메라에 정면으로 열려 있다. 양팔은 앰버의 양옆에서 안으로 굽어 몸통을 둘러싸고, 두 손은 아이의 등과 허리 쪽으로 모인다.",
        "built_space": "왼쪽 위의 헬리콥터 잔해 한 대, 오른쪽 뒤의 혼 스피커가 달린 주크박스 한 대, 왼쪽 아래의 열린 공구함 하나가 보인다. 양옆과 아래의 고철 더미가 이전 장면과 같은 장소를 형성한다. 시설 중복이나 불가능한 반사는 없다. 앰버는 중앙 왼쪽, 찰리의 얼굴은 그 위쪽 오른편에 놓이지만 찰리의 양 어깨와 가슴을 거의 정면으로 보는 구도라 지정된 사선 관찰 시점은 A보다 약하다.",
        "entities": "앰버와 찰리 외의 인물은 없다. 앰버의 어린 외형, 금발과 밝은 피부, 착용 중인 방진 마스크, 오염된 카키 작업복과 허리 공구 주머니가 유지된다. 가려진 얼굴만으로 정확한 혈통을 판별할 수는 없다. 찰리의 흰 기계 얼굴, 파란 눈 두 개, 선형 입, 낡은 베이지 장갑판, 푸른 원형 가슴 장치와 그물이 참조에 부합한다. 두 전완과 두 손이 모두 비교적 명확하게 보인다. 주크박스에는 혼이 부착되어 있고 공구도 보이나 음반과 담요의 상태는 이 구도에서 분명하지 않다. 추가 인물, 비산 파편, 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "양쪽 팔꿈치가 굽혀져 두 전완이 아이 주위를 감싸며, 손은 아이의 등과 몸통 부근에 실제로 닿아 있다. 손목과 기계 관절의 연결에 명백한 단절은 없다. 앰버의 몸은 찰리의 가슴과 팔에 기대어 있고, 발이 잘린 구도일 뿐 부유를 나타내는 자세는 아니다. 그물은 장갑판에 걸쳐 처지고 주크박스와 공구함은 고철 위에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 9,
   "A": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "약간 올려다보는 사선 미디엄 구도와 서로를 향한 시선, 몸을 밀착시킨 포옹이 더 정확하지만 먼 쪽 전완은 일부 가려진다."
   },
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "두 전완이 만드는 포옹은 명확하나 구도가 더 정면에 가깝고 찰리가 앰버에게 고개를 숙여 집중하는 정도가 A보다 약하다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh15_sel.png",
    "asset_id": "99dec37f-a58a-43c2-9a7e-1b16cce79870",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0802-e7a6-7234-b9a7-ea9be1b7fcb8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S6sh15"
  }
 },
 "S6sh34::signage": {
  "fp": "52c55556950231e9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::scrapyard_exit_road": {
  "input_fingerprint": "ec13d0acaabe0c0c",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "scrapyard_exit_road",
    "tags": [
     "S6sh34"
    ]
   },
   "context_sig": "58b4d436bba07855"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the road just outside the refugee-settlement dump, leading toward the container homes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 쓰레기장: 난민 구역 외곽에 위치한 거대한 쓰레기 산으로, 온갖 고철, 모포, 낡은 가전이 널려 있다. (특징: 녹슬고 부서진 헬리콥터 잔해 및 거대한 폐품 산; 먼지를 막는 마스크와 수선된 헌 옷을 입은 앰버(한국계 백인 혼혈 여자아이)와 라울(라틴계 흑인 혼혈 남자아이); 공구 주머니와 렌치를 들고 로봇 팔을 해체하는 앰버; 부서진 구형 주크박스와 널브러진 LP판들; 쓰레기 더미를 뚫고 일어나는 고릴라형 외형의 기계 관절 로봇 찰리 (푸른 발광 눈 장착); 외투와 모자를 걸쳐 사람처럼 위장한 찰리의 뚱뚱한 실루엣)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 위기를 벗어나서 키득거리며 페드로 집으로 향하는 앰버와 라울, 그리고 로봇.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the road just outside the refugee-settlement dump, leading toward the container homes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 쓰레기장: 난민 구역 외곽에 위치한 거대한 쓰레기 산으로, 온갖 고철, 모포, 낡은 가전이 널려 있다. (특징: 녹슬고 부서진 헬리콥터 잔해 및 거대한 폐품 산; 먼지를 막는 마스크와 수선된 헌 옷을 입은 앰버(한국계 백인 혼혈 여자아이)와 라울(라틴계 흑인 혼혈 남자아이); 공구 주머니와 렌치를 들고 로봇 팔을 해체하는 앰버; 부서진 구형 주크박스와 널브러진 LP판들; 쓰레기 더미를 뚫고 일어나는 고릴라형 외형의 기계 관절 로봇 찰리 (푸른 발광 눈 장착); 외투와 모자를 걸쳐 사람처럼 위장한 찰리의 뚱뚱한 실루엣)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 위기를 벗어나서 키득거리며 페드로 집으로 향하는 앰버와 라울, 그리고 로봇.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrapyard_exit_road_3033ca.png",
  "asset_id": "ba23ebba-28ad-4bbc-90cf-eeb40fb40e22",
  "input_asset_ids": [
   "cb051d37-bff5-44e7-940f-2024f8d465b7"
  ],
  "origin_tag": "S6sh34",
  "place_text": "On the road just outside the refugee-settlement dump, leading toward the container homes.",
  "origin_inputs": {
   "place_text": "On the road just outside the refugee-settlement dump, leading toward the container homes.",
   "time_of_day_en": "day",
   "conti_asset_id": "cb051d37-bff5-44e7-940f-2024f8d465b7"
  }
 },
 "S6sh34::bgfirst_bg": {
  "input_fingerprint": "9665c662ffeb73b0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기장 밖 도로에서 장난스럽게 웃으며 나란히 걷는 도중, 각자 한쪽 다리를 내디딘 앰버, 라울, 위장한 찰리의 전신 구도.\n\nLOCATION (lock): On the road just outside the refugee-settlement dump, leading toward the container homes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track directly alongside the trio at the children's chest height in the inherited lateral three-quarter view, remaining slightly behind their forward axes and framing all three bodies with space below their feet. Arrange 라울 at left, 앰버 near center and 찰리 at right as they advance toward the right side of the image, the children laughing toward their route and 찰리 walking in his worn coat and hat. Emphasize the wider camera distance, capturing different stride phases and shoulder angles rather than synchronized marching, and retain headroom for the following upward tilt without beginning it in this keyframe.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road continuing ahead of the trio in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Road outside the rubbish dump (The trio is walking away along it) — The visible route continues toward screen right; used as Provides foot clearance and directional space ahead of the walkers; 찰리's worn coat and hat (Worn as a disguise) — Seen mainly from the side, concealing his recognizable robot body; used as Makes the successful disguise readable at full-body scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast maintain continuity while the children's relaxed movement supplies the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기장 밖 도로에서 장난스럽게 웃으며 나란히 걷는 도중, 각자 한쪽 다리를 내디딘 앰버, 라울, 위장한 찰리의 전신 구도.\n\nLOCATION (lock): On the road just outside the refugee-settlement dump, leading toward the container homes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track directly alongside the trio at the children's chest height in the inherited lateral three-quarter view, remaining slightly behind their forward axes and framing all three bodies with space below their feet. Arrange 라울 at left, 앰버 near center and 찰리 at right as they advance toward the right side of the image, the children laughing toward their route and 찰리 walking in his worn coat and hat. Emphasize the wider camera distance, capturing different stride phases and shoulder angles rather than synchronized marching, and retain headroom for the following upward tilt without beginning it in this keyframe.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road continuing ahead of the trio in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Road outside the rubbish dump (The trio is walking away along it) — The visible route continues toward screen right; used as Provides foot clearance and directional space ahead of the walkers; 찰리's worn coat and hat (Worn as a disguise) — Seen mainly from the side, concealing his recognizable robot body; used as Makes the successful disguise readable at full-body scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast maintain continuity while the children's relaxed movement supplies the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh34__bgfirst_bg.png",
  "asset_id": "ee40dc87-434f-4e50-bffc-40732a21c99a",
  "input_asset_ids": [
   "cb051d37-bff5-44e7-940f-2024f8d465b7",
   "ba23ebba-28ad-4bbc-90cf-eeb40fb40e22"
  ]
 },
 "S6sh34": {
  "input_fingerprint": "39632d92ee4c8e3b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기장 밖 도로에서 장난스럽게 웃으며 나란히 걷는 도중, 각자 한쪽 다리를 내디딘 앰버, 라울, 위장한 찰리의 전신 구도.\n\nLOCATION (lock): On the road just outside the refugee-settlement dump, leading toward the container homes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track directly alongside the trio at the children's chest height in the inherited lateral three-quarter view, remaining slightly behind their forward axes and framing all three bodies with space below their feet. Arrange 라울 at left, 앰버 near center and 찰리 at right as they advance toward the right side of the image, the children laughing toward their route and 찰리 walking in his worn coat and hat. Emphasize the wider camera distance, capturing different stride phases and shoulder angles rather than synchronized marching, and retain headroom for the following upward tilt without beginning it in this keyframe.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road continuing ahead of the trio in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Road outside the rubbish dump (The trio is walking away along it) — The visible route continues toward screen right; used as Provides foot clearance and directional space ahead of the walkers; 찰리's worn coat and hat (Worn as a disguise) — Seen mainly from the side, concealing his recognizable robot body; used as Makes the successful disguise readable at full-body scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast maintain continuity while the children's relaxed movement supplies the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and a hat, disguising the bulky gorilla-shaped robot as a short, stout man. The dirty metal body, worn UBIC logo, and unreleased netting remain beneath the disguise. 앰버: Walks away from the dump, smiling, with her mask and waist tool pouch still in place. 라울: Walks away from the dump, smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 로봇 외형을 숨기기 위해 입은 낡고 두꺼운 회색 롱 코트와 푹 눌러쓴 챙 넓은 모자.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기장 밖 도로에서 장난스럽게 웃으며 나란히 걷는 도중, 각자 한쪽 다리를 내디딘 앰버, 라울, 위장한 찰리의 전신 구도.\n\nLOCATION (lock): On the road just outside the refugee-settlement dump, leading toward the container homes. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track directly alongside the trio at the children's chest height in the inherited lateral three-quarter view, remaining slightly behind their forward axes and framing all three bodies with space below their feet. Arrange 라울 at left, 앰버 near center and 찰리 at right as they advance toward the right side of the image, the children laughing toward their route and 찰리 walking in his worn coat and hat. Emphasize the wider camera distance, capturing different stride phases and shoulder angles rather than synchronized marching, and retain headroom for the following upward tilt without beginning it in this keyframe.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road continuing ahead of the trio in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Road outside the rubbish dump (The trio is walking away along it) — The visible route continues toward screen right; used as Provides foot clearance and directional space ahead of the walkers; 찰리's worn coat and hat (Worn as a disguise) — Seen mainly from the side, concealing his recognizable robot body; used as Makes the successful disguise readable at full-body scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast maintain continuity while the children's relaxed movement supplies the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and a hat, disguising the bulky gorilla-shaped robot as a short, stout man. The dirty metal body, worn UBIC logo, and unreleased netting remain beneath the disguise. 앰버: Walks away from the dump, smiling, with her mask and waist tool pouch still in place. 라울: Walks away from the dump, smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 로봇 외형을 숨기기 위해 입은 낡고 두꺼운 회색 롱 코트와 푹 눌러쓴 챙 넓은 모자.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰레기장 밖 도로에서 장난스럽게 웃으며 나란히 걷는 도중, 각자 한쪽 다리를 내디딘 앰버, 라울, 위장한 찰리의 전신 구도.\n\nLOCATION (lock): On the road just outside the refugee-settlement dump, leading toward the container homes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track directly alongside the trio at the children's chest height in the inherited lateral three-quarter view, remaining slightly behind their forward axes and framing all three bodies with space below their feet. Arrange 라울 at left, 앰버 near center and 찰리 at right as they advance toward the right side of the image, the children laughing toward their route and 찰리 walking in his worn coat and hat. Emphasize the wider camera distance, capturing different stride phases and shoulder angles rather than synchronized marching, and retain headroom for the following upward tilt without beginning it in this keyframe.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road continuing ahead of the trio in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Road outside the rubbish dump (The trio is walking away along it) — The visible route continues toward screen right; used as Provides foot clearance and directional space ahead of the walkers; 찰리's worn coat and hat (Worn as a disguise) — Seen mainly from the side, concealing his recognizable robot body; used as Makes the successful disguise readable at full-body scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast maintain continuity while the children's relaxed movement supplies the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie now wears an old coat and a hat, disguising the bulky gorilla-shaped robot as a short, stout man. The dirty metal body, worn UBIC logo, and unreleased netting remain beneath the disguise. 앰버: Walks away from the dump, smiling, with her mask and waist tool pouch still in place. 라울: Walks away from the dump, smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 로봇 외형을 숨기기 위해 입은 낡고 두꺼운 회색 롱 코트와 푹 눌러쓴 챙 넓은 모자.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh34__bgfirst_bg.png",
     "asset_id": "ee40dc87-434f-4e50-bffc-40732a21c99a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S6sh34.png",
     "asset_id": "cb051d37-bff5-44e7-940f-2024f8d465b7",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1126517>",
     "asset_id": "13c85edd-6f32-4e1f-bff4-67869a01ee0b",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrapyard_exit_road_3033ca.png",
     "asset_id": "ba23ebba-28ad-4bbc-90cf-eeb40fb40e22",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1126517>",
     "asset_id": "13c85edd-6f32-4e1f-bff4-67869a01ee0b",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "세 인물이 화면 우측으로 걷고 있으며 아이들은 앞을 보며 웃고 있음.",
    "built_space": "쓰레기장 밖 도로, 펜스, 쓰레기 산, 컨테이너가 올바르게 배치됨.",
    "entities": "라울과 앰버는 일치하나, 찰리의 다리와 발이 로봇이 아닌 인간의 형태임.",
    "hard_violations": [
     "[gemini-pro] 찰리의 다리와 발이 지정된 기계 몸체가 아닌 인간의 바지와 신발로 렌더링됨"
    ],
    "physics": "세 인물 모두 바닥에 발을 딛고 자연스럽게 걷는 동작을 보여줌."
   },
   {
    "label": "B",
    "direction": "세 인물이 화면 우측으로 걷고 있으며 라울과 앰버는 미소 지으며 앞을 향함.",
    "built_space": "참조 이미지의 도로, 쓰레기 더미, 우측 컨테이너가 정확히 구현됨.",
    "entities": "세 인물 모두 일치하며, 찰리의 코트 아래로 기계 팔과 발이 명확히 묘사됨.",
    "hard_violations": [],
    "physics": "세 인물의 걷는 자세와 체중 이동이 지면에 안정적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 기계 팔다리를 포함해 세 인물의 외형을 참조에 맞게 구현했으며, 요구된 구도와 배경을 정확히 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 하반신이 지정된 로봇이 아닌 바지와 신발을 착용한 인간의 형태로 묘사되어 캐릭터 설정을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물이 화면 우측으로 걷고 있으며 아이들은 앞을 보며 웃고 있음.",
        "built_space": "쓰레기장 밖 도로, 펜스, 쓰레기 산, 컨테이너가 올바르게 배치됨.",
        "entities": "라울과 앰버는 일치하나, 찰리의 다리와 발이 로봇이 아닌 인간의 형태임.",
        "hard_violations": [
         "찰리의 다리와 발이 지정된 기계 몸체가 아닌 인간의 바지와 신발로 렌더링됨"
        ],
        "physics": "세 인물 모두 바닥에 발을 딛고 자연스럽게 걷는 동작을 보여줌."
       },
       {
        "label": "B",
        "direction": "세 인물이 화면 우측으로 걷고 있으며 라울과 앰버는 미소 지으며 앞을 향함.",
        "built_space": "참조 이미지의 도로, 쓰레기 더미, 우측 컨테이너가 정확히 구현됨.",
        "entities": "세 인물 모두 일치하며, 찰리의 코트 아래로 기계 팔과 발이 명확히 묘사됨.",
        "hard_violations": [],
        "physics": "세 인물의 걷는 자세와 체중 이동이 지면에 안정적으로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 기계 팔다리를 포함해 세 인물의 외형을 참조에 맞게 구현했으며, 요구된 구도와 배경을 정확히 따랐습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 하반신이 지정된 로봇이 아닌 바지와 신발을 착용한 인간의 형태로 묘사되어 캐릭터 설정을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물이 화면 우측으로 걷고 있으며 아이들은 앞을 보며 웃고 있음.",
        "built_space": "쓰레기장 밖 도로, 펜스, 쓰레기 산, 컨테이너가 올바르게 배치됨.",
        "entities": "라울과 앰버는 일치하나, 찰리의 다리와 발이 로봇이 아닌 인간의 형태임.",
        "hard_violations": [
         "찰리의 다리와 발이 지정된 기계 몸체가 아닌 인간의 바지와 신발로 렌더링됨"
        ],
        "physics": "세 인물 모두 바닥에 발을 딛고 자연스럽게 걷는 동작을 보여줌."
       },
       {
        "label": "B",
        "direction": "세 인물이 화면 우측으로 걷고 있으며 라울과 앰버는 미소 지으며 앞을 향함.",
        "built_space": "참조 이미지의 도로, 쓰레기 더미, 우측 컨테이너가 정확히 구현됨.",
        "entities": "세 인물 모두 일치하며, 찰리의 코트 아래로 기계 팔과 발이 명확히 묘사됨.",
        "hard_violations": [],
        "physics": "세 인물의 걷는 자세와 체중 이동이 지면에 안정적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 장소는 잘 맞지만 세 인물이 카메라 앞으로 다가오고 아이들이 서로를 바라봐, 오른쪽으로 이동하는 측후방 와이드 숏이라는 핵심 연출을 놓쳤다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽을 향한 보행과 시선, 좌우 인물 순서, 찰리의 위장이 더 충실하지만 카메라는 여전히 전방 사선이며 보폭도 지나치게 비슷하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울은 오른쪽의 앰버를, 앰버는 왼쪽의 라울을 보며 웃고 찰리도 아이들 쪽으로 얼굴을 돌린다. 세 몸과 내디딘 발은 대체로 카메라 쪽을 향한다. 오른쪽으로 이어지는 배경 도로는 보이지만 인물들의 이동은 그 경로를 따르지 않는다.",
        "built_space": "왼쪽에는 골판금 담장 한 줄과 철망 구간, 그 뒤에는 폐헬기 한 대가 놓인 쓰레기 더미가 있다. 오른쪽에는 컨테이너 주택 행렬, 여러 전신주와 전선, 도로 가장자리 배수로가 있어 장소의 주요 구조가 일치한다. 세 인물은 담장 밖 도로에 라울·앰버·찰리 순으로 서 있다. 전신과 발밑 여백은 확보됐지만 인물이 화면 높이 대부분을 차지하고, 시점은 요구된 측후방이 아니라 전방이다.",
        "entities": "등장인물은 세 명뿐이다. 라울은 짙은 피부의 어린 남자아이로 뒤로 묶은 곱슬머리, 바랜 녹색 티셔츠와 베이지 반바지가 참조에 부합한다. 앰버는 밝은 피부와 금발의 어린 여자아이이며 카키 작업복과 가죽 공구 벨트를 착용했다. 방진 마스크는 목에 걸려 있다. 찰리는 흰 기계 얼굴과 베이지 장갑, 넓은 챙의 낡은 모자와 회색 긴 코트를 갖췄지만 손과 가슴의 기계 구조가 상당히 드러난다. 코트 아래 로고와 그물은 확인할 수 없다.",
        "hard_violations": [],
        "physics": "라울은 앞으로 내민 신발의 뒤꿈치가 도로에 닿는 보행 순간으로 보이고, 앰버는 앞쪽 부츠로 체중을 받는다. 찰리의 넓은 기계 발도 도로에 접촉한다. 떠 있는 발은 지지 다리를 둔 보행 동작으로 설명된다. 마스크는 목의 끈으로, 공구 주머니는 허리 벨트로 지지되며 코트와 모자는 몸에 자연스럽게 걸려 있다."
       },
       {
        "label": "B",
        "direction": "라울과 앰버는 얼굴과 시선을 화면 오른쪽의 진행 경로로 향하고 웃는다. 찰리도 오른쪽 앞을 향한다. 세 사람의 내디딘 발과 몸의 이동 방향이 오른쪽 도로와 맞는다. 다만 가슴과 얼굴 앞면이 상당히 보여 카메라가 진행축보다 약간 뒤에 있다는 조건까지 충족하지는 않는다.",
        "built_space": "왼쪽의 철망 구간과 골판금 담장 한 줄, 쓰레기 산 위 폐헬기 한 대, 오른쪽으로 이어지는 컨테이너 주택과 전신주·전선, 우측 배수로가 참조 장소와 대응한다. 세 인물은 도로 위에 라울·앰버·찰리 순으로 배치되고 오른쪽에 이동 공간이 남아 있다. 전신과 발 아래 여백, 머리 위 여백은 있지만 인물 크기가 커서 요구된 넓은 촬영 거리에는 덜 부합한다. 시점은 측후방보다 측전방에 가깝다.",
        "entities": "세 인물의 수와 순서가 맞는다. 라울의 어린 얼굴, 짙은 피부, 묶은 곱슬머리와 낡은 티셔츠·반바지가 참조와 가깝다. 앰버는 금발의 밝은 피부를 가진 어린 여자아이이며 카키 작업복, 허리 공구 주머니와 도구가 보이고 방진 마스크는 코와 입을 덮는다. 어깨에는 참조에서 두드러지지 않던 갈색 끈이 보인다. 찰리는 흰 점·선 기계 얼굴, 베이지 금속 몸체, 낡은 회색 롱코트와 챙 넓은 모자를 유지하며 손을 가려 A보다 위장이 잘 읽힌다. 가슴과 발 일부는 여전히 기계로 드러나며 로고와 그물은 가려져 확인되지 않는다.",
        "hard_violations": [],
        "physics": "라울은 뒤쪽 신발 앞부분을 도로에 대고 앞발을 내밀며, 앰버도 뒤쪽 부츠로 지지하면서 앞발을 내려놓는다. 찰리는 뒤쪽의 넓은 기계 발로 몸을 받치고 앞발을 내딛는다. 모두 지지점이 있는 가능한 보행 자세이나 세 인물의 걸음 단계가 서로 비슷해 비동기적인 보행 요구는 약하다. 마스크는 얼굴 끈으로, 공구는 허리 벨트와 주머니로 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 장소는 잘 맞지만 세 인물이 카메라 앞으로 다가오고 아이들이 서로를 바라봐, 오른쪽으로 이동하는 측후방 와이드 숏이라는 핵심 연출을 놓쳤다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽을 향한 보행과 시선, 좌우 인물 순서, 찰리의 위장이 더 충실하지만 카메라는 여전히 전방 사선이며 보폭도 지나치게 비슷하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "라울은 오른쪽의 앰버를, 앰버는 왼쪽의 라울을 보며 웃고 찰리도 아이들 쪽으로 얼굴을 돌린다. 세 몸과 내디딘 발은 대체로 카메라 쪽을 향한다. 오른쪽으로 이어지는 배경 도로는 보이지만 인물들의 이동은 그 경로를 따르지 않는다.",
        "built_space": "왼쪽에는 골판금 담장 한 줄과 철망 구간, 그 뒤에는 폐헬기 한 대가 놓인 쓰레기 더미가 있다. 오른쪽에는 컨테이너 주택 행렬, 여러 전신주와 전선, 도로 가장자리 배수로가 있어 장소의 주요 구조가 일치한다. 세 인물은 담장 밖 도로에 라울·앰버·찰리 순으로 서 있다. 전신과 발밑 여백은 확보됐지만 인물이 화면 높이 대부분을 차지하고, 시점은 요구된 측후방이 아니라 전방이다.",
        "entities": "등장인물은 세 명뿐이다. 라울은 짙은 피부의 어린 남자아이로 뒤로 묶은 곱슬머리, 바랜 녹색 티셔츠와 베이지 반바지가 참조에 부합한다. 앰버는 밝은 피부와 금발의 어린 여자아이이며 카키 작업복과 가죽 공구 벨트를 착용했다. 방진 마스크는 목에 걸려 있다. 찰리는 흰 기계 얼굴과 베이지 장갑, 넓은 챙의 낡은 모자와 회색 긴 코트를 갖췄지만 손과 가슴의 기계 구조가 상당히 드러난다. 코트 아래 로고와 그물은 확인할 수 없다.",
        "hard_violations": [],
        "physics": "라울은 앞으로 내민 신발의 뒤꿈치가 도로에 닿는 보행 순간으로 보이고, 앰버는 앞쪽 부츠로 체중을 받는다. 찰리의 넓은 기계 발도 도로에 접촉한다. 떠 있는 발은 지지 다리를 둔 보행 동작으로 설명된다. 마스크는 목의 끈으로, 공구 주머니는 허리 벨트로 지지되며 코트와 모자는 몸에 자연스럽게 걸려 있다."
       },
       {
        "label": "A",
        "direction": "라울과 앰버는 얼굴과 시선을 화면 오른쪽의 진행 경로로 향하고 웃는다. 찰리도 오른쪽 앞을 향한다. 세 사람의 내디딘 발과 몸의 이동 방향이 오른쪽 도로와 맞는다. 다만 가슴과 얼굴 앞면이 상당히 보여 카메라가 진행축보다 약간 뒤에 있다는 조건까지 충족하지는 않는다.",
        "built_space": "왼쪽의 철망 구간과 골판금 담장 한 줄, 쓰레기 산 위 폐헬기 한 대, 오른쪽으로 이어지는 컨테이너 주택과 전신주·전선, 우측 배수로가 참조 장소와 대응한다. 세 인물은 도로 위에 라울·앰버·찰리 순으로 배치되고 오른쪽에 이동 공간이 남아 있다. 전신과 발 아래 여백, 머리 위 여백은 있지만 인물 크기가 커서 요구된 넓은 촬영 거리에는 덜 부합한다. 시점은 측후방보다 측전방에 가깝다.",
        "entities": "세 인물의 수와 순서가 맞는다. 라울의 어린 얼굴, 짙은 피부, 묶은 곱슬머리와 낡은 티셔츠·반바지가 참조와 가깝다. 앰버는 금발의 밝은 피부를 가진 어린 여자아이이며 카키 작업복, 허리 공구 주머니와 도구가 보이고 방진 마스크는 코와 입을 덮는다. 어깨에는 참조에서 두드러지지 않던 갈색 끈이 보인다. 찰리는 흰 점·선 기계 얼굴, 베이지 금속 몸체, 낡은 회색 롱코트와 챙 넓은 모자를 유지하며 손을 가려 A보다 위장이 잘 읽힌다. 가슴과 발 일부는 여전히 기계로 드러나며 로고와 그물은 가려져 확인되지 않는다.",
        "hard_violations": [],
        "physics": "라울은 뒤쪽 신발 앞부분을 도로에 대고 앞발을 내밀며, 앰버도 뒤쪽 부츠로 지지하면서 앞발을 내려놓는다. 찰리는 뒤쪽의 넓은 기계 발로 몸을 받치고 앞발을 내딛는다. 모두 지지점이 있는 가능한 보행 자세이나 세 인물의 걸음 단계가 서로 비슷해 비동기적인 보행 요구는 약하다. 마스크는 얼굴 끈으로, 공구는 허리 벨트와 주머니로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.714
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리의 다리와 발이 지정된 기계 몸체가 아닌 인간의 바지와 신발로 렌더링됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1714,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "찰리의 기계 팔다리를 포함해 세 인물의 외형을 참조에 맞게 구현했으며, 요구된 구도와 배경을 정확히 따랐습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "찰리의 하반신이 지정된 로봇이 아닌 바지와 신발을 착용한 인간의 형태로 묘사되어 캐릭터 설정을 위반했습니다.  ★위반: [gemini-pro] 찰리의 다리와 발이 지정된 기계 몸체가 아닌 인간의 바지와 신발로 렌더링됨"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_scrapyard_exit_road_3033ca.png",
    "asset_id": "ba23ebba-28ad-4bbc-90cf-eeb40fb40e22",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1126517>",
    "asset_id": "13c85edd-6f32-4e1f-bff4-67869a01ee0b",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0808-35aa-7023-befa-f6787a11e6af",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S6sh34__bgfirst_bg.png",
   "bg_asset_id": "ee40dc87-434f-4e50-bffc-40732a21c99a",
   "bg_record_key": "S6sh34::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "scrapyard_exit_road",
   "groupbg_asset_id": "ba23ebba-28ad-4bbc-90cf-eeb40fb40e22"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S7sh25::signage": {
  "fp": "131ca827ba6cd73d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::189d76f17f8ec00d": {
  "subjects": [],
  "subject_text": "인천 난민촌 인공제방·보수 공사장, 제방 도로\n바다를 막아선 거대한 콘크리트 제방과 그 옆 도로. 벽 곳곳에 균열과 물샘이 보이며 지지대 주변에는 보수 자재와 돌, 웅덩이가 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L181",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::seawall_worksite": {
  "input_fingerprint": "a4541e4fc78d72c7",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "seawall_worksite",
    "tags": [
     "S33sh15",
     "S7sh25",
     "S7sh28",
     "S7sh30"
    ]
   },
   "context_sig": "93d086effec6aeb7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 인공제방·보수 공사장, 제방 도로: 바닷물을 막기 위해 세워진 거대한 콘크리트 벽과 지지대로, 금이 가고 물이 새는 노후된 구조물이다. (특징: 벽면 곳곳에 심하게 금이 가고 물줄기가 새어 나오는 거대한 콘크리트 인공제방; 자재를 옮기는 소형 지게차를 운전하는 미연(현우와 앰버의 엄마); 로만칼라 복장을 입은 60대 남자 신부; 군용 트럭과 깔끔한 정장·구두를 착용한 관리소장 및 무장 정규 경비병; 제방 벽에 다닥다닥 달라붙어 연쇄 폭발을 일으키는 자폭 드론들; 상단부가 터져나가며 쏟아져 들어오는 흙탕물과 휩쓸리는 중장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 멀리 보이는 인공제방. 보수 공사가 한창이다. 제방 벽 곳곳이 심하게 금이 가 있다.\n- 제방 근처에 있던 보수작업을 위한 중장비 기계들과 건설 장비들이 한 순간에 바닷물에 휩쓸려 간다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 인공제방·보수 공사장, 제방 도로: 바닷물을 막기 위해 세워진 거대한 콘크리트 벽과 지지대로, 금이 가고 물이 새는 노후된 구조물이다. (특징: 벽면 곳곳에 심하게 금이 가고 물줄기가 새어 나오는 거대한 콘크리트 인공제방; 자재를 옮기는 소형 지게차를 운전하는 미연(현우와 앰버의 엄마); 로만칼라 복장을 입은 60대 남자 신부; 군용 트럭과 깔끔한 정장·구두를 착용한 관리소장 및 무장 정규 경비병; 제방 벽에 다닥다닥 달라붙어 연쇄 폭발을 일으키는 자폭 드론들; 상단부가 터져나가며 쏟아져 들어오는 흙탕물과 휩쓸리는 중장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 멀리 보이는 인공제방. 보수 공사가 한창이다. 제방 벽 곳곳이 심하게 금이 가 있다.\n- 제방 근처에 있던 보수작업을 위한 중장비 기계들과 건설 장비들이 한 순간에 바닷물에 휩쓸려 간다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_worksite_44c24f.png",
  "asset_id": "fb45e62b-5030-4b51-a9e4-b701abc827ca",
  "input_asset_ids": [
   "d8b361a8-535d-42d4-ad94-db9f45c0871a"
  ],
  "origin_tag": "S7sh25",
  "place_text": "On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.",
  "origin_inputs": {
   "place_text": "On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.",
   "time_of_day_en": "day",
   "conti_asset_id": "d8b361a8-535d-42d4-ad94-db9f45c0871a"
  }
 },
 "S7sh25::bgfirst_bg": {
  "input_fingerprint": "55bc587e748f2e27",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 균열이 가고 물이 새는 방벽 쪽을 향해 한 손을 거세게 뻗은 미연의 모습.\n\nLOCATION (lock): On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the crowd's outer edge at shoulder height, observing directly from slightly behind 미연's shoulder line and diagonally across her toward 관리소장 beside the truck. Frame 미연 from the waist up at center-right, her arm thrust toward the cracked, leaking embankment at left while her head turns toward the administrator beyond the right edge; retain his retreating torso at the boundary as he directs his attention down toward the truck entrance. Emphasize her redirected arm and body position, using the extended forearm to connect the appeal with the damaged wall without forcing her to address the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Cracked and leaking embankment indicated by 미연 in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Severely cracked and leaking water) — The damaged face remains visible obliquely along the left side; used as Provides visible evidence at the end of 미연's pointing gesture; Military truck (Stopped beside the administrator) — A limited portion of its entrance side appears at the right boundary; used as Marks the administrator's intended withdrawal without occupying the center.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve the leaking cracks and the urgency of 미연's appeal without adding a dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 균열이 가고 물이 새는 방벽 쪽을 향해 한 손을 거세게 뻗은 미연의 모습.\n\nLOCATION (lock): On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the crowd's outer edge at shoulder height, observing directly from slightly behind 미연's shoulder line and diagonally across her toward 관리소장 beside the truck. Frame 미연 from the waist up at center-right, her arm thrust toward the cracked, leaking embankment at left while her head turns toward the administrator beyond the right edge; retain his retreating torso at the boundary as he directs his attention down toward the truck entrance. Emphasize her redirected arm and body position, using the extended forearm to connect the appeal with the damaged wall without forcing her to address the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Cracked and leaking embankment indicated by 미연 in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Severely cracked and leaking water) — The damaged face remains visible obliquely along the left side; used as Provides visible evidence at the end of 미연's pointing gesture; Military truck (Stopped beside the administrator) — A limited portion of its entrance side appears at the right boundary; used as Marks the administrator's intended withdrawal without occupying the center.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve the leaking cracks and the urgency of 미연's appeal without adding a dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S7sh25__bgfirst_bg.png",
  "asset_id": "3e4684d4-cfe5-45a2-8516-ee2307f166cf",
  "input_asset_ids": [
   "d8b361a8-535d-42d4-ad94-db9f45c0871a",
   "fb45e62b-5030-4b51-a9e4-b701abc827ca"
  ]
 },
 "S7sh25": {
  "input_fingerprint": "48bc259ea47ce379",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 균열이 가고 물이 새는 방벽 쪽을 향해 한 손을 거세게 뻗은 미연의 모습.\n\nLOCATION (lock): On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the crowd's outer edge at shoulder height, observing directly from slightly behind 미연's shoulder line and diagonally across her toward 관리소장 beside the truck. Frame 미연 from the waist up at center-right, her arm thrust toward the cracked, leaking embankment at left while her head turns toward the administrator beyond the right edge; retain his retreating torso at the boundary as he directs his attention down toward the truck entrance. Emphasize her redirected arm and body position, using the extended forearm to connect the appeal with the damaged wall without forcing her to address the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Cracked and leaking embankment indicated by 미연 in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Severely cracked and leaking water) — The damaged face remains visible obliquely along the left side; used as Provides visible evidence at the end of 미연's pointing gesture; Military truck (Stopped beside the administrator) — A limited portion of its entrance side appears at the right boundary; used as Marks the administrator's intended withdrawal without occupying the center.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve the leaking cracks and the urgency of 미연's appeal without adding a dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage, with unfinished repair materials and equipment still at the site. The military truck remains stopped near the forklift, and puddles remain on the ground. 미연: Has left the forklift and stands on the ground, urgently gesturing toward the leaking wall. 관리소장: Is beside the military truck with puddle-soiled shoes and the megaphone still in his possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 균열이 가고 물이 새는 방벽 쪽을 향해 한 손을 거세게 뻗은 미연의 모습.\n\nLOCATION (lock): On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the crowd's outer edge at shoulder height, observing directly from slightly behind 미연's shoulder line and diagonally across her toward 관리소장 beside the truck. Frame 미연 from the waist up at center-right, her arm thrust toward the cracked, leaking embankment at left while her head turns toward the administrator beyond the right edge; retain his retreating torso at the boundary as he directs his attention down toward the truck entrance. Emphasize her redirected arm and body position, using the extended forearm to connect the appeal with the damaged wall without forcing her to address the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Cracked and leaking embankment indicated by 미연 in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Severely cracked and leaking water) — The damaged face remains visible obliquely along the left side; used as Provides visible evidence at the end of 미연's pointing gesture; Military truck (Stopped beside the administrator) — A limited portion of its entrance side appears at the right boundary; used as Marks the administrator's intended withdrawal without occupying the center.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve the leaking cracks and the urgency of 미연's appeal without adding a dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage, with unfinished repair materials and equipment still at the site. The military truck remains stopped near the forklift, and puddles remain on the ground. 미연: Has left the forklift and stands on the ground, urgently gesturing toward the leaking wall. 관리소장: Is beside the military truck with puddle-soiled shoes and the megaphone still in his possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 균열이 가고 물이 새는 방벽 쪽을 향해 한 손을 거세게 뻗은 미연의 모습.\n\nLOCATION (lock): On the outdoor repair site at the foot of a cracked artificial seawall, beside its visibly leaking wall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the crowd's outer edge at shoulder height, observing directly from slightly behind 미연's shoulder line and diagonally across her toward 관리소장 beside the truck. Frame 미연 from the waist up at center-right, her arm thrust toward the cracked, leaking embankment at left while her head turns toward the administrator beyond the right edge; retain his retreating torso at the boundary as he directs his attention down toward the truck entrance. Emphasize her redirected arm and body position, using the extended forearm to connect the appeal with the damaged wall without forcing her to address the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Cracked and leaking embankment indicated by 미연 in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Severely cracked and leaking water) — The damaged face remains visible obliquely along the left side; used as Provides visible evidence at the end of 미연's pointing gesture; Military truck (Stopped beside the administrator) — A limited portion of its entrance side appears at the right boundary; used as Marks the administrator's intended withdrawal without occupying the center.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve the leaking cracks and the urgency of 미연's appeal without adding a dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall has severe cracks and visible water leakage, with unfinished repair materials and equipment still at the site. The military truck remains stopped near the forklift, and puddles remain on the ground. 미연: Has left the forklift and stands on the ground, urgently gesturing toward the leaking wall. 관리소장: Is beside the military truck with puddle-soiled shoes and the megaphone still in his possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S7sh25__bgfirst_bg.png",
     "asset_id": "3e4684d4-cfe5-45a2-8516-ee2307f166cf",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S7sh25.png",
     "asset_id": "d8b361a8-535d-42d4-ad94-db9f45c0871a",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 관리소장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1319720>",
     "asset_id": "20aac1ba-1f77-4ab5-b88f-b337ea3b046b",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_worksite_44c24f.png",
     "asset_id": "fb45e62b-5030-4b51-a9e4-b701abc827ca",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 관리소장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1319720>",
     "asset_id": "20aac1ba-1f77-4ab5-b88f-b337ea3b046b",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연의 뻗은 손은 왼쪽 누수 방벽을 향하고, 고개는 오른쪽을 향함. 관리소장은 우측 트럭 방향으로 시선을 둠.",
    "built_space": "좌측의 갈라진 인공 방벽, 우측 경계의 군용 트럭, 중앙 배경의 지게차 등 참조된 장소 구조가 올바르게 배치됨.",
    "entities": "미연은 지시된 회색 셔츠와 어두운 바지를 착용했으며, 관리소장은 대머리 외형에 확성기를 소지한 모습이 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 관리소장이 들고 있는 확성기의 파지 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "미연의 손은 왼쪽 방벽을 가리키나, 관리소장의 시선이 트럭 출입구가 아닌 미연을 향해 있음.",
    "built_space": "참조 이미지와 동일한 방벽, 물웅덩이, 군용 트럭, 지게차가 지시된 프레임 구도에 맞게 위치함.",
    "entities": "미연의 얼굴과 복장, 관리소장의 대머리 및 확성기 소지 상태 모두 주어진 참조 정보와 부합함.",
    "hard_violations": [],
    "physics": "인물들의 발이 지면에 닿아 있고 자세가 동작에 맞게 지지되고 있으며 물리적 오류가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '미연의 어깨선 뒤에서 바라보는 카메라 앵글'과 '트럭을 향해 몸을 돌린 관리소장의 뒷모습'을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라가 미연의 정면에 위치해 뷰포인트 지시를 어겼으며, 관리소장 역시 후퇴하지 않고 미연을 마주보고 있어 연출 의도와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 뻗은 손은 왼쪽 누수 방벽을 향하고, 고개는 오른쪽을 향함. 관리소장은 우측 트럭 방향으로 시선을 둠.",
        "built_space": "좌측의 갈라진 인공 방벽, 우측 경계의 군용 트럭, 중앙 배경의 지게차 등 참조된 장소 구조가 올바르게 배치됨.",
        "entities": "미연은 지시된 회색 셔츠와 어두운 바지를 착용했으며, 관리소장은 대머리 외형에 확성기를 소지한 모습이 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 관리소장이 들고 있는 확성기의 파지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "미연의 손은 왼쪽 방벽을 가리키나, 관리소장의 시선이 트럭 출입구가 아닌 미연을 향해 있음.",
        "built_space": "참조 이미지와 동일한 방벽, 물웅덩이, 군용 트럭, 지게차가 지시된 프레임 구도에 맞게 위치함.",
        "entities": "미연의 얼굴과 복장, 관리소장의 대머리 및 확성기 소지 상태 모두 주어진 참조 정보와 부합함.",
        "hard_violations": [],
        "physics": "인물들의 발이 지면에 닿아 있고 자세가 동작에 맞게 지지되고 있으며 물리적 오류가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 '미연의 어깨선 뒤에서 바라보는 카메라 앵글'과 '트럭을 향해 몸을 돌린 관리소장의 뒷모습'을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라가 미연의 정면에 위치해 뷰포인트 지시를 어겼으며, 관리소장 역시 후퇴하지 않고 미연을 마주보고 있어 연출 의도와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연의 뻗은 손은 왼쪽 누수 방벽을 향하고, 고개는 오른쪽을 향함. 관리소장은 우측 트럭 방향으로 시선을 둠.",
        "built_space": "좌측의 갈라진 인공 방벽, 우측 경계의 군용 트럭, 중앙 배경의 지게차 등 참조된 장소 구조가 올바르게 배치됨.",
        "entities": "미연은 지시된 회색 셔츠와 어두운 바지를 착용했으며, 관리소장은 대머리 외형에 확성기를 소지한 모습이 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 관리소장이 들고 있는 확성기의 파지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "미연의 손은 왼쪽 방벽을 가리키나, 관리소장의 시선이 트럭 출입구가 아닌 미연을 향해 있음.",
        "built_space": "참조 이미지와 동일한 방벽, 물웅덩이, 군용 트럭, 지게차가 지시된 프레임 구도에 맞게 위치함.",
        "entities": "미연의 얼굴과 복장, 관리소장의 대머리 및 확성기 소지 상태 모두 주어진 참조 정보와 부합함.",
        "hard_violations": [],
        "physics": "인물들의 발이 지면에 닿아 있고 자세가 동작에 맞게 지지되고 있으며 물리적 오류가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "방벽으로 뻗은 팔과 관리소장을 향해 돌린 머리, 트럭 쪽으로 물러나는 관리소장의 관계가 더 정확하지만, 카메라는 요구된 미연의 어깨 뒤가 아니라 몸통 앞쪽에 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "누수 지점으로 향하는 손과 관리소장을 보는 시선은 맞지만, 미연을 정면에 가깝게 관찰하며 관리소장의 트럭 쪽 퇴장 동작도 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 뻗은 검지는 왼쪽 방벽의 두 번째 큰 누수와 균열 부근을 향하고, 눈은 오른쪽 관리소장의 얼굴을 본다. 관리소장은 고개를 아래로 숙였지만 시선이 트럭 출입구보다 자기 앞 확성기 쪽으로 읽힌다. 트럭으로 돌아서 물러나는 방향은 분명하지 않다.",
        "built_space": "왼쪽에 연속된 콘크리트 방벽 하나와 큰 낙수 두 줄기, 상단 방호 블록들이 있다. 오른쪽에는 군용 트럭 한 대의 출입구 측면과 거울 하나, 그 뒤로 지게차 한 대가 보인다. 등대 하나, 모래주머니와 적재 자재, 교통콘 하나, 바닥 웅덩이가 참고 장소와 대응한다. 미연은 중앙 오른쪽에서 허리 부근까지 나오고 관리소장은 오른쪽 경계에 잘린다. 다만 미연의 셔츠 앞면과 얼굴을 넓게 보는 전방 사선 시점으로, 어깨선 뒤에서 관찰하라는 카메라 조건은 충족하지 않는다.",
        "entities": "등장 인물은 미연과 관리소장 두 명이다. 미연은 중년 동아시아계 여성으로 보이며 검은 단발, 얼굴 윤곽과 체형, 먼지 묻고 해진 회색 셔츠가 참고와 가깝다. 관리소장은 중년의 대머리 동아시아계 남성으로, 보이는 옆얼굴과 짙은 남색 상의가 참고에 가깝다. 흰색·붉은색 확성기를 소지한다. 군용 트럭, 지게차, 심하게 갈라져 물이 새는 인공 방벽과 미완성 보수 자재가 모두 식별된다. 신발은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "미연의 팔은 어깨에서 자연스럽게 이어지고 팔꿈치를 펴 방벽을 가리키는 동작으로 성립한다. 두 사람의 발은 보이지 않지만 몸통이 아래 프레임으로 이어져 공중에 뜬 정황은 없다. 확성기 아래에는 관리소장의 손이 보여 지지가 확인된다. 자재와 차량은 지면에 놓이고, 누수는 균열에서 아래로 떨어져 웅덩이로 이어진다."
       },
       {
        "label": "B",
        "direction": "미연의 검지는 왼쪽 방벽의 균열과 두 번째 낙수 상부를 향한다. 머리는 오른쪽으로 크게 돌아가 트럭 옆 관리소장을 향하며 렌즈를 보지 않는다. 관리소장은 등을 보인 채 트럭 출입구 방향으로 몸을 돌리고 고개를 아래로 기울여, 물러나는 동선이 A보다 분명하다.",
        "built_space": "왼쪽에 긴 콘크리트 방벽 하나와 큰 누수 두 곳, 반복되는 상단 방호 블록이 보인다. 오른쪽의 군용 트럭 한 대와 거울 하나, 그 옆 지게차 한 대, 먼 등대 하나, 모래주머니 적재물과 교통콘 하나 및 웅덩이가 참고 장소의 공간 관계를 유지한다. 미연은 중앙 오른쪽의 허리 부근까지 잡히고 트럭은 오른쪽 경계에 한정된다. 관리소장은 경계에 있지만 몸통뿐 아니라 허벅지까지 비교적 많이 보인다. 미연의 얼굴은 돌아가 있으나 셔츠 앞면이 정면으로 보여, 카메라 자체는 요구된 어깨 뒤 시점이 아니다.",
        "entities": "미연과 관리소장 두 명만 식별된다. 미연은 검은 단발의 중년 동아시아계 여성으로, 회색의 해지고 더러운 셔츠와 어두운 안쪽 옷이 참고에 부합한다. 얼굴 대부분이 가려져 정확한 얼굴 일치는 제한적으로만 판단된다. 관리소장은 대머리의 중년 동아시아계 남성으로 보이나 얼굴은 거의 숨겨져 있고, 참고의 니트 대신 짙은 작업 재킷을 입었다. 오른쪽 아래에 확성기가 있으며 군용 트럭, 지게차, 누수 방벽과 보수 자재도 확인된다. 신발의 오염 여부는 프레임 밖이라 판단할 수 없다.",
        "hard_violations": [],
        "physics": "미연은 몸통을 유지하면서 한 팔을 왼쪽으로 뻗고 목을 오른쪽으로 돌린 현실적으로 가능한 자세다. 관리소장의 몸과 다리는 프레임 아래로 이어지며 부유 징후는 없다. 확성기는 내려간 팔 끝의 손 위치에 붙어 있으나 손잡이 접촉은 오른쪽 경계에 가려져 세부 확인이 어렵다. 차량과 적재 자재는 바닥에 지지되고 물은 벽에서 중력 방향으로 흘러내린다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "방벽으로 뻗은 팔과 관리소장을 향해 돌린 머리, 트럭 쪽으로 물러나는 관리소장의 관계가 더 정확하지만, 카메라는 요구된 미연의 어깨 뒤가 아니라 몸통 앞쪽에 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "누수 지점으로 향하는 손과 관리소장을 보는 시선은 맞지만, 미연을 정면에 가깝게 관찰하며 관리소장의 트럭 쪽 퇴장 동작도 약하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연의 뻗은 검지는 왼쪽 방벽의 두 번째 큰 누수와 균열 부근을 향하고, 눈은 오른쪽 관리소장의 얼굴을 본다. 관리소장은 고개를 아래로 숙였지만 시선이 트럭 출입구보다 자기 앞 확성기 쪽으로 읽힌다. 트럭으로 돌아서 물러나는 방향은 분명하지 않다.",
        "built_space": "왼쪽에 연속된 콘크리트 방벽 하나와 큰 낙수 두 줄기, 상단 방호 블록들이 있다. 오른쪽에는 군용 트럭 한 대의 출입구 측면과 거울 하나, 그 뒤로 지게차 한 대가 보인다. 등대 하나, 모래주머니와 적재 자재, 교통콘 하나, 바닥 웅덩이가 참고 장소와 대응한다. 미연은 중앙 오른쪽에서 허리 부근까지 나오고 관리소장은 오른쪽 경계에 잘린다. 다만 미연의 셔츠 앞면과 얼굴을 넓게 보는 전방 사선 시점으로, 어깨선 뒤에서 관찰하라는 카메라 조건은 충족하지 않는다.",
        "entities": "등장 인물은 미연과 관리소장 두 명이다. 미연은 중년 동아시아계 여성으로 보이며 검은 단발, 얼굴 윤곽과 체형, 먼지 묻고 해진 회색 셔츠가 참고와 가깝다. 관리소장은 중년의 대머리 동아시아계 남성으로, 보이는 옆얼굴과 짙은 남색 상의가 참고에 가깝다. 흰색·붉은색 확성기를 소지한다. 군용 트럭, 지게차, 심하게 갈라져 물이 새는 인공 방벽과 미완성 보수 자재가 모두 식별된다. 신발은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "미연의 팔은 어깨에서 자연스럽게 이어지고 팔꿈치를 펴 방벽을 가리키는 동작으로 성립한다. 두 사람의 발은 보이지 않지만 몸통이 아래 프레임으로 이어져 공중에 뜬 정황은 없다. 확성기 아래에는 관리소장의 손이 보여 지지가 확인된다. 자재와 차량은 지면에 놓이고, 누수는 균열에서 아래로 떨어져 웅덩이로 이어진다."
       },
       {
        "label": "A",
        "direction": "미연의 검지는 왼쪽 방벽의 균열과 두 번째 낙수 상부를 향한다. 머리는 오른쪽으로 크게 돌아가 트럭 옆 관리소장을 향하며 렌즈를 보지 않는다. 관리소장은 등을 보인 채 트럭 출입구 방향으로 몸을 돌리고 고개를 아래로 기울여, 물러나는 동선이 A보다 분명하다.",
        "built_space": "왼쪽에 긴 콘크리트 방벽 하나와 큰 누수 두 곳, 반복되는 상단 방호 블록이 보인다. 오른쪽의 군용 트럭 한 대와 거울 하나, 그 옆 지게차 한 대, 먼 등대 하나, 모래주머니 적재물과 교통콘 하나 및 웅덩이가 참고 장소의 공간 관계를 유지한다. 미연은 중앙 오른쪽의 허리 부근까지 잡히고 트럭은 오른쪽 경계에 한정된다. 관리소장은 경계에 있지만 몸통뿐 아니라 허벅지까지 비교적 많이 보인다. 미연의 얼굴은 돌아가 있으나 셔츠 앞면이 정면으로 보여, 카메라 자체는 요구된 어깨 뒤 시점이 아니다.",
        "entities": "미연과 관리소장 두 명만 식별된다. 미연은 검은 단발의 중년 동아시아계 여성으로, 회색의 해지고 더러운 셔츠와 어두운 안쪽 옷이 참고에 부합한다. 얼굴 대부분이 가려져 정확한 얼굴 일치는 제한적으로만 판단된다. 관리소장은 대머리의 중년 동아시아계 남성으로 보이나 얼굴은 거의 숨겨져 있고, 참고의 니트 대신 짙은 작업 재킷을 입었다. 오른쪽 아래에 확성기가 있으며 군용 트럭, 지게차, 누수 방벽과 보수 자재도 확인된다. 신발의 오염 여부는 프레임 밖이라 판단할 수 없다.",
        "hard_violations": [],
        "physics": "미연은 몸통을 유지하면서 한 팔을 왼쪽으로 뻗고 목을 오른쪽으로 돌린 현실적으로 가능한 자세다. 관리소장의 몸과 다리는 프레임 아래로 이어지며 부유 징후는 없다. 확성기는 내려간 팔 끝의 손 위치에 붙어 있으나 손잡이 접촉은 오른쪽 경계에 가려져 세부 확인이 어렵다. 차량과 적재 자재는 바닥에 지지되고 물은 벽에서 중력 방향으로 흘러내린다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.429
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.429
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 '미연의 어깨선 뒤에서 바라보는 카메라 앵글'과 '트럭을 향해 몸을 돌린 관리소장의 뒷모습'을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "카메라가 미연의 정면에 위치해 뷰포인트 지시를 어겼으며, 관리소장 역시 후퇴하지 않고 미연을 마주보고 있어 연출 의도와 다릅니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_worksite_44c24f.png",
    "asset_id": "fb45e62b-5030-4b51-a9e4-b701abc827ca",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 관리소장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1319720>",
    "asset_id": "20aac1ba-1f77-4ab5-b88f-b337ea3b046b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0810-7e71-7822-b0dc-d875ef8571ea",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S7sh25__bgfirst_bg.png",
   "bg_asset_id": "3e4684d4-cfe5-45a2-8516-ee2307f166cf",
   "bg_record_key": "S7sh25::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "seawall_worksite",
   "groupbg_asset_id": "fb45e62b-5030-4b51-a9e4-b701abc827ca"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C44"
  ]
 },
 "S7sh28::signage": {
  "fp": "d1f914fddec8cbc4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh28": {
  "input_fingerprint": "4531bf28d405536e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 땅바닥에 쓰러진 난민 남자의 곁에 무릎을 꿇고 두 팔로 그의 어깨를 감싸 쥔 미연의 전신.\n\nLOCATION (lock): On the wet ground of the seawall repair site, in the confrontation area beside the stopped military truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the crane settle below standing shoulder height on the same side of the action, easing back into a slightly downward three-quarter view that directly observes 미연's entire kneeling body and both arms. Place her left of center, folding over the fallen man's shoulders as she looks down to support him; he lies diagonally across the lower middle and looks upward, while 관리소장 recedes toward the truck at the far right. Emphasize 미연's lowered position, leaving the surrounding workers at irregular depths and different phases of crowding inward, their attention dropping toward the fallen man rather than forming a uniform row.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground beside the repair site (Supporting the fallen man and kneeling 미연); used as Keeps both full bodies and their points of contact readable; Military truck (The administrator is returning to it) — Its entrance side remains oblique at the far-right background; used as Contrasts the withdrawal in the background with assistance in the foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daylight and the preceding contrast level, preserving clear separation between the supporting arms and the fallen body.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 관리소장, 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked, leaking barrier and the wet construction ground from the reference. Exclude materials from the separate rubbish dump and any active construction operation, since work has been halted.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues to leak, and the repair equipment remains at the unfinished site. The military truck is still at the puddled stopping place near the forklift. 미연: Is low to the ground with her arms extended in a supporting posture. 관리소장: Is hurriedly boarding the military truck, with his shoes still soiled by puddle water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 땅바닥에 쓰러진 난민 남자의 곁에 무릎을 꿇고 두 팔로 그의 어깨를 감싸 쥔 미연의 전신.\n\nLOCATION (lock): On the wet ground of the seawall repair site, in the confrontation area beside the stopped military truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the crane settle below standing shoulder height on the same side of the action, easing back into a slightly downward three-quarter view that directly observes 미연's entire kneeling body and both arms. Place her left of center, folding over the fallen man's shoulders as she looks down to support him; he lies diagonally across the lower middle and looks upward, while 관리소장 recedes toward the truck at the far right. Emphasize 미연's lowered position, leaving the surrounding workers at irregular depths and different phases of crowding inward, their attention dropping toward the fallen man rather than forming a uniform row.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground beside the repair site (Supporting the fallen man and kneeling 미연); used as Keeps both full bodies and their points of contact readable; Military truck (The administrator is returning to it) — Its entrance side remains oblique at the far-right background; used as Contrasts the withdrawal in the background with assistance in the foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daylight and the preceding contrast level, preserving clear separation between the supporting arms and the fallen body.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 관리소장, 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked, leaking barrier and the wet construction ground from the reference. Exclude materials from the separate rubbish dump and any active construction operation, since work has been halted.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues to leak, and the repair equipment remains at the unfinished site. The military truck is still at the puddled stopping place near the forklift. 미연: Is low to the ground with her arms extended in a supporting posture. 관리소장: Is hurriedly boarding the military truck, with his shoes still soiled by puddle water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 땅바닥에 쓰러진 난민 남자의 곁에 무릎을 꿇고 두 팔로 그의 어깨를 감싸 쥔 미연의 전신.\n\nLOCATION (lock): On the wet ground of the seawall repair site, in the confrontation area beside the stopped military truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the crane settle below standing shoulder height on the same side of the action, easing back into a slightly downward three-quarter view that directly observes 미연's entire kneeling body and both arms. Place her left of center, folding over the fallen man's shoulders as she looks down to support him; he lies diagonally across the lower middle and looks upward, while 관리소장 recedes toward the truck at the far right. Emphasize 미연's lowered position, leaving the surrounding workers at irregular depths and different phases of crowding inward, their attention dropping toward the fallen man rather than forming a uniform row.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ground beside the repair site (Supporting the fallen man and kneeling 미연); used as Keeps both full bodies and their points of contact readable; Military truck (The administrator is returning to it) — Its entrance side remains oblique at the far-right background; used as Contrasts the withdrawal in the background with assistance in the foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daylight and the preceding contrast level, preserving clear separation between the supporting arms and the fallen body.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 관리소장, 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked, leaking barrier and the wet construction ground from the reference. Exclude materials from the separate rubbish dump and any active construction operation, since work has been halted.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cracked seawall continues to leak, and the repair equipment remains at the unfinished site. The military truck is still at the puddled stopping place near the forklift. 미연: Is low to the ground with her arms extended in a supporting posture. 관리소장: Is hurriedly boarding the military truck, with his shoes still soiled by puddle water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 관리소장 (한국인 남성, 50대의 얼굴, 대머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "미연은 쓰러진 남자를 내려다보고 있으며, 남자는 미연을 올려다봄. 관리소장은 우측 군용 트럭 운전석을 향하고 있음.",
    "built_space": "좌측에 물이 새는 방파제가 있고, 우측에 지게차와 군용 트럭이 배치되어 기준 이미지의 공간 구조를 정확히 따름.",
    "entities": "미연의 얼굴과 복장이 기준 이미지와 정확히 일치함. 쓰러진 난민 남자가 묘사되었으며, 관리소장 역시 어두운 작업복 차림으로 기준 사진의 뒷모습과 부합함.",
    "hard_violations": [],
    "physics": "미연은 무릎을 꿇고 지면의 지지를 받으며, 남자는 땅에 누워 있음. 관리소장은 왼발을 땅에 딛고 오른발은 트럭 발판을 디딘 채 손잡이를 잡고 있어 물리적으로 안정적임."
   },
   {
    "label": "B",
    "direction": "미연은 쓰러진 남자를 향해 시선을 두고, 남자도 미연을 바라봄. 관리소장은 트럭 적재함 쪽을 향해 몸을 기울이고 있음.",
    "built_space": "방파제, 지게차, 군용 트럭 등 주요 구조물이 기준 사진과 동일한 위치에 배치됨.",
    "entities": "미연의 측면 얼굴과 의상이 기준과 일치함. 난민 남자가 있으며, 관리소장도 지정된 인상착의를 갖춤.",
    "hard_violations": [],
    "physics": "미연과 남자는 지면의 지지를 받으나, 남자 왼쪽 어깨 부근에 누구의 것인지 불분명한 손가락들이 엉켜 있어 신체 구조가 부정확함. 관리소장은 적재함에 오르고 있으나 왼다리의 지지 상태가 시각적으로 불분명함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "미연과 난민 남자의 상호작용이 지문에 맞게 잘 표현되었으며, 관리소장이 트럭에 오르는 자세와 물리적 접촉이 매우 자연스럽고 영화적인 연출이 돋보입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 구도와 인물 배치는 준수하나, 쓰러진 남자 어깨 부근의 손 해부학적 구조가 엉켜 보이고 관리소장의 탑승 자세가 다소 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 쓰러진 남자를 내려다보고 있으며, 남자는 미연을 올려다봄. 관리소장은 우측 군용 트럭 운전석을 향하고 있음.",
        "built_space": "좌측에 물이 새는 방파제가 있고, 우측에 지게차와 군용 트럭이 배치되어 기준 이미지의 공간 구조를 정확히 따름.",
        "entities": "미연의 얼굴과 복장이 기준 이미지와 정확히 일치함. 쓰러진 난민 남자가 묘사되었으며, 관리소장 역시 어두운 작업복 차림으로 기준 사진의 뒷모습과 부합함.",
        "hard_violations": [],
        "physics": "미연은 무릎을 꿇고 지면의 지지를 받으며, 남자는 땅에 누워 있음. 관리소장은 왼발을 땅에 딛고 오른발은 트럭 발판을 디딘 채 손잡이를 잡고 있어 물리적으로 안정적임."
       },
       {
        "label": "B",
        "direction": "미연은 쓰러진 남자를 향해 시선을 두고, 남자도 미연을 바라봄. 관리소장은 트럭 적재함 쪽을 향해 몸을 기울이고 있음.",
        "built_space": "방파제, 지게차, 군용 트럭 등 주요 구조물이 기준 사진과 동일한 위치에 배치됨.",
        "entities": "미연의 측면 얼굴과 의상이 기준과 일치함. 난민 남자가 있으며, 관리소장도 지정된 인상착의를 갖춤.",
        "hard_violations": [],
        "physics": "미연과 남자는 지면의 지지를 받으나, 남자 왼쪽 어깨 부근에 누구의 것인지 불분명한 손가락들이 엉켜 있어 신체 구조가 부정확함. 관리소장은 적재함에 오르고 있으나 왼다리의 지지 상태가 시각적으로 불분명함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "미연과 난민 남자의 상호작용이 지문에 맞게 잘 표현되었으며, 관리소장이 트럭에 오르는 자세와 물리적 접촉이 매우 자연스럽고 영화적인 연출이 돋보입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 구도와 인물 배치는 준수하나, 쓰러진 남자 어깨 부근의 손 해부학적 구조가 엉켜 보이고 관리소장의 탑승 자세가 다소 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연은 쓰러진 남자를 내려다보고 있으며, 남자는 미연을 올려다봄. 관리소장은 우측 군용 트럭 운전석을 향하고 있음.",
        "built_space": "좌측에 물이 새는 방파제가 있고, 우측에 지게차와 군용 트럭이 배치되어 기준 이미지의 공간 구조를 정확히 따름.",
        "entities": "미연의 얼굴과 복장이 기준 이미지와 정확히 일치함. 쓰러진 난민 남자가 묘사되었으며, 관리소장 역시 어두운 작업복 차림으로 기준 사진의 뒷모습과 부합함.",
        "hard_violations": [],
        "physics": "미연은 무릎을 꿇고 지면의 지지를 받으며, 남자는 땅에 누워 있음. 관리소장은 왼발을 땅에 딛고 오른발은 트럭 발판을 디딘 채 손잡이를 잡고 있어 물리적으로 안정적임."
       },
       {
        "label": "B",
        "direction": "미연은 쓰러진 남자를 향해 시선을 두고, 남자도 미연을 바라봄. 관리소장은 트럭 적재함 쪽을 향해 몸을 기울이고 있음.",
        "built_space": "방파제, 지게차, 군용 트럭 등 주요 구조물이 기준 사진과 동일한 위치에 배치됨.",
        "entities": "미연의 측면 얼굴과 의상이 기준과 일치함. 난민 남자가 있으며, 관리소장도 지정된 인상착의를 갖춤.",
        "hard_violations": [],
        "physics": "미연과 남자는 지면의 지지를 받으나, 남자 왼쪽 어깨 부근에 누구의 것인지 불분명한 손가락들이 엉켜 있어 신체 구조가 부정확함. 관리소장은 적재함에 오르고 있으나 왼다리의 지지 상태가 시각적으로 불분명함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "약간 내려다보는 넓은 시점과 두 사람의 전신·접촉점, 관리소장의 승차 동작이 더 정확하지만, 주변 노동자들이 없고 미연의 한 손은 어깨보다 가슴을 붙잡는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미연의 왼쪽 배치와 얼굴은 잘 맞지만, 카메라가 더 낮고 난민의 다리가 전경을 크게 차지해 지정된 하향 사선 와이드 구도에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 고개를 숙여 난민 남자의 얼굴을 보고, 남자는 얼굴과 시선을 위쪽의 미연에게 향한다. 관리소장은 등을 보인 채 오른쪽 트럭 출입구를 향해 올라간다. 주변에서 남자를 내려다보며 모여드는 노동자들은 없다.",
        "built_space": "왼쪽에는 크게 갈라지고 두 주요 지점에서 물이 흐르는 방벽 하나가 이어진다. 오른쪽에는 군용 트럭 한 대와 그 뒤편의 정지한 지게차 한 대가 있고, 모래주머니 더미·교통콘 하나·먼 등대 하나가 보인다. 젖은 작업장과 방벽의 재질은 이전 장면에 부합한다. 카메라는 두 사람을 약간 내려다보며 전신과 바닥 접촉을 보여 준다. 미연은 중앙보다 조금 왼쪽, 남자는 하단 중앙을 대각선으로 가로지르고, 관리소장은 오른쪽 뒤에 있다. 물웅덩이 반사는 배치상 자연스럽다.",
        "entities": "미연은 검은 단발머리의 중년 동아시아계 여성으로, 참고의 먼지 묻은 회색 셔츠와 어두운 작업 바지·작업화를 유지한다. 옆얼굴이라 정확한 얼굴 일치 확인에는 제한이 있다. 관리소장은 대머리 중년 남성이며 이전 장면의 어두운 작업복을 입고 있으나 얼굴은 보이지 않는다. 난민은 곱슬머리와 수염, 낡고 더러운 작업복을 가진 성인 남성이다. 미연·난민·관리소장 세 명만 있으며, 카메라 지시에 명시된 주변 노동자는 누락됐다. 낮이지만 직사광과 선명한 그림자가 있어 요청한 절제된 빛보다 밝다.",
        "hard_violations": [],
        "physics": "미연은 접힌 다리와 무릎을 젖은 땅에 대고 체중을 지탱한다. 난민의 다리·골반과 한 손은 바닥에 놓이고, 들린 상체는 미연의 몸과 어깨 뒤로 감긴 팔에 받쳐진다. 다른 손은 어깨가 아니라 가슴 쪽 옷을 쥔다. 관리소장은 출입구를 손으로 잡고 발판에 발을 걸쳐 올라가므로 지지점이 있다. 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "미연의 시선은 아래의 난민 얼굴을 향하고, 난민의 얼굴은 미연 쪽 위로 들려 있다. 눈이 가늘게 감겨 있어 난민의 실제 시선은 A보다 불분명하다. 관리소장은 트럭 출입구를 향해 몸을 돌리고 손을 뻗는다. 난민에게 주의를 돌리는 주변 노동자들은 없다.",
        "built_space": "왼쪽의 누수 방벽 하나, 오른쪽 군용 트럭 한 대와 지게차 한 대, 모래주머니 더미와 교통콘 하나, 먼 등대 하나가 보인다. 젖은 콘크리트와 파손 잔해는 같은 수리 현장으로 읽힌다. 미연은 명확히 왼쪽에 있고 난민은 하단을 대각선으로 가로지르지만, 시점이 낮아 바닥을 내려다보는 정도가 약하고 난민의 다리와 신발이 전경을 크게 점유한다. 관리소장은 오른쪽 뒤의 출입구에 있다. 물웅덩이 반사에서 불가능한 대응은 보이지 않는다.",
        "entities": "미연은 참고와 가까운 중년 동아시아계 여성의 얼굴과 검은 단발, 해진 회색 셔츠 및 어두운 작업복을 보여 준다. 관리소장은 대머리 중년 남성으로 이전 장면의 어두운 작업복을 유지하며, 뒷모습이라 얼굴 동일성은 확인하기 어렵다. 난민은 곱슬머리와 수염을 가진 성인 남성이고 더러운 작업복과 작업화를 착용한다. 등장인물은 세 명뿐이어서 지시된 노동자 군집은 없다. 밝은 햇빛과 뚜렷한 그림자는 차분한 낮 조명 요구보다 강하다.",
        "hard_violations": [],
        "physics": "미연의 무릎과 접힌 정강이가 바닥을 지지하며, 난민의 골반·다리·팔은 지면에 놓인다. 난민의 머리와 어깨는 미연의 몸과 감싼 팔에 받쳐져 있다. 미연의 앞쪽 손은 어깨 대신 가슴 아래쪽 옷을 쥔다. 관리소장은 한 발을 바닥에 두고 손으로 출입구를 잡은 채 다른 발을 들어 승차를 시작한다. 지지 없는 부유나 명백히 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "약간 내려다보는 넓은 시점과 두 사람의 전신·접촉점, 관리소장의 승차 동작이 더 정확하지만, 주변 노동자들이 없고 미연의 한 손은 어깨보다 가슴을 붙잡는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미연의 왼쪽 배치와 얼굴은 잘 맞지만, 카메라가 더 낮고 난민의 다리가 전경을 크게 차지해 지정된 하향 사선 와이드 구도에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 고개를 숙여 난민 남자의 얼굴을 보고, 남자는 얼굴과 시선을 위쪽의 미연에게 향한다. 관리소장은 등을 보인 채 오른쪽 트럭 출입구를 향해 올라간다. 주변에서 남자를 내려다보며 모여드는 노동자들은 없다.",
        "built_space": "왼쪽에는 크게 갈라지고 두 주요 지점에서 물이 흐르는 방벽 하나가 이어진다. 오른쪽에는 군용 트럭 한 대와 그 뒤편의 정지한 지게차 한 대가 있고, 모래주머니 더미·교통콘 하나·먼 등대 하나가 보인다. 젖은 작업장과 방벽의 재질은 이전 장면에 부합한다. 카메라는 두 사람을 약간 내려다보며 전신과 바닥 접촉을 보여 준다. 미연은 중앙보다 조금 왼쪽, 남자는 하단 중앙을 대각선으로 가로지르고, 관리소장은 오른쪽 뒤에 있다. 물웅덩이 반사는 배치상 자연스럽다.",
        "entities": "미연은 검은 단발머리의 중년 동아시아계 여성으로, 참고의 먼지 묻은 회색 셔츠와 어두운 작업 바지·작업화를 유지한다. 옆얼굴이라 정확한 얼굴 일치 확인에는 제한이 있다. 관리소장은 대머리 중년 남성이며 이전 장면의 어두운 작업복을 입고 있으나 얼굴은 보이지 않는다. 난민은 곱슬머리와 수염, 낡고 더러운 작업복을 가진 성인 남성이다. 미연·난민·관리소장 세 명만 있으며, 카메라 지시에 명시된 주변 노동자는 누락됐다. 낮이지만 직사광과 선명한 그림자가 있어 요청한 절제된 빛보다 밝다.",
        "hard_violations": [],
        "physics": "미연은 접힌 다리와 무릎을 젖은 땅에 대고 체중을 지탱한다. 난민의 다리·골반과 한 손은 바닥에 놓이고, 들린 상체는 미연의 몸과 어깨 뒤로 감긴 팔에 받쳐진다. 다른 손은 어깨가 아니라 가슴 쪽 옷을 쥔다. 관리소장은 출입구를 손으로 잡고 발판에 발을 걸쳐 올라가므로 지지점이 있다. 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "미연의 시선은 아래의 난민 얼굴을 향하고, 난민의 얼굴은 미연 쪽 위로 들려 있다. 눈이 가늘게 감겨 있어 난민의 실제 시선은 A보다 불분명하다. 관리소장은 트럭 출입구를 향해 몸을 돌리고 손을 뻗는다. 난민에게 주의를 돌리는 주변 노동자들은 없다.",
        "built_space": "왼쪽의 누수 방벽 하나, 오른쪽 군용 트럭 한 대와 지게차 한 대, 모래주머니 더미와 교통콘 하나, 먼 등대 하나가 보인다. 젖은 콘크리트와 파손 잔해는 같은 수리 현장으로 읽힌다. 미연은 명확히 왼쪽에 있고 난민은 하단을 대각선으로 가로지르지만, 시점이 낮아 바닥을 내려다보는 정도가 약하고 난민의 다리와 신발이 전경을 크게 점유한다. 관리소장은 오른쪽 뒤의 출입구에 있다. 물웅덩이 반사에서 불가능한 대응은 보이지 않는다.",
        "entities": "미연은 참고와 가까운 중년 동아시아계 여성의 얼굴과 검은 단발, 해진 회색 셔츠 및 어두운 작업복을 보여 준다. 관리소장은 대머리 중년 남성으로 이전 장면의 어두운 작업복을 유지하며, 뒷모습이라 얼굴 동일성은 확인하기 어렵다. 난민은 곱슬머리와 수염을 가진 성인 남성이고 더러운 작업복과 작업화를 착용한다. 등장인물은 세 명뿐이어서 지시된 노동자 군집은 없다. 밝은 햇빛과 뚜렷한 그림자는 차분한 낮 조명 요구보다 강하다.",
        "hard_violations": [],
        "physics": "미연의 무릎과 접힌 정강이가 바닥을 지지하며, 난민의 골반·다리·팔은 지면에 놓인다. 난민의 머리와 어깨는 미연의 몸과 감싼 팔에 받쳐져 있다. 미연의 앞쪽 손은 어깨 대신 가슴 아래쪽 옷을 쥔다. 관리소장은 한 발을 바닥에 두고 손으로 출입구를 잡은 채 다른 발을 들어 승차를 시작한다. 지지 없는 부유나 명백히 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "미연과 난민 남자의 상호작용이 지문에 맞게 잘 표현되었으며, 관리소장이 트럭에 오르는 자세와 물리적 접촉이 매우 자연스럽고 영화적인 연출이 돋보입니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "요구된 구도와 인물 배치는 준수하나, 쓰러진 남자 어깨 부근의 손 해부학적 구조가 엉켜 보이고 관리소장의 탑승 자세가 다소 어색합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 관리소장, 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S7sh25_sel.png",
    "asset_id": "873eda73-f02d-467f-9860-ce44d74ed3c9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 관리소장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1319720>",
    "asset_id": "20aac1ba-1f77-4ab5-b88f-b337ea3b046b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0819-79e6-7e00-8844-775a6664fbf0",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh25"
  },
  "staged_characters_added": [
   "C44"
  ]
 },
 "S7sh30::signage": {
  "fp": "489b3c20cdb2027a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh30": {
  "input_fingerprint": "e9080c04fc6ff916",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 격분한 난민들을 뒤로한 채, 배기구에서 검은 매연을 뿜으며 주행을 시작한 군용 트럭의 뒷모습.\n\nLOCATION (lock): On the access road leaving the artificial-seawall repair site, with the refugee workers behind the departing truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the lowered position behind and to the passenger side of the truck's departure line, pan along its movement while observing its rear three-quarter aspect directly. Keep the departing truck below two-fifths of the image in the upper-right midground, with angry refugees remaining along the near-left edge, bodies angled after it and eyes following it into the distance. Emphasize the vehicle's changing position while holding the camera height and exposure steady; vary the refugees' forward lean, shoulder rotation and spacing without adding new confrontations or a synchronized gesture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Military truck receding along its departure route in the upper-right of the frame, midground, moves toward far end of the departure route.\n- KEY BACKGROUND ELEMENTS: Military truck (Beginning to drive away) — Its rear and passenger-side flank are visible as it recedes; used as Supplies the outward movement opposed to the near-edge refugees; Departure route beside the embankment (Open along the truck's direction of travel) — Runs away from the near crowd toward the upper right; used as Keeps the camera visibly outside the vehicle's path.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained daylight and controlled contrast so the departing vehicle and abandoned crowd remain equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wet ground, damaged barrier, and remaining construction materials from the reference. Exclude the earlier struggle on the ground as a frozen event, and keep construction machinery inactive.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing the unfinished repair site. The seawall remains badly cracked and leaking, and the forklift and construction materials remain behind.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 격분한 난민들을 뒤로한 채, 배기구에서 검은 매연을 뿜으며 주행을 시작한 군용 트럭의 뒷모습.\n\nLOCATION (lock): On the access road leaving the artificial-seawall repair site, with the refugee workers behind the departing truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the lowered position behind and to the passenger side of the truck's departure line, pan along its movement while observing its rear three-quarter aspect directly. Keep the departing truck below two-fifths of the image in the upper-right midground, with angry refugees remaining along the near-left edge, bodies angled after it and eyes following it into the distance. Emphasize the vehicle's changing position while holding the camera height and exposure steady; vary the refugees' forward lean, shoulder rotation and spacing without adding new confrontations or a synchronized gesture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Military truck receding along its departure route in the upper-right of the frame, midground, moves toward far end of the departure route.\n- KEY BACKGROUND ELEMENTS: Military truck (Beginning to drive away) — Its rear and passenger-side flank are visible as it recedes; used as Supplies the outward movement opposed to the near-edge refugees; Departure route beside the embankment (Open along the truck's direction of travel) — Runs away from the near crowd toward the upper right; used as Keeps the camera visibly outside the vehicle's path.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained daylight and controlled contrast so the departing vehicle and abandoned crowd remain equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wet ground, damaged barrier, and remaining construction materials from the reference. Exclude the earlier struggle on the ground as a frozen event, and keep construction machinery inactive.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing the unfinished repair site. The seawall remains badly cracked and leaking, and the forklift and construction materials remain behind.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 격분한 난민들을 뒤로한 채, 배기구에서 검은 매연을 뿜으며 주행을 시작한 군용 트럭의 뒷모습.\n\nLOCATION (lock): On the access road leaving the artificial-seawall repair site, with the refugee workers behind the departing truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the lowered position behind and to the passenger side of the truck's departure line, pan along its movement while observing its rear three-quarter aspect directly. Keep the departing truck below two-fifths of the image in the upper-right midground, with angry refugees remaining along the near-left edge, bodies angled after it and eyes following it into the distance. Emphasize the vehicle's changing position while holding the camera height and exposure steady; vary the refugees' forward lean, shoulder rotation and spacing without adding new confrontations or a synchronized gesture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Military truck receding along its departure route in the upper-right of the frame, midground, moves toward far end of the departure route.\n- KEY BACKGROUND ELEMENTS: Military truck (Beginning to drive away) — Its rear and passenger-side flank are visible as it recedes; used as Supplies the outward movement opposed to the near-edge refugees; Departure route beside the embankment (Open along the truck's direction of travel) — Runs away from the near crowd toward the upper right; used as Keeps the camera visibly outside the vehicle's path.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained daylight and controlled contrast so the departing vehicle and abandoned crowd remain equally readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wet ground, damaged barrier, and remaining construction materials from the reference. Exclude the earlier struggle on the ground as a frozen event, and keep construction machinery inactive.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military truck is departing the unfinished repair site. The seawall remains badly cracked and leaking, and the forklift and construction materials remain behind.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 좌측에 있는 난민들의 시선과 손짓이 우측 상단으로 멀어지는 군용 트럭을 정확히 향하고 있습니다. 트럭은 화면 안쪽을 향해 주행 중입니다.",
    "built_space": "젖은 바닥, 물이 새는 파손된 방파제, 모래주머니, 신호수, 지게차 등 레퍼런스의 배경 요소와 공간 구조가 동일하게 유지되었습니다. 카메라는 트럭의 주행 경로 뒤쪽에서 낮은 각도로 자리 잡고 있습니다.",
    "entities": "다인종으로 보이는 난민들이 낡은 옷을 입고 격분한 표정과 몸짓을 보여줍니다. 캔버스 천을 씌운 군용 트럭이 우측 상단에 있으며 배기구에서 검은 매연을 뿜고 있습니다. 레퍼런스에 있던 기존 인물들은 제외되었습니다.",
    "hard_violations": [],
    "physics": "모든 인물은 지면에 단단히 발을 딛고 역동적인 자세를 취하고 있으며, 트럭의 바퀴 역시 지면에 정상적으로 접지되어 있습니다. 뿜어져 나오는 매연의 퍼짐도 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "난민들의 시선이 우측 상단으로 주행하는 트럭을 향하고 있습니다.",
    "built_space": "방파제와 젖은 바닥의 구조는 유지되었으나, 화면 하단 중앙을 초점이 맞지 않는 인물의 머리와 어깨가 가리고 있어 '좌측 근경에 난민을 배치하라'는 구도 지시를 방해합니다. 또한 배경의 지게차 주변에 레퍼런스에 없던 드럼통 같은 자재들이 추가되었습니다.",
    "entities": "난민들, 매연을 뿜는 군용 트럭이 존재합니다. 그러나 배경의 지게차 형태가 레퍼런스와 확연히 다르며 비정상적인 구조물(드럼통 등)이 추가되었습니다.",
    "hard_violations": [
     "[gemini-pro] 지시된 카메라 프레이밍 구도를 무시하고 화면 하단 전체에 걸쳐 초점 나간 인물들의 뒷모습을 임의로 추가함",
     "[gemini-pro] 레퍼런스에 고정되어야 할 건설 자재(지게차 및 주변 구조물)의 형태를 임의로 변경하고 드럼통을 추가함"
    ],
    "physics": "인물들과 차량은 지면에 지지되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스의 배경 요소를 정확히 유지하면서, 좌측 근경에 분노한 난민들을 배치하고 우측 상단으로 멀어지는 트럭을 포착하는 카메라 구도 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "화면 하단 중앙에 초점이 나간 인물들을 크게 배치하여 좌측 근경에 난민을 배치하라는 프레이밍 지시를 어겼으며, 배경의 지게차와 자재가 레퍼런스와 다르게 변형되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 좌측에 있는 난민들의 시선과 손짓이 우측 상단으로 멀어지는 군용 트럭을 정확히 향하고 있습니다. 트럭은 화면 안쪽을 향해 주행 중입니다.",
        "built_space": "젖은 바닥, 물이 새는 파손된 방파제, 모래주머니, 신호수, 지게차 등 레퍼런스의 배경 요소와 공간 구조가 동일하게 유지되었습니다. 카메라는 트럭의 주행 경로 뒤쪽에서 낮은 각도로 자리 잡고 있습니다.",
        "entities": "다인종으로 보이는 난민들이 낡은 옷을 입고 격분한 표정과 몸짓을 보여줍니다. 캔버스 천을 씌운 군용 트럭이 우측 상단에 있으며 배기구에서 검은 매연을 뿜고 있습니다. 레퍼런스에 있던 기존 인물들은 제외되었습니다.",
        "hard_violations": [],
        "physics": "모든 인물은 지면에 단단히 발을 딛고 역동적인 자세를 취하고 있으며, 트럭의 바퀴 역시 지면에 정상적으로 접지되어 있습니다. 뿜어져 나오는 매연의 퍼짐도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "난민들의 시선이 우측 상단으로 주행하는 트럭을 향하고 있습니다.",
        "built_space": "방파제와 젖은 바닥의 구조는 유지되었으나, 화면 하단 중앙을 초점이 맞지 않는 인물의 머리와 어깨가 가리고 있어 '좌측 근경에 난민을 배치하라'는 구도 지시를 방해합니다. 또한 배경의 지게차 주변에 레퍼런스에 없던 드럼통 같은 자재들이 추가되었습니다.",
        "entities": "난민들, 매연을 뿜는 군용 트럭이 존재합니다. 그러나 배경의 지게차 형태가 레퍼런스와 확연히 다르며 비정상적인 구조물(드럼통 등)이 추가되었습니다.",
        "hard_violations": [
         "지시된 카메라 프레이밍 구도를 무시하고 화면 하단 전체에 걸쳐 초점 나간 인물들의 뒷모습을 임의로 추가함",
         "레퍼런스에 고정되어야 할 건설 자재(지게차 및 주변 구조물)의 형태를 임의로 변경하고 드럼통을 추가함"
        ],
        "physics": "인물들과 차량은 지면에 지지되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스의 배경 요소를 정확히 유지하면서, 좌측 근경에 분노한 난민들을 배치하고 우측 상단으로 멀어지는 트럭을 포착하는 카메라 구도 지시를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "화면 하단 중앙에 초점이 나간 인물들을 크게 배치하여 좌측 근경에 난민을 배치하라는 프레이밍 지시를 어겼으며, 배경의 지게차와 자재가 레퍼런스와 다르게 변형되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 좌측에 있는 난민들의 시선과 손짓이 우측 상단으로 멀어지는 군용 트럭을 정확히 향하고 있습니다. 트럭은 화면 안쪽을 향해 주행 중입니다.",
        "built_space": "젖은 바닥, 물이 새는 파손된 방파제, 모래주머니, 신호수, 지게차 등 레퍼런스의 배경 요소와 공간 구조가 동일하게 유지되었습니다. 카메라는 트럭의 주행 경로 뒤쪽에서 낮은 각도로 자리 잡고 있습니다.",
        "entities": "다인종으로 보이는 난민들이 낡은 옷을 입고 격분한 표정과 몸짓을 보여줍니다. 캔버스 천을 씌운 군용 트럭이 우측 상단에 있으며 배기구에서 검은 매연을 뿜고 있습니다. 레퍼런스에 있던 기존 인물들은 제외되었습니다.",
        "hard_violations": [],
        "physics": "모든 인물은 지면에 단단히 발을 딛고 역동적인 자세를 취하고 있으며, 트럭의 바퀴 역시 지면에 정상적으로 접지되어 있습니다. 뿜어져 나오는 매연의 퍼짐도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "난민들의 시선이 우측 상단으로 주행하는 트럭을 향하고 있습니다.",
        "built_space": "방파제와 젖은 바닥의 구조는 유지되었으나, 화면 하단 중앙을 초점이 맞지 않는 인물의 머리와 어깨가 가리고 있어 '좌측 근경에 난민을 배치하라'는 구도 지시를 방해합니다. 또한 배경의 지게차 주변에 레퍼런스에 없던 드럼통 같은 자재들이 추가되었습니다.",
        "entities": "난민들, 매연을 뿜는 군용 트럭이 존재합니다. 그러나 배경의 지게차 형태가 레퍼런스와 확연히 다르며 비정상적인 구조물(드럼통 등)이 추가되었습니다.",
        "hard_violations": [
         "지시된 카메라 프레이밍 구도를 무시하고 화면 하단 전체에 걸쳐 초점 나간 인물들의 뒷모습을 임의로 추가함",
         "레퍼런스에 고정되어야 할 건설 자재(지게차 및 주변 구조물)의 형태를 임의로 변경하고 드럼통을 추가함"
        ],
        "physics": "인물들과 차량은 지면에 지지되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "트럭을 오른쪽 위 중경에 작게 두어 멀어지는 와이드숏에 더 충실하지만, 보이는 측면과 실제 진행축은 지정된 조수석 쪽·오른쪽 위 출발 방향과 다르다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "왼쪽 난민들의 분노와 검은 배기는 명확하지만, 트럭이 더 크고 가까우며 A와 마찬가지로 조수석 쪽 후방 시점과 출발 방향을 제대로 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "난민들의 얼굴과 상체는 오른쪽 중경의 트럭을 향하고, 든 주먹과 뻗은 손도 떠나는 차량에 대한 항의로 읽힌다. 트럭은 후면과 화면 왼쪽으로 드러난 측면이 보이며, 앞부분이 후미보다 화면 왼쪽 안쪽에 있다. 따라서 차량 위치는 오른쪽 위지만 실제 진행축은 그 위치에서 왼쪽 안쪽으로 향한다. 통상적인 좌핸들 차량 기준으로 보이는 것은 운전석 쪽 측면이며, 요구한 조수석 쪽 후방 시점과 반대다.",
        "built_space": "왼쪽에 크게 갈라져 물이 흐르는 방벽 한 줄, 중앙 후경에 흰 등대 하나, 중앙에 정지한 지게차 한 대, 모래주머니 더미와 드럼통·공사 자재가 있다. 군용 트럭은 한 대이며 화면 오른쪽 위에서 높이 약 3할을 차지한다. 젖은 노면과 잔해, 바다를 접한 공사장 재질은 참고 이미지와 이어진다. 난민들은 왼쪽 가까이에 모였으나 일부 상체가 화면 아래 중앙까지 넓게 들어온다. 물웅덩이의 차량 반사는 위치상 가능하다.",
        "entities": "성인 남녀로 보이는 난민이 최소 여덟 명 보이며, 피부색과 머리 형태가 다양하고 더러운 작업복·두건·모자·배낭을 착용한다. 개별 국적이나 정확한 나이는 판별할 수 없으며 별도 인물 신원 지정도 없다. 방수포 적재함을 갖춘 군용 화물차 한 대, 검은 배기가스, 누수 방벽, 지게차와 남겨진 자재가 모두 보인다. 참고 이미지의 바닥에서 벌어지던 사건은 재현하지 않았다.",
        "hard_violations": [],
        "physics": "트럭의 타이어는 노면에 닿고 차체를 지지한다. 짙은 배기는 차량 후부 하단에서 뒤쪽으로 퍼져 배출 상황으로 읽힌다. 난민들의 앞으로 기운 상체와 팔은 서서 항의할 수 있는 범위이며, 다수가 하체를 프레임 밖에 두었다는 이유만으로 부유한다고 볼 근거는 없다. 배낭은 어깨끈으로 지지되고 자재와 지게차는 바닥에 놓여 있다. 방벽의 물은 아래로 떨어져 노면에 고인다."
       },
       {
        "label": "B",
        "direction": "앞쪽 남성의 얼굴, 가운데 남성의 뻗은 팔, 모자를 쓴 인물의 시선은 모두 오른쪽 트럭을 향한다. 주먹과 펼친 손의 동작이 달라 동기화된 군중 동작은 아니다. 트럭 앞부분은 후미보다 화면 왼쪽 안쪽에 있어 오른쪽 위로 뻗는 출발 방향과 일치하지 않는다. 후면과 함께 드러난 측면 역시 통상적인 좌핸들 차량의 운전석 쪽으로 읽혀 지정한 조수석 쪽 관찰축을 충족하지 않는다.",
        "built_space": "왼쪽 누수 방벽 한 줄, 후경 등대 하나, 모래주머니 더미, 정지한 지게차 한 대와 젖은 공사장 도로가 보인다. 오른쪽에는 공사용 차단물도 있다. 군용 트럭 한 대는 오른쪽에 있으며 높이가 화면의 약 4할에 가까워 A보다 크고 가까이 보인다. 난민들은 왼쪽 가장자리에 더 잘 묶여 있고 중앙 도로는 열려 있다. 노면의 차량과 후미등 반사는 차량 아래쪽에 형성되어 광학적으로 가능하다.",
        "entities": "최소 여섯 명의 성인 난민이 보이고, 전경은 남성 중심이며 다양한 피부색과 머리 형태를 보인다. 낡고 오염된 작업복, 모자, 두건, 배낭은 난민 노동자 설정에 부합한다. 군용 방수포 화물차와 검은 배기, 심하게 손상되어 누수하는 방벽, 지게차, 공사 자재가 식별된다. 이전 장면의 바닥에 누운 인물이나 부축 동작은 없다.",
        "hard_violations": [],
        "physics": "트럭은 타이어로 노면에 지지되고 후부 하단의 검은 연기가 뒤쪽으로 퍼진다. 난민들은 몸을 앞으로 기울이되 지면에 선 자세로 읽히며, 뒤쪽 인물의 신발은 노면과 접촉한다. 뻗은 팔과 쥔 주먹에 명백히 불가능한 관절 자세는 없다. 지게차와 자재는 지면에 있고 방벽 누수도 중력 방향으로 흐른다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "트럭을 오른쪽 위 중경에 작게 두어 멀어지는 와이드숏에 더 충실하지만, 보이는 측면과 실제 진행축은 지정된 조수석 쪽·오른쪽 위 출발 방향과 다르다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "왼쪽 난민들의 분노와 검은 배기는 명확하지만, 트럭이 더 크고 가까우며 A와 마찬가지로 조수석 쪽 후방 시점과 출발 방향을 제대로 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "난민들의 얼굴과 상체는 오른쪽 중경의 트럭을 향하고, 든 주먹과 뻗은 손도 떠나는 차량에 대한 항의로 읽힌다. 트럭은 후면과 화면 왼쪽으로 드러난 측면이 보이며, 앞부분이 후미보다 화면 왼쪽 안쪽에 있다. 따라서 차량 위치는 오른쪽 위지만 실제 진행축은 그 위치에서 왼쪽 안쪽으로 향한다. 통상적인 좌핸들 차량 기준으로 보이는 것은 운전석 쪽 측면이며, 요구한 조수석 쪽 후방 시점과 반대다.",
        "built_space": "왼쪽에 크게 갈라져 물이 흐르는 방벽 한 줄, 중앙 후경에 흰 등대 하나, 중앙에 정지한 지게차 한 대, 모래주머니 더미와 드럼통·공사 자재가 있다. 군용 트럭은 한 대이며 화면 오른쪽 위에서 높이 약 3할을 차지한다. 젖은 노면과 잔해, 바다를 접한 공사장 재질은 참고 이미지와 이어진다. 난민들은 왼쪽 가까이에 모였으나 일부 상체가 화면 아래 중앙까지 넓게 들어온다. 물웅덩이의 차량 반사는 위치상 가능하다.",
        "entities": "성인 남녀로 보이는 난민이 최소 여덟 명 보이며, 피부색과 머리 형태가 다양하고 더러운 작업복·두건·모자·배낭을 착용한다. 개별 국적이나 정확한 나이는 판별할 수 없으며 별도 인물 신원 지정도 없다. 방수포 적재함을 갖춘 군용 화물차 한 대, 검은 배기가스, 누수 방벽, 지게차와 남겨진 자재가 모두 보인다. 참고 이미지의 바닥에서 벌어지던 사건은 재현하지 않았다.",
        "hard_violations": [],
        "physics": "트럭의 타이어는 노면에 닿고 차체를 지지한다. 짙은 배기는 차량 후부 하단에서 뒤쪽으로 퍼져 배출 상황으로 읽힌다. 난민들의 앞으로 기운 상체와 팔은 서서 항의할 수 있는 범위이며, 다수가 하체를 프레임 밖에 두었다는 이유만으로 부유한다고 볼 근거는 없다. 배낭은 어깨끈으로 지지되고 자재와 지게차는 바닥에 놓여 있다. 방벽의 물은 아래로 떨어져 노면에 고인다."
       },
       {
        "label": "A",
        "direction": "앞쪽 남성의 얼굴, 가운데 남성의 뻗은 팔, 모자를 쓴 인물의 시선은 모두 오른쪽 트럭을 향한다. 주먹과 펼친 손의 동작이 달라 동기화된 군중 동작은 아니다. 트럭 앞부분은 후미보다 화면 왼쪽 안쪽에 있어 오른쪽 위로 뻗는 출발 방향과 일치하지 않는다. 후면과 함께 드러난 측면 역시 통상적인 좌핸들 차량의 운전석 쪽으로 읽혀 지정한 조수석 쪽 관찰축을 충족하지 않는다.",
        "built_space": "왼쪽 누수 방벽 한 줄, 후경 등대 하나, 모래주머니 더미, 정지한 지게차 한 대와 젖은 공사장 도로가 보인다. 오른쪽에는 공사용 차단물도 있다. 군용 트럭 한 대는 오른쪽에 있으며 높이가 화면의 약 4할에 가까워 A보다 크고 가까이 보인다. 난민들은 왼쪽 가장자리에 더 잘 묶여 있고 중앙 도로는 열려 있다. 노면의 차량과 후미등 반사는 차량 아래쪽에 형성되어 광학적으로 가능하다.",
        "entities": "최소 여섯 명의 성인 난민이 보이고, 전경은 남성 중심이며 다양한 피부색과 머리 형태를 보인다. 낡고 오염된 작업복, 모자, 두건, 배낭은 난민 노동자 설정에 부합한다. 군용 방수포 화물차와 검은 배기, 심하게 손상되어 누수하는 방벽, 지게차, 공사 자재가 식별된다. 이전 장면의 바닥에 누운 인물이나 부축 동작은 없다.",
        "hard_violations": [],
        "physics": "트럭은 타이어로 노면에 지지되고 후부 하단의 검은 연기가 뒤쪽으로 퍼진다. 난민들은 몸을 앞으로 기울이되 지면에 선 자세로 읽히며, 뒤쪽 인물의 신발은 노면과 접촉한다. 뻗은 팔과 쥔 주먹에 명백히 불가능한 관절 자세는 없다. 지게차와 자재는 지면에 있고 방벽 누수도 중력 방향으로 흐른다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시된 카메라 프레이밍 구도를 무시하고 화면 하단 전체에 걸쳐 초점 나간 인물들의 뒷모습을 임의로 추가함",
     "[gemini-pro] 레퍼런스에 고정되어야 할 건설 자재(지게차 및 주변 구조물)의 형태를 임의로 변경하고 드럼통을 추가함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "레퍼런스의 배경 요소를 정확히 유지하면서, 좌측 근경에 분노한 난민들을 배치하고 우측 상단으로 멀어지는 트럭을 포착하는 카메라 구도 지시를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "화면 하단 중앙에 초점이 나간 인물들을 크게 배치하여 좌측 근경에 난민을 배치하라는 프레이밍 지시를 어겼으며, 배경의 지게차와 자재가 레퍼런스와 다르게 변형되었습니다.  ★위반: [gemini-pro] 지시된 카메라 프레이밍 구도를 무시하고 화면 하단 전체에 걸쳐 초점 나간 인물들의 뒷모습을 임의로 추가함 / [gemini-pro] 레퍼런스에 고정되어야 할 건설 자재(지게차 및 주변 구조물)의 형태를 임의로 변경하고 드럼통을 추가함"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S7sh28_sel.png",
    "asset_id": "5d0f5002-1c6e-49f2-98cb-20d8fc1a5a58",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab081e-7181-7659-a463-ad97fdf728ea",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh28"
  }
 },
 "S8sh12::signage": {
  "fp": "ecfa77cd65fd0809",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::36725db12208d000": {
  "subjects": [],
  "subject_text": "난민 재판소 내부\n허름한 실내에 낡은 탁자와 의자가 놓여 있다. 탁자 위에는 서류가 여러 겹 쌓여 있고 앞쪽 바닥은 비어 있다.",
  "identity": "canonical",
  "scope_id": "L19",
  "scope_role": "location_interior",
  "scope_sha": "c8dc06668e039dce"
 },
 "S8sh12::bgfirst_bg": {
  "input_fingerprint": "4ef5e9b095fdff50",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 바닥에 넘어진 재판관의 위로 올라탄 채 양손으로 멱살을 꽉 움켜쥔 이현우의 피 끓는 표정.\n\nLOCATION (lock): On the floor beside the magistrate's shabby table and fallen chair inside a makeshift refugee tribunal. The room is lit for the daytime hearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the handheld camera beside the fallen judge's shoulder, close to floor level and looking upward at 이현우 from the inherited three-quarter side angle rather than his frontal axis. Observe the action directly, placing 이현우's feverish, pleading face in the upper middle and both collar-gripping hands in the lower middle, with the judge's shoulder and frightened partial profile along the lower edge establishing whom he is addressing. 이현우 pitches down toward the judge while the judge looks up at him, and the close camera distance—not a new lighting effect—makes the desperation legible before the baton strike.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Courtroom floor (The judge has fallen onto it); used as A narrow visible strip anchors the struggle at floor level; Shabby courtroom table (Remaining beside the struggle) — Only a peripheral side portion is visible behind the bodies; used as Maintains courtroom geography without obstructing either gripping hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the courtroom, with controlled contrast preserving the feverish expression and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 바닥에 넘어진 재판관의 위로 올라탄 채 양손으로 멱살을 꽉 움켜쥔 이현우의 피 끓는 표정.\n\nLOCATION (lock): On the floor beside the magistrate's shabby table and fallen chair inside a makeshift refugee tribunal. The room is lit for the daytime hearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the handheld camera beside the fallen judge's shoulder, close to floor level and looking upward at 이현우 from the inherited three-quarter side angle rather than his frontal axis. Observe the action directly, placing 이현우's feverish, pleading face in the upper middle and both collar-gripping hands in the lower middle, with the judge's shoulder and frightened partial profile along the lower edge establishing whom he is addressing. 이현우 pitches down toward the judge while the judge looks up at him, and the close camera distance—not a new lighting effect—makes the desperation legible before the baton strike.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Courtroom floor (The judge has fallen onto it); used as A narrow visible strip anchors the struggle at floor level; Shabby courtroom table (Remaining beside the struggle) — Only a peripheral side portion is visible behind the bodies; used as Maintains courtroom geography without obstructing either gripping hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the courtroom, with controlled contrast preserving the feverish expression and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh12__bgfirst_bg.png",
  "asset_id": "f7b04f1e-1a1c-4410-a6ad-46de366459e9",
  "input_asset_ids": [
   "069597c7-28c2-4114-b511-412526581ad0",
   "752ad9ff-4178-49ce-b7c2-33467e12b0c5"
  ]
 },
 "S8sh12": {
  "input_fingerprint": "f4d5afc30b8bb970",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 넘어진 재판관의 위로 올라탄 채 양손으로 멱살을 꽉 움켜쥔 이현우의 피 끓는 표정.\n\nLOCATION (lock): On the floor beside the magistrate's shabby table and fallen chair inside a makeshift refugee tribunal. The room is lit for the daytime hearing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the handheld camera beside the fallen judge's shoulder, close to floor level and looking upward at 이현우 from the inherited three-quarter side angle rather than his frontal axis. Observe the action directly, placing 이현우's feverish, pleading face in the upper middle and both collar-gripping hands in the lower middle, with the judge's shoulder and frightened partial profile along the lower edge establishing whom he is addressing. 이현우 pitches down toward the judge while the judge looks up at him, and the close camera distance—not a new lighting effect—makes the desperation legible before the baton strike.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Courtroom floor (The judge has fallen onto it); used as A narrow visible strip anchors the struggle at floor level; Shabby courtroom table (Remaining beside the struggle) — Only a peripheral side portion is visible behind the bodies; used as Maintains courtroom geography without obstructing either gripping hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the courtroom, with controlled contrast preserving the feverish expression and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby tribunal table holds a backlog of documents, with the judge's chair beside it. 이현우: Still has the untreated dog-bite injury to his leg and is feverish and physically weak. He is now down at floor level with both hands clenched forward.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 넘어진 재판관의 위로 올라탄 채 양손으로 멱살을 꽉 움켜쥔 이현우의 피 끓는 표정.\n\nLOCATION (lock): On the floor beside the magistrate's shabby table and fallen chair inside a makeshift refugee tribunal. The room is lit for the daytime hearing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the handheld camera beside the fallen judge's shoulder, close to floor level and looking upward at 이현우 from the inherited three-quarter side angle rather than his frontal axis. Observe the action directly, placing 이현우's feverish, pleading face in the upper middle and both collar-gripping hands in the lower middle, with the judge's shoulder and frightened partial profile along the lower edge establishing whom he is addressing. 이현우 pitches down toward the judge while the judge looks up at him, and the close camera distance—not a new lighting effect—makes the desperation legible before the baton strike.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Courtroom floor (The judge has fallen onto it); used as A narrow visible strip anchors the struggle at floor level; Shabby courtroom table (Remaining beside the struggle) — Only a peripheral side portion is visible behind the bodies; used as Maintains courtroom geography without obstructing either gripping hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the courtroom, with controlled contrast preserving the feverish expression and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby tribunal table holds a backlog of documents, with the judge's chair beside it. 이현우: Still has the untreated dog-bite injury to his leg and is feverish and physically weak. He is now down at floor level with both hands clenched forward.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 넘어진 재판관의 위로 올라탄 채 양손으로 멱살을 꽉 움켜쥔 이현우의 피 끓는 표정.\n\nLOCATION (lock): On the floor beside the magistrate's shabby table and fallen chair inside a makeshift refugee tribunal. The room is lit for the daytime hearing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the handheld camera beside the fallen judge's shoulder, close to floor level and looking upward at 이현우 from the inherited three-quarter side angle rather than his frontal axis. Observe the action directly, placing 이현우's feverish, pleading face in the upper middle and both collar-gripping hands in the lower middle, with the judge's shoulder and frightened partial profile along the lower edge establishing whom he is addressing. 이현우 pitches down toward the judge while the judge looks up at him, and the close camera distance—not a new lighting effect—makes the desperation legible before the baton strike.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Courtroom floor (The judge has fallen onto it); used as A narrow visible strip anchors the struggle at floor level; Shabby courtroom table (Remaining beside the struggle) — Only a peripheral side portion is visible behind the bodies; used as Maintains courtroom geography without obstructing either gripping hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the courtroom, with controlled contrast preserving the feverish expression and gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shabby tribunal table holds a backlog of documents, with the judge's chair beside it. 이현우: Still has the untreated dog-bite injury to his leg and is feverish and physically weak. He is now down at floor level with both hands clenched forward.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh12__bgfirst_bg.png",
     "asset_id": "f7b04f1e-1a1c-4410-a6ad-46de366459e9",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S8sh12.png",
     "asset_id": "069597c7-28c2-4114-b511-412526581ad0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L19B01.png",
     "asset_id": "752ad9ff-4178-49ce-b7c2-33467e12b0c5",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 바닥에 누운 재판관을 향해 시선을 고정하고, 재판관은 그를 올려다봄.",
    "built_space": "법정 바닥, 뒷배경의 책상과 태극기, 우측 문이 참조 이미지와 동일하게 배치됨.",
    "entities": "이현우의 외모, 의상, 무전기가 일치하며 양손으로 재판관의 멱살을 정확히 쥠.",
    "hard_violations": [],
    "physics": "이현우는 무릎을 꿇은 채 양손으로 체중을 지탱하며, 재판관은 바닥에 완전히 밀착됨."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 재판관의 얼굴을 향하고 재판관도 위를 응시함.",
    "built_space": "바닥과 배경의 책상 일부, 태극기가 공간 설정에 맞게 등장함.",
    "entities": "이현우의 인상착의는 일치하나, 멱살을 쥐어야 할 왼손의 형태가 상실됨.",
    "hard_violations": [
     "[gemini-pro] 재판관의 안경테가 감긴 눈꺼풀 안쪽으로 파고들어간 해부학적 오류",
     "[gemini-pro] 이현우의 왼손이 뭉개져 재판관의 옷감과 융합된 신체 구조적 오류"
    ],
    "physics": "오른손은 옷깃을 쥐고 지탱하나, 왼손은 형태가 없어 지지점 역할이 불가능함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 클로즈업 대신 샷을 넓혀 프레이밍 우선순위를 어겼으나, 치명적인 신체 왜곡이나 물리적 오류가 없어 승리함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 프레이밍 척도와 앵글을 정확히 구현했으나, 안경과 손의 형태가 무너진 치명적 오류로 인해 실격됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 바닥에 누운 재판관을 향해 시선을 고정하고, 재판관은 그를 올려다봄.",
        "built_space": "법정 바닥, 뒷배경의 책상과 태극기, 우측 문이 참조 이미지와 동일하게 배치됨.",
        "entities": "이현우의 외모, 의상, 무전기가 일치하며 양손으로 재판관의 멱살을 정확히 쥠.",
        "hard_violations": [],
        "physics": "이현우는 무릎을 꿇은 채 양손으로 체중을 지탱하며, 재판관은 바닥에 완전히 밀착됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 재판관의 얼굴을 향하고 재판관도 위를 응시함.",
        "built_space": "바닥과 배경의 책상 일부, 태극기가 공간 설정에 맞게 등장함.",
        "entities": "이현우의 인상착의는 일치하나, 멱살을 쥐어야 할 왼손의 형태가 상실됨.",
        "hard_violations": [
         "재판관의 안경테가 감긴 눈꺼풀 안쪽으로 파고들어간 해부학적 오류",
         "이현우의 왼손이 뭉개져 재판관의 옷감과 융합된 신체 구조적 오류"
        ],
        "physics": "오른손은 옷깃을 쥐고 지탱하나, 왼손은 형태가 없어 지지점 역할이 불가능함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 클로즈업 대신 샷을 넓혀 프레이밍 우선순위를 어겼으나, 치명적인 신체 왜곡이나 물리적 오류가 없어 승리함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 프레이밍 척도와 앵글을 정확히 구현했으나, 안경과 손의 형태가 무너진 치명적 오류로 인해 실격됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 바닥에 누운 재판관을 향해 시선을 고정하고, 재판관은 그를 올려다봄.",
        "built_space": "법정 바닥, 뒷배경의 책상과 태극기, 우측 문이 참조 이미지와 동일하게 배치됨.",
        "entities": "이현우의 외모, 의상, 무전기가 일치하며 양손으로 재판관의 멱살을 정확히 쥠.",
        "hard_violations": [],
        "physics": "이현우는 무릎을 꿇은 채 양손으로 체중을 지탱하며, 재판관은 바닥에 완전히 밀착됨."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 재판관의 얼굴을 향하고 재판관도 위를 응시함.",
        "built_space": "바닥과 배경의 책상 일부, 태극기가 공간 설정에 맞게 등장함.",
        "entities": "이현우의 인상착의는 일치하나, 멱살을 쥐어야 할 왼손의 형태가 상실됨.",
        "hard_violations": [
         "재판관의 안경테가 감긴 눈꺼풀 안쪽으로 파고들어간 해부학적 오류",
         "이현우의 왼손이 뭉개져 재판관의 옷감과 융합된 신체 구조적 오류"
        ],
        "physics": "오른손은 옷깃을 쥐고 지탱하나, 왼손은 형태가 없어 지지점 역할이 불가능함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴과 멱살을 쥔 두 손을 크게 잡은 바닥 높이 클로즈업이 우세하지만, 정면에 가까운 각도와 세워진 재판관 의자는 지시와 다르다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "올라타서 양손으로 멱살을 잡는 동작은 명확하지만, 무릎과 넓은 바닥·책상까지 보여 주는 확장된 구도가 지정된 클로즈업을 크게 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽 아래의 재판관 얼굴을 내려다보고, 재판관은 고개를 들어 이현우 쪽을 바라본다. 두 손은 재판관 목 아래의 옷깃을 향해 뻗어 실제로 움켜쥐고 있다. 시선의 상대는 맞지만 이현우 얼굴은 지정된 사선 측면보다 정면에 가깝게 보인다.",
        "built_space": "왼쪽 창과 난방기 각 하나, 태극기 하나, 서류가 놓인 낡은 재판관 책상 하나, 그 뒤 검은 의자 하나, 오른쪽 문 하나가 보인다. 회백색 이색 벽과 닳은 바닥은 장소 참고와 부합한다. 책상은 인물 뒤 왼쪽 가장자리로 제한되지만 바닥은 좁은 띠보다 조금 넓다. 검은 재판관 의자는 여전히 세워져 있어 넘어진 의자라는 현장 상태를 충족하지 않는다. 반사는 없다.",
        "entities": "등장인물은 젊은 동아시아계 남성 이현우와 나이 든 남성 재판관 두 명이다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 흙먼지와 핏자국이 묻은 어두운 셔츠, 귀의 작은 검은 인이어가 요구와 부합한다. 얼굴은 참고 인물과 대체로 유사하며 땀과 찡그린 표정으로 열기와 절박함을 표현한다. 재판관은 안경과 검은 겉옷을 착용한 실물 인물로 보인다. 다리 상처는 구도 밖이므로 평가하지 않는다. 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "재판관의 몸은 화면 아래 바닥에 누워 지지되고, 이현우는 왼쪽 아래에 보이는 굽힌 다리와 무릎으로 바닥을 짚으며 그 위로 몸을 기울인다. 한 손은 전경에서 옷깃을 잡고 다른 손은 그 뒤에서 반대쪽 옷깃을 잡아 천을 당긴다. 뒤쪽 손 일부는 가려져 있지만 접촉은 보인다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 아래 재판관을 내려다보며 상체를 숙이고, 재판관은 그를 올려다본다. 이현우의 두 주먹은 재판관의 양쪽 옷깃을 잡고 있다. 상대를 향한 시선과 손의 작용 방향은 맞는다.",
        "built_space": "왼쪽 창과 난방기, 태극기, 중앙 뒤의 서류 쌓인 책상과 검은 의자, 벽의 금색 표장, 오른쪽 문이 각각 보인다. 왼쪽 전경에는 별도의 목제 가구 일부와 바닥 쪽으로 뻗은 부재도 보인다. 장소의 기본 재료와 배치는 참고와 유사하지만, 책상 정면과 넓은 바닥이 드러나 주변부만 보여 달라는 지시를 벗어난다. 검은 재판관 의자는 여전히 바로 서 있으며, 왼쪽 목제 부재를 그 의자가 넘어진 모습으로 볼 수는 없다. 반사는 없다.",
        "entities": "이현우와 나이 든 남성 재판관 두 명만 등장한다. 이현우의 검은 머리, 젊은 동아시아계 외모, 마른 체격, 오염된 어두운 셔츠와 바지, 작은 인이어는 요구에 대체로 맞는다. 바지의 찢어진 상처 부위도 보이지만 개에게 물린 상처인지는 영상만으로 확정할 수 없다. 재판관의 검은 옷과 흰 셔츠, 방어적으로 든 손이 보인다. 이현우의 표정은 격앙되어 있으나 애원보다는 분노가 강하다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "이현우의 굽힌 무릎과 정강이가 바닥에 닿아 체중을 지탱하며 재판관 몸 위에 올라탄 자세를 만든다. 재판관은 등과 하체를 바닥에 댄 채 머리와 어깨를 조금 들고 있다. 두 손이 각각 옷깃을 쥐어 천을 팽팽하게 당기며, 재판관의 한 손도 가슴 부근에서 저항한다. 보이는 신체와 가구에 지지 없는 부유나 불가능한 관절 구조는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴과 멱살을 쥔 두 손을 크게 잡은 바닥 높이 클로즈업이 우세하지만, 정면에 가까운 각도와 세워진 재판관 의자는 지시와 다르다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "올라타서 양손으로 멱살을 잡는 동작은 명확하지만, 무릎과 넓은 바닥·책상까지 보여 주는 확장된 구도가 지정된 클로즈업을 크게 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽 아래의 재판관 얼굴을 내려다보고, 재판관은 고개를 들어 이현우 쪽을 바라본다. 두 손은 재판관 목 아래의 옷깃을 향해 뻗어 실제로 움켜쥐고 있다. 시선의 상대는 맞지만 이현우 얼굴은 지정된 사선 측면보다 정면에 가깝게 보인다.",
        "built_space": "왼쪽 창과 난방기 각 하나, 태극기 하나, 서류가 놓인 낡은 재판관 책상 하나, 그 뒤 검은 의자 하나, 오른쪽 문 하나가 보인다. 회백색 이색 벽과 닳은 바닥은 장소 참고와 부합한다. 책상은 인물 뒤 왼쪽 가장자리로 제한되지만 바닥은 좁은 띠보다 조금 넓다. 검은 재판관 의자는 여전히 세워져 있어 넘어진 의자라는 현장 상태를 충족하지 않는다. 반사는 없다.",
        "entities": "등장인물은 젊은 동아시아계 남성 이현우와 나이 든 남성 재판관 두 명이다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 흙먼지와 핏자국이 묻은 어두운 셔츠, 귀의 작은 검은 인이어가 요구와 부합한다. 얼굴은 참고 인물과 대체로 유사하며 땀과 찡그린 표정으로 열기와 절박함을 표현한다. 재판관은 안경과 검은 겉옷을 착용한 실물 인물로 보인다. 다리 상처는 구도 밖이므로 평가하지 않는다. 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "재판관의 몸은 화면 아래 바닥에 누워 지지되고, 이현우는 왼쪽 아래에 보이는 굽힌 다리와 무릎으로 바닥을 짚으며 그 위로 몸을 기울인다. 한 손은 전경에서 옷깃을 잡고 다른 손은 그 뒤에서 반대쪽 옷깃을 잡아 천을 당긴다. 뒤쪽 손 일부는 가려져 있지만 접촉은 보인다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 아래 재판관을 내려다보며 상체를 숙이고, 재판관은 그를 올려다본다. 이현우의 두 주먹은 재판관의 양쪽 옷깃을 잡고 있다. 상대를 향한 시선과 손의 작용 방향은 맞는다.",
        "built_space": "왼쪽 창과 난방기, 태극기, 중앙 뒤의 서류 쌓인 책상과 검은 의자, 벽의 금색 표장, 오른쪽 문이 각각 보인다. 왼쪽 전경에는 별도의 목제 가구 일부와 바닥 쪽으로 뻗은 부재도 보인다. 장소의 기본 재료와 배치는 참고와 유사하지만, 책상 정면과 넓은 바닥이 드러나 주변부만 보여 달라는 지시를 벗어난다. 검은 재판관 의자는 여전히 바로 서 있으며, 왼쪽 목제 부재를 그 의자가 넘어진 모습으로 볼 수는 없다. 반사는 없다.",
        "entities": "이현우와 나이 든 남성 재판관 두 명만 등장한다. 이현우의 검은 머리, 젊은 동아시아계 외모, 마른 체격, 오염된 어두운 셔츠와 바지, 작은 인이어는 요구에 대체로 맞는다. 바지의 찢어진 상처 부위도 보이지만 개에게 물린 상처인지는 영상만으로 확정할 수 없다. 재판관의 검은 옷과 흰 셔츠, 방어적으로 든 손이 보인다. 이현우의 표정은 격앙되어 있으나 애원보다는 분노가 강하다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "이현우의 굽힌 무릎과 정강이가 바닥에 닿아 체중을 지탱하며 재판관 몸 위에 올라탄 자세를 만든다. 재판관은 등과 하체를 바닥에 댄 채 머리와 어깨를 조금 들고 있다. 두 손이 각각 옷깃을 쥐어 천을 팽팽하게 당기며, 재판관의 한 손도 가슴 부근에서 저항한다. 보이는 신체와 가구에 지지 없는 부유나 불가능한 관절 구조는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.625,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.625,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gemini-pro] 재판관의 안경테가 감긴 눈꺼풀 안쪽으로 파고들어간 해부학적 오류",
     "[gemini-pro] 이현우의 왼손이 뭉개져 재판관의 옷감과 융합된 신체 구조적 오류"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1625,
   "B": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1625,
    "verdict_ko": "지시된 클로즈업 대신 샷을 넓혀 프레이밍 우선순위를 어겼으나, 치명적인 신체 왜곡이나 물리적 오류가 없어 승리함."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "지시된 프레이밍 척도와 앵글을 정확히 구현했으나, 안경과 손의 형태가 무너진 치명적 오류로 인해 실격됨.  ★위반: [gemini-pro] 재판관의 안경테가 감긴 눈꺼풀 안쪽으로 파고들어간 해부학적 오류 / [gemini-pro] 이현우의 왼손이 뭉개져 재판관의 옷감과 융합된 신체 구조적 오류"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L19B01.png",
    "asset_id": "752ad9ff-4178-49ce-b7c2-33467e12b0c5",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0823-6d11-718c-a035-fc25889d2d53",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh12__bgfirst_bg.png",
   "bg_asset_id": "f7b04f1e-1a1c-4410-a6ad-46de366459e9",
   "bg_record_key": "S8sh12::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh18::signage": {
  "fp": "59ea170d6aad61e9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh18::bgfirst_bg": {
  "input_fingerprint": "bc98b684f8d258c8",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흑백 화면 속, 무릎 꿇은 어린 이현우와 앰버의 머리를 양팔로 감싸 안은 채 공포에 질린 젊은 미연의 얼굴.\n\nLOCATION (lock): On a city street at night during a violent roundup, rendered as a monochrome memory.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 이현우's explicitly black-and-white dream of the past, continue the gentle push from kneeling face height, remaining oblique to 미연 and just outside the family–attacker axis rather than adopting the attacker's eyes. Place the younger 미연's frightened face in the upper center and the younger 이현우 and 앰버 tucked beneath her arms at the lower left and right, preserving both protective forearms and the children's heads within the close frame. 미연 looks upward toward the attacker outside the frame while the children lower their faces into her shelter; emphasize the closing camera distance and retain the dream's discontinuous frame cadence without adding visual hallucinations.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: American city street (The family is kneeling during the attack); used as A limited, unfocused strip grounds the enclosed family within the remembered street.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime dream in black and white with restrained brightness and controlled facial contrast, without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흑백 화면 속, 무릎 꿇은 어린 이현우와 앰버의 머리를 양팔로 감싸 안은 채 공포에 질린 젊은 미연의 얼굴.\n\nLOCATION (lock): On a city street at night during a violent roundup, rendered as a monochrome memory.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 이현우's explicitly black-and-white dream of the past, continue the gentle push from kneeling face height, remaining oblique to 미연 and just outside the family–attacker axis rather than adopting the attacker's eyes. Place the younger 미연's frightened face in the upper center and the younger 이현우 and 앰버 tucked beneath her arms at the lower left and right, preserving both protective forearms and the children's heads within the close frame. 미연 looks upward toward the attacker outside the frame while the children lower their faces into her shelter; emphasize the closing camera distance and retain the dream's discontinuous frame cadence without adding visual hallucinations.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: American city street (The family is kneeling during the attack); used as A limited, unfocused strip grounds the enclosed family within the remembered street.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime dream in black and white with restrained brightness and controlled facial contrast, without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh18__bgfirst_bg.png",
  "asset_id": "2d730d67-7714-4d44-87a6-02cd7c574a48",
  "input_asset_ids": [
   "a46259a7-1888-4810-a974-90bac21a6e29",
   "17f89a7f-3d0b-4d77-89c2-e3191cabdeb7"
  ]
 },
 "S8sh18": {
  "input_fingerprint": "571ca9d6e0930ef2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 무릎 꿇은 어린 이현우와 앰버의 머리를 양팔로 감싸 안은 채 공포에 질린 젊은 미연의 얼굴.\n\nLOCATION (lock): On a city street at night during a violent roundup, rendered as a monochrome memory. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 이현우's explicitly black-and-white dream of the past, continue the gentle push from kneeling face height, remaining oblique to 미연 and just outside the family–attacker axis rather than adopting the attacker's eyes. Place the younger 미연's frightened face in the upper center and the younger 이현우 and 앰버 tucked beneath her arms at the lower left and right, preserving both protective forearms and the children's heads within the close frame. 미연 looks upward toward the attacker outside the frame while the children lower their faces into her shelter; emphasize the closing camera distance and retain the dream's discontinuous frame cadence without adding visual hallucinations.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: American city street (The family is kneeling during the attack); used as A limited, unfocused strip grounds the enclosed family within the remembered street.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime dream in black and white with restrained brightness and controlled facial contrast, without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This is a black-and-white nighttime memory in an American city street, with intermittent gunfire flashes and a large American flag among the attackers. 미연: Appears younger, kneeling in terror with both arms held protectively around her sides. 이현우: Appears as a young child, kneeling in fear. 앰버: Appears as a younger child, kneeling in fear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 무릎 꿇은 어린 이현우와 앰버의 머리를 양팔로 감싸 안은 채 공포에 질린 젊은 미연의 얼굴.\n\nLOCATION (lock): On a city street at night during a violent roundup, rendered as a monochrome memory. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 이현우's explicitly black-and-white dream of the past, continue the gentle push from kneeling face height, remaining oblique to 미연 and just outside the family–attacker axis rather than adopting the attacker's eyes. Place the younger 미연's frightened face in the upper center and the younger 이현우 and 앰버 tucked beneath her arms at the lower left and right, preserving both protective forearms and the children's heads within the close frame. 미연 looks upward toward the attacker outside the frame while the children lower their faces into her shelter; emphasize the closing camera distance and retain the dream's discontinuous frame cadence without adding visual hallucinations.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: American city street (The family is kneeling during the attack); used as A limited, unfocused strip grounds the enclosed family within the remembered street.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime dream in black and white with restrained brightness and controlled facial contrast, without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This is a black-and-white nighttime memory in an American city street, with intermittent gunfire flashes and a large American flag among the attackers. 미연: Appears younger, kneeling in terror with both arms held protectively around her sides. 이현우: Appears as a young child, kneeling in fear. 앰버: Appears as a younger child, kneeling in fear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 무릎 꿇은 어린 이현우와 앰버의 머리를 양팔로 감싸 안은 채 공포에 질린 젊은 미연의 얼굴.\n\nLOCATION (lock): On a city street at night during a violent roundup, rendered as a monochrome memory. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 이현우's explicitly black-and-white dream of the past, continue the gentle push from kneeling face height, remaining oblique to 미연 and just outside the family–attacker axis rather than adopting the attacker's eyes. Place the younger 미연's frightened face in the upper center and the younger 이현우 and 앰버 tucked beneath her arms at the lower left and right, preserving both protective forearms and the children's heads within the close frame. 미연 looks upward toward the attacker outside the frame while the children lower their faces into her shelter; emphasize the closing camera distance and retain the dream's discontinuous frame cadence without adding visual hallucinations.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: American city street (The family is kneeling during the attack); used as A limited, unfocused strip grounds the enclosed family within the remembered street.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime dream in black and white with restrained brightness and controlled facial contrast, without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This is a black-and-white nighttime memory in an American city street, with intermittent gunfire flashes and a large American flag among the attackers. 미연: Appears younger, kneeling in terror with both arms held protectively around her sides. 이현우: Appears as a young child, kneeling in fear. 앰버: Appears as a younger child, kneeling in fear.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh18__bgfirst_bg.png",
     "asset_id": "2d730d67-7714-4d44-87a6-02cd7c574a48",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S8sh18.png",
     "asset_id": "a46259a7-1888-4810-a974-90bac21a6e29",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L19B02.png",
     "asset_id": "17f89a7f-3d0b-4d77-89c2-e3191cabdeb7",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연은 화면 밖 위쪽을 응시하고, 두 아이는 그녀의 양팔 아래로 얼굴을 깊게 파묻고 있다.",
    "built_space": "레퍼런스 이미지의 건물 외부 계단, 조명, 철조망, 소화전이 정확한 위치에 배치된 도로 한가운데 인물들이 있다.",
    "entities": "미연은 레퍼런스 얼굴을 유지한 채 젊은 모습이며, 현우와 앰버도 어린 모습으로 묘사되었고 앰버의 작업복과 마스크도 잘 표현되었다.",
    "hard_violations": [],
    "physics": "세 명 모두 아스팔트 바닥에 무릎을 꿇고 체중을 싣고 있으며, 미연이 양팔로 아이들을 안고 있는 자세가 물리적으로 자연스럽다."
   },
   {
    "label": "B",
    "direction": "미연은 위쪽을 보지만, 두 아이는 얼굴을 숙이지 않고 앞과 옆을 멍하니 바라보고 있다.",
    "built_space": "레퍼런스와 전혀 일치하지 않는 낯선 벽돌 건물들이 배경으로 그려진 거리이다.",
    "entities": "미연의 얼굴이 레퍼런스와 크게 다르며, 앰버는 마스크를 목에 걸지 않고 얼굴 전체에 착용하고 있다.",
    "hard_violations": [
     "[gemini-pro] 지정된 로케이션 레퍼런스를 완전히 무시하고 구조와 재질이 다른 엉뚱한 배경 건물을 렌더링함."
    ],
    "physics": "바닥에 무릎을 꿇고 있으나 아이들과 미연이 엉켜있는 팔과 손의 해부학적 연결과 지지 상태가 모호하다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로케이션을 완벽히 재현하고 아이들이 얼굴을 숙인 자세를 잘 표현했으나, 요구된 클로즈업 숏보다 프레이밍이 넓게 잡혔습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 장소를 완전히 무시하고 다른 배경을 생성했으며, 아이들이 얼굴을 파묻는 동작과 캐릭터의 외모도 제대로 반영하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 화면 밖 위쪽을 응시하고, 두 아이는 그녀의 양팔 아래로 얼굴을 깊게 파묻고 있다.",
        "built_space": "레퍼런스 이미지의 건물 외부 계단, 조명, 철조망, 소화전이 정확한 위치에 배치된 도로 한가운데 인물들이 있다.",
        "entities": "미연은 레퍼런스 얼굴을 유지한 채 젊은 모습이며, 현우와 앰버도 어린 모습으로 묘사되었고 앰버의 작업복과 마스크도 잘 표현되었다.",
        "hard_violations": [],
        "physics": "세 명 모두 아스팔트 바닥에 무릎을 꿇고 체중을 싣고 있으며, 미연이 양팔로 아이들을 안고 있는 자세가 물리적으로 자연스럽다."
       },
       {
        "label": "B",
        "direction": "미연은 위쪽을 보지만, 두 아이는 얼굴을 숙이지 않고 앞과 옆을 멍하니 바라보고 있다.",
        "built_space": "레퍼런스와 전혀 일치하지 않는 낯선 벽돌 건물들이 배경으로 그려진 거리이다.",
        "entities": "미연의 얼굴이 레퍼런스와 크게 다르며, 앰버는 마스크를 목에 걸지 않고 얼굴 전체에 착용하고 있다.",
        "hard_violations": [
         "지정된 로케이션 레퍼런스를 완전히 무시하고 구조와 재질이 다른 엉뚱한 배경 건물을 렌더링함."
        ],
        "physics": "바닥에 무릎을 꿇고 있으나 아이들과 미연이 엉켜있는 팔과 손의 해부학적 연결과 지지 상태가 모호하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로케이션을 완벽히 재현하고 아이들이 얼굴을 숙인 자세를 잘 표현했으나, 요구된 클로즈업 숏보다 프레이밍이 넓게 잡혔습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 장소를 완전히 무시하고 다른 배경을 생성했으며, 아이들이 얼굴을 파묻는 동작과 캐릭터의 외모도 제대로 반영하지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연은 화면 밖 위쪽을 응시하고, 두 아이는 그녀의 양팔 아래로 얼굴을 깊게 파묻고 있다.",
        "built_space": "레퍼런스 이미지의 건물 외부 계단, 조명, 철조망, 소화전이 정확한 위치에 배치된 도로 한가운데 인물들이 있다.",
        "entities": "미연은 레퍼런스 얼굴을 유지한 채 젊은 모습이며, 현우와 앰버도 어린 모습으로 묘사되었고 앰버의 작업복과 마스크도 잘 표현되었다.",
        "hard_violations": [],
        "physics": "세 명 모두 아스팔트 바닥에 무릎을 꿇고 체중을 싣고 있으며, 미연이 양팔로 아이들을 안고 있는 자세가 물리적으로 자연스럽다."
       },
       {
        "label": "B",
        "direction": "미연은 위쪽을 보지만, 두 아이는 얼굴을 숙이지 않고 앞과 옆을 멍하니 바라보고 있다.",
        "built_space": "레퍼런스와 전혀 일치하지 않는 낯선 벽돌 건물들이 배경으로 그려진 거리이다.",
        "entities": "미연의 얼굴이 레퍼런스와 크게 다르며, 앰버는 마스크를 목에 걸지 않고 얼굴 전체에 착용하고 있다.",
        "hard_violations": [
         "지정된 로케이션 레퍼런스를 완전히 무시하고 구조와 재질이 다른 엉뚱한 배경 건물을 렌더링함."
        ],
        "physics": "바닥에 무릎을 꿇고 있으나 아이들과 미연이 엉켜있는 팔과 손의 해부학적 연결과 지지 상태가 모호하다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 대신 배경이 넓은 중경이며, 아이들이 얼굴을 숨기지 않고 미연도 머리 대신 어깨를 감싸며, 벽돌 거리와 주간 조명이 지정된 밤의 장소와 다릅니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미연의 화면 밖 위쪽 시선, 양손으로 감싼 아이들의 숙인 머리, 흑백 야간 장소는 충실하지만, 무릎과 신발까지 담은 넓은 구도는 필수 클로즈업을 놓칩니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 화면 왼쪽 위를 약간 올려다보지만 시선이 거의 정면에 가깝고, 화면 밖 높은 공격자를 향한다는 관계가 약합니다. 현우는 카메라 가까운 앞쪽을, 앰버는 화면 왼쪽 앞을 봅니다. 둘 다 미연의 품으로 얼굴을 낮추지 않습니다. 무기나 방향을 판정할 휴대 물체는 없습니다.",
        "built_space": "벽돌 건물의 문과 창문들, 오른쪽 상단의 철창 창문 하나, 왼쪽 차양 하나, 오른쪽 가장자리의 소화전 일부가 보입니다. 가족은 차도에 모여 있으나, 참조의 콘크리트 공장 외벽과 철망 울타리가 확인되지 않아 같은 장소로 읽히지 않습니다. 도로와 건물이 화면 대부분에 선명하게 남아 제한된 흐릿한 배경 띠라는 요구도 어깁니다.",
        "entities": "세 사람만 보입니다. 미연은 검은 머리의 젊은 동아시아계 여성으로 보이나 참조의 단발과 달리 머리를 뒤로 묶었습니다. 현우는 짧은 검은 머리의 어린 동아시아계 남자아이이며 어두운 셔츠를 입었습니다. 앰버는 밝은 머리와 창백한 피부의 여자아이로, 방진 마스크를 얼굴에 착용하고 작업복과 공구 벨트를 갖췄습니다. 혼혈 정체성이나 현우의 인이어는 이 화면에서 확정하기 어렵습니다. 흑백이지만 밝은 주간 거리로 보여 반복 지정된 야간 기억과 다릅니다. 성조기와 공격자는 프레임 밖이므로 누락으로 보지 않습니다.",
        "hard_violations": [],
        "physics": "미연은 한쪽 무릎을 올리고 다른 쪽을 낮춘 자세이며, 아이들의 굽힌 하체는 도로 쪽으로 이어집니다. 하단 접지 일부는 잘렸으나 부유하는 모습은 아닙니다. 미연의 두 손은 아이들의 위팔에 닿고 현우의 팔도 앰버 쪽에 닿아 포옹 자체는 가능합니다. 다만 보호하는 접점이 머리가 아니라 어깨와 팔입니다. 마스크는 얼굴의 끈으로, 공구 주머니는 허리 벨트로 지지됩니다."
       },
       {
        "label": "B",
        "direction": "미연의 눈과 얼굴은 화면 오른쪽 위의 보이지 않는 대상을 향해 있어 높은 위치의 화면 밖 공격자를 바라보는 지시와 맞습니다. 현우와 앰버는 모두 고개와 시선을 아래로 내리고 미연의 몸 안쪽으로 얼굴을 숨깁니다. 미연의 양손은 각각 두 아이의 머리를 감쌉니다. 보이는 무기는 없습니다.",
        "built_space": "뒤쪽에 콘크리트 공장 건물 하나, 오른쪽 철망 울타리 구간과 적재물, 소화전 하나, 오른쪽 하단 배수구 하나가 보입니다. 왼쪽 가까운 전신주와 점등된 곡선형 가로등 하나도 참조의 배치와 부합합니다. 외부 계단은 가족 뒤에 가려져 확인하기 어렵습니다. 가족은 차도에 무릎 꿇고 있어 공간 관계가 자연스럽지만, 건물과 도로를 넓게 드러내어 배경을 작은 흐릿한 띠로 제한하지 못했습니다.",
        "entities": "지정된 세 사람만 등장합니다. 미연은 참조와 유사한 검은 단발의 젊은 동아시아계 여성이고, 해진 회색 계열 셔츠를 입었습니다. 현우는 헝클어진 검은 머리의 어린 남자아이로 어두운 작업성 셔츠를 입고 있습니다. 앰버는 밝은 머리의 어린 여자아이이며 작업복과 공구 벨트를 착용합니다. 방진 마스크는 참조처럼 목 아래에 걸려 있지만 얼굴을 덮지는 않습니다. 숙인 얼굴 때문에 아이들의 정확한 얼굴 동일성 및 혼혈 특징은 제한적으로만 확인됩니다. 흑백 야간 분위기와 참조에 있는 가로등은 맞으며, 프레임 밖 성조기와 공격자는 평가상 누락이 아닙니다.",
        "hard_violations": [],
        "physics": "미연의 양 무릎이 아스팔트에 닿고, 아이들은 접힌 무릎과 정강이로 도로에 몸을 지탱합니다. 아이들의 머리는 미연의 가슴과 팔에 기대며, 미연의 손바닥이 각 머리 위와 뒤에 실제로 접촉합니다. 앰버의 마스크는 목의 끈과 가슴으로, 공구 주머니는 벨트로 지지됩니다. 젖은 노면의 밝은 반사는 주변 조명과 양립하며, 지지 없는 신체나 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업 대신 배경이 넓은 중경이며, 아이들이 얼굴을 숨기지 않고 미연도 머리 대신 어깨를 감싸며, 벽돌 거리와 주간 조명이 지정된 밤의 장소와 다릅니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미연의 화면 밖 위쪽 시선, 양손으로 감싼 아이들의 숙인 머리, 흑백 야간 장소는 충실하지만, 무릎과 신발까지 담은 넓은 구도는 필수 클로즈업을 놓칩니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 화면 왼쪽 위를 약간 올려다보지만 시선이 거의 정면에 가깝고, 화면 밖 높은 공격자를 향한다는 관계가 약합니다. 현우는 카메라 가까운 앞쪽을, 앰버는 화면 왼쪽 앞을 봅니다. 둘 다 미연의 품으로 얼굴을 낮추지 않습니다. 무기나 방향을 판정할 휴대 물체는 없습니다.",
        "built_space": "벽돌 건물의 문과 창문들, 오른쪽 상단의 철창 창문 하나, 왼쪽 차양 하나, 오른쪽 가장자리의 소화전 일부가 보입니다. 가족은 차도에 모여 있으나, 참조의 콘크리트 공장 외벽과 철망 울타리가 확인되지 않아 같은 장소로 읽히지 않습니다. 도로와 건물이 화면 대부분에 선명하게 남아 제한된 흐릿한 배경 띠라는 요구도 어깁니다.",
        "entities": "세 사람만 보입니다. 미연은 검은 머리의 젊은 동아시아계 여성으로 보이나 참조의 단발과 달리 머리를 뒤로 묶었습니다. 현우는 짧은 검은 머리의 어린 동아시아계 남자아이이며 어두운 셔츠를 입었습니다. 앰버는 밝은 머리와 창백한 피부의 여자아이로, 방진 마스크를 얼굴에 착용하고 작업복과 공구 벨트를 갖췄습니다. 혼혈 정체성이나 현우의 인이어는 이 화면에서 확정하기 어렵습니다. 흑백이지만 밝은 주간 거리로 보여 반복 지정된 야간 기억과 다릅니다. 성조기와 공격자는 프레임 밖이므로 누락으로 보지 않습니다.",
        "hard_violations": [],
        "physics": "미연은 한쪽 무릎을 올리고 다른 쪽을 낮춘 자세이며, 아이들의 굽힌 하체는 도로 쪽으로 이어집니다. 하단 접지 일부는 잘렸으나 부유하는 모습은 아닙니다. 미연의 두 손은 아이들의 위팔에 닿고 현우의 팔도 앰버 쪽에 닿아 포옹 자체는 가능합니다. 다만 보호하는 접점이 머리가 아니라 어깨와 팔입니다. 마스크는 얼굴의 끈으로, 공구 주머니는 허리 벨트로 지지됩니다."
       },
       {
        "label": "A",
        "direction": "미연의 눈과 얼굴은 화면 오른쪽 위의 보이지 않는 대상을 향해 있어 높은 위치의 화면 밖 공격자를 바라보는 지시와 맞습니다. 현우와 앰버는 모두 고개와 시선을 아래로 내리고 미연의 몸 안쪽으로 얼굴을 숨깁니다. 미연의 양손은 각각 두 아이의 머리를 감쌉니다. 보이는 무기는 없습니다.",
        "built_space": "뒤쪽에 콘크리트 공장 건물 하나, 오른쪽 철망 울타리 구간과 적재물, 소화전 하나, 오른쪽 하단 배수구 하나가 보입니다. 왼쪽 가까운 전신주와 점등된 곡선형 가로등 하나도 참조의 배치와 부합합니다. 외부 계단은 가족 뒤에 가려져 확인하기 어렵습니다. 가족은 차도에 무릎 꿇고 있어 공간 관계가 자연스럽지만, 건물과 도로를 넓게 드러내어 배경을 작은 흐릿한 띠로 제한하지 못했습니다.",
        "entities": "지정된 세 사람만 등장합니다. 미연은 참조와 유사한 검은 단발의 젊은 동아시아계 여성이고, 해진 회색 계열 셔츠를 입었습니다. 현우는 헝클어진 검은 머리의 어린 남자아이로 어두운 작업성 셔츠를 입고 있습니다. 앰버는 밝은 머리의 어린 여자아이이며 작업복과 공구 벨트를 착용합니다. 방진 마스크는 참조처럼 목 아래에 걸려 있지만 얼굴을 덮지는 않습니다. 숙인 얼굴 때문에 아이들의 정확한 얼굴 동일성 및 혼혈 특징은 제한적으로만 확인됩니다. 흑백 야간 분위기와 참조에 있는 가로등은 맞으며, 프레임 밖 성조기와 공격자는 평가상 누락이 아닙니다.",
        "hard_violations": [],
        "physics": "미연의 양 무릎이 아스팔트에 닿고, 아이들은 접힌 무릎과 정강이로 도로에 몸을 지탱합니다. 아이들의 머리는 미연의 가슴과 팔에 기대며, 미연의 손바닥이 각 머리 위와 뒤에 실제로 접촉합니다. 앰버의 마스크는 목의 끈과 가슴으로, 공구 주머니는 벨트로 지지됩니다. 젖은 노면의 밝은 반사는 주변 조명과 양립하며, 지지 없는 신체나 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.857
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.607
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 로케이션 레퍼런스를 완전히 무시하고 구조와 재질이 다른 엉뚱한 배경 건물을 렌더링함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 로케이션을 완벽히 재현하고 아이들이 얼굴을 숙인 자세를 잘 표현했으나, 요구된 클로즈업 숏보다 프레이밍이 넓게 잡혔습니다."
   },
   {
    "label": "B",
    "score": 607,
    "verdict_ko": "요구된 장소를 완전히 무시하고 다른 배경을 생성했으며, 아이들이 얼굴을 파묻는 동작과 캐릭터의 외모도 제대로 반영하지 못했습니다.  ★위반: [gemini-pro] 지정된 로케이션 레퍼런스를 완전히 무시하고 구조와 재질이 다른 엉뚱한 배경 건물을 렌더링함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L19B02.png",
    "asset_id": "17f89a7f-3d0b-4d77-89c2-e3191cabdeb7",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab082b-15c3-7ba3-a7f1-2e5b73032c89",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh18__bgfirst_bg.png",
   "bg_asset_id": "2d730d67-7714-4d44-87a6-02cd7c574a48",
   "bg_record_key": "S8sh18::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh22::signage": {
  "fp": "77510ea9d4fe9ef8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh22": {
  "input_fingerprint": "c9e437e725a7b136",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 군복 입은 아버지의 등 뒤로 또 다른 복면 남자의 총구가 불을 뿜는 결정적 순간.\n\nLOCATION (lock): On the same nighttime city street in the monochrome memory, among the masked gunmen and captive families. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track at a low, oblique rear-quarter position beside 현우의 아버지, keeping his uniformed back near center and the first masked attacker partially visible at left within their grapple. Reveal the second masked attacker at right with his firing muzzle clearly separated from the father's silhouette, while remaining outside the firing line; the father calls toward his family beyond the left edge as both attackers concentrate on him. Capture the flash as a discontinuous nightmare image, emphasizing only the second attacker's newly revealed position rather than adding a camera jolt.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: City street (The location of the attack in the nightmare); used as Retain a narrow strip beneath the figures to ground their shared positions; Second masked attacker's gun (Firing toward the father's back) — Seen obliquely from the side, with the muzzle pointing toward the father rather than toward the camera; used as Keep the muzzle clear of the father's silhouette as the decisive spatial evidence.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime nightmare in black and white, with the gun's brief muzzle flash interrupting restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street remains in black-and-white nighttime imagery, with gunfire flashes and the large American flag present. 현우의 아버지: Still wears his military uniform and is lunging forward.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 또 다른 복면 남자 right now, so 또 다른 복면 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 또 다른 복면 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 현우의 아버지 (미국인 남성, 성인의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 군복 입은 아버지의 등 뒤로 또 다른 복면 남자의 총구가 불을 뿜는 결정적 순간.\n\nLOCATION (lock): On the same nighttime city street in the monochrome memory, among the masked gunmen and captive families. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track at a low, oblique rear-quarter position beside 현우의 아버지, keeping his uniformed back near center and the first masked attacker partially visible at left within their grapple. Reveal the second masked attacker at right with his firing muzzle clearly separated from the father's silhouette, while remaining outside the firing line; the father calls toward his family beyond the left edge as both attackers concentrate on him. Capture the flash as a discontinuous nightmare image, emphasizing only the second attacker's newly revealed position rather than adding a camera jolt.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: City street (The location of the attack in the nightmare); used as Retain a narrow strip beneath the figures to ground their shared positions; Second masked attacker's gun (Firing toward the father's back) — Seen obliquely from the side, with the muzzle pointing toward the father rather than toward the camera; used as Keep the muzzle clear of the father's silhouette as the decisive spatial evidence.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime nightmare in black and white, with the gun's brief muzzle flash interrupting restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street remains in black-and-white nighttime imagery, with gunfire flashes and the large American flag present. 현우의 아버지: Still wears his military uniform and is lunging forward.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 또 다른 복면 남자 right now, so 또 다른 복면 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 또 다른 복면 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 현우의 아버지 (미국인 남성, 성인의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 군복 입은 아버지의 등 뒤로 또 다른 복면 남자의 총구가 불을 뿜는 결정적 순간.\n\nLOCATION (lock): On the same nighttime city street in the monochrome memory, among the masked gunmen and captive families. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track at a low, oblique rear-quarter position beside 현우의 아버지, keeping his uniformed back near center and the first masked attacker partially visible at left within their grapple. Reveal the second masked attacker at right with his firing muzzle clearly separated from the father's silhouette, while remaining outside the firing line; the father calls toward his family beyond the left edge as both attackers concentrate on him. Capture the flash as a discontinuous nightmare image, emphasizing only the second attacker's newly revealed position rather than adding a camera jolt.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: City street (The location of the attack in the nightmare); used as Retain a narrow strip beneath the figures to ground their shared positions; Second masked attacker's gun (Firing toward the father's back) — Seen obliquely from the side, with the muzzle pointing toward the father rather than toward the camera; used as Keep the muzzle clear of the father's silhouette as the decisive spatial evidence.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the nighttime nightmare in black and white, with the gun's brief muzzle flash interrupting restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The American street remains in black-and-white nighttime imagery, with gunfire flashes and the large American flag present. 현우의 아버지: Still wears his military uniform and is lunging forward.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 또 다른 복면 남자 right now, so 또 다른 복면 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 또 다른 복면 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 현우의 아버지 (미국인 남성, 성인의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "아버지는 화면 왼쪽 밖을 향해 시선을 두고 외치며, 오른쪽 복면 남자는 아버지의 몸통을 향해 총을 겨누고 있음.",
    "built_space": "레퍼런스와 유사한 젖은 야간 도로와 배경 건물이 보이나, 프롬프트에서 명시한 대형 미국 국기는 없음.",
    "entities": "아버지는 인물 레퍼런스와 일치하며 군복을 착용함. 두 명의 복면 남자가 묘사됨.",
    "hard_violations": [],
    "physics": "아버지와 첫 번째 습격자가 서서 몸싸움을 벌이고 있으며, 두 번째 습격자는 손으로 총을 자연스럽게 파지하고 있음."
   },
   {
    "label": "B",
    "direction": "아버지는 왼쪽으로 고개를 돌려 외치고, 오른쪽 복면 남자는 아버지의 등을 향해 총을 겨누며 발포함.",
    "built_space": "야간 거리와 건물 배경이 레퍼런스와 일치하며, 프롬프트에서 요구한 대형 미국 국기가 건물 벽에 나타남.",
    "entities": "아버지는 레퍼런스와 일치하는 외모에 군복을 입고 있으며, 두 복면 남자가 적절히 배치됨.",
    "hard_violations": [],
    "physics": "아버지와 첫 번째 습격자가 몸을 밀착해 체중을 싣고 있고, 두 번째 습격자는 총을 올바르게 손에 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "아버지의 앞모습과 측면을 보여주어 '등이 보이는 후방 측면 구도' 지시를 크게 위반했으며, 배경에 명시된 대형 국기가 누락됨."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "'등이 보이는 후방 측면 구도'를 정확히 구현했고, 총구의 위치 및 배경의 대형 미국 국기 등 지시사항을 충실히 반영함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "아버지는 화면 왼쪽 밖을 향해 시선을 두고 외치며, 오른쪽 복면 남자는 아버지의 몸통을 향해 총을 겨누고 있음.",
        "built_space": "레퍼런스와 유사한 젖은 야간 도로와 배경 건물이 보이나, 프롬프트에서 명시한 대형 미국 국기는 없음.",
        "entities": "아버지는 인물 레퍼런스와 일치하며 군복을 착용함. 두 명의 복면 남자가 묘사됨.",
        "hard_violations": [],
        "physics": "아버지와 첫 번째 습격자가 서서 몸싸움을 벌이고 있으며, 두 번째 습격자는 손으로 총을 자연스럽게 파지하고 있음."
       },
       {
        "label": "B",
        "direction": "아버지는 왼쪽으로 고개를 돌려 외치고, 오른쪽 복면 남자는 아버지의 등을 향해 총을 겨누며 발포함.",
        "built_space": "야간 거리와 건물 배경이 레퍼런스와 일치하며, 프롬프트에서 요구한 대형 미국 국기가 건물 벽에 나타남.",
        "entities": "아버지는 레퍼런스와 일치하는 외모에 군복을 입고 있으며, 두 복면 남자가 적절히 배치됨.",
        "hard_violations": [],
        "physics": "아버지와 첫 번째 습격자가 몸을 밀착해 체중을 싣고 있고, 두 번째 습격자는 총을 올바르게 손에 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "아버지의 앞모습과 측면을 보여주어 '등이 보이는 후방 측면 구도' 지시를 크게 위반했으며, 배경에 명시된 대형 국기가 누락됨."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "'등이 보이는 후방 측면 구도'를 정확히 구현했고, 총구의 위치 및 배경의 대형 미국 국기 등 지시사항을 충실히 반영함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "아버지는 화면 왼쪽 밖을 향해 시선을 두고 외치며, 오른쪽 복면 남자는 아버지의 몸통을 향해 총을 겨누고 있음.",
        "built_space": "레퍼런스와 유사한 젖은 야간 도로와 배경 건물이 보이나, 프롬프트에서 명시한 대형 미국 국기는 없음.",
        "entities": "아버지는 인물 레퍼런스와 일치하며 군복을 착용함. 두 명의 복면 남자가 묘사됨.",
        "hard_violations": [],
        "physics": "아버지와 첫 번째 습격자가 서서 몸싸움을 벌이고 있으며, 두 번째 습격자는 손으로 총을 자연스럽게 파지하고 있음."
       },
       {
        "label": "B",
        "direction": "아버지는 왼쪽으로 고개를 돌려 외치고, 오른쪽 복면 남자는 아버지의 등을 향해 총을 겨누며 발포함.",
        "built_space": "야간 거리와 건물 배경이 레퍼런스와 일치하며, 프롬프트에서 요구한 대형 미국 국기가 건물 벽에 나타남.",
        "entities": "아버지는 레퍼런스와 일치하는 외모에 군복을 입고 있으며, 두 복면 남자가 적절히 배치됨.",
        "hard_violations": [],
        "physics": "아버지와 첫 번째 습격자가 몸을 밀착해 체중을 싣고 있고, 두 번째 습격자는 총을 올바르게 손에 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙의 군복 입은 등, 왼쪽 몸싸움, 등과 분리된 오른쪽 발사 총구 및 대형 성조기를 함께 지켰으며, 전방 돌진의 기울기는 다소 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 후방 사선 구도와 전방 돌진, 등을 향한 발사는 충실하지만, 지속 상태로 명시된 거리의 대형 성조기가 보이지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "아버지는 고개를 왼쪽으로 돌리고 입을 벌려 화면 밖 가족을 부르는 모습이다. 왼쪽 복면 남자는 아버지의 몸통에 밀착해 붙잡고 있으며 눈은 보이지 않는다. 오른쪽 남자는 아버지 쪽으로 머리와 팔을 향한다. 권총 총열은 카메라가 아니라 왼쪽의 아버지 등 상부를 향하고, 총구와 섬광은 등 윤곽 바깥에 분리되어 있다.",
        "built_space": "젖은 도로와 연석, 오른쪽 철망 울타리, 뒤편 콘크리트 건물의 가로 창들이 참조 장소의 재료와 배치를 잇는다. 왼쪽 가까운 가로등 한 개와 멀어지는 가로등 열, 오른쪽 건물의 대형 성조기 한 장이 보인다. 아버지는 중앙, 몸싸움 상대는 왼쪽에 일부 잘리고 사수는 오른쪽 전경에 배치되어 있다. 하단과 인물 사이의 도로가 같은 지면 위의 관계를 보여 주며, 불가능한 반사는 없다.",
        "entities": "보이는 인물은 아버지와 복면 남자 두 명으로, 카메라 지시의 세 인물에 해당한다. 아버지는 참조와 유사한 중년 백인 남성의 얼굴, 짧게 정돈한 머리와 수염을 갖고 위장 군복 및 전술 조끼를 입었다. 국적 자체는 외모만으로 확정할 수 없지만 미국 국기 패치가 있다. 두 공격자는 검은 복면과 어두운 옷을 착용했고, 오른쪽 공격자의 장갑 낀 손과 소매가 발사 중인 권총에 연결된다. 가족을 화면 안에 추가하지 않았으며 흑백 야간 표현과 대형 성조기가 유지된다.",
        "hard_violations": [],
        "physics": "오른쪽 남자의 손이 권총 손잡이를 잡고 손목과 팔이 총을 지지한다. 섬광은 총구에서 발생한다. 아버지와 왼쪽 남자는 팔과 몸통을 맞대고 힘을 겨루며, 아버지의 몸은 약간 왼쪽으로 향하지만 강한 전방 돌진보다는 버티며 비트는 동작에 가깝다. 발은 프레임 밖이라 접지는 직접 확인할 수 없으나, 몸이 공중에 떠 있다는 증거나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "아버지는 왼쪽 화면 밖을 향해 입을 벌리고, 몸통도 몸싸움 상대 쪽으로 기울인다. 왼쪽 복면 남자의 드러난 눈은 아버지 쪽을 향한다. 오른쪽 사수는 아버지를 바라보며 권총을 왼쪽으로 겨눈다. 총열의 연장 방향은 아버지의 등 상부에 닿으며 카메라를 향하지 않는다. 발사 총구와 섬광은 아버지 윤곽에서 떨어져 보인다.",
        "built_space": "콘크리트 건물의 상층 가로 창군, 수직 배관, 철망 울타리, 연석과 젖은 도로가 참조 거리와 잘 대응한다. 오른쪽 보도에 소화전 한 개가 있고, 왼쪽 가까운 곡선형 가로등 한 개와 뒤쪽 조명 및 원거리 가로등 열이 보인다. 중앙 아버지, 왼쪽 몸싸움 상대, 오른쪽 사수의 배치는 지시와 맞고 후방 사선의 낮은 시점도 읽힌다. 거리의 대형 성조기는 보이지 않는다. 도로의 빛 반사는 가능한 위치에 있다.",
        "entities": "아버지와 복면 공격자 두 명만 보인다. 아버지의 짧은 머리, 중년 백인 남성으로 보이는 옆얼굴과 체격은 인물 참조에 부합하며 위장 군복과 미국 국기 패치를 착용했다. 두 공격자는 검은 복면과 어두운 옷을 입었다. 오른쪽 남자의 맨손, 손목과 검은 소매가 권총을 쥔 팔에 자연스럽게 이어진다. 권총의 발사 섬광과 흑백 야간 분위기는 맞지만, 어깨 패치는 요구된 거리의 대형 성조기를 대신하지 못한다.",
        "hard_violations": [],
        "physics": "사수의 양손이 권총을 지지하고 팔이 몸으로 이어져 발사 자세가 성립한다. 아버지는 상체를 왼쪽 앞으로 기울이며 상대와 팔 및 몸통으로 접촉해, 몸싸움 중 전방으로 밀고 나가는 동작이 드러난다. 인물들의 발은 잘려 직접 접지를 확인할 수 없지만 공중 부양으로 보이지 않는다. 총구 섬광 외에 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙의 군복 입은 등, 왼쪽 몸싸움, 등과 분리된 오른쪽 발사 총구 및 대형 성조기를 함께 지켰으며, 전방 돌진의 기울기는 다소 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 후방 사선 구도와 전방 돌진, 등을 향한 발사는 충실하지만, 지속 상태로 명시된 거리의 대형 성조기가 보이지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "아버지는 고개를 왼쪽으로 돌리고 입을 벌려 화면 밖 가족을 부르는 모습이다. 왼쪽 복면 남자는 아버지의 몸통에 밀착해 붙잡고 있으며 눈은 보이지 않는다. 오른쪽 남자는 아버지 쪽으로 머리와 팔을 향한다. 권총 총열은 카메라가 아니라 왼쪽의 아버지 등 상부를 향하고, 총구와 섬광은 등 윤곽 바깥에 분리되어 있다.",
        "built_space": "젖은 도로와 연석, 오른쪽 철망 울타리, 뒤편 콘크리트 건물의 가로 창들이 참조 장소의 재료와 배치를 잇는다. 왼쪽 가까운 가로등 한 개와 멀어지는 가로등 열, 오른쪽 건물의 대형 성조기 한 장이 보인다. 아버지는 중앙, 몸싸움 상대는 왼쪽에 일부 잘리고 사수는 오른쪽 전경에 배치되어 있다. 하단과 인물 사이의 도로가 같은 지면 위의 관계를 보여 주며, 불가능한 반사는 없다.",
        "entities": "보이는 인물은 아버지와 복면 남자 두 명으로, 카메라 지시의 세 인물에 해당한다. 아버지는 참조와 유사한 중년 백인 남성의 얼굴, 짧게 정돈한 머리와 수염을 갖고 위장 군복 및 전술 조끼를 입었다. 국적 자체는 외모만으로 확정할 수 없지만 미국 국기 패치가 있다. 두 공격자는 검은 복면과 어두운 옷을 착용했고, 오른쪽 공격자의 장갑 낀 손과 소매가 발사 중인 권총에 연결된다. 가족을 화면 안에 추가하지 않았으며 흑백 야간 표현과 대형 성조기가 유지된다.",
        "hard_violations": [],
        "physics": "오른쪽 남자의 손이 권총 손잡이를 잡고 손목과 팔이 총을 지지한다. 섬광은 총구에서 발생한다. 아버지와 왼쪽 남자는 팔과 몸통을 맞대고 힘을 겨루며, 아버지의 몸은 약간 왼쪽으로 향하지만 강한 전방 돌진보다는 버티며 비트는 동작에 가깝다. 발은 프레임 밖이라 접지는 직접 확인할 수 없으나, 몸이 공중에 떠 있다는 증거나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "아버지는 왼쪽 화면 밖을 향해 입을 벌리고, 몸통도 몸싸움 상대 쪽으로 기울인다. 왼쪽 복면 남자의 드러난 눈은 아버지 쪽을 향한다. 오른쪽 사수는 아버지를 바라보며 권총을 왼쪽으로 겨눈다. 총열의 연장 방향은 아버지의 등 상부에 닿으며 카메라를 향하지 않는다. 발사 총구와 섬광은 아버지 윤곽에서 떨어져 보인다.",
        "built_space": "콘크리트 건물의 상층 가로 창군, 수직 배관, 철망 울타리, 연석과 젖은 도로가 참조 거리와 잘 대응한다. 오른쪽 보도에 소화전 한 개가 있고, 왼쪽 가까운 곡선형 가로등 한 개와 뒤쪽 조명 및 원거리 가로등 열이 보인다. 중앙 아버지, 왼쪽 몸싸움 상대, 오른쪽 사수의 배치는 지시와 맞고 후방 사선의 낮은 시점도 읽힌다. 거리의 대형 성조기는 보이지 않는다. 도로의 빛 반사는 가능한 위치에 있다.",
        "entities": "아버지와 복면 공격자 두 명만 보인다. 아버지의 짧은 머리, 중년 백인 남성으로 보이는 옆얼굴과 체격은 인물 참조에 부합하며 위장 군복과 미국 국기 패치를 착용했다. 두 공격자는 검은 복면과 어두운 옷을 입었다. 오른쪽 남자의 맨손, 손목과 검은 소매가 권총을 쥔 팔에 자연스럽게 이어진다. 권총의 발사 섬광과 흑백 야간 분위기는 맞지만, 어깨 패치는 요구된 거리의 대형 성조기를 대신하지 못한다.",
        "hard_violations": [],
        "physics": "사수의 양손이 권총을 지지하고 팔이 몸으로 이어져 발사 자세가 성립한다. 아버지는 상체를 왼쪽 앞으로 기울이며 상대와 팔 및 몸통으로 접촉해, 몸싸움 중 전방으로 밀고 나가는 동작이 드러난다. 인물들의 발은 잘려 직접 접지를 확인할 수 없지만 공중 부양으로 보이지 않는다. 총구 섬광 외에 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.46,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.46,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1460,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1460,
    "verdict_ko": "아버지의 앞모습과 측면을 보여주어 '등이 보이는 후방 측면 구도' 지시를 크게 위반했으며, 배경에 명시된 대형 국기가 누락됨."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "'등이 보이는 후방 측면 구도'를 정확히 구현했고, 총구의 위치 및 배경의 대형 미국 국기 등 지시사항을 충실히 반영함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S8sh18_sel.png",
    "asset_id": "8e0e4c5f-b482-4ea3-8e2c-cea46a65a96a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 현우의 아버지: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1216694>",
    "asset_id": "9c3735f8-7c23-402d-be8b-c6e3c53d9e79",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab083f-2202-703e-b219-bf3d5966c974",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S8sh18"
  }
 },
 "S9sh1::signage": {
  "fp": "172e20d877fa77c8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::49e8f10872dfabe9": {
  "subjects": [],
  "subject_text": "난민 재판소 외부 쓰레기 더미\n허름한 재판소 건물 바깥에 잡다한 폐기물이 쌓인 황폐한 공간. 쓰레기 더미 사이로 거친 바닥이 드러나고 주변 조명은 희미하다.",
  "identity": "canonical",
  "scope_id": "L21",
  "scope_role": "location_exterior",
  "scope_sha": "88ed78f83dfe5d52"
 },
 "S9sh1::bgfirst_bg": {
  "input_fingerprint": "a131b0e321175247",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 위에 쓰러져 있던 이현우가 입을 크게 벌리고 비명을 지르며 상체를 튕기듯 일으키는 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin low and diagonally beside 이현우's reclining body, looking slightly downward from outside his frontal axis before the crane begins to rise. His abruptly lifting torso occupies the central half of the frame, his mouth open and shoulders contracting, while rubbish remains visible beneath him and along both lower corners; his unfocused gaze passes beyond the frame toward his surroundings. Hold the entry camera position for this instant, letting his bodily displacement supply the change from the nightmare rather than anticipating his later rise to his feet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heaps (Surrounding and beneath 이현우 outside the refugee court); used as Frame the lower torso and establish where he has awakened without introducing individual invented objects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination with readable facial detail and restrained contrast, returning to an undistorted view of the present.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쓰레기 더미 위에 쓰러져 있던 이현우가 입을 크게 벌리고 비명을 지르며 상체를 튕기듯 일으키는 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin low and diagonally beside 이현우's reclining body, looking slightly downward from outside his frontal axis before the crane begins to rise. His abruptly lifting torso occupies the central half of the frame, his mouth open and shoulders contracting, while rubbish remains visible beneath him and along both lower corners; his unfocused gaze passes beyond the frame toward his surroundings. Hold the entry camera position for this instant, letting his bodily displacement supply the change from the nightmare rather than anticipating his later rise to his feet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heaps (Surrounding and beneath 이현우 outside the refugee court); used as Frame the lower torso and establish where he has awakened without introducing individual invented objects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination with readable facial detail and restrained contrast, returning to an undistorted view of the present.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S9sh1__bgfirst_bg.png",
  "asset_id": "11d5e9db-0a62-45ca-96ac-0e59a23f04c3",
  "input_asset_ids": [
   "d049d6dd-fa75-48bc-b6ad-3d345925157e",
   "37d09030-d38c-44ef-bfe0-ff791db954d2"
  ]
 },
 "S9sh1": {
  "input_fingerprint": "0c4e45a00a8ac716",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 위에 쓰러져 있던 이현우가 입을 크게 벌리고 비명을 지르며 상체를 튕기듯 일으키는 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin low and diagonally beside 이현우's reclining body, looking slightly downward from outside his frontal axis before the crane begins to rise. His abruptly lifting torso occupies the central half of the frame, his mouth open and shoulders contracting, while rubbish remains visible beneath him and along both lower corners; his unfocused gaze passes beyond the frame toward his surroundings. Hold the entry camera position for this instant, letting his bodily displacement supply the change from the nightmare rather than anticipating his later rise to his feet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heaps (Surrounding and beneath 이현우 outside the refugee court); used as Frame the lower torso and establish where he has awakened without introducing individual invented objects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination with readable facial detail and restrained contrast, returning to an undistorted view of the present.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the shabby refugee tribunal, surrounded by desolate heaps of rubbish. 이현우: Wakes from unconsciousness with his dog-bitten leg and baton-struck head still injured. The judge's wallet is already in his pocket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 위에 쓰러져 있던 이현우가 입을 크게 벌리고 비명을 지르며 상체를 튕기듯 일으키는 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin low and diagonally beside 이현우's reclining body, looking slightly downward from outside his frontal axis before the crane begins to rise. His abruptly lifting torso occupies the central half of the frame, his mouth open and shoulders contracting, while rubbish remains visible beneath him and along both lower corners; his unfocused gaze passes beyond the frame toward his surroundings. Hold the entry camera position for this instant, letting his bodily displacement supply the change from the nightmare rather than anticipating his later rise to his feet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heaps (Surrounding and beneath 이현우 outside the refugee court); used as Frame the lower torso and establish where he has awakened without introducing individual invented objects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination with readable facial detail and restrained contrast, returning to an undistorted view of the present.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the shabby refugee tribunal, surrounded by desolate heaps of rubbish. 이현우: Wakes from unconsciousness with his dog-bitten leg and baton-struck head still injured. The judge's wallet is already in his pocket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 쓰레기 더미 위에 쓰러져 있던 이현우가 입을 크게 벌리고 비명을 지르며 상체를 튕기듯 일으키는 찰나.\n\nLOCATION (lock): On a desolate rubbish heap immediately outside the refugee tribunal at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin low and diagonally beside 이현우's reclining body, looking slightly downward from outside his frontal axis before the crane begins to rise. His abruptly lifting torso occupies the central half of the frame, his mouth open and shoulders contracting, while rubbish remains visible beneath him and along both lower corners; his unfocused gaze passes beyond the frame toward his surroundings. Hold the entry camera position for this instant, letting his bodily displacement supply the change from the nightmare rather than anticipating his later rise to his feet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubbish heaps (Surrounding and beneath 이현우 outside the refugee court); used as Frame the lower torso and establish where he has awakened without introducing individual invented objects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination with readable facial detail and restrained contrast, returning to an undistorted view of the present.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night outside the shabby refugee tribunal, surrounded by desolate heaps of rubbish. 이현우: Wakes from unconsciousness with his dog-bitten leg and baton-struck head still injured. The judge's wallet is already in his pocket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S9sh1__bgfirst_bg.png",
     "asset_id": "11d5e9db-0a62-45ca-96ac-0e59a23f04c3",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S9sh1.png",
     "asset_id": "d049d6dd-fa75-48bc-b6ad-3d345925157e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L21B01.png",
     "asset_id": "37d09030-d38c-44ef-bfe0-ff791db954d2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 밖 상단을 향하고 있으며, 입을 크게 벌려 허공을 향해 비명을 지르고 있습니다.",
    "built_space": "레퍼런스의 야외 건물과 조명이 일치하며, 바닥에 흩어진 쓰레기 더미 위에 인물이 위치해 지정된 공간 스케일을 유지합니다.",
    "entities": "이현우의 얼굴, 피와 흙이 묻은 어두운 셔츠와 바지, 귀에 꽂힌 인이어 무전기 모두 참조 이미지와 일치합니다.",
    "hard_violations": [],
    "physics": "하반신은 쓰레기 더미에 닿아 있고 오른손으로 바닥을 짚은 채 상체를 강하게 들어 올리는 역동적인 힘이 묘사되어 있습니다."
   },
   {
    "label": "B",
    "direction": "시선은 화면 좌측 밖을 향하고 있으며, 입을 벌리고 놀란 표정을 짓고 있습니다.",
    "built_space": "배경 건물은 레퍼런스와 일치하나, 인물이 바닥의 쓰레기 더미가 아닌 배경에 있어야 할 철제 케이지 위에 걸터앉아 있어 스케일과 배치가 변경되었습니다.",
    "entities": "이현우의 외모, 의상, 무전기 등이 일치하며, 바지의 찢어진 틈으로 다리의 상처가 보입니다.",
    "hard_violations": [],
    "physics": "철제 케이지 모서리에 엉덩이를 대고 체중을 지지하고 있어, 누워있다가 튕기듯 일어나는 것이 아니라 앉아서 뒤로 물러나는 자세에 가깝습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "비명을 지르며 상체를 튕기듯 일으키는 역동적인 찰나와 바닥의 쓰레기 더미라는 공간적 설정을 프롬프트에 맞게 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "상체를 강하게 일으키는 동작 대신 놀란 표정으로 기대어 앉아 있으며, 바닥이 아닌 철제 케이지 위에 위치해 상황 묘사와 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 밖 상단을 향하고 있으며, 입을 크게 벌려 허공을 향해 비명을 지르고 있습니다.",
        "built_space": "레퍼런스의 야외 건물과 조명이 일치하며, 바닥에 흩어진 쓰레기 더미 위에 인물이 위치해 지정된 공간 스케일을 유지합니다.",
        "entities": "이현우의 얼굴, 피와 흙이 묻은 어두운 셔츠와 바지, 귀에 꽂힌 인이어 무전기 모두 참조 이미지와 일치합니다.",
        "hard_violations": [],
        "physics": "하반신은 쓰레기 더미에 닿아 있고 오른손으로 바닥을 짚은 채 상체를 강하게 들어 올리는 역동적인 힘이 묘사되어 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 좌측 밖을 향하고 있으며, 입을 벌리고 놀란 표정을 짓고 있습니다.",
        "built_space": "배경 건물은 레퍼런스와 일치하나, 인물이 바닥의 쓰레기 더미가 아닌 배경에 있어야 할 철제 케이지 위에 걸터앉아 있어 스케일과 배치가 변경되었습니다.",
        "entities": "이현우의 외모, 의상, 무전기 등이 일치하며, 바지의 찢어진 틈으로 다리의 상처가 보입니다.",
        "hard_violations": [],
        "physics": "철제 케이지 모서리에 엉덩이를 대고 체중을 지지하고 있어, 누워있다가 튕기듯 일어나는 것이 아니라 앉아서 뒤로 물러나는 자세에 가깝습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "비명을 지르며 상체를 튕기듯 일으키는 역동적인 찰나와 바닥의 쓰레기 더미라는 공간적 설정을 프롬프트에 맞게 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "상체를 강하게 일으키는 동작 대신 놀란 표정으로 기대어 앉아 있으며, 바닥이 아닌 철제 케이지 위에 위치해 상황 묘사와 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 밖 상단을 향하고 있으며, 입을 크게 벌려 허공을 향해 비명을 지르고 있습니다.",
        "built_space": "레퍼런스의 야외 건물과 조명이 일치하며, 바닥에 흩어진 쓰레기 더미 위에 인물이 위치해 지정된 공간 스케일을 유지합니다.",
        "entities": "이현우의 얼굴, 피와 흙이 묻은 어두운 셔츠와 바지, 귀에 꽂힌 인이어 무전기 모두 참조 이미지와 일치합니다.",
        "hard_violations": [],
        "physics": "하반신은 쓰레기 더미에 닿아 있고 오른손으로 바닥을 짚은 채 상체를 강하게 들어 올리는 역동적인 힘이 묘사되어 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 좌측 밖을 향하고 있으며, 입을 벌리고 놀란 표정을 짓고 있습니다.",
        "built_space": "배경 건물은 레퍼런스와 일치하나, 인물이 바닥의 쓰레기 더미가 아닌 배경에 있어야 할 철제 케이지 위에 걸터앉아 있어 스케일과 배치가 변경되었습니다.",
        "entities": "이현우의 외모, 의상, 무전기 등이 일치하며, 바지의 찢어진 틈으로 다리의 상처가 보입니다.",
        "hard_violations": [],
        "physics": "철제 케이지 모서리에 엉덩이를 대고 체중을 지지하고 있어, 누워있다가 튕기듯 일어나는 것이 아니라 앉아서 뒤로 물러나는 자세에 가깝습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장소와 인물은 대체로 맞지만, 어깨를 수축하며 비명을 지르는 순간보다는 이미 몸을 일으켜 놀란 채 기대앉은 모습에 가깝다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "수축한 어깨와 크게 벌린 입, 흐트러진 머리가 상체를 튕겨 일으키는 찰나를 더 충실히 구현하지만, 카메라는 지정된 약한 내려다보기보다 올려다보기에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 정상적인 눈동자는 화면 오른쪽 바깥의 주변 공간을 향한다. 렌즈를 응시하지는 않지만 눈이 특정 대상을 또렷하게 좇는 듯하여 초점 없는 시선은 약하다. 상체는 골반에서 위로 세워져 있으며 뒤로 기울어 있다. 겨누는 무기나 사용 중인 도구는 없다.",
        "built_space": "뒤쪽의 낡은 콘크리트 건물에는 상층 창 네 구획이 있고 일부가 머리에 가려진다. 하층 창은 몸과 폐기물에 상당 부분 가려져 있다. 오른쪽에는 출입문 위 차양 하나와 그 아래 켜진 조명 하나, 벽 배관과 건물 사이 통로가 보인다. 인물은 전경 철망 폐기물 수거함의 내용물 위에 기대앉아 있다. 참조의 건물 배치와 재료는 유지되며 명백한 고정 시설 중복이나 불가능한 반사는 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며 짧은 검은 머리, 마른 체격과 얼굴은 인물 참조에 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 어두운 긴팔 셔츠와 바지에 먼지가 묻어 있고 바지의 찢김과 붉은 상처 흔적, 귀의 검은 인이어 장치가 보인다. 머리의 타격 상처와 셔츠의 핏자국은 뚜렷하지 않다. 주머니 속 지갑은 확인되지 않으며 별도 소품으로 노출되지도 않는다. 쓰레기는 몸 아래와 양쪽 아래 모서리에 있다.",
        "hard_violations": [],
        "physics": "골반과 허벅지가 수거함 안에 쌓인 쓰레기에 얹혀 있고, 화면 왼쪽 팔과 손도 내용물에 닿아 있다. 몸을 받치는 지지가 있으며 공중에 뜬 신체는 없다. 다만 상체와 머리가 비교적 안정되어 있어 갑작스러운 복근 수축보다는 몸을 일으킨 뒤의 자세로 읽힌다."
       },
       {
        "label": "B",
        "direction": "고개가 뒤로 젖혀지고 시선은 화면 오른쪽 위 바깥을 향한다. 특정 사물이나 카메라를 응시하지 않아 주변을 향하는 흐트러진 시선으로 읽힌다. 입을 크게 벌린 채 어깨가 올라가고 상체가 골반에서 위로 들리는 방향이 분명하다. 겨누는 물체는 없다.",
        "built_space": "배경의 콘크리트 건물 상층 창 네 구획과 세 줄의 수직 배관이 참조와 대응한다. 하층은 인물과 폐기물에 가려지고 오른쪽 창 구획이 드러난다. 오른쪽 벽에는 출입문 차양 하나와 켜진 조명 하나, 배관 및 뒤로 이어지는 통로가 있다. 인물은 마당의 쓰레기 더미 위에 놓였고 양쪽 아래 모서리에도 쓰레기가 이어진다. 배경 시설의 크기도 거리와 대체로 맞는다. 다만 얼굴과 몸을 보는 각도는 약한 내려다보기보다 낮은 올려다보기에 가깝다.",
        "entities": "참조와 대체로 일치하는 젊고 마른 동아시아계 남성 한 명이다. 헝클어진 짧은 검은 머리, 정상적인 눈, 귀의 소형 인이어 장치가 보인다. 어두운 긴팔 셔츠와 바지는 흙먼지로 더럽고 셔츠에는 붉은 얼룩이 있다. 이마와 관자 부근에도 상처 흔적이 보여 머리 부상 상태를 전달한다. 개에게 물린 다리 상처와 주머니 속 지갑은 이 화면에서 식별되지 않는다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "골반과 다리는 쓰레기 더미에 놓이고 화면 왼쪽 손바닥은 아래의 폐기물을 짚는다. 이 접촉과 팔의 버팀으로 기울어진 상체를 지탱할 수 있다. 올라간 어깨와 목의 긴장, 들린 머리카락은 갑자기 상체를 일으키는 동작과 양립한다. 몸 전체가 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장소와 인물은 대체로 맞지만, 어깨를 수축하며 비명을 지르는 순간보다는 이미 몸을 일으켜 놀란 채 기대앉은 모습에 가깝다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "수축한 어깨와 크게 벌린 입, 흐트러진 머리가 상체를 튕겨 일으키는 찰나를 더 충실히 구현하지만, 카메라는 지정된 약한 내려다보기보다 올려다보기에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 정상적인 눈동자는 화면 오른쪽 바깥의 주변 공간을 향한다. 렌즈를 응시하지는 않지만 눈이 특정 대상을 또렷하게 좇는 듯하여 초점 없는 시선은 약하다. 상체는 골반에서 위로 세워져 있으며 뒤로 기울어 있다. 겨누는 무기나 사용 중인 도구는 없다.",
        "built_space": "뒤쪽의 낡은 콘크리트 건물에는 상층 창 네 구획이 있고 일부가 머리에 가려진다. 하층 창은 몸과 폐기물에 상당 부분 가려져 있다. 오른쪽에는 출입문 위 차양 하나와 그 아래 켜진 조명 하나, 벽 배관과 건물 사이 통로가 보인다. 인물은 전경 철망 폐기물 수거함의 내용물 위에 기대앉아 있다. 참조의 건물 배치와 재료는 유지되며 명백한 고정 시설 중복이나 불가능한 반사는 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며 짧은 검은 머리, 마른 체격과 얼굴은 인물 참조에 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 어두운 긴팔 셔츠와 바지에 먼지가 묻어 있고 바지의 찢김과 붉은 상처 흔적, 귀의 검은 인이어 장치가 보인다. 머리의 타격 상처와 셔츠의 핏자국은 뚜렷하지 않다. 주머니 속 지갑은 확인되지 않으며 별도 소품으로 노출되지도 않는다. 쓰레기는 몸 아래와 양쪽 아래 모서리에 있다.",
        "hard_violations": [],
        "physics": "골반과 허벅지가 수거함 안에 쌓인 쓰레기에 얹혀 있고, 화면 왼쪽 팔과 손도 내용물에 닿아 있다. 몸을 받치는 지지가 있으며 공중에 뜬 신체는 없다. 다만 상체와 머리가 비교적 안정되어 있어 갑작스러운 복근 수축보다는 몸을 일으킨 뒤의 자세로 읽힌다."
       },
       {
        "label": "A",
        "direction": "고개가 뒤로 젖혀지고 시선은 화면 오른쪽 위 바깥을 향한다. 특정 사물이나 카메라를 응시하지 않아 주변을 향하는 흐트러진 시선으로 읽힌다. 입을 크게 벌린 채 어깨가 올라가고 상체가 골반에서 위로 들리는 방향이 분명하다. 겨누는 물체는 없다.",
        "built_space": "배경의 콘크리트 건물 상층 창 네 구획과 세 줄의 수직 배관이 참조와 대응한다. 하층은 인물과 폐기물에 가려지고 오른쪽 창 구획이 드러난다. 오른쪽 벽에는 출입문 차양 하나와 켜진 조명 하나, 배관 및 뒤로 이어지는 통로가 있다. 인물은 마당의 쓰레기 더미 위에 놓였고 양쪽 아래 모서리에도 쓰레기가 이어진다. 배경 시설의 크기도 거리와 대체로 맞는다. 다만 얼굴과 몸을 보는 각도는 약한 내려다보기보다 낮은 올려다보기에 가깝다.",
        "entities": "참조와 대체로 일치하는 젊고 마른 동아시아계 남성 한 명이다. 헝클어진 짧은 검은 머리, 정상적인 눈, 귀의 소형 인이어 장치가 보인다. 어두운 긴팔 셔츠와 바지는 흙먼지로 더럽고 셔츠에는 붉은 얼룩이 있다. 이마와 관자 부근에도 상처 흔적이 보여 머리 부상 상태를 전달한다. 개에게 물린 다리 상처와 주머니 속 지갑은 이 화면에서 식별되지 않는다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "골반과 다리는 쓰레기 더미에 놓이고 화면 왼쪽 손바닥은 아래의 폐기물을 짚는다. 이 접촉과 팔의 버팀으로 기울어진 상체를 지탱할 수 있다. 올라간 어깨와 목의 긴장, 들린 머리카락은 갑자기 상체를 일으키는 동작과 양립한다. 몸 전체가 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.446
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.446
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1446
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "비명을 지르며 상체를 튕기듯 일으키는 역동적인 찰나와 바닥의 쓰레기 더미라는 공간적 설정을 프롬프트에 맞게 잘 구현했습니다."
   },
   {
    "label": "B",
    "score": 1446,
    "verdict_ko": "상체를 강하게 일으키는 동작 대신 놀란 표정으로 기대어 앉아 있으며, 바닥이 아닌 철제 케이지 위에 위치해 상황 묘사와 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L21B01.png",
    "asset_id": "37d09030-d38c-44ef-bfe0-ff791db954d2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0845-6fa9-7f38-a052-0c463a2be944",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S9sh1__bgfirst_bg.png",
   "bg_asset_id": "11d5e9db-0a62-45ca-96ac-0e59a23f04c3",
   "bg_record_key": "S9sh1::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S9sh6::signage": {
  "fp": "b6e82912f3166c19",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "재판관의 신분증"
   }
  ],
  "dropped": []
 },
 "S9sh6": {
  "input_fingerprint": "07b980f6f19bbe4f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펼쳐진 지갑의 투명한 칸 너머로 재판관의 신분증 사진이 드러난 근접 구도.\n\nLOCATION (lock): In the rubbish-strewn outdoor ground beside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the approach from just behind and outside 이현우's shoulder on the established side, tilting obliquely downward until the photograph is unobstructed. Place the open wallet in the lower center at less than two-fifths of the image, with its transparent compartment facing the lens and the rubbish beyond softly readable; its supporting hands remain immediately below the crop, leaving no person visible. Emphasize the reduced camera distance, treating the photograph as a physical object seen directly through the compartment rather than as a cut into the pictured man's world.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Opened to reveal the identification compartment) — The interior faces upward and obliquely toward the camera; used as Provide the physical border around the photograph while preserving environmental context; Transparent wallet compartment (The judge's identification is visible through it) — The camera looks through the front of the compartment at the identification underneath; used as Establish that the photograph is inside the wallet rather than a separate image; Judge's identification (Inside the open wallet) — The photograph-bearing face is visible to the camera; do not invent additional readable personal details; used as Make the judge's photograph the focal detail; Rubbish heaps (Present around 이현우); used as Supply a softly resolved background around the small wallet.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, allowing the photograph to remain legible without introducing a special light or reflective glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open wallet contains the judge's identification card and money; the card identifies the judge. The tribunal exterior remains dark and surrounded by rubbish.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펼쳐진 지갑의 투명한 칸 너머로 재판관의 신분증 사진이 드러난 근접 구도.\n\nLOCATION (lock): In the rubbish-strewn outdoor ground beside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the approach from just behind and outside 이현우's shoulder on the established side, tilting obliquely downward until the photograph is unobstructed. Place the open wallet in the lower center at less than two-fifths of the image, with its transparent compartment facing the lens and the rubbish beyond softly readable; its supporting hands remain immediately below the crop, leaving no person visible. Emphasize the reduced camera distance, treating the photograph as a physical object seen directly through the compartment rather than as a cut into the pictured man's world.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Opened to reveal the identification compartment) — The interior faces upward and obliquely toward the camera; used as Provide the physical border around the photograph while preserving environmental context; Transparent wallet compartment (The judge's identification is visible through it) — The camera looks through the front of the compartment at the identification underneath; used as Establish that the photograph is inside the wallet rather than a separate image; Judge's identification (Inside the open wallet) — The photograph-bearing face is visible to the camera; do not invent additional readable personal details; used as Make the judge's photograph the focal detail; Rubbish heaps (Present around 이현우); used as Supply a softly resolved background around the small wallet.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, allowing the photograph to remain legible without introducing a special light or reflective glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open wallet contains the judge's identification card and money; the card identifies the judge. The tribunal exterior remains dark and surrounded by rubbish.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펼쳐진 지갑의 투명한 칸 너머로 재판관의 신분증 사진이 드러난 근접 구도.\n\nLOCATION (lock): In the rubbish-strewn outdoor ground beside the refugee tribunal at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the approach from just behind and outside 이현우's shoulder on the established side, tilting obliquely downward until the photograph is unobstructed. Place the open wallet in the lower center at less than two-fifths of the image, with its transparent compartment facing the lens and the rubbish beyond softly readable; its supporting hands remain immediately below the crop, leaving no person visible. Emphasize the reduced camera distance, treating the photograph as a physical object seen directly through the compartment rather than as a cut into the pictured man's world.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open wallet (Opened to reveal the identification compartment) — The interior faces upward and obliquely toward the camera; used as Provide the physical border around the photograph while preserving environmental context; Transparent wallet compartment (The judge's identification is visible through it) — The camera looks through the front of the compartment at the identification underneath; used as Establish that the photograph is inside the wallet rather than a separate image; Judge's identification (Inside the open wallet) — The photograph-bearing face is visible to the camera; do not invent additional readable personal details; used as Make the judge's photograph the focal detail; Rubbish heaps (Present around 이현우); used as Supply a softly resolved background around the small wallet.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, allowing the photograph to remain legible without introducing a special light or reflective glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open wallet contains the judge's identification card and money; the card identifies the judge. The tribunal exterior remains dark and surrounded by rubbish.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
    "built_space": "야외 쓰레기 더미가 있는 바닥이며, 다양한 형태의 쓰레기들이 아웃포커싱되어 배경에 위치합니다.",
    "entities": "펼쳐진 가죽 지갑, 한국어 지폐(5만원권), 카드류, 투명 칸 안에 들어있는 법복 차림 남성의 사진이 있는 신분증, 흙이 묻은 두 손이 확인됩니다.",
    "hard_violations": [],
    "physics": "지갑은 양쪽 하단 모서리를 잡고 있는 두 손에 의해 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
    "built_space": "야외 쓰레기 더미가 있는 바닥이며, 젖은 질감과 파편들이 부드러운 심도로 배경을 채우고 있습니다.",
    "entities": "펼쳐진 낡고 오염된 가죽 지갑, 한국어 지폐(5만원권), 투명 칸 안에 들어있는 정장 차림 남성의 사진과 문양이 있는 신분증, 흙과 피가 묻은 손이 확인됩니다.",
    "hard_violations": [],
    "physics": "지갑은 프레임 하단에서 올라온 두 손에 의해 단단히 파지되어 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "야간의 어두운 조명과 주변 쓰레기 더미의 젖은 질감을 훌륭하게 재현했으며, 지갑을 잡은 손을 화면 하단에 바짝 붙여 크롭하여 프레이밍 지시를 매우 충실히 이행했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지갑을 잡고 있는 손이 화면에 너무 많이 드러나 '프레임 바로 아래에 배치'하라는 지시에서 벗어났으며, 지갑 내부가 주변 환경의 오염도에 비해 다소 이질적으로 깨끗합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
        "built_space": "야외 쓰레기 더미가 있는 바닥이며, 젖은 질감과 파편들이 부드러운 심도로 배경을 채우고 있습니다.",
        "entities": "펼쳐진 낡고 오염된 가죽 지갑, 한국어 지폐(5만원권), 투명 칸 안에 들어있는 정장 차림 남성의 사진과 문양이 있는 신분증, 흙과 피가 묻은 손이 확인됩니다.",
        "hard_violations": [],
        "physics": "지갑은 프레임 하단에서 올라온 두 손에 의해 단단히 파지되어 지탱되고 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
        "built_space": "야외 쓰레기 더미가 있는 바닥이며, 다양한 형태의 쓰레기들이 아웃포커싱되어 배경에 위치합니다.",
        "entities": "펼쳐진 가죽 지갑, 한국어 지폐(5만원권), 카드류, 투명 칸 안에 들어있는 법복 차림 남성의 사진이 있는 신분증, 흙이 묻은 두 손이 확인됩니다.",
        "hard_violations": [],
        "physics": "지갑은 양쪽 하단 모서리를 잡고 있는 두 손에 의해 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "야간의 어두운 조명과 주변 쓰레기 더미의 젖은 질감을 훌륭하게 재현했으며, 지갑을 잡은 손을 화면 하단에 바짝 붙여 크롭하여 프레이밍 지시를 매우 충실히 이행했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지갑을 잡고 있는 손이 화면에 너무 많이 드러나 '프레임 바로 아래에 배치'하라는 지시에서 벗어났으며, 지갑 내부가 주변 환경의 오염도에 비해 다소 이질적으로 깨끗합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
        "built_space": "야외 쓰레기 더미가 있는 바닥이며, 젖은 질감과 파편들이 부드러운 심도로 배경을 채우고 있습니다.",
        "entities": "펼쳐진 낡고 오염된 가죽 지갑, 한국어 지폐(5만원권), 투명 칸 안에 들어있는 정장 차림 남성의 사진과 문양이 있는 신분증, 흙과 피가 묻은 손이 확인됩니다.",
        "hard_violations": [],
        "physics": "지갑은 프레임 하단에서 올라온 두 손에 의해 단단히 파지되어 지탱되고 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 아래를 향해 비스듬히 기울어져 양손에 들린 펼쳐진 지갑과 그 안의 신분증을 바라보고 있습니다.",
        "built_space": "야외 쓰레기 더미가 있는 바닥이며, 다양한 형태의 쓰레기들이 아웃포커싱되어 배경에 위치합니다.",
        "entities": "펼쳐진 가죽 지갑, 한국어 지폐(5만원권), 카드류, 투명 칸 안에 들어있는 법복 차림 남성의 사진이 있는 신분증, 흙이 묻은 두 손이 확인됩니다.",
        "hard_violations": [],
        "physics": "지갑은 양쪽 하단 모서리를 잡고 있는 두 손에 의해 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지갑이 화면 하단에 더 낮고 작게 놓이고 비스듬한 하향 시점과 쓰레기 배경이 살아 있어 지정 구도에 더 가깝지만, 지지하는 손은 프레임 안에 일부 드러난다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "투명 칸 속 신분증과 손의 지지는 명확하지만, 지갑이 화면 중앙까지 크게 올라오고 양손도 넓게 보여 하단의 작은 지갑을 강조한 지정 구도에서 더 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "펼친 지갑의 안쪽과 오른쪽 투명 칸이 위쪽 및 카메라를 향한다. 카메라는 비스듬히 내려다보며, 신분증 속 남성의 정면 얼굴은 손이나 지갑 테두리에 가려지지 않는다. 실제 인물의 눈이나 시선은 보이지 않는다.",
        "built_space": "젖은 야외 바닥 위에 금속 조각, 투명 비닐, 파란 방수포가 쌓여 있고 오른쪽 가장자리에 철망 수거함 일부가 보인다. 참고 장면의 폐기물과 재질 및 야간 분위기에 부합한다. 건물과 고정 조명은 근접 구도 밖이므로 개수나 배치는 확인할 수 없다. 지갑은 하단 중앙에 치우쳐 화면 면적의 대략 3분의 1을 차지하며 아래쪽 일부가 잘린다.",
        "entities": "낡은 검은색 계열의 펼친 지갑 한 개, 오른쪽 투명 수납 칸 한 개, 그 안의 사진 부착 신분증 한 장, 왼쪽의 여러 지폐가 보인다. 지폐에는 50000으로 보이는 숫자가 있으나 국가 표기는 판독하기 어렵다. 증명사진은 정장과 넥타이를 착용한 중년 동아시아계 남성으로 보이며, 별도 인물 참고가 없어 재판관의 정확한 외모 일치는 검증할 수 없다. 카드에는 건물형 문장과 바코드가 있으나 읽을 수 있는 추가 개인정보는 보이지 않는다. 실제 사람의 얼굴이나 몸통은 없고 하단에 손 일부만 보인다.",
        "hard_violations": [],
        "physics": "하단에서 들어온 손가락과 오른쪽 손이 지갑 아래쪽을 받치고 가장자리를 잡는다. 지갑은 공중에 무지지 상태로 떠 있지 않다. 신분증은 투명 칸에, 지폐는 왼쪽 수납부에 끼워져 있다. 주변 폐기물은 바닥과 다른 폐기물 위에 놓여 있으며, 투명 칸의 약한 반사는 사진을 가리지 않는다."
       },
       {
        "label": "B",
        "direction": "지갑 내부와 오른쪽 신분증의 사진 면이 카메라를 향한다. 사진 속 남성은 정면을 바라보며 얼굴은 가려지지 않는다. 지갑 면을 비교적 정면에 가깝게 보여 주어 A보다 비스듬한 하향 관찰의 느낌이 약하다.",
        "built_space": "배경에는 젖은 바닥, 파란 방수포, 투명 비닐, 폐금속과 상단 오른쪽의 팔레트 일부가 보인다. 참고 장소의 폐기물 재질과 어두운 조명은 유지한다. 건물의 고정 설비는 보이지 않아 검증할 수 없다. 지갑은 하단부터 화면 중간 위까지 걸쳐 약 5분의 2 안팎을 차지하며, 양손이 하단과 왼쪽 가장자리에서 크게 드러난다.",
        "entities": "펼친 가죽 지갑 한 개와 오른쪽 투명 칸 한 개, 사진 부착 신분증 한 장, 왼쪽 수납부의 지폐들이 보인다. 앞쪽 지폐에는 50000 숫자가 보이지만 정확한 발행국 표기는 확인하기 어렵다. 신분증 사진은 정장과 넥타이를 착용한 중년 동아시아계 남성으로 보인다. 저울형 문장과 바코드가 있지만 추가 개인정보는 읽히지 않는다. 인물 식별 참고가 없으므로 재판관의 구체적 얼굴 일치는 확인할 수 없다. 실제 인물의 얼굴이나 몸통은 없고 양손만 보인다.",
        "hard_violations": [],
        "physics": "왼손과 오른손의 엄지가 지갑 앞면 하단에 닿고 나머지 손가락들이 뒤와 아래를 받쳐, 열린 지갑을 안정적으로 지지한다. 신분증과 지폐도 각각 수납부에 고정되어 있다. 배경 폐기물은 바닥이나 더미에 지지된다. 투명 칸 가장자리의 반사는 물리적으로 가능하며 사진의 얼굴을 가리지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지갑이 화면 하단에 더 낮고 작게 놓이고 비스듬한 하향 시점과 쓰레기 배경이 살아 있어 지정 구도에 더 가깝지만, 지지하는 손은 프레임 안에 일부 드러난다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "투명 칸 속 신분증과 손의 지지는 명확하지만, 지갑이 화면 중앙까지 크게 올라오고 양손도 넓게 보여 하단의 작은 지갑을 강조한 지정 구도에서 더 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "펼친 지갑의 안쪽과 오른쪽 투명 칸이 위쪽 및 카메라를 향한다. 카메라는 비스듬히 내려다보며, 신분증 속 남성의 정면 얼굴은 손이나 지갑 테두리에 가려지지 않는다. 실제 인물의 눈이나 시선은 보이지 않는다.",
        "built_space": "젖은 야외 바닥 위에 금속 조각, 투명 비닐, 파란 방수포가 쌓여 있고 오른쪽 가장자리에 철망 수거함 일부가 보인다. 참고 장면의 폐기물과 재질 및 야간 분위기에 부합한다. 건물과 고정 조명은 근접 구도 밖이므로 개수나 배치는 확인할 수 없다. 지갑은 하단 중앙에 치우쳐 화면 면적의 대략 3분의 1을 차지하며 아래쪽 일부가 잘린다.",
        "entities": "낡은 검은색 계열의 펼친 지갑 한 개, 오른쪽 투명 수납 칸 한 개, 그 안의 사진 부착 신분증 한 장, 왼쪽의 여러 지폐가 보인다. 지폐에는 50000으로 보이는 숫자가 있으나 국가 표기는 판독하기 어렵다. 증명사진은 정장과 넥타이를 착용한 중년 동아시아계 남성으로 보이며, 별도 인물 참고가 없어 재판관의 정확한 외모 일치는 검증할 수 없다. 카드에는 건물형 문장과 바코드가 있으나 읽을 수 있는 추가 개인정보는 보이지 않는다. 실제 사람의 얼굴이나 몸통은 없고 하단에 손 일부만 보인다.",
        "hard_violations": [],
        "physics": "하단에서 들어온 손가락과 오른쪽 손이 지갑 아래쪽을 받치고 가장자리를 잡는다. 지갑은 공중에 무지지 상태로 떠 있지 않다. 신분증은 투명 칸에, 지폐는 왼쪽 수납부에 끼워져 있다. 주변 폐기물은 바닥과 다른 폐기물 위에 놓여 있으며, 투명 칸의 약한 반사는 사진을 가리지 않는다."
       },
       {
        "label": "A",
        "direction": "지갑 내부와 오른쪽 신분증의 사진 면이 카메라를 향한다. 사진 속 남성은 정면을 바라보며 얼굴은 가려지지 않는다. 지갑 면을 비교적 정면에 가깝게 보여 주어 A보다 비스듬한 하향 관찰의 느낌이 약하다.",
        "built_space": "배경에는 젖은 바닥, 파란 방수포, 투명 비닐, 폐금속과 상단 오른쪽의 팔레트 일부가 보인다. 참고 장소의 폐기물 재질과 어두운 조명은 유지한다. 건물의 고정 설비는 보이지 않아 검증할 수 없다. 지갑은 하단부터 화면 중간 위까지 걸쳐 약 5분의 2 안팎을 차지하며, 양손이 하단과 왼쪽 가장자리에서 크게 드러난다.",
        "entities": "펼친 가죽 지갑 한 개와 오른쪽 투명 칸 한 개, 사진 부착 신분증 한 장, 왼쪽 수납부의 지폐들이 보인다. 앞쪽 지폐에는 50000 숫자가 보이지만 정확한 발행국 표기는 확인하기 어렵다. 신분증 사진은 정장과 넥타이를 착용한 중년 동아시아계 남성으로 보인다. 저울형 문장과 바코드가 있지만 추가 개인정보는 읽히지 않는다. 인물 식별 참고가 없으므로 재판관의 구체적 얼굴 일치는 확인할 수 없다. 실제 인물의 얼굴이나 몸통은 없고 양손만 보인다.",
        "hard_violations": [],
        "physics": "왼손과 오른손의 엄지가 지갑 앞면 하단에 닿고 나머지 손가락들이 뒤와 아래를 받쳐, 열린 지갑을 안정적으로 지지한다. 신분증과 지폐도 각각 수납부에 고정되어 있다. 배경 폐기물은 바닥이나 더미에 지지된다. 투명 칸 가장자리의 반사는 물리적으로 가능하며 사진의 얼굴을 가리지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "야간의 어두운 조명과 주변 쓰레기 더미의 젖은 질감을 훌륭하게 재현했으며, 지갑을 잡은 손을 화면 하단에 바짝 붙여 크롭하여 프레이밍 지시를 매우 충실히 이행했습니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "지갑을 잡고 있는 손이 화면에 너무 많이 드러나 '프레임 바로 아래에 배치'하라는 지시에서 벗어났으며, 지갑 내부가 주변 환경의 오염도에 비해 다소 이질적으로 깨끗합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S9sh1_sel.png",
    "asset_id": "b8303a7d-6271-435a-800b-be1528495e20",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab084d-bcc7-7778-87ba-b79ab3e13e18",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S9sh8::signage": {
  "fp": "d2e00c36dbdd3f64",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S9sh8": {
  "input_fingerprint": "080aa043da452c9d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 내용물이 빈 지갑을 허공으로 무심하게 툭 던져 지갑이 손끝에서 막 떨어져 나간 순간의 이현우 측면.\n\nLOCATION (lock): On the desolate ground outside the refugee tribunal, surrounded by heaps of rubbish. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the outward lateral track at 이현우's side, using chest-height placement and a slight upward angle to show his profile, face, and throwing hand together. Keep him in the left half, with the wallet just clear of his fingertips near center-right and open space beyond it; his weight remains guarded over one leg as he looks past the discarded wallet toward the rubbish outside the right edge. Emphasize the widened camera distance and retain his face instead of panning after the wallet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Discarded wallet (Released from his hand after the money has been removed) — Seen obliquely just beyond his fingertips, without requiring its contents to face the camera; used as Create a small separation between hand and wallet that makes the discard unmistakable; Rubbish heaps (Surrounding the area outside the refugee court); used as Continue the setting beneath and behind the release gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same subdued nighttime exposure and restrained contrast used during the wallet inspection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the bleak rubbish piles and nighttime colors from the reference. Exclude the sudden awakening as a repeated event and any courtroom furnishings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has had its money removed, but the judge's identification card remains inside. Rubbish heaps surround the nighttime tribunal exterior. 이현우: Has kept the money and released the wallet. He remains alert but limps on the injured leg, with the head injury still unresolved.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 내용물이 빈 지갑을 허공으로 무심하게 툭 던져 지갑이 손끝에서 막 떨어져 나간 순간의 이현우 측면.\n\nLOCATION (lock): On the desolate ground outside the refugee tribunal, surrounded by heaps of rubbish. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the outward lateral track at 이현우's side, using chest-height placement and a slight upward angle to show his profile, face, and throwing hand together. Keep him in the left half, with the wallet just clear of his fingertips near center-right and open space beyond it; his weight remains guarded over one leg as he looks past the discarded wallet toward the rubbish outside the right edge. Emphasize the widened camera distance and retain his face instead of panning after the wallet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Discarded wallet (Released from his hand after the money has been removed) — Seen obliquely just beyond his fingertips, without requiring its contents to face the camera; used as Create a small separation between hand and wallet that makes the discard unmistakable; Rubbish heaps (Surrounding the area outside the refugee court); used as Continue the setting beneath and behind the release gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same subdued nighttime exposure and restrained contrast used during the wallet inspection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the bleak rubbish piles and nighttime colors from the reference. Exclude the sudden awakening as a repeated event and any courtroom furnishings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has had its money removed, but the judge's identification card remains inside. Rubbish heaps surround the nighttime tribunal exterior. 이현우: Has kept the money and released the wallet. He remains alert but limps on the injured leg, with the head injury still unresolved.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 내용물이 빈 지갑을 허공으로 무심하게 툭 던져 지갑이 손끝에서 막 떨어져 나간 순간의 이현우 측면.\n\nLOCATION (lock): On the desolate ground outside the refugee tribunal, surrounded by heaps of rubbish. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the outward lateral track at 이현우's side, using chest-height placement and a slight upward angle to show his profile, face, and throwing hand together. Keep him in the left half, with the wallet just clear of his fingertips near center-right and open space beyond it; his weight remains guarded over one leg as he looks past the discarded wallet toward the rubbish outside the right edge. Emphasize the widened camera distance and retain his face instead of panning after the wallet.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Discarded wallet (Released from his hand after the money has been removed) — Seen obliquely just beyond his fingertips, without requiring its contents to face the camera; used as Create a small separation between hand and wallet that makes the discard unmistakable; Rubbish heaps (Surrounding the area outside the refugee court); used as Continue the setting beneath and behind the release gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same subdued nighttime exposure and restrained contrast used during the wallet inspection.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the bleak rubbish piles and nighttime colors from the reference. Exclude the sudden awakening as a repeated event and any courtroom furnishings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The discarded wallet has had its money removed, but the judge's identification card remains inside. Rubbish heaps surround the nighttime tribunal exterior. 이현우: Has kept the money and released the wallet. He remains alert but limps on the injured leg, with the head injury still unresolved.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 우측 프레임 바깥을 응시하고 있으며, 던져진 지갑은 손끝을 막 벗어나 우측으로 향하고 있습니다.",
    "built_space": "이전 샷과 동일한 야외 공간으로, 쓰레기 더미와 낡은 건물이 올바른 위치에 배치되어 있습니다.",
    "entities": "이현우의 외모, 피 묻은 셔츠, 인이어 무전기가 레퍼런스와 일치하며, 공중에 떠 있는 검은 지갑 내부에는 신분증이 보입니다.",
    "hard_violations": [],
    "physics": "인물의 팔과 손이 지갑을 막 던진 직후의 포즈를 취하고 있으며, 지갑은 이 던지는 힘에 의해 공중에 체공 중입니다."
   },
   {
    "label": "B",
    "direction": "인물이 우측을 향해 시선을 두고 있으며, 뻗은 손끝에서 지갑이 날아가고 있습니다.",
    "built_space": "배경의 건물과 쓰레기 더미가 이전 샷과 일치하게 잘 유지되어 있습니다.",
    "entities": "이현우의 외모와 복장이 레퍼런스와 일치하지만, 날아가는 검은 지갑 내부에 신분증이 명확히 묘사되지 않았습니다.",
    "hard_violations": [],
    "physics": "손을 뻗은 동작에서 지갑이 방금 떨어져 나간 발사 순간의 물리적 상태가 설득력 있게 표현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 구도와 측면 앵글을 잘 구현했으며, 공중으로 던져진 지갑 내부에 신분증이 남아있는 디테일까지 정확하게 표현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 액션은 일치하나, 던져진 지갑 내부에 명시된 신분증이 명확히 보이지 않아 세부 묘사에서 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 프레임 바깥을 응시하고 있으며, 던져진 지갑은 손끝을 막 벗어나 우측으로 향하고 있습니다.",
        "built_space": "이전 샷과 동일한 야외 공간으로, 쓰레기 더미와 낡은 건물이 올바른 위치에 배치되어 있습니다.",
        "entities": "이현우의 외모, 피 묻은 셔츠, 인이어 무전기가 레퍼런스와 일치하며, 공중에 떠 있는 검은 지갑 내부에는 신분증이 보입니다.",
        "hard_violations": [],
        "physics": "인물의 팔과 손이 지갑을 막 던진 직후의 포즈를 취하고 있으며, 지갑은 이 던지는 힘에 의해 공중에 체공 중입니다."
       },
       {
        "label": "B",
        "direction": "인물이 우측을 향해 시선을 두고 있으며, 뻗은 손끝에서 지갑이 날아가고 있습니다.",
        "built_space": "배경의 건물과 쓰레기 더미가 이전 샷과 일치하게 잘 유지되어 있습니다.",
        "entities": "이현우의 외모와 복장이 레퍼런스와 일치하지만, 날아가는 검은 지갑 내부에 신분증이 명확히 묘사되지 않았습니다.",
        "hard_violations": [],
        "physics": "손을 뻗은 동작에서 지갑이 방금 떨어져 나간 발사 순간의 물리적 상태가 설득력 있게 표현되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 구도와 측면 앵글을 잘 구현했으며, 공중으로 던져진 지갑 내부에 신분증이 남아있는 디테일까지 정확하게 표현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 액션은 일치하나, 던져진 지갑 내부에 명시된 신분증이 명확히 보이지 않아 세부 묘사에서 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 프레임 바깥을 응시하고 있으며, 던져진 지갑은 손끝을 막 벗어나 우측으로 향하고 있습니다.",
        "built_space": "이전 샷과 동일한 야외 공간으로, 쓰레기 더미와 낡은 건물이 올바른 위치에 배치되어 있습니다.",
        "entities": "이현우의 외모, 피 묻은 셔츠, 인이어 무전기가 레퍼런스와 일치하며, 공중에 떠 있는 검은 지갑 내부에는 신분증이 보입니다.",
        "hard_violations": [],
        "physics": "인물의 팔과 손이 지갑을 막 던진 직후의 포즈를 취하고 있으며, 지갑은 이 던지는 힘에 의해 공중에 체공 중입니다."
       },
       {
        "label": "B",
        "direction": "인물이 우측을 향해 시선을 두고 있으며, 뻗은 손끝에서 지갑이 날아가고 있습니다.",
        "built_space": "배경의 건물과 쓰레기 더미가 이전 샷과 일치하게 잘 유지되어 있습니다.",
        "entities": "이현우의 외모와 복장이 레퍼런스와 일치하지만, 날아가는 검은 지갑 내부에 신분증이 명확히 묘사되지 않았습니다.",
        "hard_violations": [],
        "physics": "손을 뻗은 동작에서 지갑이 방금 떨어져 나간 발사 순간의 물리적 상태가 설득력 있게 표현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 인물의 측면과 얼굴·던지는 손을 함께 담고 지갑을 비스듬히 보여 지시된 구도에 더 충실하지만, 손끝과 지갑 간격은 막 놓친 순간치고 다소 넓다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "야간 장소와 인물·소품은 잘 이어지지만, 몸통이 카메라 쪽으로 더 열리고 지갑 내부도 정면에 가까워 측면에서 포착한 무심한 투척이라는 지시에는 A보다 덜 맞는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 오른쪽을 향하며 지갑 자체보다 그 너머 화면 밖을 본다. 화면 밖 쓰레기를 실제로 응시하는지는 확인할 수 없지만 방향은 맞는다. 뻗은 팔과 풀린 손가락의 연장선 오른쪽에 지갑이 있어 바깥으로 던지는 동작으로 읽힌다.",
        "built_space": "뒤쪽의 낡은 콘크리트 건물 한 동과 오른쪽 건물 벽이 통로를 만든다. 뒤 건물 상층에는 창군 세 곳이 드러나고 일부는 인물에게 가려진다. 하층 창과 폐기물 용기들은 팔과 쓰레기 뒤로 부분적으로 보인다. 오른쪽에는 출입구 차양 하나와 그 아래 점등된 벽등 하나가 있으며, 양쪽 쓰레기 더미와 금속 적재함이 이전 장소를 이어간다. 인물은 건물 밖 통로 왼쪽에 서 있다. 젖은 바닥의 조명 반사도 가능한 배치다.",
        "entities": "보이는 사람은 이현우 한 명뿐이다. 십대 후반으로 보이는 동아시아계 남성의 얼굴, 마른 체격, 헝클어진 짧은 검은 머리, 검은 인이어, 얼굴의 상처와 오염, 피와 흙이 묻은 어두운 셔츠·바지가 참고와 부합한다. 국적은 외관만으로 확인할 수 없다. 검은 접이식 가죽 지갑 하나가 비스듬히 열려 있으며 내부 카드 일부가 보이고 지폐는 보이지 않는다. 카드가 판사의 신분증인지는 이 크기에서 판독되지 않는다. 보관한 돈은 노출되지 않아 확인할 수 없다.",
        "hard_violations": [],
        "physics": "몸통은 세워져 있고 골반이 팔의 움직임과 반대쪽에 남아 있어 서서 가볍게 던지는 자세로 가능하다. 발과 바닥의 접점은 프레임 밖이므로 어느 다리가 체중을 받는지는 확인할 수 없다. 지갑은 손에 닿지 않지만 바로 왼쪽의 뻗은 팔과 놓인 손가락이 발사 동작을 설명하며, 오른쪽 아래 바닥과 쓰레기 쪽으로 떨어질 수 있다. 근거 없는 공중 부양은 아니다. 다만 손끝과 지갑 사이가 손 하나 정도 벌어져 있어 손에서 막 떨어진 찰나보다는 조금 뒤처럼 보인다."
       },
       {
        "label": "B",
        "direction": "얼굴과 눈은 오른쪽 지갑 너머를 향한다. 던지는 손도 오른쪽으로 열려 있고 지갑은 그 연장선에 있다. 다만 얼굴은 측면인 반면 몸통은 카메라를 향해 더 돌아와 있다. 지갑의 카드 수납면은 관객에게 상당히 직접적으로 보인다.",
        "built_space": "뒤쪽 콘크리트 건물 한 동, 오른쪽 벽, 두 건물 사이 통로가 참고의 공간을 유지한다. 상층 창군 세 곳이 드러나며 인물 뒤 구간은 가려져 있다. 하층에는 창 일부와 폐기물 용기들이 보인다. 오른쪽 출입구에는 차양 하나와 켜진 벽등 하나가 있고 통로 끝에도 작은 광원이 보인다. 금속 적재함과 쓰레기가 왼쪽 및 아래쪽에 쌓여 있다. 인물은 왼쪽 통로에 서 있으며 고정 시설의 명백한 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 한 명이며 젊은 동아시아계 남성의 얼굴과 체격, 헝클어진 검은 머리, 인이어, 머리 주변 상처, 오염된 어두운 셔츠와 바지가 참고와 대체로 일치한다. 검은 가죽 접이식 지갑 한 개에는 왼쪽 카드 슬롯과 오른쪽 사진 카드 창이 보여 소품 참고의 구조를 잘 따른다. 지폐는 보이지 않지만 신분증의 소유자와 돈의 보관 여부는 확인되지 않는다. 추가 인물이나 법정 가구는 없다.",
        "hard_violations": [],
        "physics": "인물의 몸통과 골반은 서 있는 자세로 연결되며, 한 팔은 아래로 내려오고 다른 팔은 팔꿈치를 약간 굽힌 채 오른쪽으로 펼쳐져 있다. 발은 프레임 밖이므로 부상 다리를 보호하는 체중 배분은 확정할 수 없다. 지갑의 공중 위치는 열린 손에서 오른쪽으로 던진 직후의 궤적으로 설명 가능하고 아래 통로로 낙하할 공간도 있다. 물리적으로 불가능한 부유는 아니지만 손과 지갑의 간격이 넓어 바로 손끝을 떠난 순간의 정밀도는 떨어진다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 인물의 측면과 얼굴·던지는 손을 함께 담고 지갑을 비스듬히 보여 지시된 구도에 더 충실하지만, 손끝과 지갑 간격은 막 놓친 순간치고 다소 넓다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "야간 장소와 인물·소품은 잘 이어지지만, 몸통이 카메라 쪽으로 더 열리고 지갑 내부도 정면에 가까워 측면에서 포착한 무심한 투척이라는 지시에는 A보다 덜 맞는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 오른쪽을 향하며 지갑 자체보다 그 너머 화면 밖을 본다. 화면 밖 쓰레기를 실제로 응시하는지는 확인할 수 없지만 방향은 맞는다. 뻗은 팔과 풀린 손가락의 연장선 오른쪽에 지갑이 있어 바깥으로 던지는 동작으로 읽힌다.",
        "built_space": "뒤쪽의 낡은 콘크리트 건물 한 동과 오른쪽 건물 벽이 통로를 만든다. 뒤 건물 상층에는 창군 세 곳이 드러나고 일부는 인물에게 가려진다. 하층 창과 폐기물 용기들은 팔과 쓰레기 뒤로 부분적으로 보인다. 오른쪽에는 출입구 차양 하나와 그 아래 점등된 벽등 하나가 있으며, 양쪽 쓰레기 더미와 금속 적재함이 이전 장소를 이어간다. 인물은 건물 밖 통로 왼쪽에 서 있다. 젖은 바닥의 조명 반사도 가능한 배치다.",
        "entities": "보이는 사람은 이현우 한 명뿐이다. 십대 후반으로 보이는 동아시아계 남성의 얼굴, 마른 체격, 헝클어진 짧은 검은 머리, 검은 인이어, 얼굴의 상처와 오염, 피와 흙이 묻은 어두운 셔츠·바지가 참고와 부합한다. 국적은 외관만으로 확인할 수 없다. 검은 접이식 가죽 지갑 하나가 비스듬히 열려 있으며 내부 카드 일부가 보이고 지폐는 보이지 않는다. 카드가 판사의 신분증인지는 이 크기에서 판독되지 않는다. 보관한 돈은 노출되지 않아 확인할 수 없다.",
        "hard_violations": [],
        "physics": "몸통은 세워져 있고 골반이 팔의 움직임과 반대쪽에 남아 있어 서서 가볍게 던지는 자세로 가능하다. 발과 바닥의 접점은 프레임 밖이므로 어느 다리가 체중을 받는지는 확인할 수 없다. 지갑은 손에 닿지 않지만 바로 왼쪽의 뻗은 팔과 놓인 손가락이 발사 동작을 설명하며, 오른쪽 아래 바닥과 쓰레기 쪽으로 떨어질 수 있다. 근거 없는 공중 부양은 아니다. 다만 손끝과 지갑 사이가 손 하나 정도 벌어져 있어 손에서 막 떨어진 찰나보다는 조금 뒤처럼 보인다."
       },
       {
        "label": "A",
        "direction": "얼굴과 눈은 오른쪽 지갑 너머를 향한다. 던지는 손도 오른쪽으로 열려 있고 지갑은 그 연장선에 있다. 다만 얼굴은 측면인 반면 몸통은 카메라를 향해 더 돌아와 있다. 지갑의 카드 수납면은 관객에게 상당히 직접적으로 보인다.",
        "built_space": "뒤쪽 콘크리트 건물 한 동, 오른쪽 벽, 두 건물 사이 통로가 참고의 공간을 유지한다. 상층 창군 세 곳이 드러나며 인물 뒤 구간은 가려져 있다. 하층에는 창 일부와 폐기물 용기들이 보인다. 오른쪽 출입구에는 차양 하나와 켜진 벽등 하나가 있고 통로 끝에도 작은 광원이 보인다. 금속 적재함과 쓰레기가 왼쪽 및 아래쪽에 쌓여 있다. 인물은 왼쪽 통로에 서 있으며 고정 시설의 명백한 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 한 명이며 젊은 동아시아계 남성의 얼굴과 체격, 헝클어진 검은 머리, 인이어, 머리 주변 상처, 오염된 어두운 셔츠와 바지가 참고와 대체로 일치한다. 검은 가죽 접이식 지갑 한 개에는 왼쪽 카드 슬롯과 오른쪽 사진 카드 창이 보여 소품 참고의 구조를 잘 따른다. 지폐는 보이지 않지만 신분증의 소유자와 돈의 보관 여부는 확인되지 않는다. 추가 인물이나 법정 가구는 없다.",
        "hard_violations": [],
        "physics": "인물의 몸통과 골반은 서 있는 자세로 연결되며, 한 팔은 아래로 내려오고 다른 팔은 팔꿈치를 약간 굽힌 채 오른쪽으로 펼쳐져 있다. 발은 프레임 밖이므로 부상 다리를 보호하는 체중 배분은 확정할 수 없다. 지갑의 공중 위치는 열린 손에서 오른쪽으로 던진 직후의 궤적으로 설명 가능하고 아래 통로로 낙하할 공간도 있다. 물리적으로 불가능한 부유는 아니지만 손과 지갑의 간격이 넓어 바로 손끝을 떠난 순간의 정밀도는 떨어진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 구도와 측면 앵글을 잘 구현했으며, 공중으로 던져진 지갑 내부에 신분증이 남아있는 디테일까지 정확하게 표현했습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "전반적인 구도와 액션은 일치하나, 던져진 지갑 내부에 명시된 신분증이 명확히 보이지 않아 세부 묘사에서 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S9sh1_sel.png",
    "asset_id": "b8303a7d-6271-435a-800b-be1528495e20",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 재판관의 검은 가죽 지갑: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1108579>",
    "asset_id": "035c6f34-f225-4515-ab55-77d1aebf6eeb",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0852-385a-7c47-a575-d8025c6000fa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh1"
  }
 },
 "S10sh3::signage": {
  "fp": "f3802b94c0e6a705",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::de1bac0a1ef4d930": {
  "subjects": [],
  "subject_text": "허름한 골목 식당 내부\n골목 안의 허름한 식당. 낡은 식탁과 의자가 놓여 있고, 안쪽에는 출입구에서 떨어진 외진 좌석이 있다.",
  "identity": "canonical",
  "scope_id": "L22",
  "scope_role": "location_interior",
  "scope_sha": "133e7e591da17b8c"
 },
 "S10sh3::bgfirst_bg": {
  "input_fingerprint": "039f6184166c718f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식탁 위 돈 봉투를 가운데 두고 마주 앉은 이현우와 40대 남성의 측면.\n\nLOCATION (lock): At a secluded dining table inside a shabby alley restaurant at night, under modest restaurant lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach beside the table at seated eye height, perpendicular to the line between the men, and settle into a broadside two-shot. Place 이현우 at left in right-facing profile, his forearm extended after sliding the envelope forward, and 40대 남성 at right in left-facing profile with his torso angled toward the exchange; both attend downward to the envelope between them. Keep the tabletop across the lower third and the envelope small at its center, making the completed seating position the endpoint of the approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Between the two seated men in a secluded part of the restaurant) — The near edge runs across the lower frame, with the men seated at opposite sides; used as Establish the negotiation axis and the distance separating the men; Money envelope (Just pushed across the table by 이현우) — Lying on the tabletop between the men, viewed at a shallow oblique angle; used as Anchor the exchange at the center without exaggerating the envelope's size.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued interior ambient illumination with controlled contrast, keeping both profiles and the envelope readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식탁 위 돈 봉투를 가운데 두고 마주 앉은 이현우와 40대 남성의 측면.\n\nLOCATION (lock): At a secluded dining table inside a shabby alley restaurant at night, under modest restaurant lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach beside the table at seated eye height, perpendicular to the line between the men, and settle into a broadside two-shot. Place 이현우 at left in right-facing profile, his forearm extended after sliding the envelope forward, and 40대 남성 at right in left-facing profile with his torso angled toward the exchange; both attend downward to the envelope between them. Keep the tabletop across the lower third and the envelope small at its center, making the completed seating position the endpoint of the approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Between the two seated men in a secluded part of the restaurant) — The near edge runs across the lower frame, with the men seated at opposite sides; used as Establish the negotiation axis and the distance separating the men; Money envelope (Just pushed across the table by 이현우) — Lying on the tabletop between the men, viewed at a shallow oblique angle; used as Anchor the exchange at the center without exaggerating the envelope's size.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued interior ambient illumination with controlled contrast, keeping both profiles and the envelope readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S10sh3__bgfirst_bg.png",
  "asset_id": "3e5e5b24-fcfc-47b3-ba6d-6344a0850901",
  "input_asset_ids": [
   "2c931ded-5806-4190-b2ad-631d44d09a46",
   "79dc7dbe-12c6-4708-9f00-03b8a7cc7073"
  ]
 },
 "S10sh3": {
  "input_fingerprint": "c5aeed3c4efacb5f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 돈 봉투를 가운데 두고 마주 앉은 이현우와 40대 남성의 측면.\n\nLOCATION (lock): At a secluded dining table inside a shabby alley restaurant at night, under modest restaurant lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach beside the table at seated eye height, perpendicular to the line between the men, and settle into a broadside two-shot. Place 이현우 at left in right-facing profile, his forearm extended after sliding the envelope forward, and 40대 남성 at right in left-facing profile with his torso angled toward the exchange; both attend downward to the envelope between them. Keep the tabletop across the lower third and the envelope small at its center, making the completed seating position the endpoint of the approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Between the two seated men in a secluded part of the restaurant) — The near edge runs across the lower frame, with the men seated at opposite sides; used as Establish the negotiation axis and the distance separating the men; Money envelope (Just pushed across the table by 이현우) — Lying on the tabletop between the men, viewed at a shallow oblique angle; used as Anchor the exchange at the center without exaggerating the envelope's size.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued interior ambient illumination with controlled contrast, keeping both profiles and the envelope readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A money-filled envelope lies on the table in a secluded part of the shabby restaurant at night. 이현우: Sits at the table with his dog-bite leg wound and injuries from the tribunal still unresolved. 40대 남성: Sits at the secluded restaurant table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 돈 봉투를 가운데 두고 마주 앉은 이현우와 40대 남성의 측면.\n\nLOCATION (lock): At a secluded dining table inside a shabby alley restaurant at night, under modest restaurant lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach beside the table at seated eye height, perpendicular to the line between the men, and settle into a broadside two-shot. Place 이현우 at left in right-facing profile, his forearm extended after sliding the envelope forward, and 40대 남성 at right in left-facing profile with his torso angled toward the exchange; both attend downward to the envelope between them. Keep the tabletop across the lower third and the envelope small at its center, making the completed seating position the endpoint of the approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Between the two seated men in a secluded part of the restaurant) — The near edge runs across the lower frame, with the men seated at opposite sides; used as Establish the negotiation axis and the distance separating the men; Money envelope (Just pushed across the table by 이현우) — Lying on the tabletop between the men, viewed at a shallow oblique angle; used as Anchor the exchange at the center without exaggerating the envelope's size.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued interior ambient illumination with controlled contrast, keeping both profiles and the envelope readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A money-filled envelope lies on the table in a secluded part of the shabby restaurant at night. 이현우: Sits at the table with his dog-bite leg wound and injuries from the tribunal still unresolved. 40대 남성: Sits at the secluded restaurant table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 돈 봉투를 가운데 두고 마주 앉은 이현우와 40대 남성의 측면.\n\nLOCATION (lock): At a secluded dining table inside a shabby alley restaurant at night, under modest restaurant lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach beside the table at seated eye height, perpendicular to the line between the men, and settle into a broadside two-shot. Place 이현우 at left in right-facing profile, his forearm extended after sliding the envelope forward, and 40대 남성 at right in left-facing profile with his torso angled toward the exchange; both attend downward to the envelope between them. Keep the tabletop across the lower third and the envelope small at its center, making the completed seating position the endpoint of the approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Between the two seated men in a secluded part of the restaurant) — The near edge runs across the lower frame, with the men seated at opposite sides; used as Establish the negotiation axis and the distance separating the men; Money envelope (Just pushed across the table by 이현우) — Lying on the tabletop between the men, viewed at a shallow oblique angle; used as Anchor the exchange at the center without exaggerating the envelope's size.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued interior ambient illumination with controlled contrast, keeping both profiles and the envelope readable without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A money-filled envelope lies on the table in a secluded part of the shabby restaurant at night. 이현우: Sits at the table with his dog-bite leg wound and injuries from the tribunal still unresolved. 40대 남성: Sits at the secluded restaurant table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S10sh3__bgfirst_bg.png",
     "asset_id": "3e5e5b24-fcfc-47b3-ba6d-6344a0850901",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S10sh3.png",
     "asset_id": "2c931ded-5806-4190-b2ad-631d44d09a46",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 40대 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860407>",
     "asset_id": "256a220f-4188-4285-966c-65d0e2e61307",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L22B01.png",
     "asset_id": "79dc7dbe-12c6-4708-9f00-03b8a7cc7073",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 40대 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860407>",
     "asset_id": "256a220f-4188-4285-966c-65d0e2e61307",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물의 시선이 모두 테이블 중앙에 놓인 돈 봉투를 향하고 있음.",
    "built_space": "레퍼런스와 동일한 식당 내부이며, 테이블과 의자 배치가 일치하고 두 인물 모두 정상적으로 착석함.",
    "entities": "이현우와 40대 남성의 의상과 외양은 레퍼런스와 일치하나, 이현우 목에 부자연스러운 검은 선이 그려져 있음.",
    "hard_violations": [
     "[gemini-pro] 이현우의 목과 턱선에 걸쳐 2D 낙서 같은 검은 선이 물리적으로 불가능한 형태로 피부 위에 렌더링됨."
    ],
    "physics": "두 사람 모두 의자에 안정적으로 앉아 있으며 이현우의 팔과 손은 테이블 위 봉투에 닿아 지지됨."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 시선이 테이블 중앙의 돈 봉투에 정확히 머물고 있음.",
    "built_space": "지정된 식당의 가구 배치와 배경 요소가 일치하며 인물들이 의자에 바르게 착석함.",
    "entities": "프롬프트가 요구한 두 인물의 외양, 의상, 소품(돈 봉투, 소형 인이어)이 오류 없이 묘사됨.",
    "hard_violations": [],
    "physics": "의자와 테이블에 의해 신체 및 팔이 자연스럽게 지지되고 있으며 자세의 오류가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "구도와 배경은 일치하나 이현우의 목에 낙서 같은 선이 렌더링되어 재질 현실성을 크게 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도, 인물 외양, 장소 및 소품을 사실적이고 정확하게 구현한 우수한 결과물임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 모두 테이블 중앙에 놓인 돈 봉투를 향하고 있음.",
        "built_space": "레퍼런스와 동일한 식당 내부이며, 테이블과 의자 배치가 일치하고 두 인물 모두 정상적으로 착석함.",
        "entities": "이현우와 40대 남성의 의상과 외양은 레퍼런스와 일치하나, 이현우 목에 부자연스러운 검은 선이 그려져 있음.",
        "hard_violations": [
         "이현우의 목과 턱선에 걸쳐 2D 낙서 같은 검은 선이 물리적으로 불가능한 형태로 피부 위에 렌더링됨."
        ],
        "physics": "두 사람 모두 의자에 안정적으로 앉아 있으며 이현우의 팔과 손은 테이블 위 봉투에 닿아 지지됨."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 테이블 중앙의 돈 봉투에 정확히 머물고 있음.",
        "built_space": "지정된 식당의 가구 배치와 배경 요소가 일치하며 인물들이 의자에 바르게 착석함.",
        "entities": "프롬프트가 요구한 두 인물의 외양, 의상, 소품(돈 봉투, 소형 인이어)이 오류 없이 묘사됨.",
        "hard_violations": [],
        "physics": "의자와 테이블에 의해 신체 및 팔이 자연스럽게 지지되고 있으며 자세의 오류가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "구도와 배경은 일치하나 이현우의 목에 낙서 같은 선이 렌더링되어 재질 현실성을 크게 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도, 인물 외양, 장소 및 소품을 사실적이고 정확하게 구현한 우수한 결과물임."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 모두 테이블 중앙에 놓인 돈 봉투를 향하고 있음.",
        "built_space": "레퍼런스와 동일한 식당 내부이며, 테이블과 의자 배치가 일치하고 두 인물 모두 정상적으로 착석함.",
        "entities": "이현우와 40대 남성의 의상과 외양은 레퍼런스와 일치하나, 이현우 목에 부자연스러운 검은 선이 그려져 있음.",
        "hard_violations": [
         "이현우의 목과 턱선에 걸쳐 2D 낙서 같은 검은 선이 물리적으로 불가능한 형태로 피부 위에 렌더링됨."
        ],
        "physics": "두 사람 모두 의자에 안정적으로 앉아 있으며 이현우의 팔과 손은 테이블 위 봉투에 닿아 지지됨."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 테이블 중앙의 돈 봉투에 정확히 머물고 있음.",
        "built_space": "지정된 식당의 가구 배치와 배경 요소가 일치하며 인물들이 의자에 바르게 착석함.",
        "entities": "프롬프트가 요구한 두 인물의 외양, 의상, 소품(돈 봉투, 소형 인이어)이 오류 없이 묘사됨.",
        "hard_violations": [],
        "physics": "의자와 테이블에 의해 신체 및 팔이 자연스럽게 지지되고 있으며 자세의 오류가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 이현우의 뻗은 팔과 두 사람의 봉투를 향한 하향 시선이 명확하며, 더 낮게 놓인 식탁과 상대적으로 밀착된 측면 구도가 지정된 미디엄 투숏에 더 가깝다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "봉투를 사이에 둔 측면 교환 동작과 장소는 충실하지만, 식탁이 더 높고 넓게 화면을 차지하며 이현우의 상체가 작아져 A보다 미디엄 투숏의 집중도가 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 이현우는 오른쪽을 향한 측면으로 앉아 중앙 봉투 쪽으로 눈을 내리고, 뻗은 손의 손끝도 봉투 왼쪽 끝에 닿아 있다. 오른쪽 남성은 왼쪽으로 몸을 향하고 고개와 눈을 봉투 방향으로 내린다. 두 사람 모두 카메라를 보지 않는다.",
        "built_space": "중앙 식탁 한 개의 양편에 각자 의자 하나씩을 사용한다. 뒤쪽에는 낮은 금속 수납장 한 개, 산과 호수 그림 한 개, 왼쪽 통로와 유리문 냉장고 한 개가 보인다. 왼쪽 달력과 오른쪽 메뉴 및 안내문, 벗겨진 회색 하단 벽도 장소 사진과 대응한다. 상판과 앞 모서리는 화면 하단에 놓이고, 카메라는 두 사람을 잇는 축에 거의 수직이다. 불가능한 반사나 고정 설비의 중복은 보이지 않는다.",
        "entities": "인물은 지정된 남성 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리, 앳된 동아시아계 얼굴, 마른 체격이며 참고 인물과 대체로 부합한다. 귀의 소형 인이어 장치와 피·먼지가 묻은 어두운 긴소매 셔츠가 보인다. 오른쪽은 중년 동아시아계 남성으로 참고의 얼굴과 검은 머리, 갈색 모자, 낡은 갈색 양복과 넥타이에 부합한다. 작은 갈색 봉투 하나가 식탁 중앙에 있으며 내용물은 보이지 않는다. 다리의 개 물림 상처는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 각각 의자 좌판에 놓이고 등받이는 몸 뒤에 있다. 이현우의 뻗은 팔과 손은 상판에 기대며 다른 손은 허벅지에 놓인다. 오른쪽 남성은 모은 손과 팔을 식탁 위에 안정적으로 둔다. 봉투는 상판에 얹혀 있고, 손끝의 접촉은 막 밀어 놓은 순간으로 자연스럽다. 지지 없이 떠 있는 물체나 인체는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 측면으로 고개를 숙이고 손끝과 봉투 쪽을 바라본다. 오른쪽 남성은 왼쪽을 향해 상체를 기울이며 시선을 교환 영역으로 낮추지만, A보다 고개 숙임이 약하고 정확한 주시점은 덜 뚜렷하다. 이현우의 팔은 중앙 봉투까지 뻗어 있다.",
        "built_space": "식탁 한 개와 양쪽 의자 두 개가 보이며 두 사람은 서로 반대편에 정상적으로 앉아 있다. 뒤쪽 금속 수납장 한 개, 풍경 그림 한 개, 왼쪽 냉장고와 통로, 달력, 오른쪽 메뉴판이 참고 장소의 배치를 대체로 유지한다. 벽의 낡은 재질과 수납장 위 휴지함 및 용기들도 대응한다. A보다 상판이 화면 위쪽에서 시작하고 더 넓게 보이며, 인물의 허벅지와 의자까지 포함된다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "지정된 두 남성만 등장한다. 이현우의 앳된 얼굴, 헝클어진 검은 머리, 마른 체격, 인이어 장치와 피 묻은 어두운 옷은 설정에 부합한다. 다만 참고와 달리 셔츠 소매가 팔꿈치 부근까지 걷혀 있다. 오른쪽 남성의 중년 얼굴과 갈색 모자, 구겨지고 더러운 갈색 양복은 참고에 대체로 부합한다. 중앙의 작은 종이 봉투는 적절한 크기이며 내부 돈은 확인할 수 없다. 다리 상처의 구체적 상태는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 좌판에 체중을 싣고 앉아 있으며 등받이 방향도 정상이다. 이현우의 노출된 아래팔과 손은 상판에 놓이고 손끝은 봉투 가장자리와 접촉한다. 오른쪽 남성의 모은 손과 팔도 식탁에 지지된다. 봉투는 상판에 평평하게 놓여 있어 밀기를 마친 동작으로 가능하다. 지지 없는 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 이현우의 뻗은 팔과 두 사람의 봉투를 향한 하향 시선이 명확하며, 더 낮게 놓인 식탁과 상대적으로 밀착된 측면 구도가 지정된 미디엄 투숏에 더 가깝다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "봉투를 사이에 둔 측면 교환 동작과 장소는 충실하지만, 식탁이 더 높고 넓게 화면을 차지하며 이현우의 상체가 작아져 A보다 미디엄 투숏의 집중도가 떨어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 이현우는 오른쪽을 향한 측면으로 앉아 중앙 봉투 쪽으로 눈을 내리고, 뻗은 손의 손끝도 봉투 왼쪽 끝에 닿아 있다. 오른쪽 남성은 왼쪽으로 몸을 향하고 고개와 눈을 봉투 방향으로 내린다. 두 사람 모두 카메라를 보지 않는다.",
        "built_space": "중앙 식탁 한 개의 양편에 각자 의자 하나씩을 사용한다. 뒤쪽에는 낮은 금속 수납장 한 개, 산과 호수 그림 한 개, 왼쪽 통로와 유리문 냉장고 한 개가 보인다. 왼쪽 달력과 오른쪽 메뉴 및 안내문, 벗겨진 회색 하단 벽도 장소 사진과 대응한다. 상판과 앞 모서리는 화면 하단에 놓이고, 카메라는 두 사람을 잇는 축에 거의 수직이다. 불가능한 반사나 고정 설비의 중복은 보이지 않는다.",
        "entities": "인물은 지정된 남성 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리, 앳된 동아시아계 얼굴, 마른 체격이며 참고 인물과 대체로 부합한다. 귀의 소형 인이어 장치와 피·먼지가 묻은 어두운 긴소매 셔츠가 보인다. 오른쪽은 중년 동아시아계 남성으로 참고의 얼굴과 검은 머리, 갈색 모자, 낡은 갈색 양복과 넥타이에 부합한다. 작은 갈색 봉투 하나가 식탁 중앙에 있으며 내용물은 보이지 않는다. 다리의 개 물림 상처는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 각각 의자 좌판에 놓이고 등받이는 몸 뒤에 있다. 이현우의 뻗은 팔과 손은 상판에 기대며 다른 손은 허벅지에 놓인다. 오른쪽 남성은 모은 손과 팔을 식탁 위에 안정적으로 둔다. 봉투는 상판에 얹혀 있고, 손끝의 접촉은 막 밀어 놓은 순간으로 자연스럽다. 지지 없이 떠 있는 물체나 인체는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 측면으로 고개를 숙이고 손끝과 봉투 쪽을 바라본다. 오른쪽 남성은 왼쪽을 향해 상체를 기울이며 시선을 교환 영역으로 낮추지만, A보다 고개 숙임이 약하고 정확한 주시점은 덜 뚜렷하다. 이현우의 팔은 중앙 봉투까지 뻗어 있다.",
        "built_space": "식탁 한 개와 양쪽 의자 두 개가 보이며 두 사람은 서로 반대편에 정상적으로 앉아 있다. 뒤쪽 금속 수납장 한 개, 풍경 그림 한 개, 왼쪽 냉장고와 통로, 달력, 오른쪽 메뉴판이 참고 장소의 배치를 대체로 유지한다. 벽의 낡은 재질과 수납장 위 휴지함 및 용기들도 대응한다. A보다 상판이 화면 위쪽에서 시작하고 더 넓게 보이며, 인물의 허벅지와 의자까지 포함된다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "지정된 두 남성만 등장한다. 이현우의 앳된 얼굴, 헝클어진 검은 머리, 마른 체격, 인이어 장치와 피 묻은 어두운 옷은 설정에 부합한다. 다만 참고와 달리 셔츠 소매가 팔꿈치 부근까지 걷혀 있다. 오른쪽 남성의 중년 얼굴과 갈색 모자, 구겨지고 더러운 갈색 양복은 참고에 대체로 부합한다. 중앙의 작은 종이 봉투는 적절한 크기이며 내부 돈은 확인할 수 없다. 다리 상처의 구체적 상태는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 좌판에 체중을 싣고 앉아 있으며 등받이 방향도 정상이다. 이현우의 노출된 아래팔과 손은 상판에 놓이고 손끝은 봉투 가장자리와 접촉한다. 오른쪽 남성의 모은 손과 팔도 식탁에 지지된다. 봉투는 상판에 평평하게 놓여 있어 밀기를 마친 동작으로 가능하다. 지지 없는 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.317,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.067,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 이현우의 목과 턱선에 걸쳐 2D 낙서 같은 검은 선이 물리적으로 불가능한 형태로 피부 위에 렌더링됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1067,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1067,
    "verdict_ko": "구도와 배경은 일치하나 이현우의 목에 낙서 같은 선이 렌더링되어 재질 현실성을 크게 위반함.  ★위반: [gemini-pro] 이현우의 목과 턱선에 걸쳐 2D 낙서 같은 검은 선이 물리적으로 불가능한 형태로 피부 위에 렌더링됨."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 구도, 인물 외양, 장소 및 소품을 사실적이고 정확하게 구현한 우수한 결과물임."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L22B01.png",
    "asset_id": "79dc7dbe-12c6-4708-9f00-03b8a7cc7073",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 40대 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860407>",
    "asset_id": "256a220f-4188-4285-966c-65d0e2e61307",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0856-a291-7d36-8946-fbbeaa49f3ee",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S10sh3__bgfirst_bg.png",
   "bg_asset_id": "3e5e5b24-fcfc-47b3-ba6d-6344a0850901",
   "bg_record_key": "S10sh3::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S10sh4::signage": {
  "fp": "4d913f6bdd4a3a71",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh4": {
  "input_fingerprint": "b07a1e035d573474",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈 봉투를 향해 뻗은 40대 남성의 손등 위를 이현우의 손이 강하게 덮어 쥔 근접 구도.\n\nLOCATION (lock): At the tabletop in a secluded corner of a shabby alley restaurant, lit for nighttime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Move to the near table edge on the same side of the dialogue axis and look obliquely downward, making the closer distance the emphasized change from the two-shot. 이현우's forearm enters from left and his hand clamps over the back of 40대 남성's reaching hand from right, with the envelope immediately underneath and the contact centered in the lower half. Crop both faces out, leaving their opposing forearms and small portions of their seated torsos as scale references; both men's attention remains on the contested envelope below.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Money envelope (Immediately beneath the overlapping hands) — Its upward-facing surface is partly obscured by the men's hands; used as Keep an exposed edge visible to identify what the hand contact is preventing; Dining table (Supporting the envelope and the men's reaching arms) — The tabletop is viewed obliquely downward from its near edge; used as Provide continuous spatial context around the hands rather than isolating them against an empty field.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior exposure and controlled contrast so the pressure of the overlapping hands reads without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the restaurant table, the money envelope, and the established nighttime lighting from the reference. Exclude outdoor rubbish and any furnishings belonging to the tribunal.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The envelope remains filled with the offered money at the restaurant table. 이현우: Remains seated and has taken hold of the money envelope again. His facial injuries and bitten leg remain untreated. 40대 남성: Remains seated with a hand extended over the table to take the envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈 봉투를 향해 뻗은 40대 남성의 손등 위를 이현우의 손이 강하게 덮어 쥔 근접 구도.\n\nLOCATION (lock): At the tabletop in a secluded corner of a shabby alley restaurant, lit for nighttime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Move to the near table edge on the same side of the dialogue axis and look obliquely downward, making the closer distance the emphasized change from the two-shot. 이현우's forearm enters from left and his hand clamps over the back of 40대 남성's reaching hand from right, with the envelope immediately underneath and the contact centered in the lower half. Crop both faces out, leaving their opposing forearms and small portions of their seated torsos as scale references; both men's attention remains on the contested envelope below.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Money envelope (Immediately beneath the overlapping hands) — Its upward-facing surface is partly obscured by the men's hands; used as Keep an exposed edge visible to identify what the hand contact is preventing; Dining table (Supporting the envelope and the men's reaching arms) — The tabletop is viewed obliquely downward from its near edge; used as Provide continuous spatial context around the hands rather than isolating them against an empty field.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior exposure and controlled contrast so the pressure of the overlapping hands reads without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the restaurant table, the money envelope, and the established nighttime lighting from the reference. Exclude outdoor rubbish and any furnishings belonging to the tribunal.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The envelope remains filled with the offered money at the restaurant table. 이현우: Remains seated and has taken hold of the money envelope again. His facial injuries and bitten leg remain untreated. 40대 남성: Remains seated with a hand extended over the table to take the envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 돈 봉투를 향해 뻗은 40대 남성의 손등 위를 이현우의 손이 강하게 덮어 쥔 근접 구도.\n\nLOCATION (lock): At the tabletop in a secluded corner of a shabby alley restaurant, lit for nighttime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Move to the near table edge on the same side of the dialogue axis and look obliquely downward, making the closer distance the emphasized change from the two-shot. 이현우's forearm enters from left and his hand clamps over the back of 40대 남성's reaching hand from right, with the envelope immediately underneath and the contact centered in the lower half. Crop both faces out, leaving their opposing forearms and small portions of their seated torsos as scale references; both men's attention remains on the contested envelope below.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Money envelope (Immediately beneath the overlapping hands) — Its upward-facing surface is partly obscured by the men's hands; used as Keep an exposed edge visible to identify what the hand contact is preventing; Dining table (Supporting the envelope and the men's reaching arms) — The tabletop is viewed obliquely downward from its near edge; used as Provide continuous spatial context around the hands rather than isolating them against an empty field.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding interior exposure and controlled contrast so the pressure of the overlapping hands reads without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the restaurant table, the money envelope, and the established nighttime lighting from the reference. Exclude outdoor rubbish and any furnishings belonging to the tribunal.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The envelope remains filled with the offered money at the restaurant table. 이현우: Remains seated and has taken hold of the money envelope again. His facial injuries and bitten leg remain untreated. 40대 남성: Remains seated with a hand extended over the table to take the envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — wearing: 뇌물을 받는 하급 관료의 구겨지고 때 묻은 칙칙한 브라운 계열 양복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
    "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
    "entities": "이현우의 팔은 흙먼지가 묻은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복과 흰 소매를 보여주나, 돈 봉투는 납작하지 않고 부풀어 오른 종이봉투 형태임.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 뻗은 오른손 검지손가락이 관절 없이 기형적으로 길게 늘어나 있음)"
    ],
    "physics": "손은 돈 봉투와 테이블에 의해 지지되며, 인물의 팔은 화면 밖 몸통에서 자연스럽게 뻗어 나옴."
   },
   {
    "label": "B",
    "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
    "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
    "entities": "이현우의 팔은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복을 입고 있으며, 돈 봉투는 레퍼런스와 유사한 납작한 황갈색 형태임.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)"
    ],
    "physics": "손은 돈 봉투와 테이블 표면에 의해 지지되며, 팔의 무게 중심이 테이블 쪽으로 향해 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "돈 봉투의 형태와 질감이 레퍼런스에 부합하고 전반적인 구도를 잘 따랐으나, 두 인물 모두 오른팔에 왼손이 달려 있는 심각한 해부학적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 근접 구도는 훌륭하게 구현되었으나, 40대 남성의 뻗은 손가락이 기형적으로 뭉개지고 양측 모두 팔과 손의 좌우가 맞지 않는 치명적인 인체 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
        "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
        "entities": "이현우의 팔은 흙먼지가 묻은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복과 흰 소매를 보여주나, 돈 봉투는 납작하지 않고 부풀어 오른 종이봉투 형태임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 뻗은 오른손 검지손가락이 관절 없이 기형적으로 길게 늘어나 있음)"
        ],
        "physics": "손은 돈 봉투와 테이블에 의해 지지되며, 인물의 팔은 화면 밖 몸통에서 자연스럽게 뻗어 나옴."
       },
       {
        "label": "B",
        "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
        "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
        "entities": "이현우의 팔은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복을 입고 있으며, 돈 봉투는 레퍼런스와 유사한 납작한 황갈색 형태임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)"
        ],
        "physics": "손은 돈 봉투와 테이블 표면에 의해 지지되며, 팔의 무게 중심이 테이블 쪽으로 향해 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "돈 봉투의 형태와 질감이 레퍼런스에 부합하고 전반적인 구도를 잘 따랐으나, 두 인물 모두 오른팔에 왼손이 달려 있는 심각한 해부학적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 근접 구도는 훌륭하게 구현되었으나, 40대 남성의 뻗은 손가락이 기형적으로 뭉개지고 양측 모두 팔과 손의 좌우가 맞지 않는 치명적인 인체 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
        "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
        "entities": "이현우의 팔은 흙먼지가 묻은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복과 흰 소매를 보여주나, 돈 봉투는 납작하지 않고 부풀어 오른 종이봉투 형태임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 뻗은 오른손 검지손가락이 관절 없이 기형적으로 길게 늘어나 있음)"
        ],
        "physics": "손은 돈 봉투와 테이블에 의해 지지되며, 인물의 팔은 화면 밖 몸통에서 자연스럽게 뻗어 나옴."
       },
       {
        "label": "B",
        "direction": "이현우의 손은 40대 남성의 손등을 향해 뻗어 덮고 있으며, 40대 남성의 손은 그 아래의 돈 봉투를 향해 있음.",
        "built_space": "나무 테이블의 가까운 가장자리에서 비스듬히 내려다보는 시점이며, 두 인물의 위치가 기존 대화 축과 일치함.",
        "entities": "이현우의 팔은 어두운 셔츠를 입고 있고, 40대 남성의 팔은 갈색 양복을 입고 있으며, 돈 봉투는 레퍼런스와 유사한 납작한 황갈색 형태임.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
         "물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)"
        ],
        "physics": "손은 돈 봉투와 테이블 표면에 의해 지지되며, 팔의 무게 중심이 테이블 쪽으로 향해 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "좌우에서 뻗은 손과 봉투의 관계는 맞지만, 접촉부가 화면 중앙 높이에 걸리고 손을 강하게 움켜쥐기보다 덮어 누르는 모습이라 B보다 핵심 순간의 재현이 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 근접 구도에서 하단 중앙의 손등을 굽힌 손가락으로 붙잡고, 바로 아래 돈 봉투의 가장자리를 남겨 제지 동작을 더 정확히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 팔이 왼쪽에서 오른쪽으로 들어와, 오른쪽에서 봉투를 향해 뻗은 남성의 손등을 덮는다. 두 손의 목표는 같은 봉투로 일치한다. 이현우의 손가락은 비교적 펴져 있어 움켜쥐기보다는 누르기에 가깝다. 얼굴은 모두 잘려 시선 자체는 확인할 수 없다.",
        "built_space": "마모된 목재 식탁 한 개와 그 위 종이봉투 한 개가 보인다. 두 사람의 몸통 일부는 각각 좌우에 있고, 카메라는 가까운 식탁 가장자리 위에서 비스듬히 내려다본다. 식탁 앞 모서리까지 포함되어 접촉부 주변의 상판 여백이 비교적 넓다. 배경의 낡은 벽과 어두운 실내 조명은 이전 장면과 양립하며, 중복된 설비나 거울 반사는 없다.",
        "entities": "왼쪽에는 피와 먼지가 묻은 어두운 셔츠 소매와 비교적 매끈한 손, 오른쪽에는 때 묻은 갈색 양복과 밝은 셔츠 소맷단, 더 주름진 손이 보인다. 지정된 두 인물의 의상과 상대적 연령 표현에 부합한다. 얼굴·머리·인이어는 구도 밖이어서 신원 세부는 검증할 수 없다. 봉투는 참고처럼 갈색 종이 재질이고 약간 부풀어 있으나 내용물은 보이지 않는다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "봉투는 식탁 위에 놓여 있고, 아래쪽 남성의 손은 봉투와 상판에 지지된다. 이현우의 손바닥은 그 손등에 접촉하며 팔은 몸통으로 이어진다. 오른쪽 남성의 다른 손도 식탁에 놓여 있어 세 번째 손이 추가 인물을 뜻하지 않는다. 손의 겹침과 지지는 가능하며 공중에 뜬 물체는 없다."
       },
       {
        "label": "B",
        "direction": "오른쪽 남성의 손이 왼쪽 아래의 봉투를 향해 뻗고, 왼쪽에서 들어온 이현우의 손이 그 손등 위를 덮어 손가락을 굽혀 붙잡는다. 봉투를 가져가려는 손을 막는 방향 관계가 분명하다. 얼굴은 모두 프레임 밖이므로 봉투를 보는 시선은 직접 확인되지 않는다.",
        "built_space": "마모된 목재 식탁 한 개 위에 봉투 한 개가 있으며 두 사람은 좌우에 자리한다. 가까운 식탁 쪽에서 비스듬히 내려다보는 근접 구도이고, 접촉부는 하단 중앙에 놓인다. 상판은 손 주변으로 연속해서 보이며 배경에는 낡은 벽과 수납장 일부가 흐리게 남는다. 기존의 어두운 실내 노출을 유지하고, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "왼쪽의 오염되고 핏자국 있는 어두운 셔츠, 오른쪽의 낡은 갈색 양복과 밝은 소맷단은 참고 의상과 맞는다. 왼쪽 손은 상대적으로 젊고 가늘며 오른쪽 손은 더 굵고 주름져 두 역할과 양립한다. 정확한 얼굴·민족적 정체성·머리 모양은 의도된 크롭으로 확인할 수 없다. 손 바로 아래에는 두께가 있는 갈색 종이봉투가 있고 노출된 가장자리로 식별된다. 돈 자체는 가려져 있다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "봉투는 상판에 안정적으로 놓이고, 남성의 뻗은 손은 봉투 위에 닿아 있다. 이현우의 손바닥과 굽힌 손가락이 손등에 밀착하여 아래로 누르고 붙잡는 동작을 만들며, 양쪽 팔은 각자의 소매와 몸통으로 자연스럽게 이어진다. 남성의 다른 손은 뒤쪽 상판에 지지된다. 지지 없는 신체나 물체, 불가능한 관절 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "좌우에서 뻗은 손과 봉투의 관계는 맞지만, 접촉부가 화면 중앙 높이에 걸리고 손을 강하게 움켜쥐기보다 덮어 누르는 모습이라 B보다 핵심 순간의 재현이 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 근접 구도에서 하단 중앙의 손등을 굽힌 손가락으로 붙잡고, 바로 아래 돈 봉투의 가장자리를 남겨 제지 동작을 더 정확히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 팔이 왼쪽에서 오른쪽으로 들어와, 오른쪽에서 봉투를 향해 뻗은 남성의 손등을 덮는다. 두 손의 목표는 같은 봉투로 일치한다. 이현우의 손가락은 비교적 펴져 있어 움켜쥐기보다는 누르기에 가깝다. 얼굴은 모두 잘려 시선 자체는 확인할 수 없다.",
        "built_space": "마모된 목재 식탁 한 개와 그 위 종이봉투 한 개가 보인다. 두 사람의 몸통 일부는 각각 좌우에 있고, 카메라는 가까운 식탁 가장자리 위에서 비스듬히 내려다본다. 식탁 앞 모서리까지 포함되어 접촉부 주변의 상판 여백이 비교적 넓다. 배경의 낡은 벽과 어두운 실내 조명은 이전 장면과 양립하며, 중복된 설비나 거울 반사는 없다.",
        "entities": "왼쪽에는 피와 먼지가 묻은 어두운 셔츠 소매와 비교적 매끈한 손, 오른쪽에는 때 묻은 갈색 양복과 밝은 셔츠 소맷단, 더 주름진 손이 보인다. 지정된 두 인물의 의상과 상대적 연령 표현에 부합한다. 얼굴·머리·인이어는 구도 밖이어서 신원 세부는 검증할 수 없다. 봉투는 참고처럼 갈색 종이 재질이고 약간 부풀어 있으나 내용물은 보이지 않는다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "봉투는 식탁 위에 놓여 있고, 아래쪽 남성의 손은 봉투와 상판에 지지된다. 이현우의 손바닥은 그 손등에 접촉하며 팔은 몸통으로 이어진다. 오른쪽 남성의 다른 손도 식탁에 놓여 있어 세 번째 손이 추가 인물을 뜻하지 않는다. 손의 겹침과 지지는 가능하며 공중에 뜬 물체는 없다."
       },
       {
        "label": "A",
        "direction": "오른쪽 남성의 손이 왼쪽 아래의 봉투를 향해 뻗고, 왼쪽에서 들어온 이현우의 손이 그 손등 위를 덮어 손가락을 굽혀 붙잡는다. 봉투를 가져가려는 손을 막는 방향 관계가 분명하다. 얼굴은 모두 프레임 밖이므로 봉투를 보는 시선은 직접 확인되지 않는다.",
        "built_space": "마모된 목재 식탁 한 개 위에 봉투 한 개가 있으며 두 사람은 좌우에 자리한다. 가까운 식탁 쪽에서 비스듬히 내려다보는 근접 구도이고, 접촉부는 하단 중앙에 놓인다. 상판은 손 주변으로 연속해서 보이며 배경에는 낡은 벽과 수납장 일부가 흐리게 남는다. 기존의 어두운 실내 노출을 유지하고, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "왼쪽의 오염되고 핏자국 있는 어두운 셔츠, 오른쪽의 낡은 갈색 양복과 밝은 소맷단은 참고 의상과 맞는다. 왼쪽 손은 상대적으로 젊고 가늘며 오른쪽 손은 더 굵고 주름져 두 역할과 양립한다. 정확한 얼굴·민족적 정체성·머리 모양은 의도된 크롭으로 확인할 수 없다. 손 바로 아래에는 두께가 있는 갈색 종이봉투가 있고 노출된 가장자리로 식별된다. 돈 자체는 가려져 있다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "봉투는 상판에 안정적으로 놓이고, 남성의 뻗은 손은 봉투 위에 닿아 있다. 이현우의 손바닥과 굽힌 손가락이 손등에 밀착하여 아래로 누르고 붙잡는 동작을 만들며, 양쪽 팔은 각자의 소매와 몸통으로 자연스럽게 이어진다. 남성의 다른 손은 뒤쪽 상판에 지지된다. 지지 없는 신체나 물체, 불가능한 관절 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.889
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.639
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 뻗은 오른손 검지손가락이 관절 없이 기형적으로 길게 늘어나 있음)"
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1639,
   "A": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1639,
    "verdict_ko": "돈 봉투의 형태와 질감이 레퍼런스에 부합하고 전반적인 구도를 잘 따랐으나, 두 인물 모두 오른팔에 왼손이 달려 있는 심각한 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음) / [gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음)"
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "지정된 근접 구도는 훌륭하게 구현되었으나, 40대 남성의 뻗은 손가락이 기형적으로 뭉개지고 양측 모두 팔과 손의 좌우가 맞지 않는 치명적인 인체 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (이현우의 화면 앞쪽 오른팔에 새끼손가락이 카메라 쪽을 향하는 왼손이 달려 있음) / [gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 배경 쪽 왼팔에 엄지가 카메라 쪽을 향하는 오른손이 달려 있음) / [gemini-pro] 물리적으로 불가능한 해부학 (40대 남성의 뻗은 오른손 검지손가락이 관절 없이 기형적으로 길게 늘어나 있음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S10sh3_sel.png",
    "asset_id": "c5f7fb28-2ec6-414f-890e-1ac0e7b88621",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 40대 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860407>",
    "asset_id": "256a220f-4188-4285-966c-65d0e2e61307",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab085c-9ea4-7b87-9a37-7efca7a1231d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh3"
  }
 },
 "S10sh6::signage": {
  "fp": "1e59a50150f880e3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh6": {
  "input_fingerprint": "8b947b7cf05ffcf2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 남성을 향해 목에 핏대를 세운 채 강하게 소리치듯 입을 크게 벌린 이현우의 분노한 얼굴.\n\nLOCATION (lock): At the secluded table inside the shabby alley restaurant, with modest nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle just behind and outside 40대 남성's shoulder at 이현우's seated eye height, retaining the established side of the dialogue axis and an oblique three-quarter view. Keep 이현우's face and taut neck across the center-left, his mouth wide in protest and his torso pitched forward, while only a narrow shoulder edge of 40대 남성 enters at right. 이현우 looks toward the man's face just beyond that edge, not into the lens, and the man remains turned toward him; emphasize the raised line of attention from the disputed money to the opponent.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Restaurant interior (The same secluded seating area); used as Leave only a softly resolved background behind 이현우, with no newly introduced furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the restaurant's subdued ambient illumination, allowing controlled facial modeling to carry the anger without an added dramatic light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same restaurant interior, table finish, and nighttime lighting from the reference. Exclude the earlier hand-over-hand struggle over the envelope as a repeated moment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope has been opened for inspection, with the money still inside. 이현우: Remains seated at the table, visibly angry, with his facial injuries and dog-bitten leg unchanged. 40대 남성: Remains seated with possession of the money envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 남성을 향해 목에 핏대를 세운 채 강하게 소리치듯 입을 크게 벌린 이현우의 분노한 얼굴.\n\nLOCATION (lock): At the secluded table inside the shabby alley restaurant, with modest nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle just behind and outside 40대 남성's shoulder at 이현우's seated eye height, retaining the established side of the dialogue axis and an oblique three-quarter view. Keep 이현우's face and taut neck across the center-left, his mouth wide in protest and his torso pitched forward, while only a narrow shoulder edge of 40대 남성 enters at right. 이현우 looks toward the man's face just beyond that edge, not into the lens, and the man remains turned toward him; emphasize the raised line of attention from the disputed money to the opponent.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Restaurant interior (The same secluded seating area); used as Leave only a softly resolved background behind 이현우, with no newly introduced furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the restaurant's subdued ambient illumination, allowing controlled facial modeling to carry the anger without an added dramatic light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same restaurant interior, table finish, and nighttime lighting from the reference. Exclude the earlier hand-over-hand struggle over the envelope as a repeated moment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope has been opened for inspection, with the money still inside. 이현우: Remains seated at the table, visibly angry, with his facial injuries and dog-bitten leg unchanged. 40대 남성: Remains seated with possession of the money envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 40대 남성을 향해 목에 핏대를 세운 채 강하게 소리치듯 입을 크게 벌린 이현우의 분노한 얼굴.\n\nLOCATION (lock): At the secluded table inside the shabby alley restaurant, with modest nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle just behind and outside 40대 남성's shoulder at 이현우's seated eye height, retaining the established side of the dialogue axis and an oblique three-quarter view. Keep 이현우's face and taut neck across the center-left, his mouth wide in protest and his torso pitched forward, while only a narrow shoulder edge of 40대 남성 enters at right. 이현우 looks toward the man's face just beyond that edge, not into the lens, and the man remains turned toward him; emphasize the raised line of attention from the disputed money to the opponent.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Restaurant interior (The same secluded seating area); used as Leave only a softly resolved background behind 이현우, with no newly introduced furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the restaurant's subdued ambient illumination, allowing controlled facial modeling to carry the anger without an added dramatic light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same restaurant interior, table finish, and nighttime lighting from the reference. Exclude the earlier hand-over-hand struggle over the envelope as a repeated moment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The money envelope has been opened for inspection, with the money still inside. 이현우: Remains seated at the table, visibly angry, with his facial injuries and dog-bitten leg unchanged. 40대 남성: Remains seated with possession of the money envelope.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 40대 남성 (한국인 남성, 40대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 화면 우측에 걸친 40대 남성의 얼굴 쪽을 강하게 향하고 있음.",
    "built_space": "우측 남성의 어깨 뒤에서 좌측을 바라보는 앵글에 맞춰, 레퍼런스의 좌측 배경(벽걸이 달력, 복도 입구, 기둥)이 이현우 뒤로 정확하게 렌더링됨.",
    "entities": "이현우(피와 흙이 묻은 셔츠, 인이어 무전기, 분노하여 입을 벌린 표정)와 40대 남성(중절모, 재킷 뒷모습) 모두 프롬프트와 레퍼런스에 일치함.",
    "hard_violations": [],
    "physics": "이현우가 앞쪽으로 상체를 강하게 기울인 채 테이블과 의자에 의지하여 안정적으로 앉아 있음."
   },
   {
    "label": "B",
    "direction": "이현우가 우측의 40대 남성을 향해 시선을 고정하고 소리침.",
    "built_space": "레퍼런스 샷에서 우측 남성의 등 뒤 벽면에 걸려 있던 풍경화 액자가, 카메라 앵글상 보이지 않아야 할 좌측 이현우의 등 뒤 배경으로 잘못 이동되어 나타남.",
    "entities": "이현우(인이어, 피 묻은 셔츠, 크게 벌린 입)와 40대 남성(중절모, 재킷)의 외형 및 복장 일치함.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 고정된 구조물(풍경화 액자)의 위치가 반대편 벽면으로 임의 이동하여 공간 연속성(LOCATION lock)을 심각하게 위반함."
    ],
    "physics": "이현우가 몸을 앞으로 숙이고 앉아 있으며 지지 상태에 이상 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "카메라 앵글 변화에 따른 배경의 공간적 연속성(달력, 복도 위치)을 완벽히 구현했으며, 이현우의 분노한 표정 연기를 지시대로 잘 표현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "표정 연기는 우수하나, 레퍼런스에서 우측 남성 뒤에 있던 풍경화가 이현우 뒤로 순간이동하는 치명적인 공간 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 우측에 걸친 40대 남성의 얼굴 쪽을 강하게 향하고 있음.",
        "built_space": "우측 남성의 어깨 뒤에서 좌측을 바라보는 앵글에 맞춰, 레퍼런스의 좌측 배경(벽걸이 달력, 복도 입구, 기둥)이 이현우 뒤로 정확하게 렌더링됨.",
        "entities": "이현우(피와 흙이 묻은 셔츠, 인이어 무전기, 분노하여 입을 벌린 표정)와 40대 남성(중절모, 재킷 뒷모습) 모두 프롬프트와 레퍼런스에 일치함.",
        "hard_violations": [],
        "physics": "이현우가 앞쪽으로 상체를 강하게 기울인 채 테이블과 의자에 의지하여 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 우측의 40대 남성을 향해 시선을 고정하고 소리침.",
        "built_space": "레퍼런스 샷에서 우측 남성의 등 뒤 벽면에 걸려 있던 풍경화 액자가, 카메라 앵글상 보이지 않아야 할 좌측 이현우의 등 뒤 배경으로 잘못 이동되어 나타남.",
        "entities": "이현우(인이어, 피 묻은 셔츠, 크게 벌린 입)와 40대 남성(중절모, 재킷)의 외형 및 복장 일치함.",
        "hard_violations": [
         "레퍼런스에 고정된 구조물(풍경화 액자)의 위치가 반대편 벽면으로 임의 이동하여 공간 연속성(LOCATION lock)을 심각하게 위반함."
        ],
        "physics": "이현우가 몸을 앞으로 숙이고 앉아 있으며 지지 상태에 이상 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "카메라 앵글 변화에 따른 배경의 공간적 연속성(달력, 복도 위치)을 완벽히 구현했으며, 이현우의 분노한 표정 연기를 지시대로 잘 표현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "표정 연기는 우수하나, 레퍼런스에서 우측 남성 뒤에 있던 풍경화가 이현우 뒤로 순간이동하는 치명적인 공간 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 우측에 걸친 40대 남성의 얼굴 쪽을 강하게 향하고 있음.",
        "built_space": "우측 남성의 어깨 뒤에서 좌측을 바라보는 앵글에 맞춰, 레퍼런스의 좌측 배경(벽걸이 달력, 복도 입구, 기둥)이 이현우 뒤로 정확하게 렌더링됨.",
        "entities": "이현우(피와 흙이 묻은 셔츠, 인이어 무전기, 분노하여 입을 벌린 표정)와 40대 남성(중절모, 재킷 뒷모습) 모두 프롬프트와 레퍼런스에 일치함.",
        "hard_violations": [],
        "physics": "이현우가 앞쪽으로 상체를 강하게 기울인 채 테이블과 의자에 의지하여 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 우측의 40대 남성을 향해 시선을 고정하고 소리침.",
        "built_space": "레퍼런스 샷에서 우측 남성의 등 뒤 벽면에 걸려 있던 풍경화 액자가, 카메라 앵글상 보이지 않아야 할 좌측 이현우의 등 뒤 배경으로 잘못 이동되어 나타남.",
        "entities": "이현우(인이어, 피 묻은 셔츠, 크게 벌린 입)와 40대 남성(중절모, 재킷)의 외형 및 복장 일치함.",
        "hard_violations": [
         "레퍼런스에 고정된 구조물(풍경화 액자)의 위치가 반대편 벽면으로 임의 이동하여 공간 연속성(LOCATION lock)을 심각하게 위반함."
        ],
        "physics": "이현우가 몸을 앞으로 숙이고 앉아 있으며 지지 상태에 이상 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "얼굴과 팽팽한 목을 더 크게 잡아 분노의 클로즈업에 가깝지만, 오른쪽에 어깨 가장자리만 보여야 한다는 지시와 달리 상대의 머리와 등까지 크게 드러난다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "상대를 향한 고함과 식당의 재질·조명은 유지하지만, 상반신과 탁자까지 담은 넓은 구도와 크게 들어온 상대의 머리·어깨가 지정된 얼굴 클로즈업에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 눈과 벌어진 입은 화면 오른쪽 남성의 얼굴을 향하며 렌즈를 보지 않는다. 남성도 이현우 쪽으로 고개를 돌리고 있다. 돈은 화면 밖이므로 돈에서 상대에게 시선이 올라간 과정 자체는 확인할 수 없다.",
        "built_space": "왼쪽에 출입구와 흐릿한 냉장고 하나, 위쪽에 조명 하나, 오른쪽 벽에 풍경 그림 하나가 보인다. 오른쪽 아래에는 수납장 일부와 붉은 뚜껑 용기 하나, 금속 원통 하나가 보이며, 기존 식당의 낡은 회벽과 회청색 하단 벽을 유지한다. 두 인물 사이로 의자 등받이 일부가 보인다. 고정 시설의 명백한 중복이나 불가능한 반사는 없다. 얼굴과 목은 중앙 왼쪽에 있지만 상대의 머리와 넓은 등까지 들어와 좁은 어깨 가장자리라는 구도 조건을 벗어난다.",
        "entities": "이현우는 한국계의 10대 후반으로 보이는 젊은 남성으로, 헝클어진 검은 머리, 마른 체격, 검은 인이어, 볼의 상처, 피와 먼지가 묻은 어두운 셔츠가 참조와 대체로 맞는다. 상대는 짧은 검은 머리와 이전 장면의 갈색 모자·재킷을 유지하지만 얼굴이 대부분 가려져 정확한 나이와 얼굴 일치 여부는 제한적으로만 판단된다. 추가 인물은 없다. 봉투와 돈, 다리 부상은 화면 밖이므로 상태를 판정하지 않는다.",
        "hard_violations": [],
        "physics": "이현우는 몸통을 앞으로 숙이고 목 근육을 긴장시켜 입을 크게 벌린 자연스러운 고함 자세다. 엉덩이와 좌판 접촉은 잘려 있지만 의자 등받이 일부와 몸통 배치는 착석 상태와 양립한다. 상대의 하체와 지지점도 화면 밖이다. 지지 없이 떠 있는 신체나 물체, 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 남성의 얼굴을 똑바로 보며 입을 크게 벌리고 있고, 남성 역시 그를 향해 돌아앉아 있다. 시선은 카메라를 향하지 않는다. 봉투가 보이지 않아 돈에서 상대 얼굴로 주의가 옮겨가는 관계는 표정과 고개 방향으로만 암시된다.",
        "built_space": "왼쪽 벽의 달력 하나, 그 아래 작업대와 금속 용기 하나, 중앙 뒤의 출입구 하나와 흐릿한 냉장고 하나, 상단 조명 하나가 보인다. 오른쪽에는 낡은 기둥과 세로 배관 하나가 있으며, 하단에는 나무 탁자 한 개의 가장자리가 들어온다. 기존 식당의 재질과 시설에 대응하며 명백한 중복이나 불가능한 반사는 없다. 다만 이현우의 상반신과 탁자를 더 넓게 보여 주고 상대의 머리와 어깨가 오른쪽을 크게 차지해 지정된 클로즈업보다 넓다.",
        "entities": "이현우의 젊은 한국계 남성 외모, 짧고 흐트러진 검은 머리, 마른 체격, 인이어, 얼굴 상처와 오염된 어두운 셔츠가 참조에 대체로 부합한다. 남성은 이전 장면의 갈색 모자와 갈색 재킷을 유지하며 옆머리와 귀, 볼 일부만 보여 얼굴 동일성은 충분히 확인할 수 없다. 두 사람 외의 인물은 없다. 돈봉투와 돈, 바지와 다리 부상은 구도 밖이다.",
        "hard_violations": [],
        "physics": "이현우의 몸통은 탁자 쪽으로 기울어 있고 목의 긴장과 턱의 벌어짐은 실제로 소리치는 동작으로 가능하다. 왼쪽 뒤에 의자 일부가 보이지만 좌판과 엉덩이 접촉은 화면 밖이다. 팔은 아래로 이어져 잘리며 탁자를 짚는 손은 확인되지 않는다. 그렇다고 몸이 떠 있는 정황은 없으며 두 사람의 착석 자세와 충돌하는 물리적 이상도 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "얼굴과 팽팽한 목을 더 크게 잡아 분노의 클로즈업에 가깝지만, 오른쪽에 어깨 가장자리만 보여야 한다는 지시와 달리 상대의 머리와 등까지 크게 드러난다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "상대를 향한 고함과 식당의 재질·조명은 유지하지만, 상반신과 탁자까지 담은 넓은 구도와 크게 들어온 상대의 머리·어깨가 지정된 얼굴 클로즈업에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 눈과 벌어진 입은 화면 오른쪽 남성의 얼굴을 향하며 렌즈를 보지 않는다. 남성도 이현우 쪽으로 고개를 돌리고 있다. 돈은 화면 밖이므로 돈에서 상대에게 시선이 올라간 과정 자체는 확인할 수 없다.",
        "built_space": "왼쪽에 출입구와 흐릿한 냉장고 하나, 위쪽에 조명 하나, 오른쪽 벽에 풍경 그림 하나가 보인다. 오른쪽 아래에는 수납장 일부와 붉은 뚜껑 용기 하나, 금속 원통 하나가 보이며, 기존 식당의 낡은 회벽과 회청색 하단 벽을 유지한다. 두 인물 사이로 의자 등받이 일부가 보인다. 고정 시설의 명백한 중복이나 불가능한 반사는 없다. 얼굴과 목은 중앙 왼쪽에 있지만 상대의 머리와 넓은 등까지 들어와 좁은 어깨 가장자리라는 구도 조건을 벗어난다.",
        "entities": "이현우는 한국계의 10대 후반으로 보이는 젊은 남성으로, 헝클어진 검은 머리, 마른 체격, 검은 인이어, 볼의 상처, 피와 먼지가 묻은 어두운 셔츠가 참조와 대체로 맞는다. 상대는 짧은 검은 머리와 이전 장면의 갈색 모자·재킷을 유지하지만 얼굴이 대부분 가려져 정확한 나이와 얼굴 일치 여부는 제한적으로만 판단된다. 추가 인물은 없다. 봉투와 돈, 다리 부상은 화면 밖이므로 상태를 판정하지 않는다.",
        "hard_violations": [],
        "physics": "이현우는 몸통을 앞으로 숙이고 목 근육을 긴장시켜 입을 크게 벌린 자연스러운 고함 자세다. 엉덩이와 좌판 접촉은 잘려 있지만 의자 등받이 일부와 몸통 배치는 착석 상태와 양립한다. 상대의 하체와 지지점도 화면 밖이다. 지지 없이 떠 있는 신체나 물체, 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 남성의 얼굴을 똑바로 보며 입을 크게 벌리고 있고, 남성 역시 그를 향해 돌아앉아 있다. 시선은 카메라를 향하지 않는다. 봉투가 보이지 않아 돈에서 상대 얼굴로 주의가 옮겨가는 관계는 표정과 고개 방향으로만 암시된다.",
        "built_space": "왼쪽 벽의 달력 하나, 그 아래 작업대와 금속 용기 하나, 중앙 뒤의 출입구 하나와 흐릿한 냉장고 하나, 상단 조명 하나가 보인다. 오른쪽에는 낡은 기둥과 세로 배관 하나가 있으며, 하단에는 나무 탁자 한 개의 가장자리가 들어온다. 기존 식당의 재질과 시설에 대응하며 명백한 중복이나 불가능한 반사는 없다. 다만 이현우의 상반신과 탁자를 더 넓게 보여 주고 상대의 머리와 어깨가 오른쪽을 크게 차지해 지정된 클로즈업보다 넓다.",
        "entities": "이현우의 젊은 한국계 남성 외모, 짧고 흐트러진 검은 머리, 마른 체격, 인이어, 얼굴 상처와 오염된 어두운 셔츠가 참조에 대체로 부합한다. 남성은 이전 장면의 갈색 모자와 갈색 재킷을 유지하며 옆머리와 귀, 볼 일부만 보여 얼굴 동일성은 충분히 확인할 수 없다. 두 사람 외의 인물은 없다. 돈봉투와 돈, 바지와 다리 부상은 구도 밖이다.",
        "hard_violations": [],
        "physics": "이현우의 몸통은 탁자 쪽으로 기울어 있고 목의 긴장과 턱의 벌어짐은 실제로 소리치는 동작으로 가능하다. 왼쪽 뒤에 의자 일부가 보이지만 좌판과 엉덩이 접촉은 화면 밖이다. 팔은 아래로 이어져 잘리며 탁자를 짚는 손은 확인되지 않는다. 그렇다고 몸이 떠 있는 정황은 없으며 두 사람의 착석 자세와 충돌하는 물리적 이상도 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.375
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.125
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스에 고정된 구조물(풍경화 액자)의 위치가 반대편 벽면으로 임의 이동하여 공간 연속성(LOCATION lock)을 심각하게 위반함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1125
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "카메라 앵글 변화에 따른 배경의 공간적 연속성(달력, 복도 위치)을 완벽히 구현했으며, 이현우의 분노한 표정 연기를 지시대로 잘 표현했습니다."
   },
   {
    "label": "B",
    "score": 1125,
    "verdict_ko": "표정 연기는 우수하나, 레퍼런스에서 우측 남성 뒤에 있던 풍경화가 이현우 뒤로 순간이동하는 치명적인 공간 오류가 발생했습니다.  ★위반: [gemini-pro] 레퍼런스에 고정된 구조물(풍경화 액자)의 위치가 반대편 벽면으로 임의 이동하여 공간 연속성(LOCATION lock)을 심각하게 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 40대 남성, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S10sh3_sel.png",
    "asset_id": "c5f7fb28-2ec6-414f-890e-1ac0e7b88621",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 40대 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1231201>",
    "asset_id": "7d739b40-eb04-4934-8e26-d981d89a7320",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0863-c6ed-7d8b-940b-24df24d87ed5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh3"
  },
  "staged_characters_added": [
   "C46"
  ]
 },
 "S11sh2::signage": {
  "fp": "57f695f3682cc57c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ee2187977b7fd681": {
  "subjects": [],
  "subject_text": "인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리\n거대한 콘크리트 방벽 아래 녹슨 컨테이너가 조밀한 거리. 좁은 시장 골목과 비탈길, 천막 공터, 쓰레기 집하장과 맨홀, 진입 표지판이 이어진다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L169",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::camp_gate": {
  "input_fingerprint": "af48fedfc72dac60",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camp_gate",
    "tags": [
     "S11sh2",
     "S42sh18",
     "S42sh2",
     "S42sh27"
    ]
   },
   "context_sig": "e2c61a831cfb31ce"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 출입문 닫기 전에 헐레벌떡 뛰어오는 난민들, 그 틈에 현우도 끼여서 간신히 들어온다.\n- 42. 난민촌 입구 – N\n- 진입로마다 검문소 설치했습니다. 감시 드론도 띄웠구요.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 출입문 닫기 전에 헐레벌떡 뛰어오는 난민들, 그 틈에 현우도 끼여서 간신히 들어온다.\n- 42. 난민촌 입구 – N\n- 진입로마다 검문소 설치했습니다. 감시 드론도 띄웠구요.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camp_gate_b39179.png",
  "asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407",
  "input_asset_ids": [
   "8908b3ff-e0a4-41a9-9d9b-917a87166eb9"
  ],
  "origin_tag": "S11sh2",
  "place_text": "At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.",
  "origin_inputs": {
   "place_text": "At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.",
   "time_of_day_en": "night",
   "conti_asset_id": "8908b3ff-e0a4-41a9-9d9b-917a87166eb9"
  }
 },
 "S11sh2::bgfirst_bg": {
  "input_fingerprint": "ac023b479494cdb0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫히는 중인 수용소 철제 출입문 틈으로 이현우와 사람들이 몸을 반쯤 들이민 다급한 자세.\n\nLOCATION (lock): At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin inside the entrance at a low torso-height position, offset from the incoming path, looking obliquely through the narrowing gate before starting the retreat. Frame the gate opening near center and 이현우 slightly right of it, his body halfway through with a shoulder turned to fit, while the other refugees enter at different stride phases and torso angles rather than in a uniform row. Their attention is on the route into the camp past the camera, not the lens, and their advance through the opening supplies the dominant positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrowing camp gate opening in the middle-center of the frame, midground; Interior arrival path toward the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing while refugees squeeze through) — Seen from inside the camp at an oblique angle, with the narrowing opening visible between its edges; used as Bracket the incoming bodies without allowing a gate panel to obscure the decisive passage; Entrance passage (Occupied by refugees arriving before closure) — Extends from the opening toward the camera's offset position; used as Keep a readable path into the foreground, with irregular spacing and different footfall phases among the arrivals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the entrance dim following the camp blackout, with restrained residual ambient visibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫히는 중인 수용소 철제 출입문 틈으로 이현우와 사람들이 몸을 반쯤 들이민 다급한 자세.\n\nLOCATION (lock): At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin inside the entrance at a low torso-height position, offset from the incoming path, looking obliquely through the narrowing gate before starting the retreat. Frame the gate opening near center and 이현우 slightly right of it, his body halfway through with a shoulder turned to fit, while the other refugees enter at different stride phases and torso angles rather than in a uniform row. Their attention is on the route into the camp past the camera, not the lens, and their advance through the opening supplies the dominant positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrowing camp gate opening in the middle-center of the frame, midground; Interior arrival path toward the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing while refugees squeeze through) — Seen from inside the camp at an oblique angle, with the narrowing opening visible between its edges; used as Bracket the incoming bodies without allowing a gate panel to obscure the decisive passage; Entrance passage (Occupied by refugees arriving before closure) — Extends from the opening toward the camera's offset position; used as Keep a readable path into the foreground, with irregular spacing and different footfall phases among the arrivals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the entrance dim following the camp blackout, with restrained residual ambient visibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh2__bgfirst_bg.png",
  "asset_id": "7ac67526-a684-4b16-bada-2c1148aacc6f",
  "input_asset_ids": [
   "8908b3ff-e0a4-41a9-9d9b-917a87166eb9",
   "cf1ff170-e216-45ef-956e-1c0eea9b4407"
  ]
 },
 "S11sh2": {
  "input_fingerprint": "c8a874851f80813c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫히는 중인 수용소 철제 출입문 틈으로 이현우와 사람들이 몸을 반쯤 들이민 다급한 자세.\n\nLOCATION (lock): At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin inside the entrance at a low torso-height position, offset from the incoming path, looking obliquely through the narrowing gate before starting the retreat. Frame the gate opening near center and 이현우 slightly right of it, his body halfway through with a shoulder turned to fit, while the other refugees enter at different stride phases and torso angles rather than in a uniform row. Their attention is on the route into the camp past the camera, not the lens, and their advance through the opening supplies the dominant positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrowing camp gate opening in the middle-center of the frame, midground; Interior arrival path toward the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing while refugees squeeze through) — Seen from inside the camp at an oblique angle, with the narrowing opening visible between its edges; used as Bracket the incoming bodies without allowing a gate panel to obscure the decisive passage; Entrance passage (Occupied by refugees arriving before closure) — Extends from the opening toward the camera's offset position; used as Keep a readable path into the foreground, with irregular spacing and different footfall phases among the arrivals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the entrance dim following the camp blackout, with restrained residual ambient visibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Curfew shutdown is darkening the camp: shops are closing, container-home lights and streetlights are going out, and the entrance gate is closing. 이현우: Is hurrying through the camp entrance, still carrying the facial injuries and painful dog-bite wound in his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫히는 중인 수용소 철제 출입문 틈으로 이현우와 사람들이 몸을 반쯤 들이민 다급한 자세.\n\nLOCATION (lock): At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin inside the entrance at a low torso-height position, offset from the incoming path, looking obliquely through the narrowing gate before starting the retreat. Frame the gate opening near center and 이현우 slightly right of it, his body halfway through with a shoulder turned to fit, while the other refugees enter at different stride phases and torso angles rather than in a uniform row. Their attention is on the route into the camp past the camera, not the lens, and their advance through the opening supplies the dominant positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrowing camp gate opening in the middle-center of the frame, midground; Interior arrival path toward the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing while refugees squeeze through) — Seen from inside the camp at an oblique angle, with the narrowing opening visible between its edges; used as Bracket the incoming bodies without allowing a gate panel to obscure the decisive passage; Entrance passage (Occupied by refugees arriving before closure) — Extends from the opening toward the camera's offset position; used as Keep a readable path into the foreground, with irregular spacing and different footfall phases among the arrivals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the entrance dim following the camp blackout, with restrained residual ambient visibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Curfew shutdown is darkening the camp: shops are closing, container-home lights and streetlights are going out, and the entrance gate is closing. 이현우: Is hurrying through the camp entrance, still carrying the facial injuries and painful dog-bite wound in his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫히는 중인 수용소 철제 출입문 틈으로 이현우와 사람들이 몸을 반쯤 들이민 다급한 자세.\n\nLOCATION (lock): At the narrowing opening of the refugee camp's outdoor entrance gate during nighttime curfew. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin inside the entrance at a low torso-height position, offset from the incoming path, looking obliquely through the narrowing gate before starting the retreat. Frame the gate opening near center and 이현우 slightly right of it, his body halfway through with a shoulder turned to fit, while the other refugees enter at different stride phases and torso angles rather than in a uniform row. Their attention is on the route into the camp past the camera, not the lens, and their advance through the opening supplies the dominant positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrowing camp gate opening in the middle-center of the frame, midground; Interior arrival path toward the camera in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camp entrance gate (Closing while refugees squeeze through) — Seen from inside the camp at an oblique angle, with the narrowing opening visible between its edges; used as Bracket the incoming bodies without allowing a gate panel to obscure the decisive passage; Entrance passage (Occupied by refugees arriving before closure) — Extends from the opening toward the camera's offset position; used as Keep a readable path into the foreground, with irregular spacing and different footfall phases among the arrivals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the entrance dim following the camp blackout, with restrained residual ambient visibility and no added visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Curfew shutdown is darkening the camp: shops are closing, container-home lights and streetlights are going out, and the entrance gate is closing. 이현우: Is hurrying through the camp entrance, still carrying the facial injuries and painful dog-bite wound in his leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh2__bgfirst_bg.png",
     "asset_id": "7ac67526-a684-4b16-bada-2c1148aacc6f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S11sh2.png",
     "asset_id": "8908b3ff-e0a4-41a9-9d9b-917a87166eb9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camp_gate_b39179.png",
     "asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 카메라를 지나 수용소 내부를 향해 다급히 이동하며 시선도 그 경로를 향함.",
    "built_space": "표지판, 철제 문, 감시탑 등 위치 레퍼런스의 구조와 일치함. 카메라 앵글도 올바름.",
    "entities": "이현우의 외모와 복장은 일치하나 인이어 무전기가 보이지 않음. 주변 인물들에서 다민족/다국적 특징이 잘 드러나지 않음.",
    "hard_violations": [
     "[gpt-high] 수용소 안쪽에서 바깥을 보아야 하는 카메라를 반대로 외부에 배치해, 문 너머 수용소 내부를 배경으로 사람들이 퇴장하는 공간 관계를 만들었습니다."
    ],
    "physics": "달려 들어오는 인물들의 체중 지지와 지면 접촉이 자연스러움."
   },
   {
    "label": "B",
    "direction": "인물들이 닫히는 문틈을 지나 카메라가 있는 수용소 안쪽 경로로 다급하게 들어오고 있음.",
    "built_space": "레퍼런스와 동일한 철제 출입문, 좌측 표지판, 후경의 구조물이 요구된 카메라 앵글에 맞게 배치됨.",
    "entities": "문틈에 몸을 들이민 이현우의 외상과 복장이 잘 표현되었으며, 히잡을 두른 여성 등 다민족 난민의 특징이 명확함.",
    "hard_violations": [
     "[gpt-high] 카메라를 수용소 입구 안쪽에 두라는 명시적 공간 배치와 반대로, 외부에서 내부 건물들을 바라보는 시점으로 연출하여 사람들의 통과 방향을 퇴장으로 뒤집었습니다."
    ],
    "physics": "무리 지어 밀려들어오는 인물들의 발 딛기와 자세가 지면에 안정적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라 구도와 이현우의 배치는 좋으나, 프롬프트가 요구한 다민족/다국적 난민의 묘사가 부족하여 아쉬움."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 카메라 구도를 정확히 따르며 다급하게 좁은 틈을 통과하는 이현우와 다민족 난민들의 모습을 훌륭하게 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 카메라를 지나 수용소 내부를 향해 다급히 이동하며 시선도 그 경로를 향함.",
        "built_space": "표지판, 철제 문, 감시탑 등 위치 레퍼런스의 구조와 일치함. 카메라 앵글도 올바름.",
        "entities": "이현우의 외모와 복장은 일치하나 인이어 무전기가 보이지 않음. 주변 인물들에서 다민족/다국적 특징이 잘 드러나지 않음.",
        "hard_violations": [],
        "physics": "달려 들어오는 인물들의 체중 지지와 지면 접촉이 자연스러움."
       },
       {
        "label": "B",
        "direction": "인물들이 닫히는 문틈을 지나 카메라가 있는 수용소 안쪽 경로로 다급하게 들어오고 있음.",
        "built_space": "레퍼런스와 동일한 철제 출입문, 좌측 표지판, 후경의 구조물이 요구된 카메라 앵글에 맞게 배치됨.",
        "entities": "문틈에 몸을 들이민 이현우의 외상과 복장이 잘 표현되었으며, 히잡을 두른 여성 등 다민족 난민의 특징이 명확함.",
        "hard_violations": [],
        "physics": "무리 지어 밀려들어오는 인물들의 발 딛기와 자세가 지면에 안정적으로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라 구도와 이현우의 배치는 좋으나, 프롬프트가 요구한 다민족/다국적 난민의 묘사가 부족하여 아쉬움."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 카메라 구도를 정확히 따르며 다급하게 좁은 틈을 통과하는 이현우와 다민족 난민들의 모습을 훌륭하게 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 카메라를 지나 수용소 내부를 향해 다급히 이동하며 시선도 그 경로를 향함.",
        "built_space": "표지판, 철제 문, 감시탑 등 위치 레퍼런스의 구조와 일치함. 카메라 앵글도 올바름.",
        "entities": "이현우의 외모와 복장은 일치하나 인이어 무전기가 보이지 않음. 주변 인물들에서 다민족/다국적 특징이 잘 드러나지 않음.",
        "hard_violations": [],
        "physics": "달려 들어오는 인물들의 체중 지지와 지면 접촉이 자연스러움."
       },
       {
        "label": "B",
        "direction": "인물들이 닫히는 문틈을 지나 카메라가 있는 수용소 안쪽 경로로 다급하게 들어오고 있음.",
        "built_space": "레퍼런스와 동일한 철제 출입문, 좌측 표지판, 후경의 구조물이 요구된 카메라 앵글에 맞게 배치됨.",
        "entities": "문틈에 몸을 들이민 이현우의 외상과 복장이 잘 표현되었으며, 히잡을 두른 여성 등 다민족 난민의 특징이 명확함.",
        "hard_violations": [],
        "physics": "무리 지어 밀려들어오는 인물들의 발 딛기와 자세가 지면에 안정적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "어깨를 틀어 문틈을 통과하는 동작은 맞지만, 수용소 밖에서 안을 보는 역방향 공간 배치이며 켜진 조명과 전경 인파가 정전 분위기와 진입로 구도를 훼손합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안팎이 뒤집힌 카메라 위치는 동일하게 실패했으나, 중앙 문틈·오른쪽 이현우·비워진 전경 통로와 상대적으로 억제된 조명이 A보다 요구에 가깝습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽 문 가장자리를 잡고 화면 오른쪽 앞을 바라보며 카메라 쪽으로 나옵니다. 앞선 여성은 왼쪽 아래 진행로를 보고, 오른쪽 전경 인물은 화면 오른쪽으로 빠집니다. 렌즈를 응시하는 일렬 행진은 아니지만, 컨테이너와 감시탑이 인물들 뒤에 있어 수용소 내부로 들어오기보다 내부에서 밖으로 나오는 방향으로 읽힙니다.",
        "built_space": "녹슨 철제 문짝 두 개, 좌우 난간 두 구간, 철조망 울타리, 왼쪽 표지판 한 개와 오른쪽 문 가장자리의 제어함 한 개가 보입니다. 재료와 주요 설비는 장소 사진에 가깝습니다. 그러나 표지판의 정면과 문 너머 수용소 건물들이 함께 보여 참조 사진의 외부 시점을 유지합니다. 중앙 문틈 오른쪽에 이현우가 있지만, 큰 전경 인물들이 하단 중앙 진입로를 상당히 가립니다. 가로등과 여러 창문, 통로 조명이 켜져 있습니다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 어두운 셔츠·바지가 참조와 대체로 맞습니다. 얼굴 상처와 옷의 오염이 보이나 소형 인이어와 개에게 물린 다리 상처는 명확히 식별되지 않습니다. 주변에는 두건을 쓴 여성, 곱슬머리 남성 등 다양한 외양의 난민들이 있으며, 이들은 숏 텍스트가 요구한 사람들에 해당합니다. 철문과 한국어 표지판도 참조의 대상과 일치합니다.",
        "hard_violations": [
         "카메라를 수용소 입구 안쪽에 두라는 명시적 공간 배치와 반대로, 외부에서 내부 건물들을 바라보는 시점으로 연출하여 사람들의 통과 방향을 퇴장으로 뒤집었습니다."
        ],
        "physics": "이현우는 앞쪽 신발로 젖은 바닥을 딛고 한 손으로 문 가장자리를 잡으며 상체를 기울입니다. 다른 인물들도 지면에 닿은 발과 굽힌 무릎으로 이동을 지탱합니다. 가방은 어깨끈이나 손에 지지되어 있으며, 근거 없이 떠 있는 몸이나 물체는 보이지 않습니다. 문짝은 경첩과 기둥에 지지되지만 정지 화면만으로 실제 닫힘 운동까지 확정할 수는 없습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 중앙 문틈의 오른쪽에서 몸을 비틀어 화면 오른쪽 앞을 바라보며 전진합니다. 왼쪽 전경 여성은 왼쪽 아래 이동 경로를 보고, 뒤의 난민들은 서로 다른 보폭으로 카메라 쪽을 향합니다. 시선은 대체로 렌즈를 피하지만, 배경의 수용소 건물에서 카메라가 있는 외부로 나오는 동선이므로 요구한 입장 방향과 반대입니다.",
        "built_space": "철제 문짝 두 개와 좌우 난간 두 구간, 왼쪽 표지판 한 개, 오른쪽 문의 제어함 한 개, 오른쪽 울타리 쪽 상자 한 개 및 상단 감시카메라 설비가 보입니다. 중앙의 좁은 틈과 하단 중앙의 젖은 통로가 읽히고 이현우는 틈의 약간 오른쪽에 있습니다. 다만 장소 사진처럼 표지판 정면과 문 너머 컨테이너·감시탑을 보여, 내부의 비스듬한 시점이 아니라 외부의 거의 정면 시점입니다. A보다 가로등 발광은 줄었지만 일부 창문과 배경 조명은 여전히 켜져 있습니다.",
        "entities": "이현우는 참조와 유사한 짧은 검은 머리, 젊은 동아시아계 남성 얼굴, 마른 체격과 낡은 어두운 셔츠·바지를 갖췄습니다. 뺨의 상처와 옷의 흙먼지가 보입니다. 인이어는 확실히 식별되지 않고, 바지의 손상만으로 개 물림 상처를 확인하기는 어렵습니다. 앞뒤의 남녀 난민들은 숏 텍스트에 허용된 군중이며 복제된 동일 인물은 뚜렷하지 않습니다. 철문, 철조망, 한국어 표지판은 장소 참조와 부합합니다.",
        "hard_violations": [
         "수용소 안쪽에서 바깥을 보아야 하는 카메라를 반대로 외부에 배치해, 문 너머 수용소 내부를 배경으로 사람들이 퇴장하는 공간 관계를 만들었습니다."
        ],
        "physics": "이현우는 앞발을 바닥에 딛고 뒷발을 들어 다음 보폭으로 옮기며 손으로 문 가장자리를 짚습니다. 몸의 기울기와 굽힌 다리는 달리기 동작으로 설명됩니다. 왼쪽 전경 여성과 뒤따르는 난민들도 지지발 또는 가려진 하체로 지면과 연결되며 공중에 매달린 인물은 없습니다. 가방은 어깨끈으로 지지되고 문짝은 고정 기둥의 경첩에 연결됩니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "어깨를 틀어 문틈을 통과하는 동작은 맞지만, 수용소 밖에서 안을 보는 역방향 공간 배치이며 켜진 조명과 전경 인파가 정전 분위기와 진입로 구도를 훼손합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안팎이 뒤집힌 카메라 위치는 동일하게 실패했으나, 중앙 문틈·오른쪽 이현우·비워진 전경 통로와 상대적으로 억제된 조명이 A보다 요구에 가깝습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽 문 가장자리를 잡고 화면 오른쪽 앞을 바라보며 카메라 쪽으로 나옵니다. 앞선 여성은 왼쪽 아래 진행로를 보고, 오른쪽 전경 인물은 화면 오른쪽으로 빠집니다. 렌즈를 응시하는 일렬 행진은 아니지만, 컨테이너와 감시탑이 인물들 뒤에 있어 수용소 내부로 들어오기보다 내부에서 밖으로 나오는 방향으로 읽힙니다.",
        "built_space": "녹슨 철제 문짝 두 개, 좌우 난간 두 구간, 철조망 울타리, 왼쪽 표지판 한 개와 오른쪽 문 가장자리의 제어함 한 개가 보입니다. 재료와 주요 설비는 장소 사진에 가깝습니다. 그러나 표지판의 정면과 문 너머 수용소 건물들이 함께 보여 참조 사진의 외부 시점을 유지합니다. 중앙 문틈 오른쪽에 이현우가 있지만, 큰 전경 인물들이 하단 중앙 진입로를 상당히 가립니다. 가로등과 여러 창문, 통로 조명이 켜져 있습니다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 어두운 셔츠·바지가 참조와 대체로 맞습니다. 얼굴 상처와 옷의 오염이 보이나 소형 인이어와 개에게 물린 다리 상처는 명확히 식별되지 않습니다. 주변에는 두건을 쓴 여성, 곱슬머리 남성 등 다양한 외양의 난민들이 있으며, 이들은 숏 텍스트가 요구한 사람들에 해당합니다. 철문과 한국어 표지판도 참조의 대상과 일치합니다.",
        "hard_violations": [
         "카메라를 수용소 입구 안쪽에 두라는 명시적 공간 배치와 반대로, 외부에서 내부 건물들을 바라보는 시점으로 연출하여 사람들의 통과 방향을 퇴장으로 뒤집었습니다."
        ],
        "physics": "이현우는 앞쪽 신발로 젖은 바닥을 딛고 한 손으로 문 가장자리를 잡으며 상체를 기울입니다. 다른 인물들도 지면에 닿은 발과 굽힌 무릎으로 이동을 지탱합니다. 가방은 어깨끈이나 손에 지지되어 있으며, 근거 없이 떠 있는 몸이나 물체는 보이지 않습니다. 문짝은 경첩과 기둥에 지지되지만 정지 화면만으로 실제 닫힘 운동까지 확정할 수는 없습니다."
       },
       {
        "label": "A",
        "direction": "이현우는 중앙 문틈의 오른쪽에서 몸을 비틀어 화면 오른쪽 앞을 바라보며 전진합니다. 왼쪽 전경 여성은 왼쪽 아래 이동 경로를 보고, 뒤의 난민들은 서로 다른 보폭으로 카메라 쪽을 향합니다. 시선은 대체로 렌즈를 피하지만, 배경의 수용소 건물에서 카메라가 있는 외부로 나오는 동선이므로 요구한 입장 방향과 반대입니다.",
        "built_space": "철제 문짝 두 개와 좌우 난간 두 구간, 왼쪽 표지판 한 개, 오른쪽 문의 제어함 한 개, 오른쪽 울타리 쪽 상자 한 개 및 상단 감시카메라 설비가 보입니다. 중앙의 좁은 틈과 하단 중앙의 젖은 통로가 읽히고 이현우는 틈의 약간 오른쪽에 있습니다. 다만 장소 사진처럼 표지판 정면과 문 너머 컨테이너·감시탑을 보여, 내부의 비스듬한 시점이 아니라 외부의 거의 정면 시점입니다. A보다 가로등 발광은 줄었지만 일부 창문과 배경 조명은 여전히 켜져 있습니다.",
        "entities": "이현우는 참조와 유사한 짧은 검은 머리, 젊은 동아시아계 남성 얼굴, 마른 체격과 낡은 어두운 셔츠·바지를 갖췄습니다. 뺨의 상처와 옷의 흙먼지가 보입니다. 인이어는 확실히 식별되지 않고, 바지의 손상만으로 개 물림 상처를 확인하기는 어렵습니다. 앞뒤의 남녀 난민들은 숏 텍스트에 허용된 군중이며 복제된 동일 인물은 뚜렷하지 않습니다. 철문, 철조망, 한국어 표지판은 장소 참조와 부합합니다.",
        "hard_violations": [
         "수용소 안쪽에서 바깥을 보아야 하는 카메라를 반대로 외부에 배치해, 문 너머 수용소 내부를 배경으로 사람들이 퇴장하는 공간 관계를 만들었습니다."
        ],
        "physics": "이현우는 앞발을 바닥에 딛고 뒷발을 들어 다음 보폭으로 옮기며 손으로 문 가장자리를 짚습니다. 몸의 기울기와 굽힌 다리는 달리기 동작으로 설명됩니다. 왼쪽 전경 여성과 뒤따르는 난민들도 지지발 또는 가려진 하체로 지면과 연결되며 공중에 매달린 인물은 없습니다. 가방은 어깨끈으로 지지되고 문짝은 고정 기둥의 경첩에 연결됩니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.417
   },
   "violations": {
    "B": [
     "[gpt-high] 카메라를 수용소 입구 안쪽에 두라는 명시적 공간 배치와 반대로, 외부에서 내부 건물들을 바라보는 시점으로 연출하여 사람들의 통과 방향을 퇴장으로 뒤집었습니다."
    ],
    "A": [
     "[gpt-high] 수용소 안쪽에서 바깥을 보아야 하는 카메라를 반대로 외부에 배치해, 문 너머 수용소 내부를 배경으로 사람들이 퇴장하는 공간 관계를 만들었습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "카메라 구도와 이현우의 배치는 좋으나, 프롬프트가 요구한 다민족/다국적 난민의 묘사가 부족하여 아쉬움.  ★위반: [gpt-high] 수용소 안쪽에서 바깥을 보아야 하는 카메라를 반대로 외부에 배치해, 문 너머 수용소 내부를 배경으로 사람들이 퇴장하는 공간 관계를 만들었습니다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "지정된 카메라 구도를 정확히 따르며 다급하게 좁은 틈을 통과하는 이현우와 다민족 난민들의 모습을 훌륭하게 구현함.  ★위반: [gpt-high] 카메라를 수용소 입구 안쪽에 두라는 명시적 공간 배치와 반대로, 외부에서 내부 건물들을 바라보는 시점으로 연출하여 사람들의 통과 방향을 퇴장으로 뒤집었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camp_gate_b39179.png",
    "asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0867-e636-70e0-ac1b-7e70729c333d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh2__bgfirst_bg.png",
   "bg_asset_id": "7ac67526-a684-4b16-bada-2c1148aacc6f",
   "bg_record_key": "S11sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "camp_gate",
   "groupbg_asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S11sh5::signage": {
  "fp": "40f33e793121811c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::family_container_front": {
  "input_fingerprint": "9bf7b4b6b31804ed",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "family_container_front",
    "tags": [
     "S11sh5",
     "S20sh2",
     "S20sh7"
    ]
   },
   "context_sig": "8f31ee23f91c2d68"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the dark alley directly outside a container home's closed front door in the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우, 컨테이너들 사이의 골목을 걸어와 자신의 집 (컨테이너) 앞에 도착한다. 컨테이너 번호 7-31, 문고리를 잡고 잠시 머뭇거리다 들어간다.\n- 현우와 페드로, 실실 웃으며 왔다가 박철진 수하들이 현우 컨테이너 안을 뒤지는 것을 보고 놀라 숨는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the dark alley directly outside a container home's closed front door in the refugee settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우, 컨테이너들 사이의 골목을 걸어와 자신의 집 (컨테이너) 앞에 도착한다. 컨테이너 번호 7-31, 문고리를 잡고 잠시 머뭇거리다 들어간다.\n- 현우와 페드로, 실실 웃으며 왔다가 박철진 수하들이 현우 컨테이너 안을 뒤지는 것을 보고 놀라 숨는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_front_d0287f.png",
  "asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b",
  "input_asset_ids": [
   "eef6d95f-a24d-47a6-9e1e-d74ccd586566"
  ],
  "origin_tag": "S11sh5",
  "place_text": "In the dark alley directly outside a container home's closed front door in the refugee settlement.",
  "origin_inputs": {
   "place_text": "In the dark alley directly outside a container home's closed front door in the refugee settlement.",
   "time_of_day_en": "night",
   "conti_asset_id": "eef6d95f-a24d-47a6-9e1e-d74ccd586566"
  }
 },
 "S11sh5::bgfirst_bg": {
  "input_fingerprint": "9cabd64fc0fc4cb2",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫힌 컨테이너 문 앞에서 고개를 약간 숙인 채 망설이는 표정을 지은 이현우의 측면.\n\nLOCATION (lock): In the dark alley directly outside a container home's closed front door in the refugee settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the side approach beside 이현우 at approximately face height, with the viewing axis nearly perpendicular to his direction toward the door. Hold his lowered profile in the left-center, his bent arm leading to the handle at lower right, and a narrow section of the still-closed door along the right edge; his gaze drops toward his hand as he hesitates before turning it. Let the reduced camera distance isolate this pause, preserving the established darkness and avoiding an anticipatory view into the home.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container home's entrance door (Still closed before 이현우 turns the handle) — A narrow oblique section of the exterior face appears at the right edge; used as Create a firm boundary in front of his hesitant profile without revealing the interior; Door handle (Held but not yet turned) — Seen from the side beneath his hand; used as Link his lowered gaze and paused body to the interrupted action.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness with subdued ambient visibility, keeping his profile readable without light spilling from the closed home.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫힌 컨테이너 문 앞에서 고개를 약간 숙인 채 망설이는 표정을 지은 이현우의 측면.\n\nLOCATION (lock): In the dark alley directly outside a container home's closed front door in the refugee settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the side approach beside 이현우 at approximately face height, with the viewing axis nearly perpendicular to his direction toward the door. Hold his lowered profile in the left-center, his bent arm leading to the handle at lower right, and a narrow section of the still-closed door along the right edge; his gaze drops toward his hand as he hesitates before turning it. Let the reduced camera distance isolate this pause, preserving the established darkness and avoiding an anticipatory view into the home.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container home's entrance door (Still closed before 이현우 turns the handle) — A narrow oblique section of the exterior face appears at the right edge; used as Create a firm boundary in front of his hesitant profile without revealing the interior; Door handle (Held but not yet turned) — Seen from the side beneath his hand; used as Link his lowered gaze and paused body to the interrupted action.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness with subdued ambient visibility, keeping his profile readable without light spilling from the closed home.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh5__bgfirst_bg.png",
  "asset_id": "0066b2ae-7cdb-48ab-b970-306f96a8a6c4",
  "input_asset_ids": [
   "eef6d95f-a24d-47a6-9e1e-d74ccd586566",
   "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b"
  ]
 },
 "S11sh5": {
  "input_fingerprint": "d32a5f23bf66074c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫힌 컨테이너 문 앞에서 고개를 약간 숙인 채 망설이는 표정을 지은 이현우의 측면.\n\nLOCATION (lock): In the dark alley directly outside a container home's closed front door in the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the side approach beside 이현우 at approximately face height, with the viewing axis nearly perpendicular to his direction toward the door. Hold his lowered profile in the left-center, his bent arm leading to the handle at lower right, and a narrow section of the still-closed door along the right edge; his gaze drops toward his hand as he hesitates before turning it. Let the reduced camera distance isolate this pause, preserving the established darkness and avoiding an anticipatory view into the home.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container home's entrance door (Still closed before 이현우 turns the handle) — A narrow oblique section of the exterior face appears at the right edge; used as Create a firm boundary in front of his hesitant profile without revealing the interior; Door handle (Held but not yet turned) — Seen from the side beneath his hand; used as Link his lowered gaze and paused body to the interrupted action.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness with subdued ambient visibility, keeping his profile readable without light spilling from the closed home.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp streets remain dark after curfew shutdown. The home container is marked “7-31,” and its entrance door is still closed. 이현우: Stands hesitantly at the container entrance with a hand on the door handle. His facial injuries and dog-bitten leg remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫힌 컨테이너 문 앞에서 고개를 약간 숙인 채 망설이는 표정을 지은 이현우의 측면.\n\nLOCATION (lock): In the dark alley directly outside a container home's closed front door in the refugee settlement. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the side approach beside 이현우 at approximately face height, with the viewing axis nearly perpendicular to his direction toward the door. Hold his lowered profile in the left-center, his bent arm leading to the handle at lower right, and a narrow section of the still-closed door along the right edge; his gaze drops toward his hand as he hesitates before turning it. Let the reduced camera distance isolate this pause, preserving the established darkness and avoiding an anticipatory view into the home.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container home's entrance door (Still closed before 이현우 turns the handle) — A narrow oblique section of the exterior face appears at the right edge; used as Create a firm boundary in front of his hesitant profile without revealing the interior; Door handle (Held but not yet turned) — Seen from the side beneath his hand; used as Link his lowered gaze and paused body to the interrupted action.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness with subdued ambient visibility, keeping his profile readable without light spilling from the closed home.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp streets remain dark after curfew shutdown. The home container is marked “7-31,” and its entrance door is still closed. 이현우: Stands hesitantly at the container entrance with a hand on the door handle. His facial injuries and dog-bitten leg remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닫힌 컨테이너 문 앞에서 고개를 약간 숙인 채 망설이는 표정을 지은 이현우의 측면.\n\nLOCATION (lock): In the dark alley directly outside a container home's closed front door in the refugee settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the side approach beside 이현우 at approximately face height, with the viewing axis nearly perpendicular to his direction toward the door. Hold his lowered profile in the left-center, his bent arm leading to the handle at lower right, and a narrow section of the still-closed door along the right edge; his gaze drops toward his hand as he hesitates before turning it. Let the reduced camera distance isolate this pause, preserving the established darkness and avoiding an anticipatory view into the home.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container home's entrance door (Still closed before 이현우 turns the handle) — A narrow oblique section of the exterior face appears at the right edge; used as Create a firm boundary in front of his hesitant profile without revealing the interior; Door handle (Held but not yet turned) — Seen from the side beneath his hand; used as Link his lowered gaze and paused body to the interrupted action.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the camp's post-blackout darkness with subdued ambient visibility, keeping his profile readable without light spilling from the closed home.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp streets remain dark after curfew shutdown. The home container is marked “7-31,” and its entrance door is still closed. 이현우: Stands hesitantly at the container entrance with a hand on the door handle. His facial injuries and dog-bitten leg remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh5__bgfirst_bg.png",
     "asset_id": "0066b2ae-7cdb-48ab-b970-306f96a8a6c4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S11sh5.png",
     "asset_id": "eef6d95f-a24d-47a6-9e1e-d74ccd586566",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_front_d0287f.png",
     "asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선이 아래로 향하여 문을 잡고 있는 자신의 손과 문손잡이에 머물고 있음.",
    "built_space": "로케이션 레퍼런스의 좁은 컨테이너 골목. 우측에 닫힌 문, '7-31' 번호판, 창문, 그리고 레퍼런스와 정확히 일치하는 형태의 문손잡이(백플레이트 포함)가 올바른 위치에 배치됨.",
    "entities": "이현우의 얼굴, 헝클어진 머리, 얼굴의 상처, 어둡고 오염된 셔츠, 소형 인이어 무전기 모두 정확히 묘사됨. '7-31' 표지판의 숫자도 명확함.",
    "hard_violations": [],
    "physics": "오른손이 문손잡이를 자연스럽고 단단하게 쥐고 지지하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
   },
   {
    "label": "B",
    "direction": "인물의 시선이 아래로 향해 자신의 손과 손잡이를 바라보고 있음.",
    "built_space": "컨테이너 배경의 골목. 우측에 닫힌 문과 '7-31' 번호판이 있으나, 문의 손잡이 부착부(백플레이트)가 레퍼런스와 다르게 생략되어 컨테이너 표면에 바로 꽂혀 있음.",
    "entities": "이현우의 인상, 상처, 인이어 무전기, 오염된 의상은 잘 표현됨. 하지만 손잡이를 잡고 있는 손가락들이 뭉개져 형태가 불분명함.",
    "hard_violations": [],
    "physics": "손이 손잡이에 닿아 있으나 손가락이 융합된 것처럼 보여 물리적인 쥠 동작의 현실성이 떨어짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 프레이밍을 완벽히 따르며, 로케이션 레퍼런스의 문손잡이 디테일과 캐릭터의 상태를 가장 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍 지시를 잘 따랐으나, 손잡이를 잡은 손가락의 해부학적 왜곡과 레퍼런스와 다른 문손잡이 형태가 아쉬움."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 아래로 향하여 문을 잡고 있는 자신의 손과 문손잡이에 머물고 있음.",
        "built_space": "로케이션 레퍼런스의 좁은 컨테이너 골목. 우측에 닫힌 문, '7-31' 번호판, 창문, 그리고 레퍼런스와 정확히 일치하는 형태의 문손잡이(백플레이트 포함)가 올바른 위치에 배치됨.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 얼굴의 상처, 어둡고 오염된 셔츠, 소형 인이어 무전기 모두 정확히 묘사됨. '7-31' 표지판의 숫자도 명확함.",
        "hard_violations": [],
        "physics": "오른손이 문손잡이를 자연스럽고 단단하게 쥐고 지지하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
       },
       {
        "label": "B",
        "direction": "인물의 시선이 아래로 향해 자신의 손과 손잡이를 바라보고 있음.",
        "built_space": "컨테이너 배경의 골목. 우측에 닫힌 문과 '7-31' 번호판이 있으나, 문의 손잡이 부착부(백플레이트)가 레퍼런스와 다르게 생략되어 컨테이너 표면에 바로 꽂혀 있음.",
        "entities": "이현우의 인상, 상처, 인이어 무전기, 오염된 의상은 잘 표현됨. 하지만 손잡이를 잡고 있는 손가락들이 뭉개져 형태가 불분명함.",
        "hard_violations": [],
        "physics": "손이 손잡이에 닿아 있으나 손가락이 융합된 것처럼 보여 물리적인 쥠 동작의 현실성이 떨어짐."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 프레이밍을 완벽히 따르며, 로케이션 레퍼런스의 문손잡이 디테일과 캐릭터의 상태를 가장 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍 지시를 잘 따랐으나, 손잡이를 잡은 손가락의 해부학적 왜곡과 레퍼런스와 다른 문손잡이 형태가 아쉬움."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 아래로 향하여 문을 잡고 있는 자신의 손과 문손잡이에 머물고 있음.",
        "built_space": "로케이션 레퍼런스의 좁은 컨테이너 골목. 우측에 닫힌 문, '7-31' 번호판, 창문, 그리고 레퍼런스와 정확히 일치하는 형태의 문손잡이(백플레이트 포함)가 올바른 위치에 배치됨.",
        "entities": "이현우의 얼굴, 헝클어진 머리, 얼굴의 상처, 어둡고 오염된 셔츠, 소형 인이어 무전기 모두 정확히 묘사됨. '7-31' 표지판의 숫자도 명확함.",
        "hard_violations": [],
        "physics": "오른손이 문손잡이를 자연스럽고 단단하게 쥐고 지지하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
       },
       {
        "label": "B",
        "direction": "인물의 시선이 아래로 향해 자신의 손과 손잡이를 바라보고 있음.",
        "built_space": "컨테이너 배경의 골목. 우측에 닫힌 문과 '7-31' 번호판이 있으나, 문의 손잡이 부착부(백플레이트)가 레퍼런스와 다르게 생략되어 컨테이너 표면에 바로 꽂혀 있음.",
        "entities": "이현우의 인상, 상처, 인이어 무전기, 오염된 의상은 잘 표현됨. 하지만 손잡이를 잡고 있는 손가락들이 뭉개져 형태가 불분명함.",
        "hard_violations": [],
        "physics": "손이 손잡이에 닿아 있으나 손가락이 융합된 것처럼 보여 물리적인 쥠 동작의 현실성이 떨어짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "B보다 얼굴에 가까이 접근하여 숙인 옆얼굴과 손잡이를 쥔 채 멈춘 순간을 잘 연결하지만, 요구된 클로즈업보다 넓고 정전 이후의 어둠도 부족하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "옆얼굴과 손잡이 접촉, 장소의 구조는 충실하지만 상반신과 골목을 넓게 보여 주어 얼굴 중심 클로즈업 지시에서 더 멀어지며, 정전 상태도 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 오른쪽 문을 향하고 고개와 눈을 아래로 내려 손잡이를 잡은 손 쪽을 본다. 굽힌 팔이 화면 오른쪽 아래의 손잡이로 이어지고 손잡이는 수평으로 유지되어, 돌리기 직전의 정지 동작으로 읽힌다. 카메라는 거의 측면이지만 얼굴 앞면도 조금 보인다.",
        "built_space": "오른쪽에 닫힌 금속 출입문 하나, 레버 손잡이와 잠금판 한 세트, 그 오른쪽 벽에 ‘7-31’ 표지 하나가 보인다. 문 왼쪽에는 창 하나와 전기함 하나, 수직 배관이 있어 참고 장소의 주요 배치를 따른다. 인물은 문 바깥에서 손이 닿는 거리에 서 있다. 다만 문 외면과 골목이 상당히 넓게 보여 오른쪽 가장자리의 좁고 비스듬한 문 조각이라는 구도와 다르다. 여러 가로등과 창 불빛이 켜져 있어 정전 후의 어둠도 약하다.",
        "entities": "보이는 사람은 한 명이며, 짧고 헝클어진 검은 머리의 마른 동아시아계 십대 후반 남성으로 참고 인물과 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 볼의 찰과상, 검은 소형 인이어 장치, 먼지와 얼룩이 묻은 낡은 어두운 셔츠가 보인다. 바지와 개에게 물린 다리는 프레임 밖이므로 평가하지 않는다. 금속 문과 손잡이, ‘7-31’ 표지는 요구된 대상이며 추가 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "손가락이 문에 고정된 레버를 감싸고 있으며 손목, 팔꿈치, 어깨의 연결과 굽힘이 자연스럽다. 손잡이는 문에 부착되어 지지된다. 하체와 발은 화면 밖이지만 상체는 정상적인 기립 자세로 이어지며 공중에 떠 있다는 징후는 없다. 정지한 채 손잡이를 잡고 망설이는 동작이 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "인물의 옆얼굴은 오른쪽 문을 향하며 눈은 아래쪽 손과 손잡이 방향으로 내려가 있다. 카메라 축은 진행 방향에 거의 수직이고, 굽힌 팔 끝의 손이 오른쪽 아래 수평 레버를 잡는다. 문을 당기거나 손잡이를 이미 돌리는 움직임은 보이지 않는다.",
        "built_space": "오른쪽의 닫힌 문 하나에 손잡이와 잠금판 한 세트가 있고, 오른쪽 벽에는 ‘7-31’ 표지 하나가 있다. 문 왼쪽의 창 하나, 전기함 하나와 수직 배관도 참고 장소와 대응한다. 인물은 문 외부의 손잡이에 접근 가능한 위치에 있다. 그러나 문 전면, 인물의 허리 부근까지의 상반신, 깊은 골목을 넓게 보여 주어 지정된 가까운 측면 구도가 아니다. 가로등 여러 개와 창 조명이 켜져 있어 통금 정전 뒤의 낮은 주변광 조건과 어긋난다.",
        "entities": "인물은 한 명이고, 참고와 대체로 일치하는 젊은 동아시아계 남성의 얼굴, 짧은 검은 머리, 마른 체격이 보인다. 한국계 미국인이라는 국적은 영상만으로 판별할 수 없다. 볼의 상처, 귀의 소형 검은 인이어 장치, 피와 흙먼지 얼룩이 있는 어두운 셔츠가 확인된다. 하의와 다리 상처는 구도 밖이므로 누락으로 판단하지 않는다. 문, 금속 레버, 호수 표지는 요청된 종류이며 추가 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "손이 레버에 직접 닿아 감싸고 있고 레버는 문에 고정되어 있다. 팔꿈치를 굽힌 자세와 손잡이까지의 거리는 자연스럽다. 발은 잘렸지만 몸통은 서 있는 하체로 이어지는 자세이며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다. 닫힌 문 앞에서 동작을 멈춘 상태로 성립한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "B보다 얼굴에 가까이 접근하여 숙인 옆얼굴과 손잡이를 쥔 채 멈춘 순간을 잘 연결하지만, 요구된 클로즈업보다 넓고 정전 이후의 어둠도 부족하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "옆얼굴과 손잡이 접촉, 장소의 구조는 충실하지만 상반신과 골목을 넓게 보여 주어 얼굴 중심 클로즈업 지시에서 더 멀어지며, 정전 상태도 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물은 오른쪽 문을 향하고 고개와 눈을 아래로 내려 손잡이를 잡은 손 쪽을 본다. 굽힌 팔이 화면 오른쪽 아래의 손잡이로 이어지고 손잡이는 수평으로 유지되어, 돌리기 직전의 정지 동작으로 읽힌다. 카메라는 거의 측면이지만 얼굴 앞면도 조금 보인다.",
        "built_space": "오른쪽에 닫힌 금속 출입문 하나, 레버 손잡이와 잠금판 한 세트, 그 오른쪽 벽에 ‘7-31’ 표지 하나가 보인다. 문 왼쪽에는 창 하나와 전기함 하나, 수직 배관이 있어 참고 장소의 주요 배치를 따른다. 인물은 문 바깥에서 손이 닿는 거리에 서 있다. 다만 문 외면과 골목이 상당히 넓게 보여 오른쪽 가장자리의 좁고 비스듬한 문 조각이라는 구도와 다르다. 여러 가로등과 창 불빛이 켜져 있어 정전 후의 어둠도 약하다.",
        "entities": "보이는 사람은 한 명이며, 짧고 헝클어진 검은 머리의 마른 동아시아계 십대 후반 남성으로 참고 인물과 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 볼의 찰과상, 검은 소형 인이어 장치, 먼지와 얼룩이 묻은 낡은 어두운 셔츠가 보인다. 바지와 개에게 물린 다리는 프레임 밖이므로 평가하지 않는다. 금속 문과 손잡이, ‘7-31’ 표지는 요구된 대상이며 추가 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "손가락이 문에 고정된 레버를 감싸고 있으며 손목, 팔꿈치, 어깨의 연결과 굽힘이 자연스럽다. 손잡이는 문에 부착되어 지지된다. 하체와 발은 화면 밖이지만 상체는 정상적인 기립 자세로 이어지며 공중에 떠 있다는 징후는 없다. 정지한 채 손잡이를 잡고 망설이는 동작이 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "인물의 옆얼굴은 오른쪽 문을 향하며 눈은 아래쪽 손과 손잡이 방향으로 내려가 있다. 카메라 축은 진행 방향에 거의 수직이고, 굽힌 팔 끝의 손이 오른쪽 아래 수평 레버를 잡는다. 문을 당기거나 손잡이를 이미 돌리는 움직임은 보이지 않는다.",
        "built_space": "오른쪽의 닫힌 문 하나에 손잡이와 잠금판 한 세트가 있고, 오른쪽 벽에는 ‘7-31’ 표지 하나가 있다. 문 왼쪽의 창 하나, 전기함 하나와 수직 배관도 참고 장소와 대응한다. 인물은 문 외부의 손잡이에 접근 가능한 위치에 있다. 그러나 문 전면, 인물의 허리 부근까지의 상반신, 깊은 골목을 넓게 보여 주어 지정된 가까운 측면 구도가 아니다. 가로등 여러 개와 창 조명이 켜져 있어 통금 정전 뒤의 낮은 주변광 조건과 어긋난다.",
        "entities": "인물은 한 명이고, 참고와 대체로 일치하는 젊은 동아시아계 남성의 얼굴, 짧은 검은 머리, 마른 체격이 보인다. 한국계 미국인이라는 국적은 영상만으로 판별할 수 없다. 볼의 상처, 귀의 소형 검은 인이어 장치, 피와 흙먼지 얼룩이 있는 어두운 셔츠가 확인된다. 하의와 다리 상처는 구도 밖이므로 누락으로 판단하지 않는다. 문, 금속 레버, 호수 표지는 요청된 종류이며 추가 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "손이 레버에 직접 닿아 감싸고 있고 레버는 문에 고정되어 있다. 팔꿈치를 굽힌 자세와 손잡이까지의 거리는 자연스럽다. 발은 잘렸지만 몸통은 서 있는 하체로 이어지는 자세이며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다. 닫힌 문 앞에서 동작을 멈춘 상태로 성립한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.7
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.7
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1700
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "지시된 클로즈업 프레이밍을 완벽히 따르며, 로케이션 레퍼런스의 문손잡이 디테일과 캐릭터의 상태를 가장 사실적으로 구현함."
   },
   {
    "label": "B",
    "score": 1700,
    "verdict_ko": "프레이밍 지시를 잘 따랐으나, 손잡이를 잡은 손가락의 해부학적 왜곡과 레퍼런스와 다른 문손잡이 형태가 아쉬움."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_front_d0287f.png",
    "asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab086f-b738-7e6d-8e95-014829dab1eb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S11sh5__bgfirst_bg.png",
   "bg_asset_id": "0066b2ae-7cdb-48ab-b970-306f96a8a6c4",
   "bg_record_key": "S11sh5::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "family_container_front",
   "groupbg_asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh7::signage": {
  "fp": "b2389f2e1a4b4917",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4c7a45ece4124140": {
  "subjects": [],
  "subject_text": "이현우의 컨테이너 내부 생활 공간, 앰버의 방\n철제 주거동 안의 작은 생활 공간. 식탁과 생활용품이 놓여 있고, 방문으로 구분된 작은 방이 이어진다. 밤에는 식탁 위 촛불이 실내를 밝힌다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L160",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::family_container_room": {
  "input_fingerprint": "7759bf5e8b422681",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "family_container_room",
    "tags": [
     "S12sh13",
     "S12sh16",
     "S12sh7",
     "S14sh12",
     "S14sh3",
     "S14sh7",
     "S24sh10",
     "S24sh4",
     "S24sh7"
    ]
   },
   "context_sig": "59d930a8b080cfbb"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 컨테이너 내부 생활 공간, 앰버의 방: 전력 부족으로 촛불에 의지하는 좁고 낡은 실내 공간이다. 낡은 가구와 생활용품이 비좁게 차 있다. (특징: 테이블 위에 켜진 촛불과 소박한 한 끼 식사; 피 묻은 현우의 다리와 상처를 치료하는 구급약통; 방바닥에 떨어져 깨진 도자기 화병 조각; 민병대 완장을 찬 대원들이 들이닥쳐 살림살이를 뒤엎는 난장판) / 이현우의 컨테이너 외부·출입문 앞: 7-31 번호가 적힌 구형 컨테이너 주택의 철문 앞이다. (특징: '7-31' 페인트 글씨가 적힌 낡은 철문과 문고리; 어둠 속 좁은 컨테이너 사이 골목길)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 12. 현우의 컨테이너 안 – N\n- 14. 현우의 컨테이너 안 – D\n- 차에서 내리는 현우. 잽싸게 집안으로 들어가는데 페드로가 헐레벌떡 따라 들어온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 컨테이너 내부 생활 공간, 앰버의 방: 전력 부족으로 촛불에 의지하는 좁고 낡은 실내 공간이다. 낡은 가구와 생활용품이 비좁게 차 있다. (특징: 테이블 위에 켜진 촛불과 소박한 한 끼 식사; 피 묻은 현우의 다리와 상처를 치료하는 구급약통; 방바닥에 떨어져 깨진 도자기 화병 조각; 민병대 완장을 찬 대원들이 들이닥쳐 살림살이를 뒤엎는 난장판) / 이현우의 컨테이너 외부·출입문 앞: 7-31 번호가 적힌 구형 컨테이너 주택의 철문 앞이다. (특징: '7-31' 페인트 글씨가 적힌 낡은 철문과 문고리; 어둠 속 좁은 컨테이너 사이 골목길)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 12. 현우의 컨테이너 안 – N\n- 14. 현우의 컨테이너 안 – D\n- 차에서 내리는 현우. 잽싸게 집안으로 들어가는데 페드로가 헐레벌떡 따라 들어온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
  "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
  "input_asset_ids": [
   "c185612b-825f-4f86-98d0-32b6fb443854"
  ],
  "origin_tag": "S12sh7",
  "place_text": "At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.",
  "origin_inputs": {
   "place_text": "At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.",
   "time_of_day_en": "night",
   "conti_asset_id": "c185612b-825f-4f86-98d0-32b6fb443854"
  }
 },
 "S12sh7::bgfirst_bg": {
  "input_fingerprint": "bb2382436fae7b58",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버를 향해 미간을 좁히고 날카로운 표정으로 소리치는 이현우의 얼굴 클로즈업.\n\nLOCATION (lock): At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from beside and slightly behind 앰버, with the lens just below 이현우's seated eye line and offset from his gaze toward her. Keep 앰버's shoulder entirely outside the frame, placing 이현우's face left of center with looking room to the right as he draws his brows together and opens his mouth sharply at her. Retain his face, neck, and a small portion of his shoulders rather than an extreme facial crop, emphasizing only the final decrease in camera distance while his attention stays on 앰버 immediately offscreen.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container interior (Occupied during the family meal); used as Keep the room softly unresolved behind his face without introducing other people or furnishings into the tight crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established tabletop candlelight with subdued brightness and controlled contrast, preserving intimate facial detail without adding another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버를 향해 미간을 좁히고 날카로운 표정으로 소리치는 이현우의 얼굴 클로즈업.\n\nLOCATION (lock): At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from beside and slightly behind 앰버, with the lens just below 이현우's seated eye line and offset from his gaze toward her. Keep 앰버's shoulder entirely outside the frame, placing 이현우's face left of center with looking room to the right as he draws his brows together and opens his mouth sharply at her. Retain his face, neck, and a small portion of his shoulders rather than an extreme facial crop, emphasizing only the final decrease in camera distance while his attention stays on 앰버 immediately offscreen.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container interior (Occupied during the family meal); used as Keep the room softly unresolved behind his face without introducing other people or furnishings into the tight crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established tabletop candlelight with subdued brightness and controlled contrast, preserving intimate facial detail without adding another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S12sh7__bgfirst_bg.png",
  "asset_id": "00d208c8-0e71-44ff-81d3-3ab2ac5f0a95",
  "input_asset_ids": [
   "c185612b-825f-4f86-98d0-32b6fb443854",
   "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  ]
 },
 "S12sh7": {
  "input_fingerprint": "9f333125a1df74e6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버를 향해 미간을 좁히고 날카로운 표정으로 소리치는 이현우의 얼굴 클로즈업.\n\nLOCATION (lock): At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from beside and slightly behind 앰버, with the lens just below 이현우's seated eye line and offset from his gaze toward her. Keep 앰버's shoulder entirely outside the frame, placing 이현우's face left of center with looking room to the right as he draws his brows together and opens his mouth sharply at her. Retain his face, neck, and a small portion of his shoulders rather than an extreme facial crop, emphasizing only the final decrease in camera distance while his attention stays on 앰버 immediately offscreen.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container interior (Occupied during the family meal); used as Keep the room softly unresolved behind his face without introducing other people or furnishings into the tight crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established tabletop candlelight with subdued brightness and controlled contrast, preserving intimate facial detail without adding another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where the prepared meal and old clothing being mended are set out. The container remains lit by candlelight during the nighttime electricity curfew. 이현우: Has removed his upper garment and is at the meal table. His face remains injured, and the dog-bite wound on his leg is visibly bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버를 향해 미간을 좁히고 날카로운 표정으로 소리치는 이현우의 얼굴 클로즈업.\n\nLOCATION (lock): At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from beside and slightly behind 앰버, with the lens just below 이현우's seated eye line and offset from his gaze toward her. Keep 앰버's shoulder entirely outside the frame, placing 이현우's face left of center with looking room to the right as he draws his brows together and opens his mouth sharply at her. Retain his face, neck, and a small portion of his shoulders rather than an extreme facial crop, emphasizing only the final decrease in camera distance while his attention stays on 앰버 immediately offscreen.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container interior (Occupied during the family meal); used as Keep the room softly unresolved behind his face without introducing other people or furnishings into the tight crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established tabletop candlelight with subdued brightness and controlled contrast, preserving intimate facial detail without adding another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where the prepared meal and old clothing being mended are set out. The container remains lit by candlelight during the nighttime electricity curfew. 이현우: Has removed his upper garment and is at the meal table. His face remains injured, and the dog-bite wound on his leg is visibly bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 앰버를 향해 미간을 좁히고 날카로운 표정으로 소리치는 이현우의 얼굴 클로즈업.\n\nLOCATION (lock): At the dining area inside a refugee family's container home. A candle on the table supplies the nighttime light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from beside and slightly behind 앰버, with the lens just below 이현우's seated eye line and offset from his gaze toward her. Keep 앰버's shoulder entirely outside the frame, placing 이현우's face left of center with looking room to the right as he draws his brows together and opens his mouth sharply at her. Retain his face, neck, and a small portion of his shoulders rather than an extreme facial crop, emphasizing only the final decrease in camera distance while his attention stays on 앰버 immediately offscreen.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container interior (Occupied during the family meal); used as Keep the room softly unresolved behind his face without introducing other people or furnishings into the tight crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the established tabletop candlelight with subdued brightness and controlled contrast, preserving intimate facial detail without adding another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where the prepared meal and old clothing being mended are set out. The container remains lit by candlelight during the nighttime electricity curfew. 이현우: Has removed his upper garment and is at the meal table. His face remains injured, and the dog-bite wound on his leg is visibly bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S12sh7__bgfirst_bg.png",
     "asset_id": "00d208c8-0e71-44ff-81d3-3ab2ac5f0a95",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S12sh7.png",
     "asset_id": "c185612b-825f-4f86-98d0-32b6fb443854",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
     "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 우측 오프스크린의 대상(앰버)을 명확히 향하고 있으며, 카메라는 지시된 대로 시선에서 빗겨난 채 인물을 포착하고 있음.",
    "built_space": "식탁 위 촛불과 그릇, 뒤로 보이는 컨테이너 벽면, 냉장고, 문 등 레퍼런스의 공간적 특징이 정확한 비례와 아웃포커싱으로 묘사됨.",
    "entities": "이현우의 외모가 레퍼런스와 일치하며 핏자국과 흙먼지가 묻은 셔츠, 우측 귀의 소형 인이어, 얼굴 상처 등 세부 요구사항이 정확함. 미간을 좁히고 날카롭게 소리치는 표정이 사실적으로 표현됨.",
    "hard_violations": [],
    "physics": "식탁 앞에 앉아있는 자세가 중력에 맞게 자연스러우며, 근육과 옷의 주름 등 물리적 묘사에 오류가 없음."
   },
   {
    "label": "B",
    "direction": "시선은 화면 우측 오프스크린을 향하고 있으며 소리치는 방향과 일치함.",
    "built_space": "컨테이너 내부 구조와 식탁 위 촛불 배치 등 레퍼런스의 공간적 특징을 잘 반영하여 배치됨.",
    "entities": "인물의 기본 인상착의와 무전기, 셔츠 등은 일치하나, 프롬프트에 명시되지 않은 땀이 얼굴과 머리카락에 과도하게 묘사됨. 눈을 크게 떠서 미간을 좁힌 날카로운 인상이라기보다는 놀라거나 당황하여 소리치는 인상에 가까움.",
    "hard_violations": [],
    "physics": "인물이 앉아있는 자세와 옷의 움직임에 물리적인 오류 없이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 클로즈업 구도, 시선 처리, 촛불 조명의 무드뿐만 아니라 미간을 좁히고 날카롭게 소리치는 인물의 표정과 질감까지 가장 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 배경 요소는 잘 맞췄으나, 프롬프트에 없는 과도한 땀방울이 묘사되어 질감이 부자연스러우며 표정의 날카로움이 A에 비해 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측 오프스크린의 대상(앰버)을 명확히 향하고 있으며, 카메라는 지시된 대로 시선에서 빗겨난 채 인물을 포착하고 있음.",
        "built_space": "식탁 위 촛불과 그릇, 뒤로 보이는 컨테이너 벽면, 냉장고, 문 등 레퍼런스의 공간적 특징이 정확한 비례와 아웃포커싱으로 묘사됨.",
        "entities": "이현우의 외모가 레퍼런스와 일치하며 핏자국과 흙먼지가 묻은 셔츠, 우측 귀의 소형 인이어, 얼굴 상처 등 세부 요구사항이 정확함. 미간을 좁히고 날카롭게 소리치는 표정이 사실적으로 표현됨.",
        "hard_violations": [],
        "physics": "식탁 앞에 앉아있는 자세가 중력에 맞게 자연스러우며, 근육과 옷의 주름 등 물리적 묘사에 오류가 없음."
       },
       {
        "label": "B",
        "direction": "시선은 화면 우측 오프스크린을 향하고 있으며 소리치는 방향과 일치함.",
        "built_space": "컨테이너 내부 구조와 식탁 위 촛불 배치 등 레퍼런스의 공간적 특징을 잘 반영하여 배치됨.",
        "entities": "인물의 기본 인상착의와 무전기, 셔츠 등은 일치하나, 프롬프트에 명시되지 않은 땀이 얼굴과 머리카락에 과도하게 묘사됨. 눈을 크게 떠서 미간을 좁힌 날카로운 인상이라기보다는 놀라거나 당황하여 소리치는 인상에 가까움.",
        "hard_violations": [],
        "physics": "인물이 앉아있는 자세와 옷의 움직임에 물리적인 오류 없이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 클로즈업 구도, 시선 처리, 촛불 조명의 무드뿐만 아니라 미간을 좁히고 날카롭게 소리치는 인물의 표정과 질감까지 가장 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 배경 요소는 잘 맞췄으나, 프롬프트에 없는 과도한 땀방울이 묘사되어 질감이 부자연스러우며 표정의 날카로움이 A에 비해 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 우측 오프스크린의 대상(앰버)을 명확히 향하고 있으며, 카메라는 지시된 대로 시선에서 빗겨난 채 인물을 포착하고 있음.",
        "built_space": "식탁 위 촛불과 그릇, 뒤로 보이는 컨테이너 벽면, 냉장고, 문 등 레퍼런스의 공간적 특징이 정확한 비례와 아웃포커싱으로 묘사됨.",
        "entities": "이현우의 외모가 레퍼런스와 일치하며 핏자국과 흙먼지가 묻은 셔츠, 우측 귀의 소형 인이어, 얼굴 상처 등 세부 요구사항이 정확함. 미간을 좁히고 날카롭게 소리치는 표정이 사실적으로 표현됨.",
        "hard_violations": [],
        "physics": "식탁 앞에 앉아있는 자세가 중력에 맞게 자연스러우며, 근육과 옷의 주름 등 물리적 묘사에 오류가 없음."
       },
       {
        "label": "B",
        "direction": "시선은 화면 우측 오프스크린을 향하고 있으며 소리치는 방향과 일치함.",
        "built_space": "컨테이너 내부 구조와 식탁 위 촛불 배치 등 레퍼런스의 공간적 특징을 잘 반영하여 배치됨.",
        "entities": "인물의 기본 인상착의와 무전기, 셔츠 등은 일치하나, 프롬프트에 명시되지 않은 땀이 얼굴과 머리카락에 과도하게 묘사됨. 눈을 크게 떠서 미간을 좁힌 날카로운 인상이라기보다는 놀라거나 당황하여 소리치는 인상에 가까움.",
        "hard_violations": [],
        "physics": "인물이 앉아있는 자세와 옷의 움직임에 물리적인 오류 없이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "얼굴의 왼쪽 배치와 오른쪽 앰버를 향한 외침, 더 흐린 배경에서 우세하지만, 상반신과 식탁까지 보이는 넓은 구도와 상의 착용은 지시와 다릅니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "앰버를 향해 미간을 좁히고 외치는 연기는 맞지만, 어깨와 식탁·실내 집기를 크게 포함해 얼굴 중심의 밀착 구도를 놓쳤고 벗어야 할 상의도 입고 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 두 눈이 화면 오른쪽의 프레임 밖 상대를 향하며 렌즈를 직접 응시하지 않습니다. 앰버가 있어야 할 방향과 맞고, 미간을 강하게 좁힌 채 입을 벌려 외칩니다. 앰버의 어깨는 보이지 않습니다.",
        "built_space": "왼쪽 벽의 걸린 수건 하나, 중앙 뒤 냉장고 하나와 수납 선반, 뒤쪽 문 하나, 오른쪽 창문 하나가 보여 참고 장소의 배치와 대체로 맞습니다. 앞쪽 식탁에는 촛대에 놓인 촛불 하나, 그릇 하나, 일부 잘린 음식 접시 하나가 있습니다. 인물은 식탁 왼쪽에서 앞으로 기울어 있습니다. 배경은 흐리지만 얼굴·목·어깨 일부만 남기라는 구도보다 몸통과 집기가 많이 보입니다. 중복 설비나 반사는 보이지 않습니다.",
        "entities": "짧고 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명으로, 참고 인물의 얼굴과 마른 체형에 대체로 부합합니다. 볼과 입 주변 상처 및 귀의 검은 인이어 장치가 보입니다. 그러나 피와 먼지가 묻은 어두운 셔츠를 그대로 입어 상의를 벗은 상태와 다릅니다. 촛불과 식사는 보이며, 수선 중인 옷과 다리의 출혈은 크롭 밖이라 판정하지 않습니다. 추가 인물이나 글자는 없습니다.",
        "hard_violations": [],
        "physics": "머리와 목은 앞으로 기울어진 몸통에 자연스럽게 연결되어 있고, 떠 있는 신체나 비정상적인 관절은 없습니다. 좌면과 하체는 화면 밖이므로 착석 접촉은 확인할 수 없습니다. 촛불은 촛대에, 그릇과 접시는 식탁 위에 놓여 지지됩니다. 인이어 장치도 귀에 장착되어 있습니다."
       },
       {
        "label": "B",
        "direction": "고개를 약간 기울이고 화면 오른쪽 위의 프레임 밖 상대를 노려보며 입을 크게 벌립니다. 시선은 카메라에서 벗어나 있어 바로 밖에 있는 앰버를 향한 외침으로 읽힙니다. 앰버의 어깨나 다른 사람은 보이지 않습니다.",
        "built_space": "왼쪽 수건 하나, 중앙 뒤 냉장고 하나와 선반, 뒤쪽 문 하나, 오른쪽 창문 하나가 참고 장소와 같은 관계로 놓여 있습니다. 천장의 꺼진 등기구도 보입니다. 식탁 위에는 촛불 하나와 촛대 하나, 그릇 하나, 일부 잘린 음식 접시 하나가 있습니다. 인물 뒤 왼쪽에는 의자 등받이 일부가 보입니다. 인물이 식탁 쪽으로 숙인 배치는 가능하지만, 넓은 어깨와 몸통 및 비교적 선명한 실내 집기까지 포함해 요구한 밀착 얼굴 구도보다 넓습니다.",
        "entities": "검은 머리가 헝클어진 젊은 동아시아계 남성 한 명이며, 참고 인물과 유사한 얼굴형과 마른 체형입니다. 볼의 찰과상, 입가의 피, 검은 인이어 장치가 보입니다. 낡고 오염된 어두운 셔츠는 일반 인물 설명에는 맞지만 이 장면의 상의 탈의 지시에는 어긋납니다. 촛불과 음식은 보이고, 다리 상처와 수선 중인 옷은 화면 밖이므로 감점하지 않습니다. 추가 인물이나 문구는 없습니다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 숙이고 목을 돌린 자세는 외치는 동작으로 가능합니다. 의자 등받이 일부는 보이지만 엉덩이와 좌면 접촉은 크롭 밖입니다. 신체가 공중에 떠 있다는 증거는 없습니다. 그릇과 접시 및 촛대는 식탁이 받치고, 초는 촛대에 세워져 있습니다. 귀의 장치도 자연스럽게 고정되어 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "얼굴의 왼쪽 배치와 오른쪽 앰버를 향한 외침, 더 흐린 배경에서 우세하지만, 상반신과 식탁까지 보이는 넓은 구도와 상의 착용은 지시와 다릅니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "앰버를 향해 미간을 좁히고 외치는 연기는 맞지만, 어깨와 식탁·실내 집기를 크게 포함해 얼굴 중심의 밀착 구도를 놓쳤고 벗어야 할 상의도 입고 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 두 눈이 화면 오른쪽의 프레임 밖 상대를 향하며 렌즈를 직접 응시하지 않습니다. 앰버가 있어야 할 방향과 맞고, 미간을 강하게 좁힌 채 입을 벌려 외칩니다. 앰버의 어깨는 보이지 않습니다.",
        "built_space": "왼쪽 벽의 걸린 수건 하나, 중앙 뒤 냉장고 하나와 수납 선반, 뒤쪽 문 하나, 오른쪽 창문 하나가 보여 참고 장소의 배치와 대체로 맞습니다. 앞쪽 식탁에는 촛대에 놓인 촛불 하나, 그릇 하나, 일부 잘린 음식 접시 하나가 있습니다. 인물은 식탁 왼쪽에서 앞으로 기울어 있습니다. 배경은 흐리지만 얼굴·목·어깨 일부만 남기라는 구도보다 몸통과 집기가 많이 보입니다. 중복 설비나 반사는 보이지 않습니다.",
        "entities": "짧고 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명으로, 참고 인물의 얼굴과 마른 체형에 대체로 부합합니다. 볼과 입 주변 상처 및 귀의 검은 인이어 장치가 보입니다. 그러나 피와 먼지가 묻은 어두운 셔츠를 그대로 입어 상의를 벗은 상태와 다릅니다. 촛불과 식사는 보이며, 수선 중인 옷과 다리의 출혈은 크롭 밖이라 판정하지 않습니다. 추가 인물이나 글자는 없습니다.",
        "hard_violations": [],
        "physics": "머리와 목은 앞으로 기울어진 몸통에 자연스럽게 연결되어 있고, 떠 있는 신체나 비정상적인 관절은 없습니다. 좌면과 하체는 화면 밖이므로 착석 접촉은 확인할 수 없습니다. 촛불은 촛대에, 그릇과 접시는 식탁 위에 놓여 지지됩니다. 인이어 장치도 귀에 장착되어 있습니다."
       },
       {
        "label": "A",
        "direction": "고개를 약간 기울이고 화면 오른쪽 위의 프레임 밖 상대를 노려보며 입을 크게 벌립니다. 시선은 카메라에서 벗어나 있어 바로 밖에 있는 앰버를 향한 외침으로 읽힙니다. 앰버의 어깨나 다른 사람은 보이지 않습니다.",
        "built_space": "왼쪽 수건 하나, 중앙 뒤 냉장고 하나와 선반, 뒤쪽 문 하나, 오른쪽 창문 하나가 참고 장소와 같은 관계로 놓여 있습니다. 천장의 꺼진 등기구도 보입니다. 식탁 위에는 촛불 하나와 촛대 하나, 그릇 하나, 일부 잘린 음식 접시 하나가 있습니다. 인물 뒤 왼쪽에는 의자 등받이 일부가 보입니다. 인물이 식탁 쪽으로 숙인 배치는 가능하지만, 넓은 어깨와 몸통 및 비교적 선명한 실내 집기까지 포함해 요구한 밀착 얼굴 구도보다 넓습니다.",
        "entities": "검은 머리가 헝클어진 젊은 동아시아계 남성 한 명이며, 참고 인물과 유사한 얼굴형과 마른 체형입니다. 볼의 찰과상, 입가의 피, 검은 인이어 장치가 보입니다. 낡고 오염된 어두운 셔츠는 일반 인물 설명에는 맞지만 이 장면의 상의 탈의 지시에는 어긋납니다. 촛불과 음식은 보이고, 다리 상처와 수선 중인 옷은 화면 밖이므로 감점하지 않습니다. 추가 인물이나 문구는 없습니다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 숙이고 목을 돌린 자세는 외치는 동작으로 가능합니다. 의자 등받이 일부는 보이지만 엉덩이와 좌면 접촉은 크롭 밖입니다. 신체가 공중에 떠 있다는 증거는 없습니다. 그릇과 접시 및 촛대는 식탁이 받치고, 초는 촛대에 세워져 있습니다. 귀의 장치도 자연스럽게 고정되어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.778
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1778
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "프롬프트가 요구한 클로즈업 구도, 시선 처리, 촛불 조명의 무드뿐만 아니라 미간을 좁히고 날카롭게 소리치는 인물의 표정과 질감까지 가장 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1778,
    "verdict_ko": "지시된 구도와 배경 요소는 잘 맞췄으나, 프롬프트에 없는 과도한 땀방울이 묘사되어 질감이 부자연스러우며 표정의 날카로움이 A에 비해 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
    "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0877-0dd9-7210-8cc6-59cc516dbbfe",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S12sh7__bgfirst_bg.png",
   "bg_asset_id": "00d208c8-0e71-44ff-81d3-3ab2ac5f0a95",
   "bg_record_key": "S12sh7::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "family_container_room",
   "groupbg_asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh13::signage": {
  "fp": "9d4a65b51e3c2afa",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S12sh13": {
  "input_fingerprint": "7d8e4d126eac8144",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연을 향해 이현우가 인상을 구긴 채 답답한 듯 자신의 가슴에 꽉 쥔 주먹을 맞댄 측면 구도.\n\nLOCATION (lock): Beside the dining table in the container home's shared living area, illuminated by the table candle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral track beside the table, slightly above 이현우's seated eye line and nearly perpendicular to his exchange with 미연. Place his bare upper body in right-facing profile on the left half, keeping his clenched fist against his chest and his frustrated gaze toward 미연 together; retain her angled upper body at the right edge, attending to him above the opened medicine container. Let the table edge connect them without obscuring the fist, holding this terminal composition before his rise.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (A meal and the opened medicine container remain on the tabletop) — The near edge crosses the lower frame diagonally between the two characters; used as Connects their shared domestic space while leaving the chest gesture unobstructed; Medicine container (Opened by 미연) — Its opened side is visible near 미연's hands; used as Provides a small contextual counterpoint to 이현우's refusal of care.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The table's candlelight gives the confrontation restrained warmth, with subdued exposure and enough facial detail to preserve its underlying family concern.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where a meal and an opened medicine container remain. 이현우: His upper garment is off, exposing his injured condition; his face is battered and his bitten leg is still bleeding. 미연: She remains inside the container beside the dining area after opening the medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연을 향해 이현우가 인상을 구긴 채 답답한 듯 자신의 가슴에 꽉 쥔 주먹을 맞댄 측면 구도.\n\nLOCATION (lock): Beside the dining table in the container home's shared living area, illuminated by the table candle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral track beside the table, slightly above 이현우's seated eye line and nearly perpendicular to his exchange with 미연. Place his bare upper body in right-facing profile on the left half, keeping his clenched fist against his chest and his frustrated gaze toward 미연 together; retain her angled upper body at the right edge, attending to him above the opened medicine container. Let the table edge connect them without obscuring the fist, holding this terminal composition before his rise.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (A meal and the opened medicine container remain on the tabletop) — The near edge crosses the lower frame diagonally between the two characters; used as Connects their shared domestic space while leaving the chest gesture unobstructed; Medicine container (Opened by 미연) — Its opened side is visible near 미연's hands; used as Provides a small contextual counterpoint to 이현우's refusal of care.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The table's candlelight gives the confrontation restrained warmth, with subdued exposure and enough facial detail to preserve its underlying family concern.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where a meal and an opened medicine container remain. 이현우: His upper garment is off, exposing his injured condition; his face is battered and his bitten leg is still bleeding. 미연: She remains inside the container beside the dining area after opening the medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연을 향해 이현우가 인상을 구긴 채 답답한 듯 자신의 가슴에 꽉 쥔 주먹을 맞댄 측면 구도.\n\nLOCATION (lock): Beside the dining table in the container home's shared living area, illuminated by the table candle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the lateral track beside the table, slightly above 이현우's seated eye line and nearly perpendicular to his exchange with 미연. Place his bare upper body in right-facing profile on the left half, keeping his clenched fist against his chest and his frustrated gaze toward 미연 together; retain her angled upper body at the right edge, attending to him above the opened medicine container. Let the table edge connect them without obscuring the fist, holding this terminal composition before his rise.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (A meal and the opened medicine container remain on the tabletop) — The near edge crosses the lower frame diagonally between the two characters; used as Connects their shared domestic space while leaving the chest gesture unobstructed; Medicine container (Opened by 미연) — Its opened side is visible near 미연's hands; used as Provides a small contextual counterpoint to 이현우's refusal of care.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The table's candlelight gives the confrontation restrained warmth, with subdued exposure and enough facial detail to preserve its underlying family concern.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A candle lights the dining table, where a meal and an opened medicine container remain. 이현우: His upper garment is off, exposing his injured condition; his face is battered and his bitten leg is still bleeding. 미연: She remains inside the container beside the dining area after opening the medicine container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 미연은 서로 시선을 교환하고 있으며, 이현우의 주먹은 자신의 가슴에 닿아 있음.",
    "built_space": "컨테이너 내부의 식탁, 촛불, 배경의 냉장고와 문이 올바른 위치와 비율로 배치됨.",
    "entities": "이현우와 미연의 외형이 레퍼런스와 일치하며, 지시대로 뚜껑이 열린 약통이 식탁 위에 놓여 있음 (상의 탈의 지시와 레퍼런스 의상 유지 지시가 충돌한 상황에서 레퍼런스를 따름).",
    "hard_violations": [],
    "physics": "인물들이 의자에 정상적으로 앉아 있고, 팔과 손은 가슴과 식탁에 안정적으로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "이현우와 미연이 서로를 응시하며, 이현우의 오른손 주먹이 가슴에 닿아 있음.",
    "built_space": "컨테이너 배경과 식탁 주변의 가구 및 조명 배치가 레퍼런스의 공간과 일치함.",
    "entities": "두 인물의 외모가 일치하나, 미연 근처의 약통에 열린 뚜껑과 바닥의 뚜껑이 중복해서 존재하는 형태적 오류가 있음.",
    "hard_violations": [],
    "physics": "인물들의 자세는 안정적이나, 미연이 약통을 손에 들고 있어 '식탁 위에 남겨져 있어야 한다'는 묘사와 어긋남."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "식탁 위에 약통을 배치하라는 지시와 인물들의 구도를 정확히 따랐으며, 레퍼런스의 의상 유지 조건도 잘 충족함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "약통이 식탁 위에 있어야 한다는 지시를 어기고 손에 들고 있으며, 약통 뚜껑의 형태 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 미연은 서로 시선을 교환하고 있으며, 이현우의 주먹은 자신의 가슴에 닿아 있음.",
        "built_space": "컨테이너 내부의 식탁, 촛불, 배경의 냉장고와 문이 올바른 위치와 비율로 배치됨.",
        "entities": "이현우와 미연의 외형이 레퍼런스와 일치하며, 지시대로 뚜껑이 열린 약통이 식탁 위에 놓여 있음 (상의 탈의 지시와 레퍼런스 의상 유지 지시가 충돌한 상황에서 레퍼런스를 따름).",
        "hard_violations": [],
        "physics": "인물들이 의자에 정상적으로 앉아 있고, 팔과 손은 가슴과 식탁에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 미연이 서로를 응시하며, 이현우의 오른손 주먹이 가슴에 닿아 있음.",
        "built_space": "컨테이너 배경과 식탁 주변의 가구 및 조명 배치가 레퍼런스의 공간과 일치함.",
        "entities": "두 인물의 외모가 일치하나, 미연 근처의 약통에 열린 뚜껑과 바닥의 뚜껑이 중복해서 존재하는 형태적 오류가 있음.",
        "hard_violations": [],
        "physics": "인물들의 자세는 안정적이나, 미연이 약통을 손에 들고 있어 '식탁 위에 남겨져 있어야 한다'는 묘사와 어긋남."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "식탁 위에 약통을 배치하라는 지시와 인물들의 구도를 정확히 따랐으며, 레퍼런스의 의상 유지 조건도 잘 충족함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "약통이 식탁 위에 있어야 한다는 지시를 어기고 손에 들고 있으며, 약통 뚜껑의 형태 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 미연은 서로 시선을 교환하고 있으며, 이현우의 주먹은 자신의 가슴에 닿아 있음.",
        "built_space": "컨테이너 내부의 식탁, 촛불, 배경의 냉장고와 문이 올바른 위치와 비율로 배치됨.",
        "entities": "이현우와 미연의 외형이 레퍼런스와 일치하며, 지시대로 뚜껑이 열린 약통이 식탁 위에 놓여 있음 (상의 탈의 지시와 레퍼런스 의상 유지 지시가 충돌한 상황에서 레퍼런스를 따름).",
        "hard_violations": [],
        "physics": "인물들이 의자에 정상적으로 앉아 있고, 팔과 손은 가슴과 식탁에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 미연이 서로를 응시하며, 이현우의 오른손 주먹이 가슴에 닿아 있음.",
        "built_space": "컨테이너 배경과 식탁 주변의 가구 및 조명 배치가 레퍼런스의 공간과 일치함.",
        "entities": "두 인물의 외모가 일치하나, 미연 근처의 약통에 열린 뚜껑과 바닥의 뚜껑이 중복해서 존재하는 형태적 오류가 있음.",
        "hard_violations": [],
        "physics": "인물들의 자세는 안정적이나, 미연이 약통을 손에 들고 있어 '식탁 위에 남겨져 있어야 한다'는 묘사와 어긋남."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "왼쪽 현우의 가슴 주먹과 미연을 향한 찡그린 시선을 상반신 중심으로 더 충실하게 묶지만, 상의 탈의와 정확한 측면 구도를 지키지 않았다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "서로 향한 시선과 대각선 식탁 가장자리는 맞지만, 현우를 더 정면에 가깝게 잡고 허벅지까지 넓혀 핵심 측면 구도가 약해졌으며 상의도 벗지 않았다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 화면 왼쪽에서 오른쪽 미연의 얼굴을 바라보고, 미연도 왼쪽 현우의 얼굴로 시선을 돌린다. 현우의 쥔 주먹은 자신의 가슴에 닿아 있다. 상대를 향한 방향은 맞지만 얼굴과 몸통은 완전한 옆모습보다 비스듬한 정면에 가깝다.",
        "built_space": "식탁 하나를 사이에 두고 현우는 왼쪽 의자에 앉아 있으며 미연의 상반신은 오른쪽 가장자리에서 잘린다. 뒤쪽에는 냉장고 하나, 끝 벽의 문 하나, 오른쪽 창 하나와 수납 선반, 왼쪽 걸린 수건이 보인다. 낡은 컨테이너 벽과 좁은 통로는 이전 장면의 공간에 부합한다. 식탁은 하단을 비스듬하게 가로지르며 가슴의 주먹을 가리지 않는다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않는다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 현우와 중년 동아시아계 여성 미연 두 명뿐이다. 현우의 헝클어진 검은 머리, 얼굴 상처, 인이어와 미연의 검은 단발 및 짙은 남색 상의는 참고와 대체로 맞는다. 다만 현우는 피와 먼지가 묻은 셔츠를 그대로 입어 맨상체 지시를 어긴다. 식탁에는 촛대에 놓인 촛불 하나, 그릇과 음식 접시, 미연 손 근처의 열린 흰 약통이 있다. 약통에는 젖혀진 뚜껑처럼 보이는 부분과 별도의 나사식 뚜껑이 함께 보여 개폐 구조가 다소 모호하다. 다리의 출혈은 프레임 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 등 뒤로 의자 등받이가 보이며 앉은 몸통과 굽힌 팔의 연결이 자연스럽다. 주먹은 가슴에 실제로 접촉하고, 미연의 손은 식탁 위 약통을 잡고 있다. 촛대, 그릇, 약통과 뚜껑은 모두 식탁에 지지된다. 공중에 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "현우는 오른쪽 미연의 얼굴을 바라보며 인상을 쓰고, 미연도 왼쪽 현우를 주시한다. 현우의 주먹은 자신의 가슴에 닿아 있어 행동의 대상은 정확하다. 그러나 현우의 얼굴과 가슴이 카메라에 상당히 열려 있어 요구한 오른쪽 방향의 측면보다 정면성이 강하다.",
        "built_space": "현우는 왼쪽 의자에 앉고 미연은 식탁 오른쪽에 자리한다. 식탁 하나의 모서리와 앞 가장자리가 하단에서 뚜렷한 대각선을 만든다. 배경에는 냉장고 하나, 뒤쪽 문 하나, 오른쪽 창 하나, 천장 등기구 하나, 왼쪽 수건과 오른쪽 수납 공간이 보여 이전 컨테이너 내부의 주요 설비를 유지한다. 주먹은 식탁에 가려지지 않는다. 다만 현우의 허벅지와 넓은 전경 식탁까지 포함해 상반신 중심의 지정 구도보다 넓다.",
        "entities": "현우와 미연 두 명만 보이며, 현우의 검은 헝클어진 머리, 얼굴의 멍과 피, 인이어 및 미연의 단발과 남색 상의는 참고에 대체로 부합한다. 현우는 여전히 낡은 셔츠를 입고 있어 명시된 상의 탈의를 충족하지 않는다. 식탁에는 촛불 하나, 식사 그릇들, 열린 흰 약통 하나와 분리된 뚜껑이 보인다. 약통 입구는 미연의 손 바로 옆에 드러난다. 현우 바지의 핏자국은 보이지만 이것만으로 물린 다리의 지속 출혈을 확인할 수는 없다.",
        "hard_violations": [],
        "physics": "현우는 의자에 앉아 허벅지를 앞으로 두고 있으며, 한쪽 팔을 굽혀 주먹을 가슴에 대고 다른 손은 아래쪽에 둔다. 미연의 팔은 식탁 가장자리에 놓이고 손은 약통에 접촉한다. 식탁 다리가 보이며 음식, 약통, 촛대와 전경 금속 그릇을 받친다. 금속 그릇 안의 도구도 그릇에 기대어 있어 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "왼쪽 현우의 가슴 주먹과 미연을 향한 찡그린 시선을 상반신 중심으로 더 충실하게 묶지만, 상의 탈의와 정확한 측면 구도를 지키지 않았다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "서로 향한 시선과 대각선 식탁 가장자리는 맞지만, 현우를 더 정면에 가깝게 잡고 허벅지까지 넓혀 핵심 측면 구도가 약해졌으며 상의도 벗지 않았다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 화면 왼쪽에서 오른쪽 미연의 얼굴을 바라보고, 미연도 왼쪽 현우의 얼굴로 시선을 돌린다. 현우의 쥔 주먹은 자신의 가슴에 닿아 있다. 상대를 향한 방향은 맞지만 얼굴과 몸통은 완전한 옆모습보다 비스듬한 정면에 가깝다.",
        "built_space": "식탁 하나를 사이에 두고 현우는 왼쪽 의자에 앉아 있으며 미연의 상반신은 오른쪽 가장자리에서 잘린다. 뒤쪽에는 냉장고 하나, 끝 벽의 문 하나, 오른쪽 창 하나와 수납 선반, 왼쪽 걸린 수건이 보인다. 낡은 컨테이너 벽과 좁은 통로는 이전 장면의 공간에 부합한다. 식탁은 하단을 비스듬하게 가로지르며 가슴의 주먹을 가리지 않는다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않는다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 현우와 중년 동아시아계 여성 미연 두 명뿐이다. 현우의 헝클어진 검은 머리, 얼굴 상처, 인이어와 미연의 검은 단발 및 짙은 남색 상의는 참고와 대체로 맞는다. 다만 현우는 피와 먼지가 묻은 셔츠를 그대로 입어 맨상체 지시를 어긴다. 식탁에는 촛대에 놓인 촛불 하나, 그릇과 음식 접시, 미연 손 근처의 열린 흰 약통이 있다. 약통에는 젖혀진 뚜껑처럼 보이는 부분과 별도의 나사식 뚜껑이 함께 보여 개폐 구조가 다소 모호하다. 다리의 출혈은 프레임 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "현우의 등 뒤로 의자 등받이가 보이며 앉은 몸통과 굽힌 팔의 연결이 자연스럽다. 주먹은 가슴에 실제로 접촉하고, 미연의 손은 식탁 위 약통을 잡고 있다. 촛대, 그릇, 약통과 뚜껑은 모두 식탁에 지지된다. 공중에 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "현우는 오른쪽 미연의 얼굴을 바라보며 인상을 쓰고, 미연도 왼쪽 현우를 주시한다. 현우의 주먹은 자신의 가슴에 닿아 있어 행동의 대상은 정확하다. 그러나 현우의 얼굴과 가슴이 카메라에 상당히 열려 있어 요구한 오른쪽 방향의 측면보다 정면성이 강하다.",
        "built_space": "현우는 왼쪽 의자에 앉고 미연은 식탁 오른쪽에 자리한다. 식탁 하나의 모서리와 앞 가장자리가 하단에서 뚜렷한 대각선을 만든다. 배경에는 냉장고 하나, 뒤쪽 문 하나, 오른쪽 창 하나, 천장 등기구 하나, 왼쪽 수건과 오른쪽 수납 공간이 보여 이전 컨테이너 내부의 주요 설비를 유지한다. 주먹은 식탁에 가려지지 않는다. 다만 현우의 허벅지와 넓은 전경 식탁까지 포함해 상반신 중심의 지정 구도보다 넓다.",
        "entities": "현우와 미연 두 명만 보이며, 현우의 검은 헝클어진 머리, 얼굴의 멍과 피, 인이어 및 미연의 단발과 남색 상의는 참고에 대체로 부합한다. 현우는 여전히 낡은 셔츠를 입고 있어 명시된 상의 탈의를 충족하지 않는다. 식탁에는 촛불 하나, 식사 그릇들, 열린 흰 약통 하나와 분리된 뚜껑이 보인다. 약통 입구는 미연의 손 바로 옆에 드러난다. 현우 바지의 핏자국은 보이지만 이것만으로 물린 다리의 지속 출혈을 확인할 수는 없다.",
        "hard_violations": [],
        "physics": "현우는 의자에 앉아 허벅지를 앞으로 두고 있으며, 한쪽 팔을 굽혀 주먹을 가슴에 대고 다른 손은 아래쪽에 둔다. 미연의 팔은 식탁 가장자리에 놓이고 손은 약통에 접촉한다. 식탁 다리가 보이며 음식, 약통, 촛대와 전경 금속 그릇을 받친다. 금속 그릇 안의 도구도 그릇에 기대어 있어 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "식탁 위에 약통을 배치하라는 지시와 인물들의 구도를 정확히 따랐으며, 레퍼런스의 의상 유지 조건도 잘 충족함."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "약통이 식탁 위에 있어야 한다는 지시를 어기고 손에 들고 있으며, 약통 뚜껑의 형태 오류가 발생함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S12sh7_sel.png",
    "asset_id": "11377750-12c2-483b-bced-b5f0095931f6",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1296917>",
    "asset_id": "4355f93d-0fde-4b21-a98a-8202637de522",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0880-275f-7ec9-a722-284d17386a2c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh7"
  },
  "staged_characters_added": [
   "C08"
  ]
 },
 "S12sh16::signage": {
  "fp": "6e4bcaa8e7bb528b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S12sh16": {
  "input_fingerprint": "bef00660e667e559",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 현관문 밖으로 상반신을 빼낸 이현우가 문고리를 쥔 채 당기는 힘을 주고 있는 멈춘 순간.\n\nLOCATION (lock): Across the open entrance threshold of the container home, between the candlelit living area and the dark settlement alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the following track just inside the entrance, below 이현우's shoulder height and laterally outside the door's sweep, retaining the flow's rear three-quarter view. His clothed torso extends through the opening in the center-right while his gripping hand remains visible below it; his attention is directed toward his route outside, beyond the frame. Emphasize his changed position across the threshold rather than tightening the framing, holding as the pulling arm begins to bring the door after him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Entrance door (Open and being pulled toward closure) — Its interior face is seen obliquely beside 이현우, with the handle accessible to his returning hand; used as Creates a narrowing boundary around his departure without concealing the grip; Entrance frame (The opening is occupied by 이현우's departing upper body) — Seen diagonally from inside the container; used as Maintains the distinction between the camera remaining indoors and the character leaving.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime threshold subdued, with controlled contrast that separates his hand, clothing and doorway without introducing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit meal and opened medicine container remain in the dining area. The entrance door is still open at this point, before it is slammed shut. 이현우: He has put his upper garment back on and is leaving through the entrance. His facial injuries and bleeding bite wound on his leg remain.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 현관문 밖으로 상반신을 빼낸 이현우가 문고리를 쥔 채 당기는 힘을 주고 있는 멈춘 순간.\n\nLOCATION (lock): Across the open entrance threshold of the container home, between the candlelit living area and the dark settlement alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the following track just inside the entrance, below 이현우's shoulder height and laterally outside the door's sweep, retaining the flow's rear three-quarter view. His clothed torso extends through the opening in the center-right while his gripping hand remains visible below it; his attention is directed toward his route outside, beyond the frame. Emphasize his changed position across the threshold rather than tightening the framing, holding as the pulling arm begins to bring the door after him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Entrance door (Open and being pulled toward closure) — Its interior face is seen obliquely beside 이현우, with the handle accessible to his returning hand; used as Creates a narrowing boundary around his departure without concealing the grip; Entrance frame (The opening is occupied by 이현우's departing upper body) — Seen diagonally from inside the container; used as Maintains the distinction between the camera remaining indoors and the character leaving.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime threshold subdued, with controlled contrast that separates his hand, clothing and doorway without introducing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit meal and opened medicine container remain in the dining area. The entrance door is still open at this point, before it is slammed shut. 이현우: He has put his upper garment back on and is leaving through the entrance. His facial injuries and bleeding bite wound on his leg remain.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 현관문 밖으로 상반신을 빼낸 이현우가 문고리를 쥔 채 당기는 힘을 주고 있는 멈춘 순간.\n\nLOCATION (lock): Across the open entrance threshold of the container home, between the candlelit living area and the dark settlement alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the following track just inside the entrance, below 이현우's shoulder height and laterally outside the door's sweep, retaining the flow's rear three-quarter view. His clothed torso extends through the opening in the center-right while his gripping hand remains visible below it; his attention is directed toward his route outside, beyond the frame. Emphasize his changed position across the threshold rather than tightening the framing, holding as the pulling arm begins to bring the door after him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Entrance door (Open and being pulled toward closure) — Its interior face is seen obliquely beside 이현우, with the handle accessible to his returning hand; used as Creates a narrowing boundary around his departure without concealing the grip; Entrance frame (The opening is occupied by 이현우's departing upper body) — Seen diagonally from inside the container; used as Maintains the distinction between the camera remaining indoors and the character leaving.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime threshold subdued, with controlled contrast that separates his hand, clothing and doorway without introducing an exterior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit meal and opened medicine container remain in the dining area. The entrance door is still open at this point, before it is slammed shut. 이현우: He has put his upper garment back on and is leaving through the entrance. His facial injuries and bleeding bite wound on his leg remain.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 우측 바깥의 어두운 곳을 향하고 있습니다.",
    "built_space": "실내 좌측에 촛불과 식기가 있는 탁자가 위치하며 우측에 열린 문이 배치되어 카메라 시점과 일치합니다.",
    "entities": "이현우의 외모, 낡은 의상, 인이어 무전기 및 탁자 위의 촛불과 약통 등 지정된 요소들이 레퍼런스와 일치합니다.",
    "hard_violations": [
     "[gemini-pro] 해부학적으로 불가능한 신체: 인물의 몸은 화면 중앙 좌측에 있으나, 문고리를 잡은 오른팔이 화면 우측 밖에서 들어와 몸과 완전히 단절되어 있습니다."
    ],
    "physics": "문고리를 쥔 손이 인물의 어깨나 몸통과 연결될 수 없는 위치에 있어 물리적 지지가 불가능합니다."
   },
   {
    "label": "B",
    "direction": "인물의 고개와 시선이 우측 바깥의 어두운 골목 동선을 향하고 있습니다.",
    "built_space": "실내에서 바깥을 바라보는 구도로, 좌측의 탁자와 중앙의 열린 문, 우측의 문틀이 요구된 공간감과 정확히 부합합니다.",
    "entities": "이현우의 특징적인 외모, 핏자국이 있는 의상, 인이어 무전기 및 식사 공간의 소품들이 모두 정확히 구현되었습니다.",
    "hard_violations": [],
    "physics": "오른손으로 문고리를 단단히 쥐고 문을 당기려는 자세가 지면의 지지를 받으며 물리적으로 자연스럽게 성립합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 실내외 경계의 구도와 조명을 잘 구현했으며, 인물이 문을 당기는 동작이 자연스럽게 표현되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물의 몸 위치와 문고리를 잡은 손의 방향이 일치하지 않는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 우측 바깥의 어두운 곳을 향하고 있습니다.",
        "built_space": "실내 좌측에 촛불과 식기가 있는 탁자가 위치하며 우측에 열린 문이 배치되어 카메라 시점과 일치합니다.",
        "entities": "이현우의 외모, 낡은 의상, 인이어 무전기 및 탁자 위의 촛불과 약통 등 지정된 요소들이 레퍼런스와 일치합니다.",
        "hard_violations": [
         "해부학적으로 불가능한 신체: 인물의 몸은 화면 중앙 좌측에 있으나, 문고리를 잡은 오른팔이 화면 우측 밖에서 들어와 몸과 완전히 단절되어 있습니다."
        ],
        "physics": "문고리를 쥔 손이 인물의 어깨나 몸통과 연결될 수 없는 위치에 있어 물리적 지지가 불가능합니다."
       },
       {
        "label": "B",
        "direction": "인물의 고개와 시선이 우측 바깥의 어두운 골목 동선을 향하고 있습니다.",
        "built_space": "실내에서 바깥을 바라보는 구도로, 좌측의 탁자와 중앙의 열린 문, 우측의 문틀이 요구된 공간감과 정확히 부합합니다.",
        "entities": "이현우의 특징적인 외모, 핏자국이 있는 의상, 인이어 무전기 및 식사 공간의 소품들이 모두 정확히 구현되었습니다.",
        "hard_violations": [],
        "physics": "오른손으로 문고리를 단단히 쥐고 문을 당기려는 자세가 지면의 지지를 받으며 물리적으로 자연스럽게 성립합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 실내외 경계의 구도와 조명을 잘 구현했으며, 인물이 문을 당기는 동작이 자연스럽게 표현되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물의 몸 위치와 문고리를 잡은 손의 방향이 일치하지 않는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 우측 바깥의 어두운 곳을 향하고 있습니다.",
        "built_space": "실내 좌측에 촛불과 식기가 있는 탁자가 위치하며 우측에 열린 문이 배치되어 카메라 시점과 일치합니다.",
        "entities": "이현우의 외모, 낡은 의상, 인이어 무전기 및 탁자 위의 촛불과 약통 등 지정된 요소들이 레퍼런스와 일치합니다.",
        "hard_violations": [
         "해부학적으로 불가능한 신체: 인물의 몸은 화면 중앙 좌측에 있으나, 문고리를 잡은 오른팔이 화면 우측 밖에서 들어와 몸과 완전히 단절되어 있습니다."
        ],
        "physics": "문고리를 쥔 손이 인물의 어깨나 몸통과 연결될 수 없는 위치에 있어 물리적 지지가 불가능합니다."
       },
       {
        "label": "B",
        "direction": "인물의 고개와 시선이 우측 바깥의 어두운 골목 동선을 향하고 있습니다.",
        "built_space": "실내에서 바깥을 바라보는 구도로, 좌측의 탁자와 중앙의 열린 문, 우측의 문틀이 요구된 공간감과 정확히 부합합니다.",
        "entities": "이현우의 특징적인 외모, 핏자국이 있는 의상, 인이어 무전기 및 식사 공간의 소품들이 모두 정확히 구현되었습니다.",
        "hard_violations": [],
        "physics": "오른손으로 문고리를 단단히 쥐고 문을 당기려는 자세가 지면의 지지를 받으며 물리적으로 자연스럽게 성립합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "문밖을 향한 상반신과 문고리를 잡은 손은 맞지만, B보다 측면 얼굴이 많이 드러나고 둥글어진 식탁 형태가 이전 장면의 공간 연속성을 약화한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "실내의 낮은 후방 사선 미디엄 구도에서 문턱을 넘어가는 몸과 뒤로 뻗어 문을 당기는 손이 더 명확하며, 다만 골목의 밝은 광점은 외부 광원 금지 지시와 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 상체는 화면 오른쪽의 어두운 골목을 향하며, 시선도 카메라가 아니라 화면 밖 이동 경로로 향한다. 뒤로 내려 뻗은 손은 몸 왼쪽의 문고리를 잡고 있어 떠나는 몸을 따라 문을 당기는 관계가 보인다.",
        "built_space": "열린 현관문 한 짝, 이를 둘러싼 문틀 한 곳, 문에 달린 손잡이 한 세트가 보인다. 손잡이 양쪽 부품은 같은 문의 안팎 부품으로 읽힌다. 왼쪽 실내에는 식탁 하나, 냉장고 하나, 뒤쪽 닫힌 문 하나와 천장 등이 보이고, 오른쪽 개구부는 골목으로 이어진다. 인물은 개구부 중앙 오른쪽에 있으며 손은 문짝에 닿을 수 있는 위치다. 다만 전경 식탁의 둥근 윤곽은 참고 이미지의 직선적인 식탁과 다르다. 반사는 없다.",
        "entities": "인물은 한 명뿐이며, 헝클어진 짧은 검은 머리와 마른 체격의 젊은 동아시아계 남성으로 보인다. 보이는 얼굴의 상처, 귀의 소형 인이어 장치, 때와 얼룩이 묻은 어두운 셔츠와 바지가 참고 인물과 부합한다. 식탁에는 촛불 하나, 음식 그릇들, 열린 흰 약통 하나와 분리된 뚜껑이 있다. 다리의 물린 상처는 프레임에서 확인되지 않는다. 골목에는 작은 밝은 광점이 보인다.",
        "hard_violations": [],
        "physics": "손가락이 문고리를 감싸며 손목과 팔이 몸에 자연스럽게 이어진다. 뒤로 뻗은 팔과 앞으로 기운 몸은 문을 따라 당기며 나가는 동작으로 가능하다. 문은 경첩으로 지지되고 식기와 약통은 식탁에 놓여 있다. 발은 화면 밖이지만 허리와 다리가 아래로 이어지는 서 있는 자세이며, 공중에 뜬 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "뒷머리와 얼굴의 작은 측면이 오른쪽 골목을 향하고, 상체도 그 경로로 기울어 있다. 오른팔은 몸 뒤쪽인 화면 오른쪽 아래로 뻗어 문고리를 잡는다. 이동 방향과 되돌아온 손의 방향이 분리되어 문을 뒤따라 당기는 순간이 명확하다.",
        "built_space": "오른쪽에 열린 현관문 한 짝과 손잡이 하나가 보이고, 중앙의 문틀이 실내와 골목을 나눈다. 카메라는 인물의 어깨보다 낮은 실내 후방 사선 위치로 읽힌다. 인물의 상체는 중앙 오른쪽 개구부를 차지하며 손과 문고리는 가려지지 않는다. 왼쪽에는 직사각형 식탁 하나, 냉장고 하나, 뒤쪽 닫힌 문 하나, 오른쪽 실내 벽의 창 하나와 침구가 보여 참고 공간의 배치와 재질을 비교적 잘 유지한다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "등을 보인 인물 한 명만 등장한다. 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성의 외형, 마른 체격, 인이어 장치와 피·먼지 얼룩이 있는 어두운 셔츠 및 바지가 참고와 부합한다. 얼굴이 대부분 가려져 정확한 얼굴과 안면 상처는 확인하기 어렵지만 지시된 후방 구도에 맞는 가림이다. 식탁에는 촛불 하나, 식사 그릇들, 열린 흰 약통 하나와 뚜껑이 남아 있다. 다리 상처는 화면 밖이다. 골목의 밝은 흰 광점은 외부 광원이 있는 듯 보인다.",
        "hard_violations": [],
        "physics": "문고리를 감싼 손과 뒤로 뻗은 팔이 연속적으로 연결되고, 앞으로 이동한 어깨와 뒤에 남은 손 사이에 당기는 힘이 성립한다. 문짝은 오른쪽 경첩 부위로 지지되는 구조이며 손잡이는 문에 고정되어 있다. 하체는 화면 아래로 이어져 발 접촉은 보이지 않지만 부유를 시사하지 않는다. 촛대, 식기, 약통은 식탁에 지지되어 있고 해부학적으로 불가능한 접촉은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "문밖을 향한 상반신과 문고리를 잡은 손은 맞지만, B보다 측면 얼굴이 많이 드러나고 둥글어진 식탁 형태가 이전 장면의 공간 연속성을 약화한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "실내의 낮은 후방 사선 미디엄 구도에서 문턱을 넘어가는 몸과 뒤로 뻗어 문을 당기는 손이 더 명확하며, 다만 골목의 밝은 광점은 외부 광원 금지 지시와 어긋난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 상체는 화면 오른쪽의 어두운 골목을 향하며, 시선도 카메라가 아니라 화면 밖 이동 경로로 향한다. 뒤로 내려 뻗은 손은 몸 왼쪽의 문고리를 잡고 있어 떠나는 몸을 따라 문을 당기는 관계가 보인다.",
        "built_space": "열린 현관문 한 짝, 이를 둘러싼 문틀 한 곳, 문에 달린 손잡이 한 세트가 보인다. 손잡이 양쪽 부품은 같은 문의 안팎 부품으로 읽힌다. 왼쪽 실내에는 식탁 하나, 냉장고 하나, 뒤쪽 닫힌 문 하나와 천장 등이 보이고, 오른쪽 개구부는 골목으로 이어진다. 인물은 개구부 중앙 오른쪽에 있으며 손은 문짝에 닿을 수 있는 위치다. 다만 전경 식탁의 둥근 윤곽은 참고 이미지의 직선적인 식탁과 다르다. 반사는 없다.",
        "entities": "인물은 한 명뿐이며, 헝클어진 짧은 검은 머리와 마른 체격의 젊은 동아시아계 남성으로 보인다. 보이는 얼굴의 상처, 귀의 소형 인이어 장치, 때와 얼룩이 묻은 어두운 셔츠와 바지가 참고 인물과 부합한다. 식탁에는 촛불 하나, 음식 그릇들, 열린 흰 약통 하나와 분리된 뚜껑이 있다. 다리의 물린 상처는 프레임에서 확인되지 않는다. 골목에는 작은 밝은 광점이 보인다.",
        "hard_violations": [],
        "physics": "손가락이 문고리를 감싸며 손목과 팔이 몸에 자연스럽게 이어진다. 뒤로 뻗은 팔과 앞으로 기운 몸은 문을 따라 당기며 나가는 동작으로 가능하다. 문은 경첩으로 지지되고 식기와 약통은 식탁에 놓여 있다. 발은 화면 밖이지만 허리와 다리가 아래로 이어지는 서 있는 자세이며, 공중에 뜬 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "뒷머리와 얼굴의 작은 측면이 오른쪽 골목을 향하고, 상체도 그 경로로 기울어 있다. 오른팔은 몸 뒤쪽인 화면 오른쪽 아래로 뻗어 문고리를 잡는다. 이동 방향과 되돌아온 손의 방향이 분리되어 문을 뒤따라 당기는 순간이 명확하다.",
        "built_space": "오른쪽에 열린 현관문 한 짝과 손잡이 하나가 보이고, 중앙의 문틀이 실내와 골목을 나눈다. 카메라는 인물의 어깨보다 낮은 실내 후방 사선 위치로 읽힌다. 인물의 상체는 중앙 오른쪽 개구부를 차지하며 손과 문고리는 가려지지 않는다. 왼쪽에는 직사각형 식탁 하나, 냉장고 하나, 뒤쪽 닫힌 문 하나, 오른쪽 실내 벽의 창 하나와 침구가 보여 참고 공간의 배치와 재질을 비교적 잘 유지한다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "등을 보인 인물 한 명만 등장한다. 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성의 외형, 마른 체격, 인이어 장치와 피·먼지 얼룩이 있는 어두운 셔츠 및 바지가 참고와 부합한다. 얼굴이 대부분 가려져 정확한 얼굴과 안면 상처는 확인하기 어렵지만 지시된 후방 구도에 맞는 가림이다. 식탁에는 촛불 하나, 식사 그릇들, 열린 흰 약통 하나와 뚜껑이 남아 있다. 다리 상처는 화면 밖이다. 골목의 밝은 흰 광점은 외부 광원이 있는 듯 보인다.",
        "hard_violations": [],
        "physics": "문고리를 감싼 손과 뒤로 뻗은 팔이 연속적으로 연결되고, 앞으로 이동한 어깨와 뒤에 남은 손 사이에 당기는 힘이 성립한다. 문짝은 오른쪽 경첩 부위로 지지되는 구조이며 손잡이는 문에 고정되어 있다. 하체는 화면 아래로 이어져 발 접촉은 보이지 않지만 부유를 시사하지 않는다. 촛대, 식기, 약통은 식탁에 지지되어 있고 해부학적으로 불가능한 접촉은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.875
   },
   "violations": {
    "A": [
     "[gemini-pro] 해부학적으로 불가능한 신체: 인물의 몸은 화면 중앙 좌측에 있으나, 문고리를 잡은 오른팔이 화면 우측 밖에서 들어와 몸과 완전히 단절되어 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1875,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 실내외 경계의 구도와 조명을 잘 구현했으며, 인물이 문을 당기는 동작이 자연스럽게 표현되었습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "인물의 몸 위치와 문고리를 잡은 손의 방향이 일치하지 않는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체: 인물의 몸은 화면 중앙 좌측에 있으나, 문고리를 잡은 오른팔이 화면 우측 밖에서 들어와 몸과 완전히 단절되어 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S12sh13_sel.png",
    "asset_id": "55882559-39b9-4217-8674-cc9246aae3e5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0884-7385-7ae5-bca6-d006dc22d21e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh13"
  }
 },
 "S13sh15::signage": {
  "fp": "204a16640e4cb033",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::roadside_hideout": {
  "input_fingerprint": "05a6992d1e01f3bc",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "roadside_hideout",
    "tags": [
     "S13sh15"
    ]
   },
   "context_sig": "4bafbdcc5e7d65b7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 인공제방 바로 옆 노상에 쿠마의 아지트가 있다.\n\nTIME OF DAY (lock): night to morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 인공제방 바로 옆 노상에 쿠마의 아지트가 있다.\n\nTIME OF DAY (lock): night to morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_roadside_hideout_f5126d.png",
  "asset_id": "2426b558-ee67-4180-9aab-b4581b82636c",
  "input_asset_ids": [
   "bcb2cb6f-36f1-4be4-9112-e02d4fa88374"
  ],
  "origin_tag": "S13sh15",
  "place_text": "At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.",
  "origin_inputs": {
   "place_text": "At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.",
   "time_of_day_en": "night to morning",
   "conti_asset_id": "bcb2cb6f-36f1-4be4-9112-e02d4fa88374"
  }
 },
 "S13sh15::bgfirst_bg": {
  "input_fingerprint": "8ff4517470b71c7f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 빼앗은 총구를 쿠마의 얼굴 정면으로 흔들림 없이 겨누고 있는 찰나.\n\nLOCATION (lock): At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.\n\nTIME OF DAY (lock): night to morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly arrest the track behind and outside 이현우's left shoulder, near shoulder height, looking obliquely along rather than directly down the weapon's axis. His shoulder and extended arm occupy the lower-left foreground, while 쿠마's seated upper body appears at center-right beyond the gun; 이현우 fixes on 쿠마's face, and 쿠마 looks back toward him without facing the lens. Keep the weapon modest in scale and its muzzle-to-face alignment legible, with a narrow strip of the grilling area establishing their shared space.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the lower-left of the frame, foreground; 쿠마 in the middle-right of the frame, midground; Barbecue grill beside 쿠마 in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized gun (Held steadily and aimed at 쿠마's face) — Seen from the side behind 이현우's hand, pointing diagonally across the frame rather than toward the lens; used as Links the foreground aggressor to the seated target without exaggerated foreshortening; Barbecue grill (Used for cooking chicken skewers) — A partial side and cooking area remain visible near 쿠마; used as Anchors the threat in the interrupted meal and supplies scale behind the gun.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The hideout's barrel fire provides restrained illumination within the nighttime darkness, preserving readable faces and controlled weapon highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 빼앗은 총구를 쿠마의 얼굴 정면으로 흔들림 없이 겨누고 있는 찰나.\n\nLOCATION (lock): At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill.\n\nTIME OF DAY (lock): night to morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly arrest the track behind and outside 이현우's left shoulder, near shoulder height, looking obliquely along rather than directly down the weapon's axis. His shoulder and extended arm occupy the lower-left foreground, while 쿠마's seated upper body appears at center-right beyond the gun; 이현우 fixes on 쿠마's face, and 쿠마 looks back toward him without facing the lens. Keep the weapon modest in scale and its muzzle-to-face alignment legible, with a narrow strip of the grilling area establishing their shared space.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the lower-left of the frame, foreground; 쿠마 in the middle-right of the frame, midground; Barbecue grill beside 쿠마 in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized gun (Held steadily and aimed at 쿠마's face) — Seen from the side behind 이현우's hand, pointing diagonally across the frame rather than toward the lens; used as Links the foreground aggressor to the seated target without exaggerated foreshortening; Barbecue grill (Used for cooking chicken skewers) — A partial side and cooking area remain visible near 쿠마; used as Anchors the threat in the interrupted meal and supplies scale behind the gun.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The hideout's barrel fire provides restrained illumination within the nighttime darkness, preserving readable faces and controlled weapon highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh15__bgfirst_bg.png",
  "asset_id": "d6cf8373-98fa-4fca-af9c-4fa06c54c5a5",
  "input_asset_ids": [
   "bcb2cb6f-36f1-4be4-9112-e02d4fa88374",
   "2426b558-ee67-4180-9aab-b4581b82636c"
  ]
 },
 "S13sh15": {
  "input_fingerprint": "d16038ac64c21e25",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 총구를 쿠마의 얼굴 정면으로 흔들림 없이 겨누고 있는 찰나.\n\nLOCATION (lock): At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly arrest the track behind and outside 이현우's left shoulder, near shoulder height, looking obliquely along rather than directly down the weapon's axis. His shoulder and extended arm occupy the lower-left foreground, while 쿠마's seated upper body appears at center-right beyond the gun; 이현우 fixes on 쿠마's face, and 쿠마 looks back toward him without facing the lens. Keep the weapon modest in scale and its muzzle-to-face alignment legible, with a narrow strip of the grilling area establishing their shared space.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the lower-left of the frame, foreground; 쿠마 in the middle-right of the frame, midground; Barbecue grill beside 쿠마 in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized gun (Held steadily and aimed at 쿠마's face) — Seen from the side behind 이현우's hand, pointing diagonally across the frame rather than toward the lens; used as Links the foreground aggressor to the seated target without exaggerated foreshortening; Barbecue grill (Used for cooking chicken skewers) — A partial side and cooking area remain visible near 쿠마; used as Anchors the threat in the interrupted meal and supplies scale behind the gun.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The hideout's barrel fire provides restrained illumination within the nighttime darkness, preserving readable faces and controlled weapon highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The outdoor hideout stands beside the artificial seawall, with fires burning in metal drums and chicken skewers on a barbecue grill. The seawall already has cracks and water seepage, including water running over its supports. 이현우: He wears his upper garment and holds the seized gun raised in an aiming position. His facial injuries and untreated leg bite remain. 쿠마: He remains at the barbecue with beer and a chicken skewer, not yet having discarded the skewer or stood up.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 총구를 쿠마의 얼굴 정면으로 흔들림 없이 겨누고 있는 찰나.\n\nLOCATION (lock): At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly arrest the track behind and outside 이현우's left shoulder, near shoulder height, looking obliquely along rather than directly down the weapon's axis. His shoulder and extended arm occupy the lower-left foreground, while 쿠마's seated upper body appears at center-right beyond the gun; 이현우 fixes on 쿠마's face, and 쿠마 looks back toward him without facing the lens. Keep the weapon modest in scale and its muzzle-to-face alignment legible, with a narrow strip of the grilling area establishing their shared space.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the lower-left of the frame, foreground; 쿠마 in the middle-right of the frame, midground; Barbecue grill beside 쿠마 in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized gun (Held steadily and aimed at 쿠마's face) — Seen from the side behind 이현우's hand, pointing diagonally across the frame rather than toward the lens; used as Links the foreground aggressor to the seated target without exaggerated foreshortening; Barbecue grill (Used for cooking chicken skewers) — A partial side and cooking area remain visible near 쿠마; used as Anchors the threat in the interrupted meal and supplies scale behind the gun.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The hideout's barrel fire provides restrained illumination within the nighttime darkness, preserving readable faces and controlled weapon highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The outdoor hideout stands beside the artificial seawall, with fires burning in metal drums and chicken skewers on a barbecue grill. The seawall already has cracks and water seepage, including water running over its supports. 이현우: He wears his upper garment and holds the seized gun raised in an aiming position. His facial injuries and untreated leg bite remain. 쿠마: He remains at the barbecue with beer and a chicken skewer, not yet having discarded the skewer or stood up.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 빼앗은 총구를 쿠마의 얼굴 정면으로 흔들림 없이 겨누고 있는 찰나.\n\nLOCATION (lock): At an open-air gang gathering spot on the roadside beside the artificial seawall, near a burning barrel and barbecue grill. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly arrest the track behind and outside 이현우's left shoulder, near shoulder height, looking obliquely along rather than directly down the weapon's axis. His shoulder and extended arm occupy the lower-left foreground, while 쿠마's seated upper body appears at center-right beyond the gun; 이현우 fixes on 쿠마's face, and 쿠마 looks back toward him without facing the lens. Keep the weapon modest in scale and its muzzle-to-face alignment legible, with a narrow strip of the grilling area establishing their shared space.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the lower-left of the frame, foreground; 쿠마 in the middle-right of the frame, midground; Barbecue grill beside 쿠마 in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Seized gun (Held steadily and aimed at 쿠마's face) — Seen from the side behind 이현우's hand, pointing diagonally across the frame rather than toward the lens; used as Links the foreground aggressor to the seated target without exaggerated foreshortening; Barbecue grill (Used for cooking chicken skewers) — A partial side and cooking area remain visible near 쿠마; used as Anchors the threat in the interrupted meal and supplies scale behind the gun.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The hideout's barrel fire provides restrained illumination within the nighttime darkness, preserving readable faces and controlled weapon highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The outdoor hideout stands beside the artificial seawall, with fires burning in metal drums and chicken skewers on a barbecue grill. The seawall already has cracks and water seepage, including water running over its supports. 이현우: He wears his upper garment and holds the seized gun raised in an aiming position. His facial injuries and untreated leg bite remain. 쿠마: He remains at the barbecue with beer and a chicken skewer, not yet having discarded the skewer or stood up.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh15__bgfirst_bg.png",
     "asset_id": "d6cf8373-98fa-4fca-af9c-4fa06c54c5a5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S13sh15.png",
     "asset_id": "bcb2cb6f-36f1-4be4-9112-e02d4fa88374",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:919321>",
     "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_roadside_hideout_f5126d.png",
     "asset_id": "2426b558-ee67-4180-9aab-b4581b82636c",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:919321>",
     "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 총구는 쿠마를 빗겨나 화면 왼쪽의 바다를 향함. 쿠마는 이현우를 응시함.",
    "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 명시된 위치에 올바르게 배치됨.",
    "entities": "이현우(피 묻은 셔츠, 무전기)와 쿠마(비니, 가죽 재킷) 모두 레퍼런스의 외형 및 의상과 일치함.",
    "hard_violations": [],
    "physics": "인물들은 지면과 상자 위에 안정적으로 지지되어 있으며, 쥐고 있는 총과 맥주병, 꼬치의 형태가 물리적으로 성립함."
   },
   {
    "label": "B",
    "direction": "이현우의 총구는 쿠마의 얼굴을 정확히 겨눔. 쿠마는 이현우를 응시함.",
    "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 올바르게 배치되었으며 인물과의 거리감도 적절함.",
    "entities": "이현우와 쿠마 모두 레퍼런스의 복장과 외형을 잘 유지하고 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 구조 (이현우의 왼팔에 오른손이 연결되어 총을 쥐고 있음)"
    ],
    "physics": "이현우의 총을 쥔 손 구조가 불가능하며, 쿠마가 오른손에 쥔 꼬치가 손가락을 통과하여 융합되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "총구가 목표 대상을 빗겨나 방향성에 아쉬움이 있으나, 치명적인 오류 없이 요구된 프레이밍과 소품 상태를 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "총구의 조준 방향은 정확하나, 왼팔에 오른손이 달려있는 치명적인 해부학적 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 총구는 쿠마를 빗겨나 화면 왼쪽의 바다를 향함. 쿠마는 이현우를 응시함.",
        "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 명시된 위치에 올바르게 배치됨.",
        "entities": "이현우(피 묻은 셔츠, 무전기)와 쿠마(비니, 가죽 재킷) 모두 레퍼런스의 외형 및 의상과 일치함.",
        "hard_violations": [],
        "physics": "인물들은 지면과 상자 위에 안정적으로 지지되어 있으며, 쥐고 있는 총과 맥주병, 꼬치의 형태가 물리적으로 성립함."
       },
       {
        "label": "B",
        "direction": "이현우의 총구는 쿠마의 얼굴을 정확히 겨눔. 쿠마는 이현우를 응시함.",
        "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 올바르게 배치되었으며 인물과의 거리감도 적절함.",
        "entities": "이현우와 쿠마 모두 레퍼런스의 복장과 외형을 잘 유지하고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 구조 (이현우의 왼팔에 오른손이 연결되어 총을 쥐고 있음)"
        ],
        "physics": "이현우의 총을 쥔 손 구조가 불가능하며, 쿠마가 오른손에 쥔 꼬치가 손가락을 통과하여 융합되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "총구가 목표 대상을 빗겨나 방향성에 아쉬움이 있으나, 치명적인 오류 없이 요구된 프레이밍과 소품 상태를 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "총구의 조준 방향은 정확하나, 왼팔에 오른손이 달려있는 치명적인 해부학적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 총구는 쿠마를 빗겨나 화면 왼쪽의 바다를 향함. 쿠마는 이현우를 응시함.",
        "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 명시된 위치에 올바르게 배치됨.",
        "entities": "이현우(피 묻은 셔츠, 무전기)와 쿠마(비니, 가죽 재킷) 모두 레퍼런스의 외형 및 의상과 일치함.",
        "hard_violations": [],
        "physics": "인물들은 지면과 상자 위에 안정적으로 지지되어 있으며, 쥐고 있는 총과 맥주병, 꼬치의 형태가 물리적으로 성립함."
       },
       {
        "label": "B",
        "direction": "이현우의 총구는 쿠마의 얼굴을 정확히 겨눔. 쿠마는 이현우를 응시함.",
        "built_space": "야외 방파제, 드럼통 화로, 바비큐 그릴이 올바르게 배치되었으며 인물과의 거리감도 적절함.",
        "entities": "이현우와 쿠마 모두 레퍼런스의 복장과 외형을 잘 유지하고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 구조 (이현우의 왼팔에 오른손이 연결되어 총을 쥐고 있음)"
        ],
        "physics": "이현우의 총을 쥔 손 구조가 불가능하며, 쿠마가 오른손에 쥔 꼬치가 손가락을 통과하여 융합되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "쿠마의 상체 중심 미디엄 구도는 더 가깝지만, 총구가 쿠마가 아닌 왼쪽 전경을 향해 핵심 조준 동작을 충족하지 못한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "맥주와 꼬치를 든 상태는 정확하나, 총구 방향이 반대이고 무릎과 좌석까지 넓게 보여 지정된 상체 중심 구도에서도 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 쿠마 쪽으로 고개와 팔을 뻗고, 쿠마는 렌즈가 아니라 왼쪽의 이현우를 바라본다. 그러나 권총의 총구로 보이는 검은 개구부가 무기 왼쪽 끝, 카메라와 이현우 쪽에 드러나 있다. 쿠마 쪽에는 슬라이드 뒤쪽이 놓여 있어 총의 높이가 얼굴과 비슷하더라도 실제 총구가 쿠마 얼굴 정면을 겨누지는 않는다. 꼬치는 쿠마의 손에서 오른쪽 위로 뻗는다.",
        "built_space": "왼쪽 컨테이너와 천막, 뒤쪽으로 이어지는 콘크리트 방조제, 바다와 오른쪽 등대가 장소 참조와 대응한다. 불붙은 드럼통은 중앙 뒤쪽 하나와 왼쪽 가장자리 일부 하나, 그릴은 오른쪽 아래 하나, 쿠마의 나무 좌석은 하나 보인다. 쿠마는 그릴 옆에 앉아 있다. 이현우의 어깨와 팔이 왼쪽 전경을 차지하고 쿠마의 상체가 오른쪽 중경에 놓여 B보다 지정 미디엄 구도에 가깝다. 다만 이현우의 머리와 등이 크게 들어오며, 보이는 귀와 뺨은 왼쪽 어깨 바깥보다는 오른쪽 후방 시점으로 읽힌다. 젖은 노면의 불빛 반사는 가능하지만 방조제 지지부를 타고 흐르는 물은 명확하지 않다.",
        "entities": "두 명의 젊은 동아시아계 남성만 보인다. 이현우의 짧고 헝클어진 검은 머리, 마른 팔, 오염된 어두운 셔츠와 인이어 장치는 지시와 맞는다. 얼굴 대부분이 가려져 정확한 얼굴 일치와 얼굴 상처는 확인하기 어렵고, 다리 상처는 프레임 밖이다. 쿠마의 얼굴, 검은 비니와 낡은 검은 가죽 재킷은 참조에 대체로 부합하며 어두운 데님도 보인다. 쿠마는 닭꼬치를 들고 있고 그릴에도 고기가 있다. 옆 탁자에 병들은 있으나 쿠마가 맥주를 손에 든 상태는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "권총은 이현우의 손과 뻗은 팔로 지지되며 공중에 떠 있지 않는다. 다만 보이는 총구와 손잡이의 관계는 정상적인 전방 조준과 맞지 않는다. 쿠마의 엉덩이는 나무 좌석에 놓여 있고 닭꼬치는 손으로 잡고 있다. 그릴의 음식은 석쇠에, 탁자의 병과 컵은 상판에 놓여 있다. 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 얼굴과 뻗은 팔은 쿠마를 향하고 쿠마도 왼쪽의 이현우를 돌아본다. 그러나 권총 왼쪽 끝에 총열 구멍과 그 아래의 원형 부품이 카메라 쪽으로 드러나 있어 총구는 쿠마 반대쪽을 향한다. 오른쪽에 있는 쿠마의 얼굴로 이어지는 조준이 아니다. 쿠마가 든 꼬치는 오른쪽 위로, 맥주병 입구는 위로 향한다.",
        "built_space": "컨테이너와 천막은 왼쪽, 방조제와 바다는 뒤쪽, 등대는 오른쪽 원경에 있다. 불붙은 드럼통 두 개가 중앙 뒤쪽과 왼쪽 가장자리에 보이며, 그릴 하나는 오른쪽 아래, 나무 상자 좌석 하나는 쿠마 아래에 있다. 뒤쪽의 녹색 수납함과 오른쪽 탁자도 장소 참조와 대응한다. 이현우는 왼쪽 전경, 쿠마는 오른쪽 중경이지만 쿠마의 무릎과 좌석까지 보여 A보다 넓은 구도다. 이현우의 보이는 귀와 뺨 역시 지정된 왼쪽 어깨 바깥보다 오른쪽 후방 시점으로 읽힌다. 노면은 젖어 있으나 지지부 위를 흐르는 누수는 식별되지 않는다.",
        "entities": "추가 인물 없이 이현우와 쿠마 두 명이 보인다. 이현우의 헝클어진 검은 머리, 인이어, 피와 먼지가 묻은 어두운 셔츠, 뺨의 상처가 지시와 부합한다. 얼굴이 일부만 보여 정확한 얼굴 동일성은 제한적으로 확인된다. 쿠마는 참조와 비슷한 젊은 얼굴, 검은 비니, 낡은 검은 가죽 재킷과 데님을 갖췄다. 한 손에는 맥주병, 다른 손에는 닭꼬치를 들고 있어 식사 중의 소지 상태는 A보다 명확하다. 오른쪽 그릴에도 꼬치가 놓여 있다.",
        "hard_violations": [],
        "physics": "이현우의 손이 권총을 잡고 팔이 이를 지탱하지만, 총열 앞면이 사수 쪽에 보이므로 실제 사용 방향은 잘못되어 있다. 쿠마의 몸무게는 나무 상자 좌석에 실려 있고 맥주병과 꼬치는 각각 손으로 지지된다. 그릴은 다리로 받쳐지고 꼬치는 석쇠 위에 놓여 있다. 공중에 지지 없이 떠 있는 사람이나 소품은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "쿠마의 상체 중심 미디엄 구도는 더 가깝지만, 총구가 쿠마가 아닌 왼쪽 전경을 향해 핵심 조준 동작을 충족하지 못한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "맥주와 꼬치를 든 상태는 정확하나, 총구 방향이 반대이고 무릎과 좌석까지 넓게 보여 지정된 상체 중심 구도에서도 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 쿠마 쪽으로 고개와 팔을 뻗고, 쿠마는 렌즈가 아니라 왼쪽의 이현우를 바라본다. 그러나 권총의 총구로 보이는 검은 개구부가 무기 왼쪽 끝, 카메라와 이현우 쪽에 드러나 있다. 쿠마 쪽에는 슬라이드 뒤쪽이 놓여 있어 총의 높이가 얼굴과 비슷하더라도 실제 총구가 쿠마 얼굴 정면을 겨누지는 않는다. 꼬치는 쿠마의 손에서 오른쪽 위로 뻗는다.",
        "built_space": "왼쪽 컨테이너와 천막, 뒤쪽으로 이어지는 콘크리트 방조제, 바다와 오른쪽 등대가 장소 참조와 대응한다. 불붙은 드럼통은 중앙 뒤쪽 하나와 왼쪽 가장자리 일부 하나, 그릴은 오른쪽 아래 하나, 쿠마의 나무 좌석은 하나 보인다. 쿠마는 그릴 옆에 앉아 있다. 이현우의 어깨와 팔이 왼쪽 전경을 차지하고 쿠마의 상체가 오른쪽 중경에 놓여 B보다 지정 미디엄 구도에 가깝다. 다만 이현우의 머리와 등이 크게 들어오며, 보이는 귀와 뺨은 왼쪽 어깨 바깥보다는 오른쪽 후방 시점으로 읽힌다. 젖은 노면의 불빛 반사는 가능하지만 방조제 지지부를 타고 흐르는 물은 명확하지 않다.",
        "entities": "두 명의 젊은 동아시아계 남성만 보인다. 이현우의 짧고 헝클어진 검은 머리, 마른 팔, 오염된 어두운 셔츠와 인이어 장치는 지시와 맞는다. 얼굴 대부분이 가려져 정확한 얼굴 일치와 얼굴 상처는 확인하기 어렵고, 다리 상처는 프레임 밖이다. 쿠마의 얼굴, 검은 비니와 낡은 검은 가죽 재킷은 참조에 대체로 부합하며 어두운 데님도 보인다. 쿠마는 닭꼬치를 들고 있고 그릴에도 고기가 있다. 옆 탁자에 병들은 있으나 쿠마가 맥주를 손에 든 상태는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "권총은 이현우의 손과 뻗은 팔로 지지되며 공중에 떠 있지 않는다. 다만 보이는 총구와 손잡이의 관계는 정상적인 전방 조준과 맞지 않는다. 쿠마의 엉덩이는 나무 좌석에 놓여 있고 닭꼬치는 손으로 잡고 있다. 그릴의 음식은 석쇠에, 탁자의 병과 컵은 상판에 놓여 있다. 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 얼굴과 뻗은 팔은 쿠마를 향하고 쿠마도 왼쪽의 이현우를 돌아본다. 그러나 권총 왼쪽 끝에 총열 구멍과 그 아래의 원형 부품이 카메라 쪽으로 드러나 있어 총구는 쿠마 반대쪽을 향한다. 오른쪽에 있는 쿠마의 얼굴로 이어지는 조준이 아니다. 쿠마가 든 꼬치는 오른쪽 위로, 맥주병 입구는 위로 향한다.",
        "built_space": "컨테이너와 천막은 왼쪽, 방조제와 바다는 뒤쪽, 등대는 오른쪽 원경에 있다. 불붙은 드럼통 두 개가 중앙 뒤쪽과 왼쪽 가장자리에 보이며, 그릴 하나는 오른쪽 아래, 나무 상자 좌석 하나는 쿠마 아래에 있다. 뒤쪽의 녹색 수납함과 오른쪽 탁자도 장소 참조와 대응한다. 이현우는 왼쪽 전경, 쿠마는 오른쪽 중경이지만 쿠마의 무릎과 좌석까지 보여 A보다 넓은 구도다. 이현우의 보이는 귀와 뺨 역시 지정된 왼쪽 어깨 바깥보다 오른쪽 후방 시점으로 읽힌다. 노면은 젖어 있으나 지지부 위를 흐르는 누수는 식별되지 않는다.",
        "entities": "추가 인물 없이 이현우와 쿠마 두 명이 보인다. 이현우의 헝클어진 검은 머리, 인이어, 피와 먼지가 묻은 어두운 셔츠, 뺨의 상처가 지시와 부합한다. 얼굴이 일부만 보여 정확한 얼굴 동일성은 제한적으로 확인된다. 쿠마는 참조와 비슷한 젊은 얼굴, 검은 비니, 낡은 검은 가죽 재킷과 데님을 갖췄다. 한 손에는 맥주병, 다른 손에는 닭꼬치를 들고 있어 식사 중의 소지 상태는 A보다 명확하다. 오른쪽 그릴에도 꼬치가 놓여 있다.",
        "hard_violations": [],
        "physics": "이현우의 손이 권총을 잡고 팔이 이를 지탱하지만, 총열 앞면이 사수 쪽에 보이므로 실제 사용 방향은 잘못되어 있다. 쿠마의 몸무게는 나무 상자 좌석에 실려 있고 맥주병과 꼬치는 각각 손으로 지지된다. 그릴은 다리로 받쳐지고 꼬치는 석쇠 위에 놓여 있다. 공중에 지지 없이 떠 있는 사람이나 소품은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 구조 (이현우의 왼팔에 오른손이 연결되어 총을 쥐고 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "총구가 목표 대상을 빗겨나 방향성에 아쉬움이 있으나, 치명적인 오류 없이 요구된 프레이밍과 소품 상태를 준수함."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "총구의 조준 방향은 정확하나, 왼팔에 오른손이 달려있는 치명적인 해부학적 오류가 발생함.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 구조 (이현우의 왼팔에 오른손이 연결되어 총을 쥐고 있음)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_roadside_hideout_f5126d.png",
    "asset_id": "2426b558-ee67-4180-9aab-b4581b82636c",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:919321>",
    "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab088a-e961-737f-ad45-cfe49f6b5d2d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh15__bgfirst_bg.png",
   "bg_asset_id": "d6cf8373-98fa-4fca-af9c-4fa06c54c5a5",
   "bg_record_key": "S13sh15::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "roadside_hideout",
   "groupbg_asset_id": "2426b558-ee67-4180-9aab-b4581b82636c"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S13sh18::signage": {
  "fp": "81546a95097a9ee9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::seawall_base": {
  "input_fingerprint": "01297b71cd110844",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "seawall_base",
    "tags": [
     "S13sh18",
     "S13sh25"
    ]
   },
   "context_sig": "f80ad5b97ec5a782"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쿠마, 현우를 아랑곳하지 않고 인공제방 쪽으로 성큼성큼 앞서 걷는다.\n- 어둠 속이지만 여기저기 제방 갈라진 틈이 보인다. 지지대에도 물이 흐르고.\n\nTIME OF DAY (lock): night to morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 쿠마, 현우를 아랑곳하지 않고 인공제방 쪽으로 성큼성큼 앞서 걷는다.\n- 어둠 속이지만 여기저기 제방 갈라진 틈이 보인다. 지지대에도 물이 흐르고.\n\nTIME OF DAY (lock): night to morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_base_fb79a5.png",
  "asset_id": "e1f2cdda-70eb-46cc-accb-dc455aeb82fd",
  "input_asset_ids": [
   "5b319b3b-c9ce-48f5-9050-3a1690e77aed"
  ],
  "origin_tag": "S13sh18",
  "place_text": "Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.",
  "origin_inputs": {
   "place_text": "Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.",
   "time_of_day_en": "night to morning",
   "conti_asset_id": "5b319b3b-c9ce-48f5-9050-3a1690e77aed"
  }
 },
 "S13sh18::bgfirst_bg": {
  "input_fingerprint": "b5b94385eae4c8ba",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갈라진 틈새로 물이 흐르는 어두운 인공제방 벽면 앞에 나란히 선 쿠마와 이현우.\n\nLOCATION (lock): Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.\n\nTIME OF DAY (lock): night to morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use the arrival position on the embankment's open side, elevated above both men and angled gently down along the wall, before 쿠마 reaches toward it. Place 쿠마 left of 이현우 across the lower-middle frame, their full figures side by side but caught at different stopping phases; both lower their attention toward the water emerging from cracks near them. Make the expanded camera distance the principal change, distributing the wall, supports and open ground around the figures rather than treating the wall as a flat backdrop.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Cracked, with water flowing through gaps) — Its exposed face recedes diagonally behind the two men; used as Turns the visible structural failure into the subject of their shared attention; Embankment supports (Water is running over them) — Visible obliquely beneath and beside the wall; used as Provides depth and a second visible indication of the leakage; Ground beside the embankment (The two men have arrived on its open side); used as Leaves negative space around their feet and prevents the wall from flattening their positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the scene's nighttime darkness while keeping the cracks and flowing water just readable through restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갈라진 틈새로 물이 흐르는 어두운 인공제방 벽면 앞에 나란히 선 쿠마와 이현우.\n\nLOCATION (lock): Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures.\n\nTIME OF DAY (lock): night to morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use the arrival position on the embankment's open side, elevated above both men and angled gently down along the wall, before 쿠마 reaches toward it. Place 쿠마 left of 이현우 across the lower-middle frame, their full figures side by side but caught at different stopping phases; both lower their attention toward the water emerging from cracks near them. Make the expanded camera distance the principal change, distributing the wall, supports and open ground around the figures rather than treating the wall as a flat backdrop.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Cracked, with water flowing through gaps) — Its exposed face recedes diagonally behind the two men; used as Turns the visible structural failure into the subject of their shared attention; Embankment supports (Water is running over them) — Visible obliquely beneath and beside the wall; used as Provides depth and a second visible indication of the leakage; Ground beside the embankment (The two men have arrived on its open side); used as Leaves negative space around their feet and prevents the wall from flattening their positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the scene's nighttime darkness while keeping the cracks and flowing water just readable through restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh18__bgfirst_bg.png",
  "asset_id": "437ed841-9bfd-4784-b9f6-b75c9b1c17f2",
  "input_asset_ids": [
   "5b319b3b-c9ce-48f5-9050-3a1690e77aed",
   "e1f2cdda-70eb-46cc-accb-dc455aeb82fd"
  ]
 },
 "S13sh18": {
  "input_fingerprint": "9844b29fa7162ec7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 갈라진 틈새로 물이 흐르는 어두운 인공제방 벽면 앞에 나란히 선 쿠마와 이현우.\n\nLOCATION (lock): Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use the arrival position on the embankment's open side, elevated above both men and angled gently down along the wall, before 쿠마 reaches toward it. Place 쿠마 left of 이현우 across the lower-middle frame, their full figures side by side but caught at different stopping phases; both lower their attention toward the water emerging from cracks near them. Make the expanded camera distance the principal change, distributing the wall, supports and open ground around the figures rather than treating the wall as a flat backdrop.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Cracked, with water flowing through gaps) — Its exposed face recedes diagonally behind the two men; used as Turns the visible structural failure into the subject of their shared attention; Embankment supports (Water is running over them) — Visible obliquely beneath and beside the wall; used as Provides depth and a second visible indication of the leakage; Ground beside the embankment (The two men have arrived on its open side); used as Leaves negative space around their feet and prevents the wall from flattening their positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the scene's nighttime darkness while keeping the cracks and flowing water just readable through restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall is cracked in multiple places, with seawater seeping through and running over its supports in the darkness. The metal-drum fires and barbecue remain at the nearby hideout. 이현우: He stands by the seawall in his upper garment, retaining the seized gun. His facial injuries and untreated leg bite remain. 쿠마: He has left the barbecue and stands beside the seawall without the discarded chicken skewer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 갈라진 틈새로 물이 흐르는 어두운 인공제방 벽면 앞에 나란히 선 쿠마와 이현우.\n\nLOCATION (lock): Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use the arrival position on the embankment's open side, elevated above both men and angled gently down along the wall, before 쿠마 reaches toward it. Place 쿠마 left of 이현우 across the lower-middle frame, their full figures side by side but caught at different stopping phases; both lower their attention toward the water emerging from cracks near them. Make the expanded camera distance the principal change, distributing the wall, supports and open ground around the figures rather than treating the wall as a flat backdrop.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Cracked, with water flowing through gaps) — Its exposed face recedes diagonally behind the two men; used as Turns the visible structural failure into the subject of their shared attention; Embankment supports (Water is running over them) — Visible obliquely beneath and beside the wall; used as Provides depth and a second visible indication of the leakage; Ground beside the embankment (The two men have arrived on its open side); used as Leaves negative space around their feet and prevents the wall from flattening their positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the scene's nighttime darkness while keeping the cracks and flowing water just readable through restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall is cracked in multiple places, with seawater seeping through and running over its supports in the darkness. The metal-drum fires and barbecue remain at the nearby hideout. 이현우: He stands by the seawall in his upper garment, retaining the seized gun. His facial injuries and untreated leg bite remain. 쿠마: He has left the barbecue and stands beside the seawall without the discarded chicken skewer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 갈라진 틈새로 물이 흐르는 어두운 인공제방 벽면 앞에 나란히 선 쿠마와 이현우.\n\nLOCATION (lock): Immediately in front of the cracked artificial-seawall wall at night, where seawater seeps through the fissures. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use the arrival position on the embankment's open side, elevated above both men and angled gently down along the wall, before 쿠마 reaches toward it. Place 쿠마 left of 이현우 across the lower-middle frame, their full figures side by side but caught at different stopping phases; both lower their attention toward the water emerging from cracks near them. Make the expanded camera distance the principal change, distributing the wall, supports and open ground around the figures rather than treating the wall as a flat backdrop.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Artificial embankment wall (Cracked, with water flowing through gaps) — Its exposed face recedes diagonally behind the two men; used as Turns the visible structural failure into the subject of their shared attention; Embankment supports (Water is running over them) — Visible obliquely beneath and beside the wall; used as Provides depth and a second visible indication of the leakage; Ground beside the embankment (The two men have arrived on its open side); used as Leaves negative space around their feet and prevents the wall from flattening their positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the scene's nighttime darkness while keeping the cracks and flowing water just readable through restrained exposure and controlled contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seawall is cracked in multiple places, with seawater seeping through and running over its supports in the darkness. The metal-drum fires and barbecue remain at the nearby hideout. 이현우: He stands by the seawall in his upper garment, retaining the seized gun. His facial injuries and untreated leg bite remain. 쿠마: He has left the barbecue and stands beside the seawall without the discarded chicken skewer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh18__bgfirst_bg.png",
     "asset_id": "437ed841-9bfd-4784-b9f6-b75c9b1c17f2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S13sh18.png",
     "asset_id": "5b319b3b-c9ce-48f5-9050-3a1690e77aed",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:919321>",
     "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_base_fb79a5.png",
     "asset_id": "e1f2cdda-70eb-46cc-accb-dc455aeb82fd",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:919321>",
     "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 시선을 아래로 향해 방파제 벽면 틈새에서 나오는 물을 바라봄.",
    "built_space": "우측에 방파제, 좌측에 넓은 평지가 있는 레퍼런스 구조를 정확히 따랐으며 인물들이 평지에 안정적으로 위치함.",
    "entities": "좌측 쿠마(가죽 재킷, 비니)와 우측 이현우(오염된 옷, 총기 파지) 모두 프롬프트 및 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "인물들은 바닥을 단단히 딛고 있으며 중력이 올바르게 작용함. 손에 쥔 총기와 벽면을 타고 흐르는 물이 자연스러움."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 방파제 벽면의 틈새에서 흐르는 물을 향해 시선을 내림.",
    "built_space": "방파제와 바닥, 바다의 배치가 기준 사진과 일치하며 요구된 카메라 각도를 반영함.",
    "entities": "쿠마와 이현우의 위치와 복장은 지시를 따랐으나, 우측 이현우의 얼굴과 형태가 다소 뭉개짐.",
    "hard_violations": [],
    "physics": "인물들의 기립 자세와 파지는 정상적이나, 방파제 너머 테트라포드의 형태가 무너져 있고 물줄기가 평면적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "제시된 넓은 카메라 앵글, 인물 배치, 물이 흐르는 방파제 환경 디테일을 가장 사실적이고 영화적으로 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 앵글은 지시를 따랐으나, 인물의 얼굴과 배경 테트라포드의 사실감이 다소 떨어짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 시선을 아래로 향해 방파제 벽면 틈새에서 나오는 물을 바라봄.",
        "built_space": "우측에 방파제, 좌측에 넓은 평지가 있는 레퍼런스 구조를 정확히 따랐으며 인물들이 평지에 안정적으로 위치함.",
        "entities": "좌측 쿠마(가죽 재킷, 비니)와 우측 이현우(오염된 옷, 총기 파지) 모두 프롬프트 및 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "인물들은 바닥을 단단히 딛고 있으며 중력이 올바르게 작용함. 손에 쥔 총기와 벽면을 타고 흐르는 물이 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 방파제 벽면의 틈새에서 흐르는 물을 향해 시선을 내림.",
        "built_space": "방파제와 바닥, 바다의 배치가 기준 사진과 일치하며 요구된 카메라 각도를 반영함.",
        "entities": "쿠마와 이현우의 위치와 복장은 지시를 따랐으나, 우측 이현우의 얼굴과 형태가 다소 뭉개짐.",
        "hard_violations": [],
        "physics": "인물들의 기립 자세와 파지는 정상적이나, 방파제 너머 테트라포드의 형태가 무너져 있고 물줄기가 평면적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "제시된 넓은 카메라 앵글, 인물 배치, 물이 흐르는 방파제 환경 디테일을 가장 사실적이고 영화적으로 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 앵글은 지시를 따랐으나, 인물의 얼굴과 배경 테트라포드의 사실감이 다소 떨어짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 시선을 아래로 향해 방파제 벽면 틈새에서 나오는 물을 바라봄.",
        "built_space": "우측에 방파제, 좌측에 넓은 평지가 있는 레퍼런스 구조를 정확히 따랐으며 인물들이 평지에 안정적으로 위치함.",
        "entities": "좌측 쿠마(가죽 재킷, 비니)와 우측 이현우(오염된 옷, 총기 파지) 모두 프롬프트 및 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "인물들은 바닥을 단단히 딛고 있으며 중력이 올바르게 작용함. 손에 쥔 총기와 벽면을 타고 흐르는 물이 자연스러움."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 방파제 벽면의 틈새에서 흐르는 물을 향해 시선을 내림.",
        "built_space": "방파제와 바닥, 바다의 배치가 기준 사진과 일치하며 요구된 카메라 각도를 반영함.",
        "entities": "쿠마와 이현우의 위치와 복장은 지시를 따랐으나, 우측 이현우의 얼굴과 형태가 다소 뭉개짐.",
        "hard_violations": [],
        "physics": "인물들의 기립 자세와 파지는 정상적이나, 방파제 너머 테트라포드의 형태가 무너져 있고 물줄기가 평면적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "더 멀고 높은 개방면 시점에서 두 사람을 하단 중간에 놓고, 서로 다른 정지 자세와 누수·지지대·빈 지면의 깊이를 더 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물 순서와 누수를 내려다보는 행동은 맞지만, A보다 인물이 크고 두 사람의 정지 자세가 비슷하여 거리 확장과 서로 다른 도착 단계의 표현이 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 쿠마는 고개를 오른쪽 아래로 숙여 벽 밑 물길 쪽을 보고, 오른쪽 이현우도 가까운 균열 아래로 흐르는 물을 내려다본다. 이현우의 오른손 권총은 총구가 지면을 향하며 누구도 겨누지 않는다. 쿠마는 아직 벽에 손을 뻗지 않았다.",
        "built_space": "제방 벽 한 줄이 오른쪽 전경에서 왼쪽 원경으로 물러나고, 그 밑에 삼각형 지지대가 10개 이상 반복된다. 원경에서는 겹쳐 정확한 총수를 구분하기 어렵다. 벽 뒤에는 소파블록과 바다가 있고, 왼쪽에는 가로등 1개, 먼 등대 1개와 기존 적치물이 보인다. 두 사람은 지지대 바깥의 열린 포장면에 서 있다. 높은 사선 시점과 넓은 발밑 여백이 벽·지지대·통로의 입체 관계를 드러낸다. 젖은 바닥의 빛 반사도 광원 위치와 양립한다.",
        "entities": "인물은 두 명뿐이다. 왼쪽 쿠마는 젊은 동아시아계 남성으로 보이며 참고의 비니, 낡은 검은 가죽 재킷, 어두운 데님과 부츠를 유지한다. 오른쪽 이현우는 짧고 흐트러진 검은 머리의 마른 젊은 동아시아계 남성으로, 오염된 어두운 셔츠와 바지 및 권총이 보인다. 작은 얼굴의 정확한 동일성, 혼혈·국적, 인이어 무전기는 확정하기 어렵다. 얼굴 부상은 뚜렷하지 않고 다리 물린 상처는 바지에 가려 확인되지 않는다. 다중 균열, 흘러내리는 물, 콘크리트 지지대가 있으며 꼬치나 바비큐·불은 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 부츠가 포장면에 닿아 체중을 지탱한다. 쿠마는 몸을 벽 쪽으로 돌리고 발을 앞뒤로 놓았으며, 이현우는 더 벌어진 발로 멈춰 있어 정지 단계에 차이가 있다. 권총은 이현우의 오른손에 잡혀 있다. 물은 균열에서 아래로 떨어져 지지대를 타고 바닥에 고이며, 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "쿠마와 이현우 모두 오른쪽 아래의 누수와 벽 밑 물길을 향해 고개를 숙였다. 이현우가 오른손에 든 권총의 총구는 발 옆 지면을 향한다. 쿠마의 두 손은 내려와 있어 벽을 만지기 전이라는 조건에 맞는다.",
        "built_space": "하나의 제방 벽이 오른쪽 전경에서 왼쪽 원경으로 이어진다. 가까운 삼각 지지대 3개가 특히 분명하고 뒤로 10개 이상의 지지대가 이어지며, 먼 부분은 겹쳐 총수를 확정하기 어렵다. 벽 뒤 소파블록·바다, 왼쪽 가로등 1개와 먼 등대 1개, 통로 가장자리 적치물이 참고 장소와 대응한다. 두 사람은 열린 통로에 정상적으로 서 있다. 사선으로 내려다보는 와이드 구도이지만 A보다 인물이 크고 시점이 낮아, 확장된 거리와 벽 상면의 깊이가 덜 강조된다.",
        "entities": "추가 인물 없이 쿠마가 왼쪽, 이현우가 오른쪽에 있다. 쿠마의 비니와 낡은 가죽 재킷·데님, 이현우의 헝클어진 검은 머리와 마른 체격·피와 먼지가 묻은 셔츠·바지는 참고와 대체로 맞는다. 둘 다 젊은 동아시아계 남성으로 보이나 세부 혈통과 국적은 외형만으로 확인할 수 없다. 이현우의 권총은 보이고, 인이어와 얼굴 부상은 명확하지 않다. 다리 상처는 바지에 가려져 있다. 균열과 누수, 지지대는 있으며 버린 꼬치나 은신처의 불·바비큐는 등장하지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 모두 바닥에 접촉하고 권총은 손으로 지지된다. 두 사람 모두 발을 벌리고 팔을 아래로 둔 비슷한 자세여서 서로 다른 도착·정지 단계는 약하게 드러난다. 다만 자세 자체는 물리적으로 가능하다. 균열의 물은 중력 방향으로 벽과 지지대를 타고 흘러 바닥에 모이며, 떠 있는 신체나 무지지 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "더 멀고 높은 개방면 시점에서 두 사람을 하단 중간에 놓고, 서로 다른 정지 자세와 누수·지지대·빈 지면의 깊이를 더 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "인물 순서와 누수를 내려다보는 행동은 맞지만, A보다 인물이 크고 두 사람의 정지 자세가 비슷하여 거리 확장과 서로 다른 도착 단계의 표현이 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 쿠마는 고개를 오른쪽 아래로 숙여 벽 밑 물길 쪽을 보고, 오른쪽 이현우도 가까운 균열 아래로 흐르는 물을 내려다본다. 이현우의 오른손 권총은 총구가 지면을 향하며 누구도 겨누지 않는다. 쿠마는 아직 벽에 손을 뻗지 않았다.",
        "built_space": "제방 벽 한 줄이 오른쪽 전경에서 왼쪽 원경으로 물러나고, 그 밑에 삼각형 지지대가 10개 이상 반복된다. 원경에서는 겹쳐 정확한 총수를 구분하기 어렵다. 벽 뒤에는 소파블록과 바다가 있고, 왼쪽에는 가로등 1개, 먼 등대 1개와 기존 적치물이 보인다. 두 사람은 지지대 바깥의 열린 포장면에 서 있다. 높은 사선 시점과 넓은 발밑 여백이 벽·지지대·통로의 입체 관계를 드러낸다. 젖은 바닥의 빛 반사도 광원 위치와 양립한다.",
        "entities": "인물은 두 명뿐이다. 왼쪽 쿠마는 젊은 동아시아계 남성으로 보이며 참고의 비니, 낡은 검은 가죽 재킷, 어두운 데님과 부츠를 유지한다. 오른쪽 이현우는 짧고 흐트러진 검은 머리의 마른 젊은 동아시아계 남성으로, 오염된 어두운 셔츠와 바지 및 권총이 보인다. 작은 얼굴의 정확한 동일성, 혼혈·국적, 인이어 무전기는 확정하기 어렵다. 얼굴 부상은 뚜렷하지 않고 다리 물린 상처는 바지에 가려 확인되지 않는다. 다중 균열, 흘러내리는 물, 콘크리트 지지대가 있으며 꼬치나 바비큐·불은 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 부츠가 포장면에 닿아 체중을 지탱한다. 쿠마는 몸을 벽 쪽으로 돌리고 발을 앞뒤로 놓았으며, 이현우는 더 벌어진 발로 멈춰 있어 정지 단계에 차이가 있다. 권총은 이현우의 오른손에 잡혀 있다. 물은 균열에서 아래로 떨어져 지지대를 타고 바닥에 고이며, 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "쿠마와 이현우 모두 오른쪽 아래의 누수와 벽 밑 물길을 향해 고개를 숙였다. 이현우가 오른손에 든 권총의 총구는 발 옆 지면을 향한다. 쿠마의 두 손은 내려와 있어 벽을 만지기 전이라는 조건에 맞는다.",
        "built_space": "하나의 제방 벽이 오른쪽 전경에서 왼쪽 원경으로 이어진다. 가까운 삼각 지지대 3개가 특히 분명하고 뒤로 10개 이상의 지지대가 이어지며, 먼 부분은 겹쳐 총수를 확정하기 어렵다. 벽 뒤 소파블록·바다, 왼쪽 가로등 1개와 먼 등대 1개, 통로 가장자리 적치물이 참고 장소와 대응한다. 두 사람은 열린 통로에 정상적으로 서 있다. 사선으로 내려다보는 와이드 구도이지만 A보다 인물이 크고 시점이 낮아, 확장된 거리와 벽 상면의 깊이가 덜 강조된다.",
        "entities": "추가 인물 없이 쿠마가 왼쪽, 이현우가 오른쪽에 있다. 쿠마의 비니와 낡은 가죽 재킷·데님, 이현우의 헝클어진 검은 머리와 마른 체격·피와 먼지가 묻은 셔츠·바지는 참고와 대체로 맞는다. 둘 다 젊은 동아시아계 남성으로 보이나 세부 혈통과 국적은 외형만으로 확인할 수 없다. 이현우의 권총은 보이고, 인이어와 얼굴 부상은 명확하지 않다. 다리 상처는 바지에 가려져 있다. 균열과 누수, 지지대는 있으며 버린 꼬치나 은신처의 불·바비큐는 등장하지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 신발은 모두 바닥에 접촉하고 권총은 손으로 지지된다. 두 사람 모두 발을 벌리고 팔을 아래로 둔 비슷한 자세여서 서로 다른 도착·정지 단계는 약하게 드러난다. 다만 자세 자체는 물리적으로 가능하다. 균열의 물은 중력 방향으로 벽과 지지대를 타고 흘러 바닥에 모이며, 떠 있는 신체나 무지지 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "제시된 넓은 카메라 앵글, 인물 배치, 물이 흐르는 방파제 환경 디테일을 가장 사실적이고 영화적으로 구현함."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "전반적인 구도와 앵글은 지시를 따랐으나, 인물의 얼굴과 배경 테트라포드의 사실감이 다소 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_seawall_base_fb79a5.png",
    "asset_id": "e1f2cdda-70eb-46cc-accb-dc455aeb82fd",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:919321>",
    "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0895-0944-7988-93b3-6f7a411667cc",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh18__bgfirst_bg.png",
   "bg_asset_id": "437ed841-9bfd-4784-b9f6-b75c9b1c17f2",
   "bg_record_key": "S13sh18::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "seawall_base",
   "groupbg_asset_id": "e1f2cdda-70eb-46cc-accb-dc455aeb82fd"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S13sh25::signage": {
  "fp": "19b7ca53d3645089",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh25": {
  "input_fingerprint": "1b2a0019005da76f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 내민 돈뭉치를 이현우의 손이 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): At the foot of the leaking artificial seawall, outside the nearby roadside gathering area at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the inward dolly settle on the same open side, looking obliquely downward from below chest height toward the exchange. 이현우's hand enters from the right and closes tightly around the money at lower center, while 쿠마's offering hand remains at the left edge; retain cropped torsos and a soft portion of the embankment so the notes occupy less than a third of the image. Their faces remain above the crop as their attention drops to the exchange, making camera distance—not a new interaction axis—the emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bundle of money (Being seized by 이현우 while still touching 쿠마's offering hand) — Seen obliquely between overlapping fingers, without emphasizing denomination or printed details; used as Marks the precise transfer of possession at the center of the hand action; Artificial embankment wall (Cracked and leaking) — An oblique portion remains softly visible behind the cropped bodies; used as Maintains spatial continuity with the preceding conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the embankment's subdued nighttime exposure, separating fingers and notes gently without adding a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 쿠마 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked barrier surface, seeping water, and nighttime lighting from the reference. Exclude the barbecue grill and burning drums from the separate gathering area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water continues to seep through the cracked seawall and run over its supports in the darkness. 이현우: He now holds the cash payment and retains the seized gun. His upper garment is on, and his facial injuries and leg bite remain. 쿠마: He stands beside the seawall with the hand used to wipe its surface still wet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마, 이현우 right now, so 쿠마, 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마, 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 내민 돈뭉치를 이현우의 손이 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): At the foot of the leaking artificial seawall, outside the nearby roadside gathering area at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the inward dolly settle on the same open side, looking obliquely downward from below chest height toward the exchange. 이현우's hand enters from the right and closes tightly around the money at lower center, while 쿠마's offering hand remains at the left edge; retain cropped torsos and a soft portion of the embankment so the notes occupy less than a third of the image. Their faces remain above the crop as their attention drops to the exchange, making camera distance—not a new interaction axis—the emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bundle of money (Being seized by 이현우 while still touching 쿠마's offering hand) — Seen obliquely between overlapping fingers, without emphasizing denomination or printed details; used as Marks the precise transfer of possession at the center of the hand action; Artificial embankment wall (Cracked and leaking) — An oblique portion remains softly visible behind the cropped bodies; used as Maintains spatial continuity with the preceding conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the embankment's subdued nighttime exposure, separating fingers and notes gently without adding a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 쿠마 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked barrier surface, seeping water, and nighttime lighting from the reference. Exclude the barbecue grill and burning drums from the separate gathering area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water continues to seep through the cracked seawall and run over its supports in the darkness. 이현우: He now holds the cash payment and retains the seized gun. His upper garment is on, and his facial injuries and leg bite remain. 쿠마: He stands beside the seawall with the hand used to wipe its surface still wet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마, 이현우 right now, so 쿠마, 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마, 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to morning.\n\nSHOT TEXT (authoritative, Korean): 쿠마가 내민 돈뭉치를 이현우의 손이 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): At the foot of the leaking artificial seawall, outside the nearby roadside gathering area at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the inward dolly settle on the same open side, looking obliquely downward from below chest height toward the exchange. 이현우's hand enters from the right and closes tightly around the money at lower center, while 쿠마's offering hand remains at the left edge; retain cropped torsos and a soft portion of the embankment so the notes occupy less than a third of the image. Their faces remain above the crop as their attention drops to the exchange, making camera distance—not a new interaction axis—the emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bundle of money (Being seized by 이현우 while still touching 쿠마's offering hand) — Seen obliquely between overlapping fingers, without emphasizing denomination or printed details; used as Marks the precise transfer of possession at the center of the hand action; Artificial embankment wall (Cracked and leaking) — An oblique portion remains softly visible behind the cropped bodies; used as Maintains spatial continuity with the preceding conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the embankment's subdued nighttime exposure, separating fingers and notes gently without adding a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 쿠마 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cracked barrier surface, seeping water, and nighttime lighting from the reference. Exclude the barbecue grill and burning drums from the separate gathering area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Water continues to seep through the cracked seawall and run over its supports in the darkness. 이현우: He now holds the cash payment and retains the seized gun. His upper garment is on, and his facial injuries and leg bite remain. 쿠마: He stands beside the seawall with the hand used to wipe its surface still wet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 쿠마, 이현우 right now, so 쿠마, 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 쿠마, 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 가슴 아래 높이에서 비스듬히 아래를 향해 두 사람의 손이 교차하는 중심부를 정확히 비추고 있음.",
    "built_space": "금이 가고 물이 새어 나오는 야간의 인공 방파제가 인물들 뒤로 적절한 원근감과 함께 아웃포커싱되어 배치됨.",
    "entities": "왼쪽은 가죽 재킷을 입은 쿠마의 몸통과 물기가 남은 오른손, 오른쪽은 흙먼지가 묻은 셔츠를 입은 이현우의 몸통과 돈을 움켜쥐는 손, 중앙에는 돈뭉치가 명확히 묘사됨.",
    "hard_violations": [],
    "physics": "두 인물의 팔이 자연스럽게 뻗어 나와 허공에서 돈뭉치를 단단히 쥐고 건네는 물리적 접촉과 지탱이 사실적으로 표현됨."
   },
   {
    "label": "B",
    "direction": "카메라가 아래를 향하며 돈을 건네받는 손을 근접 촬영하여 시선 방향과 프레이밍 의도에 부합함.",
    "built_space": "야간의 방파제와 흘러내리는 물줄기가 프롬프트의 위치 설정에 맞게 배경에 올바르게 자리 잡고 있음.",
    "entities": "왼쪽 인물(쿠마)의 복장과 손, 돈뭉치는 일치하나 오른쪽 인물(이현우)의 신체 구조에 심각한 오류가 발생함.",
    "hard_violations": [
     "[gemini-pro] 오른쪽 인물(이현우)의 어깨에서 이어진 팔이 프레임 우측 하단으로 늘어져 있음에도, 가슴 부위에서 동일한 소매를 입은 또 다른 팔이 튀어나와 돈을 잡고 있는 물리적으로 불가능한 신체 구조(여분의 팔)"
    ],
    "physics": "왼쪽 인물의 손은 자연스럽게 돈을 들고 있으나, 오른쪽 인물은 비정상적인 위치에서 생성된 팔로 돈을 쥐고 있어 물리적 자세가 성립하지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항에 따라 크롭된 구도 내에서 물에 젖은 쿠마의 손과 이현우가 돈을 쥐는 순간을 정확하고 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "오른쪽 인물의 팔이 아래로 향해 있음에도 몸통 중앙에서 돈을 쥐는 또 다른 팔이 생성된 치명적인 신체 오류가 있어 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 가슴 아래 높이에서 비스듬히 아래를 향해 두 사람의 손이 교차하는 중심부를 정확히 비추고 있음.",
        "built_space": "금이 가고 물이 새어 나오는 야간의 인공 방파제가 인물들 뒤로 적절한 원근감과 함께 아웃포커싱되어 배치됨.",
        "entities": "왼쪽은 가죽 재킷을 입은 쿠마의 몸통과 물기가 남은 오른손, 오른쪽은 흙먼지가 묻은 셔츠를 입은 이현우의 몸통과 돈을 움켜쥐는 손, 중앙에는 돈뭉치가 명확히 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물의 팔이 자연스럽게 뻗어 나와 허공에서 돈뭉치를 단단히 쥐고 건네는 물리적 접촉과 지탱이 사실적으로 표현됨."
       },
       {
        "label": "B",
        "direction": "카메라가 아래를 향하며 돈을 건네받는 손을 근접 촬영하여 시선 방향과 프레이밍 의도에 부합함.",
        "built_space": "야간의 방파제와 흘러내리는 물줄기가 프롬프트의 위치 설정에 맞게 배경에 올바르게 자리 잡고 있음.",
        "entities": "왼쪽 인물(쿠마)의 복장과 손, 돈뭉치는 일치하나 오른쪽 인물(이현우)의 신체 구조에 심각한 오류가 발생함.",
        "hard_violations": [
         "오른쪽 인물(이현우)의 어깨에서 이어진 팔이 프레임 우측 하단으로 늘어져 있음에도, 가슴 부위에서 동일한 소매를 입은 또 다른 팔이 튀어나와 돈을 잡고 있는 물리적으로 불가능한 신체 구조(여분의 팔)"
        ],
        "physics": "왼쪽 인물의 손은 자연스럽게 돈을 들고 있으나, 오른쪽 인물은 비정상적인 위치에서 생성된 팔로 돈을 쥐고 있어 물리적 자세가 성립하지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항에 따라 크롭된 구도 내에서 물에 젖은 쿠마의 손과 이현우가 돈을 쥐는 순간을 정확하고 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "오른쪽 인물의 팔이 아래로 향해 있음에도 몸통 중앙에서 돈을 쥐는 또 다른 팔이 생성된 치명적인 신체 오류가 있어 실격입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 가슴 아래 높이에서 비스듬히 아래를 향해 두 사람의 손이 교차하는 중심부를 정확히 비추고 있음.",
        "built_space": "금이 가고 물이 새어 나오는 야간의 인공 방파제가 인물들 뒤로 적절한 원근감과 함께 아웃포커싱되어 배치됨.",
        "entities": "왼쪽은 가죽 재킷을 입은 쿠마의 몸통과 물기가 남은 오른손, 오른쪽은 흙먼지가 묻은 셔츠를 입은 이현우의 몸통과 돈을 움켜쥐는 손, 중앙에는 돈뭉치가 명확히 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물의 팔이 자연스럽게 뻗어 나와 허공에서 돈뭉치를 단단히 쥐고 건네는 물리적 접촉과 지탱이 사실적으로 표현됨."
       },
       {
        "label": "B",
        "direction": "카메라가 아래를 향하며 돈을 건네받는 손을 근접 촬영하여 시선 방향과 프레이밍 의도에 부합함.",
        "built_space": "야간의 방파제와 흘러내리는 물줄기가 프롬프트의 위치 설정에 맞게 배경에 올바르게 자리 잡고 있음.",
        "entities": "왼쪽 인물(쿠마)의 복장과 손, 돈뭉치는 일치하나 오른쪽 인물(이현우)의 신체 구조에 심각한 오류가 발생함.",
        "hard_violations": [
         "오른쪽 인물(이현우)의 어깨에서 이어진 팔이 프레임 우측 하단으로 늘어져 있음에도, 가슴 부위에서 동일한 소매를 입은 또 다른 팔이 튀어나와 돈을 잡고 있는 물리적으로 불가능한 신체 구조(여분의 팔)"
        ],
        "physics": "왼쪽 인물의 손은 자연스럽게 돈을 들고 있으나, 오른쪽 인물은 비정상적인 위치에서 생성된 팔로 돈을 쥐고 있어 물리적 자세가 성립하지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴을 제외한 근접 구도와 방벽의 연속성은 충실하지만, 이현우가 돈을 꽉 움켜쥐는 찰나보다는 양쪽에서 함께 잡고 있는 모습에 가깝다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "오른쪽에서 들어온 이현우의 손가락이 돈뭉치를 강하게 감싸고 쿠마의 젖은 손도 접촉해 있어, 소유권이 넘어가는 핵심 찰나를 더 정확히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "쿠마의 팔은 왼쪽에서 오른쪽 아래의 돈뭉치로, 이현우의 팔은 오른쪽에서 왼쪽 아래의 같은 돈뭉치로 향한다. 두 손은 하단 중앙에서 만난다. 얼굴과 눈은 잘려 시선 방향은 확인할 수 없고, 총구도 보이지 않는다.",
        "built_space": "두 사람은 하나의 콘크리트 방벽 앞 통로에 서 있다. 방벽은 오른쪽 가까운 곳에서 왼쪽 먼 곳으로 이어지고, 반복되는 판재와 하부 경사 지지대 일부가 보인다. 균열에서 흘러내리는 물과 젖은 포장 바닥, 멀리 등대 하나가 이전 장면과 연결된다. 별도 모임 공간의 화로나 드럼통은 없다. 얼굴을 제외한 몸통과 교환 손을 담았으나, 쿠마의 손 자체는 왼쪽 가장자리보다 중앙에 가깝다.",
        "entities": "왼쪽의 낡은 검은 가죽 재킷과 어두운 데님은 쿠마, 오른쪽의 피와 먼지가 묻은 어두운 셔츠와 갈색 벨트는 이현우의 참고 복장에 부합한다. 보이는 손과 팔에서 명백한 인물 불일치는 없지만 얼굴이 없어 정확한 나이와 민족적 정체성은 확인할 수 없다. 쿠마의 손은 젖어 있다. 고무줄로 묶인 지폐 한 뭉치는 화면의 3분의 1보다 훨씬 작으며 국가와 액면은 확실히 판독되지 않는다. 총과 인이어, 얼굴 및 다리 부상은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "돈뭉치는 쿠마의 엄지와 아래쪽 손가락, 이현우의 감싼 손에 함께 지지된다. 손목과 팔은 각 몸통으로 자연스럽게 이어지며 떠 있는 물체는 없다. 이현우의 손도 돈을 잡지만 손가락의 조임이 비교적 느슨해 강하게 움켜쥐는 동작은 덜 선명하다. 물은 방벽을 따라 아래로 흐른다."
       },
       {
        "label": "B",
        "direction": "왼쪽 쿠마의 손은 중앙 아래의 돈뭉치를 내밀고, 오른쪽 이현우의 손은 그 돈뭉치 끝을 향해 들어와 손가락을 닫는다. 두 손의 행동 대상이 같은 돈으로 명확하다. 얼굴은 프레임 위로 잘려 시선은 검증할 수 없으며 총구는 보이지 않는다.",
        "built_space": "하나의 균열 난 방벽이 두 몸통 뒤에서 비스듬히 이어진다. 하부 지지대 일부와 누수, 젖은 포장 통로가 보이고 방벽 너머 소파 구조물 여러 개와 먼 등대 하나가 남아 있다. 참고 장소의 구조와 야간 조명이 유지되며 추가 인물이나 화로는 없다. 몸통을 자른 근접 구도와 하단 중앙의 교환 위치가 맞지만, 쿠마의 손은 지정된 왼쪽 가장자리보다 안쪽에 있다.",
        "entities": "쿠마의 닳은 검은 가죽 재킷과 어두운 데님, 이현우의 얼룩진 어두운 셔츠와 갈색 벨트가 참고 이미지와 일치한다. 쿠마의 손에는 젖은 광택과 물방울이 보인다. 손과 팔의 체격은 인물 설정에 어긋나지 않으나 얼굴이 없어 정확한 신원과 나이는 판별할 수 없다. 지폐 한 뭉치는 고무줄로 묶여 있고 손가락 사이에 비스듬히 놓이며 인쇄 내용이 강조되지 않는다. 액면과 국가는 확정하기 어렵고 프롬프트에도 지정되어 있지 않다. 총과 얼굴·다리 부상은 화면 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "쿠마가 지폐 왼쪽을 받친 상태에서 이현우의 엄지와 굽힌 손가락들이 오른쪽 끝을 단단히 감싼다. 지폐 가장자리가 그립에 따라 약간 휘어 실제로 힘을 주어 빼앗듯 쥐는 동작이 읽힌다. 돈과 손 모두 지지 및 연결이 명확하고 불가능한 부유나 관절 형태는 없다. 방벽의 물과 손끝의 물방울도 아래로 떨어지는 자연스러운 상태다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴을 제외한 근접 구도와 방벽의 연속성은 충실하지만, 이현우가 돈을 꽉 움켜쥐는 찰나보다는 양쪽에서 함께 잡고 있는 모습에 가깝다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "오른쪽에서 들어온 이현우의 손가락이 돈뭉치를 강하게 감싸고 쿠마의 젖은 손도 접촉해 있어, 소유권이 넘어가는 핵심 찰나를 더 정확히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "쿠마의 팔은 왼쪽에서 오른쪽 아래의 돈뭉치로, 이현우의 팔은 오른쪽에서 왼쪽 아래의 같은 돈뭉치로 향한다. 두 손은 하단 중앙에서 만난다. 얼굴과 눈은 잘려 시선 방향은 확인할 수 없고, 총구도 보이지 않는다.",
        "built_space": "두 사람은 하나의 콘크리트 방벽 앞 통로에 서 있다. 방벽은 오른쪽 가까운 곳에서 왼쪽 먼 곳으로 이어지고, 반복되는 판재와 하부 경사 지지대 일부가 보인다. 균열에서 흘러내리는 물과 젖은 포장 바닥, 멀리 등대 하나가 이전 장면과 연결된다. 별도 모임 공간의 화로나 드럼통은 없다. 얼굴을 제외한 몸통과 교환 손을 담았으나, 쿠마의 손 자체는 왼쪽 가장자리보다 중앙에 가깝다.",
        "entities": "왼쪽의 낡은 검은 가죽 재킷과 어두운 데님은 쿠마, 오른쪽의 피와 먼지가 묻은 어두운 셔츠와 갈색 벨트는 이현우의 참고 복장에 부합한다. 보이는 손과 팔에서 명백한 인물 불일치는 없지만 얼굴이 없어 정확한 나이와 민족적 정체성은 확인할 수 없다. 쿠마의 손은 젖어 있다. 고무줄로 묶인 지폐 한 뭉치는 화면의 3분의 1보다 훨씬 작으며 국가와 액면은 확실히 판독되지 않는다. 총과 인이어, 얼굴 및 다리 부상은 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "돈뭉치는 쿠마의 엄지와 아래쪽 손가락, 이현우의 감싼 손에 함께 지지된다. 손목과 팔은 각 몸통으로 자연스럽게 이어지며 떠 있는 물체는 없다. 이현우의 손도 돈을 잡지만 손가락의 조임이 비교적 느슨해 강하게 움켜쥐는 동작은 덜 선명하다. 물은 방벽을 따라 아래로 흐른다."
       },
       {
        "label": "A",
        "direction": "왼쪽 쿠마의 손은 중앙 아래의 돈뭉치를 내밀고, 오른쪽 이현우의 손은 그 돈뭉치 끝을 향해 들어와 손가락을 닫는다. 두 손의 행동 대상이 같은 돈으로 명확하다. 얼굴은 프레임 위로 잘려 시선은 검증할 수 없으며 총구는 보이지 않는다.",
        "built_space": "하나의 균열 난 방벽이 두 몸통 뒤에서 비스듬히 이어진다. 하부 지지대 일부와 누수, 젖은 포장 통로가 보이고 방벽 너머 소파 구조물 여러 개와 먼 등대 하나가 남아 있다. 참고 장소의 구조와 야간 조명이 유지되며 추가 인물이나 화로는 없다. 몸통을 자른 근접 구도와 하단 중앙의 교환 위치가 맞지만, 쿠마의 손은 지정된 왼쪽 가장자리보다 안쪽에 있다.",
        "entities": "쿠마의 닳은 검은 가죽 재킷과 어두운 데님, 이현우의 얼룩진 어두운 셔츠와 갈색 벨트가 참고 이미지와 일치한다. 쿠마의 손에는 젖은 광택과 물방울이 보인다. 손과 팔의 체격은 인물 설정에 어긋나지 않으나 얼굴이 없어 정확한 신원과 나이는 판별할 수 없다. 지폐 한 뭉치는 고무줄로 묶여 있고 손가락 사이에 비스듬히 놓이며 인쇄 내용이 강조되지 않는다. 액면과 국가는 확정하기 어렵고 프롬프트에도 지정되어 있지 않다. 총과 얼굴·다리 부상은 화면 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "쿠마가 지폐 왼쪽을 받친 상태에서 이현우의 엄지와 굽힌 손가락들이 오른쪽 끝을 단단히 감싼다. 지폐 가장자리가 그립에 따라 약간 휘어 실제로 힘을 주어 빼앗듯 쥐는 동작이 읽힌다. 돈과 손 모두 지지 및 연결이 명확하고 불가능한 부유나 관절 형태는 없다. 방벽의 물과 손끝의 물방울도 아래로 떨어지는 자연스러운 상태다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.889
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.639
   },
   "violations": {
    "B": [
     "[gemini-pro] 오른쪽 인물(이현우)의 어깨에서 이어진 팔이 프레임 우측 하단으로 늘어져 있음에도, 가슴 부위에서 동일한 소매를 입은 또 다른 팔이 튀어나와 돈을 잡고 있는 물리적으로 불가능한 신체 구조(여분의 팔)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 639
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트의 요구사항에 따라 크롭된 구도 내에서 물에 젖은 쿠마의 손과 이현우가 돈을 쥐는 순간을 정확하고 사실적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 639,
    "verdict_ko": "오른쪽 인물의 팔이 아래로 향해 있음에도 몸통 중앙에서 돈을 쥐는 또 다른 팔이 생성된 치명적인 신체 오류가 있어 실격입니다.  ★위반: [gemini-pro] 오른쪽 인물(이현우)의 어깨에서 이어진 팔이 프레임 우측 하단으로 늘어져 있음에도, 가슴 부위에서 동일한 소매를 입은 또 다른 팔이 튀어나와 돈을 잡고 있는 물리적으로 불가능한 신체 구조(여분의 팔)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우, 쿠마 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S13sh18_sel.png",
    "asset_id": "4125385a-8675-4534-b45b-421059447318",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:919321>",
    "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab089d-0410-7bc7-99b4-6be971142318",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh18"
  }
 },
 "S14sh3::signage": {
  "fp": "6b0a7940d1d8267e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh3::bgfirst_bg": {
  "input_fingerprint": "f7288120f16636d4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 발밑에 떨어진 화병이 산산조각 나며 파편이 튀어 오르는 근접 찰나.\n\nLOCATION (lock): On the floor of the container home's shared living area, where a vase has shattered. Daylight provides the daytime interior illumination.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to the flow's static floor-level position, offset beside 찰리's toes and tilted slightly down toward the impact. Keep his feet and lower legs in the upper half, with the broken vase's impact area below center and clear space above it for the jumping fragments; the fragments collectively occupy less than a third of the frame. Hold the instant of his startled halt before any sweeping foot movement, with his head and attention toward the accident remaining outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Broken vase (Shattering beside 찰리's feet, with fragments jumping upward); used as Places the impact low in the composition while leaving room to read the fragments' upward motion; Container interior floor (Receiving the fallen vase and scattered fragments); used as Provides an uninterrupted scale reference around the feet and impact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime ambient illumination with crisp but controlled separation between 찰리's hard-surface contours, the fragments and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 발밑에 떨어진 화병이 산산조각 나며 파편이 튀어 오르는 근접 찰나.\n\nLOCATION (lock): On the floor of the container home's shared living area, where a vase has shattered. Daylight provides the daytime interior illumination.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to the flow's static floor-level position, offset beside 찰리's toes and tilted slightly down toward the impact. Keep his feet and lower legs in the upper half, with the broken vase's impact area below center and clear space above it for the jumping fragments; the fragments collectively occupy less than a third of the frame. Hold the instant of his startled halt before any sweeping foot movement, with his head and attention toward the accident remaining outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Broken vase (Shattering beside 찰리's feet, with fragments jumping upward); used as Places the impact low in the composition while leaving room to read the fragments' upward motion; Container interior floor (Receiving the fallen vase and scattered fragments); used as Provides an uninterrupted scale reference around the feet and impact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime ambient illumination with crisp but controlled separation between 찰리's hard-surface contours, the fragments and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S14sh3__bgfirst_bg.png",
  "asset_id": "c882f2d1-dce3-4861-9531-662f7c4a65cb",
  "input_asset_ids": [
   "d1ec31a3-bd5b-4bcc-bd3f-d8dbd19cc1e0",
   "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  ]
 },
 "S14sh3": {
  "input_fingerprint": "eb4140a200d44e62",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 발밑에 떨어진 화병이 산산조각 나며 파편이 튀어 오르는 근접 찰나.\n\nLOCATION (lock): On the floor of the container home's shared living area, where a vase has shattered. Daylight provides the daytime interior illumination. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to the flow's static floor-level position, offset beside 찰리's toes and tilted slightly down toward the impact. Keep his feet and lower legs in the upper half, with the broken vase's impact area below center and clear space above it for the jumping fragments; the fragments collectively occupy less than a third of the frame. Hold the instant of his startled halt before any sweeping foot movement, with his head and attention toward the accident remaining outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Broken vase (Shattering beside 찰리's feet, with fragments jumping upward); used as Places the impact low in the composition while leaving room to read the fragments' upward motion; Container interior floor (Receiving the fallen vase and scattered fragments); used as Provides an uninterrupted scale reference around the feet and impact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime ambient illumination with crisp but controlled separation between 찰리's hard-surface contours, the fragments and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The vase has broken, leaving fragments at Charlie's feet inside the container. Charlie remains a worn gorilla-shaped robot with blue-lit eyes, an old coat and hat, and an already worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 발밑에 떨어진 화병이 산산조각 나며 파편이 튀어 오르는 근접 찰나.\n\nLOCATION (lock): On the floor of the container home's shared living area, where a vase has shattered. Daylight provides the daytime interior illumination. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to the flow's static floor-level position, offset beside 찰리's toes and tilted slightly down toward the impact. Keep his feet and lower legs in the upper half, with the broken vase's impact area below center and clear space above it for the jumping fragments; the fragments collectively occupy less than a third of the frame. Hold the instant of his startled halt before any sweeping foot movement, with his head and attention toward the accident remaining outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Broken vase (Shattering beside 찰리's feet, with fragments jumping upward); used as Places the impact low in the composition while leaving room to read the fragments' upward motion; Container interior floor (Receiving the fallen vase and scattered fragments); used as Provides an uninterrupted scale reference around the feet and impact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime ambient illumination with crisp but controlled separation between 찰리's hard-surface contours, the fragments and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The vase has broken, leaving fragments at Charlie's feet inside the container. Charlie remains a worn gorilla-shaped robot with blue-lit eyes, an old coat and hat, and an already worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 발밑에 떨어진 화병이 산산조각 나며 파편이 튀어 오르는 근접 찰나.\n\nLOCATION (lock): On the floor of the container home's shared living area, where a vase has shattered. Daylight provides the daytime interior illumination. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to the flow's static floor-level position, offset beside 찰리's toes and tilted slightly down toward the impact. Keep his feet and lower legs in the upper half, with the broken vase's impact area below center and clear space above it for the jumping fragments; the fragments collectively occupy less than a third of the frame. Hold the instant of his startled halt before any sweeping foot movement, with his head and attention toward the accident remaining outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Broken vase (Shattering beside 찰리's feet, with fragments jumping upward); used as Places the impact low in the composition while leaving room to read the fragments' upward motion; Container interior floor (Receiving the fallen vase and scattered fragments); used as Provides an uninterrupted scale reference around the feet and impact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime ambient illumination with crisp but controlled separation between 찰리's hard-surface contours, the fragments and the floor.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The vase has broken, leaving fragments at Charlie's feet inside the container. Charlie remains a worn gorilla-shaped robot with blue-lit eyes, an old coat and hat, and an already worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S14sh3__bgfirst_bg.png",
     "asset_id": "c882f2d1-dce3-4861-9531-662f7c4a65cb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S14sh3.png",
     "asset_id": "d1ec31a3-bd5b-4bcc-bd3f-d8dbd19cc1e0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
     "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병을 향하고 있음.",
    "built_space": "대형 유리창과 가죽 소파가 있는 거실 구조로, 레퍼런스에 명시된 컨테이너 내부의 비품이나 벽면 구조가 전혀 존재하지 않음.",
    "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 조각난 도자기 화병.",
    "hard_violations": [
     "[gemini-pro] 제시된 로케이션 레퍼런스와 완전히 다른 창조된 공간 구조 (대형 창문, 소파 등)",
     "[gpt-high] 정확한 장소로 고정된 참조의 왼쪽 벽을 대형 창이 있는 벽으로 바꾸고, 참조에 없는 펜던트등·가죽 소파·넓은 수납장 등을 배치하여 다른 생활 공간을 만들었다."
    ],
    "physics": "화병의 파편들이 바닥에서 튀어 올라 허공에 체공 중임."
   },
   {
    "label": "B",
    "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병의 파편을 향해 아래로 약간 기울어져 있음.",
    "built_space": "컨테이너 내부. 좌측의 주름진 벽면과 가구 다리, 배경의 냉장고와 문 등 레퍼런스의 구조와 비품이 정확한 위치에 있음.",
    "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 산산조각 난 푸른 무늬 화병.",
    "hard_violations": [],
    "physics": "바닥에 부딪힌 화병 파편과 물방울들이 충격으로 인해 허공으로 튀어 오르고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글과 숏 스케일을 완벽히 따랐으며, 레퍼런스의 컨테이너 내부 구조를 정확하게 재현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 컨테이너 내부가 아닌 완전히 다른 구조의 거실을 그려내어 로케이션 조건을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병의 파편을 향해 아래로 약간 기울어져 있음.",
        "built_space": "컨테이너 내부. 좌측의 주름진 벽면과 가구 다리, 배경의 냉장고와 문 등 레퍼런스의 구조와 비품이 정확한 위치에 있음.",
        "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 산산조각 난 푸른 무늬 화병.",
        "hard_violations": [],
        "physics": "바닥에 부딪힌 화병 파편과 물방울들이 충격으로 인해 허공으로 튀어 오르고 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병을 향하고 있음.",
        "built_space": "대형 유리창과 가죽 소파가 있는 거실 구조로, 레퍼런스에 명시된 컨테이너 내부의 비품이나 벽면 구조가 전혀 존재하지 않음.",
        "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 조각난 도자기 화병.",
        "hard_violations": [
         "제시된 로케이션 레퍼런스와 완전히 다른 창조된 공간 구조 (대형 창문, 소파 등)"
        ],
        "physics": "화병의 파편들이 바닥에서 튀어 올라 허공에 체공 중임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글과 숏 스케일을 완벽히 따랐으며, 레퍼런스의 컨테이너 내부 구조를 정확하게 재현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 컨테이너 내부가 아닌 완전히 다른 구조의 거실을 그려내어 로케이션 조건을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병의 파편을 향해 아래로 약간 기울어져 있음.",
        "built_space": "컨테이너 내부. 좌측의 주름진 벽면과 가구 다리, 배경의 냉장고와 문 등 레퍼런스의 구조와 비품이 정확한 위치에 있음.",
        "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 산산조각 난 푸른 무늬 화병.",
        "hard_violations": [],
        "physics": "바닥에 부딪힌 화병 파편과 물방울들이 충격으로 인해 허공으로 튀어 오르고 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 바닥 높이에서 로봇의 발과 깨진 화병을 향하고 있음.",
        "built_space": "대형 유리창과 가죽 소파가 있는 거실 구조로, 레퍼런스에 명시된 컨테이너 내부의 비품이나 벽면 구조가 전혀 존재하지 않음.",
        "entities": "찰리(샌드 베이지색 기계 하반신 및 발), 조각난 도자기 화병.",
        "hard_violations": [
         "제시된 로케이션 레퍼런스와 완전히 다른 창조된 공간 구조 (대형 창문, 소파 등)"
        ],
        "physics": "화병의 파편들이 바닥에서 튀어 올라 허공에 체공 중임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "발끝 옆의 낮은 근접 시점과 충돌 직후 튀는 파편, 참조의 좁은 컨테이너 내부를 잘 구현했지만 발이 화면 상반부 아래까지 내려온다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "파편의 튀어 오름과 기계 발의 접지는 타당하지만, 참조에 없는 창·조명·가구로 공간을 바꾸고 양발 사이 정면 구도로 촬영해 장소와 카메라 지시를 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 기계 발의 발끝은 화면 아래쪽과 약간 오른쪽을 향하며, 화병 충돌 지점은 그 앞쪽 오른편에 있다. 파편은 화면 중앙 아래의 충돌 지점에서 위와 옆으로 퍼진다. 머리와 눈은 잘려 있어 사고를 바라보는 시선은 확인할 수 없으며, 이는 지정된 크롭과 맞는다.",
        "built_space": "왼쪽 전경에 목재 가구 다리 하나, 뒤쪽에 금속 선반 하나와 냉장고 하나, 후면 문 하나, 오른쪽 벽에 창 하나와 낮은 침구가 보인다. 참조의 좁은 골조 벽 실내와 후면 문·오른쪽 창의 관계를 대체로 유지한다. 바닥 높이에서 발 옆을 보는 구도이나, 발은 화면 중간을 넘어 아래까지 차지하고 배경 노출도 다소 많다. 바닥의 젖은 광택과 반사는 창빛 방향에서 가능한 표현이다.",
        "entities": "등장 신체는 찰리의 기계 다리 두 개와 발 두 개뿐이다. 마모된 샌드 베이지 장갑판, 검은 관절, 분절된 발가락이 참조와 부합하며 사람 피부는 없다. 얼굴·눈·모자·가슴 로고는 크롭 밖이므로 평가 대상이 아니다. 청색 꽃무늬 도자기 화병이 여러 곡면 조각으로 부서져 있고 물방울과 물이 보인다. 화병의 무늬는 브리프에서 고정하지 않았다. 파편 자체의 면적은 화면의 삼분의 일보다 작다.",
        "hard_violations": [],
        "physics": "양발의 발가락과 뒤꿈치가 바닥에 닿아 기계 몸체를 지지한다. 발로 쓸어내는 동작은 보이지 않는다. 바닥에 놓인 조각과 위로 솟은 조각, 물방울이 같은 충돌 지점에 연결되어 있어 낙하 충격으로 튀는 순간으로 해석된다. 공중 조각에는 충돌이라는 발사 원인과 아래 바닥이라는 착지면이 있으며, 근거 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 발끝은 화면 왼쪽 아래, 오른쪽 발끝은 화면 아래를 향한다. 충돌 지점은 양발 사이보다 앞쪽에 있고 파편이 그 지점에서 위로 퍼진다. 시선은 머리가 크롭 밖이라 확인할 수 없다. 카메라는 발끝 옆으로 비껴난 위치보다 벌어진 양발 사이를 정면으로 보는 위치에 가깝다.",
        "built_space": "후면 문 하나와 오른쪽 선반은 보이지만, 왼쪽에 큰 다분할 창과 뒤쪽의 좁은 창, 천장 펜던트등 하나, 가죽 소파 하나, 식탁 하나와 의자들이 추가되어 있다. 오른쪽에는 넓은 하부 수납장이 놓인다. 참조에서 막힌 왼쪽 골조 벽과 오른쪽 창으로 구성된 좁은 생활 공간을 다른 거실 배치로 바꿨다. 양발이 중앙 높이 아래까지 내려오고 배경 가구가 넓게 드러나 지정된 근접 크롭도 느슨하다.",
        "entities": "마모된 베이지색 기계 다리 두 개와 분절된 발 두 개가 보이며 사람이나 추가 신체는 없다. 보이는 재질과 관절은 찰리의 기계 정체성에 부합한다. 얼굴·눈·모자·로고는 프레임 밖이다. 화병은 짙은 테두리가 있는 밝은 도자기로, 큰 몸통 일부와 다수의 작은 파편이 남아 있다. 파편 면적은 화면의 삼분의 일보다 작지만 큰 몸통이 남아 있어 산산조각 난 정도는 A보다 약하다.",
        "hard_violations": [
         "정확한 장소로 고정된 참조의 왼쪽 벽을 대형 창이 있는 벽으로 바꾸고, 참조에 없는 펜던트등·가죽 소파·넓은 수납장 등을 배치하여 다른 생활 공간을 만들었다."
        ],
        "physics": "두 발의 발가락과 뒤꿈치가 바닥에 닿아 몸체를 지지하며, 발로 파편을 쓸고 있지는 않다. 남은 화병 몸통은 옆면으로 바닥에 놓여 있고 작은 조각들은 바닥에 분산되어 있다. 공중 파편은 바로 아래 파손 지점의 충격으로 튀어 오른 것으로 읽히며 다시 바닥으로 떨어질 수 있다. 물리적으로 지지나 발사 원인을 찾을 수 없는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "발끝 옆의 낮은 근접 시점과 충돌 직후 튀는 파편, 참조의 좁은 컨테이너 내부를 잘 구현했지만 발이 화면 상반부 아래까지 내려온다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "파편의 튀어 오름과 기계 발의 접지는 타당하지만, 참조에 없는 창·조명·가구로 공간을 바꾸고 양발 사이 정면 구도로 촬영해 장소와 카메라 지시를 어긴다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 기계 발의 발끝은 화면 아래쪽과 약간 오른쪽을 향하며, 화병 충돌 지점은 그 앞쪽 오른편에 있다. 파편은 화면 중앙 아래의 충돌 지점에서 위와 옆으로 퍼진다. 머리와 눈은 잘려 있어 사고를 바라보는 시선은 확인할 수 없으며, 이는 지정된 크롭과 맞는다.",
        "built_space": "왼쪽 전경에 목재 가구 다리 하나, 뒤쪽에 금속 선반 하나와 냉장고 하나, 후면 문 하나, 오른쪽 벽에 창 하나와 낮은 침구가 보인다. 참조의 좁은 골조 벽 실내와 후면 문·오른쪽 창의 관계를 대체로 유지한다. 바닥 높이에서 발 옆을 보는 구도이나, 발은 화면 중간을 넘어 아래까지 차지하고 배경 노출도 다소 많다. 바닥의 젖은 광택과 반사는 창빛 방향에서 가능한 표현이다.",
        "entities": "등장 신체는 찰리의 기계 다리 두 개와 발 두 개뿐이다. 마모된 샌드 베이지 장갑판, 검은 관절, 분절된 발가락이 참조와 부합하며 사람 피부는 없다. 얼굴·눈·모자·가슴 로고는 크롭 밖이므로 평가 대상이 아니다. 청색 꽃무늬 도자기 화병이 여러 곡면 조각으로 부서져 있고 물방울과 물이 보인다. 화병의 무늬는 브리프에서 고정하지 않았다. 파편 자체의 면적은 화면의 삼분의 일보다 작다.",
        "hard_violations": [],
        "physics": "양발의 발가락과 뒤꿈치가 바닥에 닿아 기계 몸체를 지지한다. 발로 쓸어내는 동작은 보이지 않는다. 바닥에 놓인 조각과 위로 솟은 조각, 물방울이 같은 충돌 지점에 연결되어 있어 낙하 충격으로 튀는 순간으로 해석된다. 공중 조각에는 충돌이라는 발사 원인과 아래 바닥이라는 착지면이 있으며, 근거 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 발끝은 화면 왼쪽 아래, 오른쪽 발끝은 화면 아래를 향한다. 충돌 지점은 양발 사이보다 앞쪽에 있고 파편이 그 지점에서 위로 퍼진다. 시선은 머리가 크롭 밖이라 확인할 수 없다. 카메라는 발끝 옆으로 비껴난 위치보다 벌어진 양발 사이를 정면으로 보는 위치에 가깝다.",
        "built_space": "후면 문 하나와 오른쪽 선반은 보이지만, 왼쪽에 큰 다분할 창과 뒤쪽의 좁은 창, 천장 펜던트등 하나, 가죽 소파 하나, 식탁 하나와 의자들이 추가되어 있다. 오른쪽에는 넓은 하부 수납장이 놓인다. 참조에서 막힌 왼쪽 골조 벽과 오른쪽 창으로 구성된 좁은 생활 공간을 다른 거실 배치로 바꿨다. 양발이 중앙 높이 아래까지 내려오고 배경 가구가 넓게 드러나 지정된 근접 크롭도 느슨하다.",
        "entities": "마모된 베이지색 기계 다리 두 개와 분절된 발 두 개가 보이며 사람이나 추가 신체는 없다. 보이는 재질과 관절은 찰리의 기계 정체성에 부합한다. 얼굴·눈·모자·로고는 프레임 밖이다. 화병은 짙은 테두리가 있는 밝은 도자기로, 큰 몸통 일부와 다수의 작은 파편이 남아 있다. 파편 면적은 화면의 삼분의 일보다 작지만 큰 몸통이 남아 있어 산산조각 난 정도는 A보다 약하다.",
        "hard_violations": [
         "정확한 장소로 고정된 참조의 왼쪽 벽을 대형 창이 있는 벽으로 바꾸고, 참조에 없는 펜던트등·가죽 소파·넓은 수납장 등을 배치하여 다른 생활 공간을 만들었다."
        ],
        "physics": "두 발의 발가락과 뒤꿈치가 바닥에 닿아 몸체를 지지하며, 발로 파편을 쓸고 있지는 않다. 남은 화병 몸통은 옆면으로 바닥에 놓여 있고 작은 조각들은 바닥에 분산되어 있다. 공중 파편은 바로 아래 파손 지점의 충격으로 튀어 오른 것으로 읽히며 다시 바닥으로 떨어질 수 있다. 물리적으로 지지나 발사 원인을 찾을 수 없는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 제시된 로케이션 레퍼런스와 완전히 다른 창조된 공간 구조 (대형 창문, 소파 등)",
     "[gpt-high] 정확한 장소로 고정된 참조의 왼쪽 벽을 대형 창이 있는 벽으로 바꾸고, 참조에 없는 펜던트등·가죽 소파·넓은 수납장 등을 배치하여 다른 생활 공간을 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 679
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 앵글과 숏 스케일을 완벽히 따랐으며, 레퍼런스의 컨테이너 내부 구조를 정확하게 재현함."
   },
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "지정된 컨테이너 내부가 아닌 완전히 다른 구조의 거실을 그려내어 로케이션 조건을 심각하게 위반함.  ★위반: [gemini-pro] 제시된 로케이션 레퍼런스와 완전히 다른 창조된 공간 구조 (대형 창문, 소파 등) / [gpt-high] 정확한 장소로 고정된 참조의 왼쪽 벽을 대형 창이 있는 벽으로 바꾸고, 참조에 없는 펜던트등·가죽 소파·넓은 수납장 등을 배치하여 다른 생활 공간을 만들었다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
    "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08a1-79e9-70cd-b0e8-2767287608f1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S14sh3__bgfirst_bg.png",
   "bg_asset_id": "c882f2d1-dce3-4861-9531-662f7c4a65cb",
   "bg_record_key": "S14sh3::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "family_container_room",
   "groupbg_asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S14sh7::signage": {
  "fp": "243487be27a5e0e9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh7": {
  "input_fingerprint": "0ef426c2b4f4a46d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 호통에 깜짝 놀란 듯 어깨를 움츠린 채 시무룩하게 고개를 푹 숙이고 있는 찰리의 굽은 상체.\n\nLOCATION (lock): Inside the daylit shared living area of the refugee family's container home, beside the household belongings. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the continuous rise beside 찰리 and hold just above his bowed head, looking gently down across his shoulders from the established three-quarter side. His hunched upper body occupies the right half as he withdraws his hands and looks toward the floor, while a smaller seated 이현우 remains at the left edge, turned toward 찰리 after the reprimand. Keep their established lateral relationship intact and let the increased distance from the earlier foot detail reveal the emotional consequence without introducing a new light cue.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior floor (Visible below 찰리's lowered head); used as Provides a simple downward destination for his withdrawn attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued daytime ambient treatment, preserving precise robot contours while allowing the bowed posture to carry tenderness rather than menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments remain on the floor where Charlie pushed them aside to hide them; they have not been cleaned up. Charlie retains the old coat and hat, worn chest logo and blue-lit eyes, and is now hunched and subdued. 이현우: He remains inside the container with facial injuries and the untreated leg bite.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 호통에 깜짝 놀란 듯 어깨를 움츠린 채 시무룩하게 고개를 푹 숙이고 있는 찰리의 굽은 상체.\n\nLOCATION (lock): Inside the daylit shared living area of the refugee family's container home, beside the household belongings. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the continuous rise beside 찰리 and hold just above his bowed head, looking gently down across his shoulders from the established three-quarter side. His hunched upper body occupies the right half as he withdraws his hands and looks toward the floor, while a smaller seated 이현우 remains at the left edge, turned toward 찰리 after the reprimand. Keep their established lateral relationship intact and let the increased distance from the earlier foot detail reveal the emotional consequence without introducing a new light cue.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior floor (Visible below 찰리's lowered head); used as Provides a simple downward destination for his withdrawn attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued daytime ambient treatment, preserving precise robot contours while allowing the bowed posture to carry tenderness rather than menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments remain on the floor where Charlie pushed them aside to hide them; they have not been cleaned up. Charlie retains the old coat and hat, worn chest logo and blue-lit eyes, and is now hunched and subdued. 이현우: He remains inside the container with facial injuries and the untreated leg bite.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우의 호통에 깜짝 놀란 듯 어깨를 움츠린 채 시무룩하게 고개를 푹 숙이고 있는 찰리의 굽은 상체.\n\nLOCATION (lock): Inside the daylit shared living area of the refugee family's container home, beside the household belongings. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the continuous rise beside 찰리 and hold just above his bowed head, looking gently down across his shoulders from the established three-quarter side. His hunched upper body occupies the right half as he withdraws his hands and looks toward the floor, while a smaller seated 이현우 remains at the left edge, turned toward 찰리 after the reprimand. Keep their established lateral relationship intact and let the increased distance from the earlier foot detail reveal the emotional consequence without introducing a new light cue.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container interior floor (Visible below 찰리's lowered head); used as Provides a simple downward destination for his withdrawn attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued daytime ambient treatment, preserving precise robot contours while allowing the bowed posture to carry tenderness rather than menace.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken vase fragments remain on the floor where Charlie pushed them aside to hide them; they have not been cleaned up. Charlie retains the old coat and hat, worn chest logo and blue-lit eyes, and is now hunched and subdued. 이현우: He remains inside the container with facial injuries and the untreated leg bite.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 찰리를 향해 시선을 두고 있으며, 찰리는 바닥의 깨진 도자기를 내려다보고 있음.",
    "built_space": "컨테이너 내부 거실 공간. 배경의 문, 러그, 선반 등 고정 요소들이 적절하게 배치되어 있음.",
    "entities": "찰리는 고릴라 형태의 기계로 낡은 코트와 모자를 착용함 (단, 눈은 지시된 푸른색이 아닌 노란색). 이현우는 왼쪽에 앉아 있으며 다리 부상 붕대가 확인됨.",
    "hard_violations": [],
    "physics": "인물들은 바닥과 가구에 안정적으로 지지되어 있음. 바닥의 도자기 파편과 물방울은 이전 샷의 충격으로 튀어오른 궤적을 띠며 공중에 떠 있음."
   },
   {
    "label": "B",
    "direction": "이현우가 찰리를 바라보고, 찰리는 바닥을 향해 고개를 숙이고 시선을 둠.",
    "built_space": "컨테이너 내부 구조와 가구, 배경 요소들의 배치가 자연스럽게 일치함.",
    "entities": "찰리는 기계 몸체로 묘사되었으나 명시된 코트와 모자를 착용하지 않음. 이현우는 왼쪽에 위치하며 다리 부상이 묘사됨.",
    "hard_violations": [],
    "physics": "두 인물의 자세와 지지점은 자연스러움. 바닥에는 충격으로 튀어오른 파편과 물방울이 공중에 떠 있는 상태로 유지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리가 낡은 코트와 모자를 착용해야 한다는 지시를 정확히 반영하여 프롬프트 충실도가 가장 높습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 낡은 코트와 모자 착용 지시를 완전히 누락하여 주요 복장 요건을 충족하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 찰리를 향해 시선을 두고 있으며, 찰리는 바닥의 깨진 도자기를 내려다보고 있음.",
        "built_space": "컨테이너 내부 거실 공간. 배경의 문, 러그, 선반 등 고정 요소들이 적절하게 배치되어 있음.",
        "entities": "찰리는 고릴라 형태의 기계로 낡은 코트와 모자를 착용함 (단, 눈은 지시된 푸른색이 아닌 노란색). 이현우는 왼쪽에 앉아 있으며 다리 부상 붕대가 확인됨.",
        "hard_violations": [],
        "physics": "인물들은 바닥과 가구에 안정적으로 지지되어 있음. 바닥의 도자기 파편과 물방울은 이전 샷의 충격으로 튀어오른 궤적을 띠며 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 찰리를 바라보고, 찰리는 바닥을 향해 고개를 숙이고 시선을 둠.",
        "built_space": "컨테이너 내부 구조와 가구, 배경 요소들의 배치가 자연스럽게 일치함.",
        "entities": "찰리는 기계 몸체로 묘사되었으나 명시된 코트와 모자를 착용하지 않음. 이현우는 왼쪽에 위치하며 다리 부상이 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물의 자세와 지지점은 자연스러움. 바닥에는 충격으로 튀어오른 파편과 물방울이 공중에 떠 있는 상태로 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리가 낡은 코트와 모자를 착용해야 한다는 지시를 정확히 반영하여 프롬프트 충실도가 가장 높습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 낡은 코트와 모자 착용 지시를 완전히 누락하여 주요 복장 요건을 충족하지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 찰리를 향해 시선을 두고 있으며, 찰리는 바닥의 깨진 도자기를 내려다보고 있음.",
        "built_space": "컨테이너 내부 거실 공간. 배경의 문, 러그, 선반 등 고정 요소들이 적절하게 배치되어 있음.",
        "entities": "찰리는 고릴라 형태의 기계로 낡은 코트와 모자를 착용함 (단, 눈은 지시된 푸른색이 아닌 노란색). 이현우는 왼쪽에 앉아 있으며 다리 부상 붕대가 확인됨.",
        "hard_violations": [],
        "physics": "인물들은 바닥과 가구에 안정적으로 지지되어 있음. 바닥의 도자기 파편과 물방울은 이전 샷의 충격으로 튀어오른 궤적을 띠며 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 찰리를 바라보고, 찰리는 바닥을 향해 고개를 숙이고 시선을 둠.",
        "built_space": "컨테이너 내부 구조와 가구, 배경 요소들의 배치가 자연스럽게 일치함.",
        "entities": "찰리는 기계 몸체로 묘사되었으나 명시된 코트와 모자를 착용하지 않음. 이현우는 왼쪽에 위치하며 다리 부상이 묘사됨.",
        "hard_violations": [],
        "physics": "두 인물의 자세와 지지점은 자연스러움. 바닥에는 충격으로 튀어오른 파편과 물방울이 공중에 떠 있는 상태로 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "고개를 숙인 동작과 좌우 관계는 맞지만, 상체 중심 미디엄 숏 대신 전신을 보여주고 이현우가 지나치게 크며 낡은 코트와 모자도 빠졌다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "찰리의 움츠린 자세와 낡은 코트·모자를 더 충실히 구현했으나, 전신 구도와 크게 잡힌 이현우는 지정된 상체 중심 촬영에 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 눈은 아래쪽 도자기 파편이 있는 바닥을 향한다. 이현우는 오른쪽의 찰리를 바라본다. 시선 관계는 맞지만, 찰리의 손은 몸 쪽으로 거두기보다 바닥 가까이 늘어져 있다.",
        "built_space": "뒤쪽 문 하나, 그 왼쪽 금속 수납 선반 하나, 오른쪽 벽의 창 구역과 침구 자리 하나, 중앙 무늬 러그 하나가 보인다. 이현우는 왼쪽 전경의 팔걸이 의자에 앉고 찰리는 오른쪽 바닥에 서 있다. 골이 진 벽과 낡은 나무 바닥은 장소 참조와 대체로 이어진다. 다만 카메라는 찰리의 어깨와 상체에 머무르지 않고 발까지 담으며, 이현우도 왼쪽 가장자리의 작은 인물이 아니라 큰 전경 인물이다.",
        "entities": "등장인물은 찰리와 이현우 둘이다. 찰리의 샌드 베이지 장갑판, 긴 팔과 짧은 다리, 흰 마스크 얼굴, 선 형태 입, 파란 원형 가슴 장치는 참조에 부합한다. 눈은 참조처럼 주황색이지만 이번 장면의 파란 눈 지시와 다르고, 요구된 코트와 모자는 없다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 남색 상의와 얼굴 상처가 확인된다. 보이는 다리에는 피 묻은 붕대가 있어 미처치 물림 상태와 다르다. 청백색 화병 파편은 바닥에 남아 있다.",
        "hard_violations": [],
        "physics": "찰리는 두 기계 발로 바닥을 딛고 허리와 목을 굽힌다. 팔과 손은 어깨 및 팔꿈치 관절에 연결되어 아래로 내려오며, 떠 있는 몸은 없다. 이현우의 골반은 의자 좌판에 지지되고 손은 무릎 부근에 놓여 있다. 큰 화병 조각들은 바닥에 놓여 있어 지지 관계가 자연스럽다."
       },
       {
        "label": "B",
        "direction": "찰리는 모자챙 아래로 고개를 깊이 숙여 앞쪽 바닥과 파편을 향한다. 이현우의 얼굴은 오른쪽 찰리에게 돌아가 있다. 두 손은 낮게 내려와 몸 가까이에 있지만, 적극적으로 손을 거두는 순간보다는 늘어뜨린 자세에 가깝다.",
        "built_space": "뒤쪽 문 하나와 그 왼쪽 금속 선반 하나, 오른쪽 창 하나, 오른쪽 침구 자리 하나, 중앙 러그 하나가 보인다. 오른쪽 가장자리에는 목재 수납장이 일부 보인다. 이현우는 왼쪽 전경 의자에 앉고 찰리는 오른쪽에 서 있어 좌우 관계와 컨테이너 생활 공간의 재질은 유지된다. 그러나 찰리의 발까지 보여주는 넓은 구도이고 이현우가 화면 왼쪽을 크게 차지해, 작은 이현우와 굽은 상체 중심의 지정 구도를 충족하지 못한다.",
        "entities": "찰리와 이현우만 등장한다. 찰리는 참조의 베이지 기계 장갑, 흰 마스크 얼굴, 점 형태 눈과 선 형태 입, 파란 가슴 장치를 유지하며 낡은 갈색 코트와 모자도 착용한다. 다만 눈빛은 요구된 파란색이 아닌 주황색이고, 가슴 장치의 마모된 로고는 명확하지 않다. 이현우는 젊은 동아시아계 남성으로 보이고 헝클어진 검은 머리, 남색 상의, 얼굴 상처가 확인된다. 드러난 다리는 피 묻은 붕대로 감겨 있어 미처치 상태와 다르다. 바닥에는 청백색 화병 파편과 물이 남아 있다.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 바닥에 닿아 몸을 지지하고, 굽힌 목과 몸통 및 늘어진 팔은 기계 관절로 연결된다. 모자는 머리에 얹혀 있고 코트는 어깨에 걸쳐 아래로 처진다. 이현우는 의자 좌판에 앉아 팔을 무릎 쪽에 기대고 있다. 큰 파편들은 바닥에 놓여 있다. 파편 사이의 작은 물방울은 아직 튀는 듯 보여, 이미 파편을 밀어 숨긴 뒤라는 정적인 시점과는 다소 어긋난다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "고개를 숙인 동작과 좌우 관계는 맞지만, 상체 중심 미디엄 숏 대신 전신을 보여주고 이현우가 지나치게 크며 낡은 코트와 모자도 빠졌다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "찰리의 움츠린 자세와 낡은 코트·모자를 더 충실히 구현했으나, 전신 구도와 크게 잡힌 이현우는 지정된 상체 중심 촬영에 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 눈은 아래쪽 도자기 파편이 있는 바닥을 향한다. 이현우는 오른쪽의 찰리를 바라본다. 시선 관계는 맞지만, 찰리의 손은 몸 쪽으로 거두기보다 바닥 가까이 늘어져 있다.",
        "built_space": "뒤쪽 문 하나, 그 왼쪽 금속 수납 선반 하나, 오른쪽 벽의 창 구역과 침구 자리 하나, 중앙 무늬 러그 하나가 보인다. 이현우는 왼쪽 전경의 팔걸이 의자에 앉고 찰리는 오른쪽 바닥에 서 있다. 골이 진 벽과 낡은 나무 바닥은 장소 참조와 대체로 이어진다. 다만 카메라는 찰리의 어깨와 상체에 머무르지 않고 발까지 담으며, 이현우도 왼쪽 가장자리의 작은 인물이 아니라 큰 전경 인물이다.",
        "entities": "등장인물은 찰리와 이현우 둘이다. 찰리의 샌드 베이지 장갑판, 긴 팔과 짧은 다리, 흰 마스크 얼굴, 선 형태 입, 파란 원형 가슴 장치는 참조에 부합한다. 눈은 참조처럼 주황색이지만 이번 장면의 파란 눈 지시와 다르고, 요구된 코트와 모자는 없다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 남색 상의와 얼굴 상처가 확인된다. 보이는 다리에는 피 묻은 붕대가 있어 미처치 물림 상태와 다르다. 청백색 화병 파편은 바닥에 남아 있다.",
        "hard_violations": [],
        "physics": "찰리는 두 기계 발로 바닥을 딛고 허리와 목을 굽힌다. 팔과 손은 어깨 및 팔꿈치 관절에 연결되어 아래로 내려오며, 떠 있는 몸은 없다. 이현우의 골반은 의자 좌판에 지지되고 손은 무릎 부근에 놓여 있다. 큰 화병 조각들은 바닥에 놓여 있어 지지 관계가 자연스럽다."
       },
       {
        "label": "A",
        "direction": "찰리는 모자챙 아래로 고개를 깊이 숙여 앞쪽 바닥과 파편을 향한다. 이현우의 얼굴은 오른쪽 찰리에게 돌아가 있다. 두 손은 낮게 내려와 몸 가까이에 있지만, 적극적으로 손을 거두는 순간보다는 늘어뜨린 자세에 가깝다.",
        "built_space": "뒤쪽 문 하나와 그 왼쪽 금속 선반 하나, 오른쪽 창 하나, 오른쪽 침구 자리 하나, 중앙 러그 하나가 보인다. 오른쪽 가장자리에는 목재 수납장이 일부 보인다. 이현우는 왼쪽 전경 의자에 앉고 찰리는 오른쪽에 서 있어 좌우 관계와 컨테이너 생활 공간의 재질은 유지된다. 그러나 찰리의 발까지 보여주는 넓은 구도이고 이현우가 화면 왼쪽을 크게 차지해, 작은 이현우와 굽은 상체 중심의 지정 구도를 충족하지 못한다.",
        "entities": "찰리와 이현우만 등장한다. 찰리는 참조의 베이지 기계 장갑, 흰 마스크 얼굴, 점 형태 눈과 선 형태 입, 파란 가슴 장치를 유지하며 낡은 갈색 코트와 모자도 착용한다. 다만 눈빛은 요구된 파란색이 아닌 주황색이고, 가슴 장치의 마모된 로고는 명확하지 않다. 이현우는 젊은 동아시아계 남성으로 보이고 헝클어진 검은 머리, 남색 상의, 얼굴 상처가 확인된다. 드러난 다리는 피 묻은 붕대로 감겨 있어 미처치 상태와 다르다. 바닥에는 청백색 화병 파편과 물이 남아 있다.",
        "hard_violations": [],
        "physics": "찰리의 두 발이 바닥에 닿아 몸을 지지하고, 굽힌 목과 몸통 및 늘어진 팔은 기계 관절로 연결된다. 모자는 머리에 얹혀 있고 코트는 어깨에 걸쳐 아래로 처진다. 이현우는 의자 좌판에 앉아 팔을 무릎 쪽에 기대고 있다. 큰 파편들은 바닥에 놓여 있다. 파편 사이의 작은 물방울은 아직 튀는 듯 보여, 이미 파편을 밀어 숨긴 뒤라는 정적인 시점과는 다소 어긋난다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.238
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.238
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1238
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "찰리가 낡은 코트와 모자를 착용해야 한다는 지시를 정확히 반영하여 프롬프트 충실도가 가장 높습니다."
   },
   {
    "label": "B",
    "score": 1238,
    "verdict_ko": "찰리의 낡은 코트와 모자 착용 지시를 완전히 누락하여 주요 복장 요건을 충족하지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S14sh3_sel.png",
    "asset_id": "768c7bd6-6825-4b71-b885-dbf8c92f1d67",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08a7-ed83-7f4e-827e-284739d70ec6",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh3"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S14sh12::signage": {
  "fp": "8370a89447439055",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh12": {
  "input_fingerprint": "90e6d990079b7050",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 현관문 쪽을 팔로 강하게 가리킨 채 찰리에게 소리치는 역동적인 상체.\n\nLOCATION (lock): In the daylit living area of the container home, facing its front entrance door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track beside 찰리 but outside his shoulder line, at 이현우's lower-chest height and tilted upward into the prescribed three-quarter view. Place 이현우's upper body center-left, facing and shouting at 찰리 along the near-left edge, while his extended arm leads across the right half toward the entrance; 찰리 angles his lowered head back toward him. Preserve the medium distance and emphasize 이현우's new position beside 찰리, leaving enough lateral room for the entire pointing gesture and a partial view of its destination.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-left of the frame, midground, points to Container entrance at screen right; Container entrance at screen right in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance (The destination indicated by 이현우's extended arm; door position is not specified) — A partial interior-side view appears at the right edge; used as Makes the command to leave visually explicit without redirecting 이현우's face away from 찰리.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained daytime illumination and controlled contrast, without a dramatic lighting change to exaggerate the reprimand.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the container's interior surfaces, household objects, and daylight appearance from the reference. Retain the broken vase pieces as settled debris where applicable. Exclude airborne shards and do not restore an intact vase.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The displaced vase fragments remain hidden aside on the container floor. Charlie still has the old coat and hat, blue-lit eyes and worn UBIK chest logo. 이현우: He has risen and moved away from his seat, preparing to leave. His facial injuries and untreated leg bite remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 현관문 쪽을 팔로 강하게 가리킨 채 찰리에게 소리치는 역동적인 상체.\n\nLOCATION (lock): In the daylit living area of the container home, facing its front entrance door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track beside 찰리 but outside his shoulder line, at 이현우's lower-chest height and tilted upward into the prescribed three-quarter view. Place 이현우's upper body center-left, facing and shouting at 찰리 along the near-left edge, while his extended arm leads across the right half toward the entrance; 찰리 angles his lowered head back toward him. Preserve the medium distance and emphasize 이현우's new position beside 찰리, leaving enough lateral room for the entire pointing gesture and a partial view of its destination.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-left of the frame, midground, points to Container entrance at screen right; Container entrance at screen right in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance (The destination indicated by 이현우's extended arm; door position is not specified) — A partial interior-side view appears at the right edge; used as Makes the command to leave visually explicit without redirecting 이현우's face away from 찰리.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained daytime illumination and controlled contrast, without a dramatic lighting change to exaggerate the reprimand.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the container's interior surfaces, household objects, and daylight appearance from the reference. Retain the broken vase pieces as settled debris where applicable. Exclude airborne shards and do not restore an intact vase.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The displaced vase fragments remain hidden aside on the container floor. Charlie still has the old coat and hat, blue-lit eyes and worn UBIK chest logo. 이현우: He has risen and moved away from his seat, preparing to leave. His facial injuries and untreated leg bite remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 이현우가 현관문 쪽을 팔로 강하게 가리킨 채 찰리에게 소리치는 역동적인 상체.\n\nLOCATION (lock): In the daylit living area of the container home, facing its front entrance door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track beside 찰리 but outside his shoulder line, at 이현우's lower-chest height and tilted upward into the prescribed three-quarter view. Place 이현우's upper body center-left, facing and shouting at 찰리 along the near-left edge, while his extended arm leads across the right half toward the entrance; 찰리 angles his lowered head back toward him. Preserve the medium distance and emphasize 이현우's new position beside 찰리, leaving enough lateral room for the entire pointing gesture and a partial view of its destination.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우 in the middle-left of the frame, midground, points to Container entrance at screen right; Container entrance at screen right in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Container entrance (The destination indicated by 이현우's extended arm; door position is not specified) — A partial interior-side view appears at the right edge; used as Makes the command to leave visually explicit without redirecting 이현우's face away from 찰리.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained daytime illumination and controlled contrast, without a dramatic lighting change to exaggerate the reprimand.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the container's interior surfaces, household objects, and daylight appearance from the reference. Retain the broken vase pieces as settled debris where applicable. Exclude airborne shards and do not restore an intact vase.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The displaced vase fragments remain hidden aside on the container floor. Charlie still has the old coat and hat, blue-lit eyes and worn UBIK chest logo. 이현우: He has risen and moved away from his seat, preparing to leave. His facial injuries and untreated leg bite remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 화면 왼쪽의 찰리를 향해 고개를 돌려 소리치며, 왼팔을 뻗어 화면 오른쪽 배경의 출입문을 가리키고 있다. 찰리는 고개를 들어 이현우를 향하고 있다.",
    "built_space": "컨테이너 내부 구조, 바닥의 무늬가 있는 카펫, 측면의 선반, 그리고 배경에 위치한 유리창이 달린 문 등 이전 샷의 환경 디자인과 디테일이 정확히 일치한다.",
    "entities": "이현우는 상처 입은 얼굴, 흙먼지가 묻은 옷, 인이어 무전기 등 참조 이미지 및 텍스트와 일치하는 인물로 묘사되었다. 찰리는 낡은 코트와 모자를 착용하고 있으며, 특유의 흰색 마스크형 얼굴 구조가 측면에서도 잘 유지되고 있다 (프롬프트의 파란 눈 대신 참조 이미지의 주황색 눈이 반영됨).",
    "hard_violations": [],
    "physics": "이현우의 서 있는 자세와 뻗은 팔의 동작, 몸의 방향이 자연스러우며, 찰리 역시 지면에 안정적으로 서서 몸을 지탱하고 있다."
   },
   {
    "label": "B",
    "direction": "이현우는 왼쪽의 찰리를 보며 소리치고, 팔을 오른쪽으로 뻗어 복도 끝 출입문을 가리킨다. 찰리는 이현우 쪽을 향해 서 있다.",
    "built_space": "내부 구조는 뼈대를 유지하고 있으나, 배경 끝에 있는 문이 유리창이 없는 밋밋한 회색 철문으로 변형되어 이전 샷의 장소와 연속성이 떨어진다.",
    "entities": "이현우의 외형은 기준에 부합하게 잘 묘사되었다. 그러나 찰리는 기계 얼굴의 측면에서 흰색 마스크 부분이 누락되고 베이지색 장갑판 위에 눈이 직접 위치하는 등 원래의 디자인이 왜곡되었다.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 안정적으로 서서 체중을 지탱하고 있으나, 화면 좌측 하단 이현우의 손에 명확하지 않은 검은색 물체가 들려 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "이현우가 찰리에게 소리치며 출입문을 가리키는 역동적인 구도를 완벽하게 소화했으며, 이전 샷의 컨테이너 내부 배경과 캐릭터 외형을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 구도와 동작은 따랐으나, 로봇 찰리의 흰색 얼굴판 구조가 훼손되었고 배경의 출입문 디자인이 이전 샷과 일치하지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽의 찰리를 향해 고개를 돌려 소리치며, 왼팔을 뻗어 화면 오른쪽 배경의 출입문을 가리키고 있다. 찰리는 고개를 들어 이현우를 향하고 있다.",
        "built_space": "컨테이너 내부 구조, 바닥의 무늬가 있는 카펫, 측면의 선반, 그리고 배경에 위치한 유리창이 달린 문 등 이전 샷의 환경 디자인과 디테일이 정확히 일치한다.",
        "entities": "이현우는 상처 입은 얼굴, 흙먼지가 묻은 옷, 인이어 무전기 등 참조 이미지 및 텍스트와 일치하는 인물로 묘사되었다. 찰리는 낡은 코트와 모자를 착용하고 있으며, 특유의 흰색 마스크형 얼굴 구조가 측면에서도 잘 유지되고 있다 (프롬프트의 파란 눈 대신 참조 이미지의 주황색 눈이 반영됨).",
        "hard_violations": [],
        "physics": "이현우의 서 있는 자세와 뻗은 팔의 동작, 몸의 방향이 자연스러우며, 찰리 역시 지면에 안정적으로 서서 몸을 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽의 찰리를 보며 소리치고, 팔을 오른쪽으로 뻗어 복도 끝 출입문을 가리킨다. 찰리는 이현우 쪽을 향해 서 있다.",
        "built_space": "내부 구조는 뼈대를 유지하고 있으나, 배경 끝에 있는 문이 유리창이 없는 밋밋한 회색 철문으로 변형되어 이전 샷의 장소와 연속성이 떨어진다.",
        "entities": "이현우의 외형은 기준에 부합하게 잘 묘사되었다. 그러나 찰리는 기계 얼굴의 측면에서 흰색 마스크 부분이 누락되고 베이지색 장갑판 위에 눈이 직접 위치하는 등 원래의 디자인이 왜곡되었다.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 서서 체중을 지탱하고 있으나, 화면 좌측 하단 이현우의 손에 명확하지 않은 검은색 물체가 들려 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "이현우가 찰리에게 소리치며 출입문을 가리키는 역동적인 구도를 완벽하게 소화했으며, 이전 샷의 컨테이너 내부 배경과 캐릭터 외형을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 구도와 동작은 따랐으나, 로봇 찰리의 흰색 얼굴판 구조가 훼손되었고 배경의 출입문 디자인이 이전 샷과 일치하지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽의 찰리를 향해 고개를 돌려 소리치며, 왼팔을 뻗어 화면 오른쪽 배경의 출입문을 가리키고 있다. 찰리는 고개를 들어 이현우를 향하고 있다.",
        "built_space": "컨테이너 내부 구조, 바닥의 무늬가 있는 카펫, 측면의 선반, 그리고 배경에 위치한 유리창이 달린 문 등 이전 샷의 환경 디자인과 디테일이 정확히 일치한다.",
        "entities": "이현우는 상처 입은 얼굴, 흙먼지가 묻은 옷, 인이어 무전기 등 참조 이미지 및 텍스트와 일치하는 인물로 묘사되었다. 찰리는 낡은 코트와 모자를 착용하고 있으며, 특유의 흰색 마스크형 얼굴 구조가 측면에서도 잘 유지되고 있다 (프롬프트의 파란 눈 대신 참조 이미지의 주황색 눈이 반영됨).",
        "hard_violations": [],
        "physics": "이현우의 서 있는 자세와 뻗은 팔의 동작, 몸의 방향이 자연스러우며, 찰리 역시 지면에 안정적으로 서서 몸을 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽의 찰리를 보며 소리치고, 팔을 오른쪽으로 뻗어 복도 끝 출입문을 가리킨다. 찰리는 이현우 쪽을 향해 서 있다.",
        "built_space": "내부 구조는 뼈대를 유지하고 있으나, 배경 끝에 있는 문이 유리창이 없는 밋밋한 회색 철문으로 변형되어 이전 샷의 장소와 연속성이 떨어진다.",
        "entities": "이현우의 외형은 기준에 부합하게 잘 묘사되었다. 그러나 찰리는 기계 얼굴의 측면에서 흰색 마스크 부분이 누락되고 베이지색 장갑판 위에 눈이 직접 위치하는 등 원래의 디자인이 왜곡되었다.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 서서 체중을 지탱하고 있으나, 화면 좌측 하단 이현우의 손에 명확하지 않은 검은색 물체가 들려 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리를 향한 고함과 현관문을 향한 손짓은 정확하지만, 상체보다 하체와 실내의 비중이 크고 현관문도 오른쪽 가장자리보다 안쪽에 놓인다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "이현우의 역동적인 상체와 찰리의 반응을 더 밀도 있게 담고 현관문도 더 오른쪽에 배치하지만, 문 전체 노출과 이전 장면의 반소매 복장 불일치는 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 얼굴과 눈을 왼쪽 전경의 찰리에게 돌리고 입을 벌려 소리친다. 오른쪽으로 뻗은 팔의 검지 연장선은 배경 현관문의 상부에 닿아 지시 대상이 분명하다. 찰리는 고개를 깊이 숙였으며 얼굴이 이현우 쪽으로 약간 돌아 있지만, 그를 되돌아보는 반응은 약하다.",
        "built_space": "뒤쪽에 닫힌 현관문 하나, 오른쪽 벽에 창 하나, 문 양옆에 수납 선반, 오른쪽 가장자리에 별도 책장과 체크무늬 좌석 일부가 보인다. 천장등 하나와 바닥 러그 하나도 보인다. 낡은 세로 골 금속 벽과 나무 바닥, 선반의 생활용품, 창으로 드는 낮빛은 이전 장소와 대체로 이어진다. 찰리는 왼쪽 전경, 이현우는 그 옆 중경에 서 있어 가구와 충돌하지 않는다. 다만 문은 가장자리의 부분 모습이 아니라 오른쪽 안쪽에 거의 전부 드러난다. 불가능한 반사는 보이지 않는다.",
        "entities": "등장 개체는 이현우와 찰리 둘뿐이다. 이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 마른 체격과 뺨의 상처를 유지하며 귀의 소형 장치도 보인다. 먼지와 핏자국이 있는 긴소매 단추 셔츠는 인물 참고에는 가깝지만, 복장이 고정된 이전 장면의 어두운 반소매 티셔츠와 다르다. 찰리는 마모된 베이지색 기계 장갑, 흰 얼굴판, 낡은 모자와 갈색 외투를 유지한다. 보이는 눈빛은 요구된 파란색이 아니라 참고처럼 주황색이다. 가슴 로고와 다리 상처는 가려지거나 프레임 밖이므로 확인할 수 없다. 온전한 꽃병이나 공중 파편은 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 몸통은 화면 아래로 이어지는 골반과 다리 위에서 앞으로 기울어 있고, 지시하는 팔은 어깨와 팔꿈치로 자연스럽게 이어진다. 발은 프레임 밖이지만 공중에 떠 있다는 징후는 없다. 찰리의 상체와 장갑 팔도 아래쪽 몸체로 이어지며, 모자는 머리에, 외투는 어깨에 걸쳐져 있다. 가구와 용기는 바닥이나 선반에 놓여 있고 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 눈과 열린 입은 왼쪽 가까이에 있는 찰리의 얼굴을 향한다. 뻗은 검지는 오른쪽 뒤 현관문의 위쪽 영역을 가리키므로 고함의 상대와 퇴거 방향이 분리되어 읽힌다. 찰리는 머리를 낮춘 상태에서 얼굴판을 이현우 쪽으로 돌리고 있어, 고개를 숙인 채 반응한다는 지시가 A보다 잘 드러난다.",
        "built_space": "닫힌 현관문 하나가 오른쪽 배경에 있고, 오른쪽 가장자리에는 창 하나가 일부 보인다. 통로 왼쪽에는 금속 수납 선반, 문 오른쪽에는 생활용품 선반이 있으며, 하단에는 러그 하나와 오른쪽 좌석·수납함 일부가 보인다. 천장에는 밝은 조명 면 두 곳이 보인다. 금속 벽의 마모, 나무 바닥, 생활용품과 자연광은 이전 장소의 특징을 대체로 유지한다. 찰리 옆의 이현우를 더 크게 담아 상체 동작이 강조되고 문도 A보다 오른쪽으로 이동했지만, 문을 가장자리에서 부분적으로만 보여 달라는 구도는 완전히 지키지 않는다. 강한 올려다보기보다는 거의 수평에 가까운 시점이며, 불가능한 반사는 없다.",
        "entities": "이현우와 찰리 외의 인물은 없다. 이현우의 젊은 동아시아계 외모, 검은 헝클어진 머리, 마른 체격, 얼굴 상처와 귀의 작은 장치는 요구에 부합한다. 긴소매 단추 셔츠와 바지는 인물 참고에 가깝지만 이전 장면의 반소매 티셔츠를 유지하지 않았다. 찰리는 베이지색 기계 장갑과 흰 얼굴판, 낡은 모자와 외투를 갖추고 있으며 인간 피부나 치아가 추가되지 않았다. 노출된 눈빛은 파란색이 아닌 주황색이다. 가슴 로고와 이현우의 다리 상처는 이 구도에서 판독할 수 없다. 온전한 꽃병과 떠다니는 파편은 없다.",
        "hard_violations": [],
        "physics": "이현우는 골반 위에서 상체를 찰리 쪽으로 기울이고 어깨에서 팔을 길게 뻗는다. 관절 연결과 손가락의 지시 자세는 가능한 동작이며, 화면 밖의 발이 보이지 않는다는 이유로 부유 상태로 읽히지는 않는다. 찰리의 숙인 머리는 목과 몸통으로 지지되고 모자와 외투도 각각 머리와 어깨에 얹혀 있다. 배경 물건들은 선반이나 바닥에 놓여 있으며, 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리를 향한 고함과 현관문을 향한 손짓은 정확하지만, 상체보다 하체와 실내의 비중이 크고 현관문도 오른쪽 가장자리보다 안쪽에 놓인다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "이현우의 역동적인 상체와 찰리의 반응을 더 밀도 있게 담고 현관문도 더 오른쪽에 배치하지만, 문 전체 노출과 이전 장면의 반소매 복장 불일치는 남는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 얼굴과 눈을 왼쪽 전경의 찰리에게 돌리고 입을 벌려 소리친다. 오른쪽으로 뻗은 팔의 검지 연장선은 배경 현관문의 상부에 닿아 지시 대상이 분명하다. 찰리는 고개를 깊이 숙였으며 얼굴이 이현우 쪽으로 약간 돌아 있지만, 그를 되돌아보는 반응은 약하다.",
        "built_space": "뒤쪽에 닫힌 현관문 하나, 오른쪽 벽에 창 하나, 문 양옆에 수납 선반, 오른쪽 가장자리에 별도 책장과 체크무늬 좌석 일부가 보인다. 천장등 하나와 바닥 러그 하나도 보인다. 낡은 세로 골 금속 벽과 나무 바닥, 선반의 생활용품, 창으로 드는 낮빛은 이전 장소와 대체로 이어진다. 찰리는 왼쪽 전경, 이현우는 그 옆 중경에 서 있어 가구와 충돌하지 않는다. 다만 문은 가장자리의 부분 모습이 아니라 오른쪽 안쪽에 거의 전부 드러난다. 불가능한 반사는 보이지 않는다.",
        "entities": "등장 개체는 이현우와 찰리 둘뿐이다. 이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 마른 체격과 뺨의 상처를 유지하며 귀의 소형 장치도 보인다. 먼지와 핏자국이 있는 긴소매 단추 셔츠는 인물 참고에는 가깝지만, 복장이 고정된 이전 장면의 어두운 반소매 티셔츠와 다르다. 찰리는 마모된 베이지색 기계 장갑, 흰 얼굴판, 낡은 모자와 갈색 외투를 유지한다. 보이는 눈빛은 요구된 파란색이 아니라 참고처럼 주황색이다. 가슴 로고와 다리 상처는 가려지거나 프레임 밖이므로 확인할 수 없다. 온전한 꽃병이나 공중 파편은 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 몸통은 화면 아래로 이어지는 골반과 다리 위에서 앞으로 기울어 있고, 지시하는 팔은 어깨와 팔꿈치로 자연스럽게 이어진다. 발은 프레임 밖이지만 공중에 떠 있다는 징후는 없다. 찰리의 상체와 장갑 팔도 아래쪽 몸체로 이어지며, 모자는 머리에, 외투는 어깨에 걸쳐져 있다. 가구와 용기는 바닥이나 선반에 놓여 있고 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 눈과 열린 입은 왼쪽 가까이에 있는 찰리의 얼굴을 향한다. 뻗은 검지는 오른쪽 뒤 현관문의 위쪽 영역을 가리키므로 고함의 상대와 퇴거 방향이 분리되어 읽힌다. 찰리는 머리를 낮춘 상태에서 얼굴판을 이현우 쪽으로 돌리고 있어, 고개를 숙인 채 반응한다는 지시가 A보다 잘 드러난다.",
        "built_space": "닫힌 현관문 하나가 오른쪽 배경에 있고, 오른쪽 가장자리에는 창 하나가 일부 보인다. 통로 왼쪽에는 금속 수납 선반, 문 오른쪽에는 생활용품 선반이 있으며, 하단에는 러그 하나와 오른쪽 좌석·수납함 일부가 보인다. 천장에는 밝은 조명 면 두 곳이 보인다. 금속 벽의 마모, 나무 바닥, 생활용품과 자연광은 이전 장소의 특징을 대체로 유지한다. 찰리 옆의 이현우를 더 크게 담아 상체 동작이 강조되고 문도 A보다 오른쪽으로 이동했지만, 문을 가장자리에서 부분적으로만 보여 달라는 구도는 완전히 지키지 않는다. 강한 올려다보기보다는 거의 수평에 가까운 시점이며, 불가능한 반사는 없다.",
        "entities": "이현우와 찰리 외의 인물은 없다. 이현우의 젊은 동아시아계 외모, 검은 헝클어진 머리, 마른 체격, 얼굴 상처와 귀의 작은 장치는 요구에 부합한다. 긴소매 단추 셔츠와 바지는 인물 참고에 가깝지만 이전 장면의 반소매 티셔츠를 유지하지 않았다. 찰리는 베이지색 기계 장갑과 흰 얼굴판, 낡은 모자와 외투를 갖추고 있으며 인간 피부나 치아가 추가되지 않았다. 노출된 눈빛은 파란색이 아닌 주황색이다. 가슴 로고와 이현우의 다리 상처는 이 구도에서 판독할 수 없다. 온전한 꽃병과 떠다니는 파편은 없다.",
        "hard_violations": [],
        "physics": "이현우는 골반 위에서 상체를 찰리 쪽으로 기울이고 어깨에서 팔을 길게 뻗는다. 관절 연결과 손가락의 지시 자세는 가능한 동작이며, 화면 밖의 발이 보이지 않는다는 이유로 부유 상태로 읽히지는 않는다. 찰리의 숙인 머리는 목과 몸통으로 지지되고 모자와 외투도 각각 머리와 어깨에 얹혀 있다. 배경 물건들은 선반이나 바닥에 놓여 있으며, 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.542
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.542
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1542
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이현우가 찰리에게 소리치며 출입문을 가리키는 역동적인 구도를 완벽하게 소화했으며, 이전 샷의 컨테이너 내부 배경과 캐릭터 외형을 매우 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1542,
    "verdict_ko": "지시된 구도와 동작은 따랐으나, 로봇 찰리의 흰색 얼굴판 구조가 훼손되었고 배경의 출입문 디자인이 이전 샷과 일치하지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S14sh7_sel.png",
    "asset_id": "8f9e6704-f959-44f6-958d-99f0e470058d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1305657>",
    "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08ac-e6b0-7f16-9360-4acfd6399198",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh7"
  },
  "staged_characters_added": [
   "C01"
  ]
 },
 "S15sh9::signage": {
  "fp": "dabe4b0bed0a2c91",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "5억 원의 현상금이 굵은 글씨로 적힌 실종 로봇 기사"
   },
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "실종 로봇 기사"
   }
  ],
  "dropped": []
 },
 "S15sh9": {
  "input_fingerprint": "15aaff81be0e5105",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 컴퓨터 모니터 화면에 5억 원의 현상금이 굵은 글씨로 적힌 실종 로봇 기사가 떠 있는 근접 구도.\n\nLOCATION (lock): At the computer workstation inside the refugee settlement's cluttered secondhand repair shop. Daylight from the shop entrance provides ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move along the line just outside 마츠다's shoulder, looking gently down at the monitor from a shallow lateral angle while leaving him entirely outside the image. Place the visible screen near center, occupying a little over a third of the frame, and focus precisely on the bold article line '유빅사, 5억원 현상금 걸어'; retain the bezel and softly rendered stacked appliances around it as physical context. Present the article as content on a real monitor, not a full-frame graphic or a subjective view through 마츠다's eyes.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Computer monitor (Displaying an internet news article about the missing robot and the 500-million-won reward) — The display face is visible at a shallow oblique angle, showing '유빅사에서 제조된 유일한 모델. 해남 인근서 실종' and the emphasized reward line '유빅사, 5억원 현상금 걸어'; used as Provides the narrative discovery through readable screen content while its bezel preserves the mediated presentation; Stacked appliances and scrap (Piled throughout the shop); used as Soft surrounding context keeps the monitor grounded in the crowded secondhand shop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the monitor's displayed text legible against restrained daytime shop ambience, without adding glare, graphic effects or an invented colored cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cluttered shop contains piles of appliances and scrap, and the computer browser displays the report identifying Charlie as UBIK's unique missing model with a 500-million-won reward. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 컴퓨터 모니터 화면에 5억 원의 현상금이 굵은 글씨로 적힌 실종 로봇 기사가 떠 있는 근접 구도.\n\nLOCATION (lock): At the computer workstation inside the refugee settlement's cluttered secondhand repair shop. Daylight from the shop entrance provides ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move along the line just outside 마츠다's shoulder, looking gently down at the monitor from a shallow lateral angle while leaving him entirely outside the image. Place the visible screen near center, occupying a little over a third of the frame, and focus precisely on the bold article line '유빅사, 5억원 현상금 걸어'; retain the bezel and softly rendered stacked appliances around it as physical context. Present the article as content on a real monitor, not a full-frame graphic or a subjective view through 마츠다's eyes.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Computer monitor (Displaying an internet news article about the missing robot and the 500-million-won reward) — The display face is visible at a shallow oblique angle, showing '유빅사에서 제조된 유일한 모델. 해남 인근서 실종' and the emphasized reward line '유빅사, 5억원 현상금 걸어'; used as Provides the narrative discovery through readable screen content while its bezel preserves the mediated presentation; Stacked appliances and scrap (Piled throughout the shop); used as Soft surrounding context keeps the monitor grounded in the crowded secondhand shop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the monitor's displayed text legible against restrained daytime shop ambience, without adding glare, graphic effects or an invented colored cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cluttered shop contains piles of appliances and scrap, and the computer browser displays the report identifying Charlie as UBIK's unique missing model with a 500-million-won reward. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 컴퓨터 모니터 화면에 5억 원의 현상금이 굵은 글씨로 적힌 실종 로봇 기사가 떠 있는 근접 구도.\n\nLOCATION (lock): At the computer workstation inside the refugee settlement's cluttered secondhand repair shop. Daylight from the shop entrance provides ambient light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move along the line just outside 마츠다's shoulder, looking gently down at the monitor from a shallow lateral angle while leaving him entirely outside the image. Place the visible screen near center, occupying a little over a third of the frame, and focus precisely on the bold article line '유빅사, 5억원 현상금 걸어'; retain the bezel and softly rendered stacked appliances around it as physical context. Present the article as content on a real monitor, not a full-frame graphic or a subjective view through 마츠다's eyes.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Computer monitor (Displaying an internet news article about the missing robot and the 500-million-won reward) — The display face is visible at a shallow oblique angle, showing '유빅사에서 제조된 유일한 모델. 해남 인근서 실종' and the emphasized reward line '유빅사, 5억원 현상금 걸어'; used as Provides the narrative discovery through readable screen content while its bezel preserves the mediated presentation; Stacked appliances and scrap (Piled throughout the shop); used as Soft surrounding context keeps the monitor grounded in the crowded secondhand shop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the monitor's displayed text legible against restrained daytime shop ambience, without adding glare, graphic effects or an invented colored cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cluttered shop contains piles of appliances and scrap, and the computer browser displays the report identifying Charlie as UBIK's unique missing model with a 500-million-won reward. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
    "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
    "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
    "hard_violations": [],
    "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
    "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
    "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
    "hard_violations": [],
    "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으나, B에 비해 주변 사물과 텍스트 디테일을 더 선명하게 유지했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으며, 모니터 베젤 등 세부 묘사에서 미세한 열화가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
        "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
        "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
        "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
        "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으나, B에 비해 주변 사물과 텍스트 디테일을 더 선명하게 유지했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으며, 모니터 베젤 등 세부 묘사에서 미세한 열화가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
        "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
        "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 약간 위에서 아래로 책상 중앙의 모니터 화면을 향해 정확히 시선을 고정하고 있습니다.",
        "built_space": "책상 위에 좌측 PC 본체, 중앙 평면 모니터, 우측 CRT TV 및 배경의 선반들이 원본과 동일하게 배치되어 있습니다.",
        "entities": "모니터 화면 기사 내용 중 프롬프트가 요구한 '유빅사'가 '유빙사'로 표기되었습니다. 지시대로 마츠다 등 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "모니터, 키보드, 본체 등 모든 사물이 나무 책상 위에 물리적으로 안정감 있게 놓여 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "화면이 더 크게 잡히고 측면 각도와 배경 흐림이 살아 있어 지정된 근접 구도에 더 가깝지만, 화면 점유율과 ‘유빅사’의 정확한 표기는 미달한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "작업대와 현상금 기사는 충실하지만 참조 사진의 구도를 거의 유지해 화면이 더 작고 정면에 가까우며, ‘유빅사’도 ‘유빙사’로 표시된다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "모니터 표시 면은 작업대 앞 독자 쪽을 향하며, 카메라는 그 면을 얕은 측면 각도에서 본다. 상판과 받침 윗면이 보여 약한 내려다보기와 양립한다. 사람이나 시선, 무기, 이동체는 없다.",
        "built_space": "중앙 평판 모니터 1대, 왼쪽 본체 1대, 오른쪽 브라운관 TV 1대, 앞쪽 키보드 1개와 오른쪽 마우스 1개가 보인다. 뒤쪽 부품 서랍과 선반, 중앙 창, 벽 안내판은 참조 장소의 구성과 대응한다. 사람은 완전히 제외되어 있다. TV의 밝은 개구부 반사는 보이는 채광 환경에서 가능하다. 화면 자체는 전체 면적의 약 28%로, 요구된 3분의 1 초과에는 못 미치지만 B보다 크다.",
        "entities": "낡은 삼성 모니터와 본체, GoldStar TV, 공구통과 잡동사니가 참조와 대응한다. 실제 모니터 안에 한국어 실종 로봇 기사와 굵은 붉은색 ‘5억원 현상금 걸어’가 표시된다. 그러나 회사명은 요구된 ‘유빅사’가 아니라 ‘유빙사’로 읽히며, 유일 모델과 해남 인근 실종 설명에도 같은 차이가 있다. 마츠다는 보이지 않고 로봇 사진도 없어 인물 외모나 찰리의 의상은 검증 대상이 아니다.",
        "hard_violations": [],
        "physics": "모니터는 목과 넓은 받침으로 작업대에 지지되고, TV는 직사각형 받침 위에 놓여 있다. 본체·키보드·마우스·공구통은 작업대에 놓였고 공구는 통에 꽂혀 있다. 떠 있는 물체나 지지 없는 신체는 없다."
       },
       {
        "label": "B",
        "direction": "모니터는 작업대 앞을 향하고 카메라에 거의 정면으로 보이되 약한 측면 원근이 있다. 작업대와 받침 윗면이 보여 약한 내려다보기는 성립한다. 인물의 시선이나 이동 방향은 없으며, 표시 면이 뒤집히거나 독자 반대쪽을 향하지 않는다.",
        "built_space": "중앙 모니터 1대, 왼쪽 본체 1대, 오른쪽 브라운관 TV 1대, 전면 키보드 1개와 오른쪽 마우스 1개가 있다. 공구통, 부품 서랍, 창과 벽 안내판까지 참조 사진과 매우 가까운 배치다. 추가 인물이나 중복 설비는 없다. TV의 창 형태 반사는 이 공간에서 가능하다. 화면 면적은 전체의 약 26%로 요구보다 작고, 참조 사진의 시점과 구도를 거의 그대로 유지한다.",
        "entities": "참조의 낡은 컴퓨터와 TV, 공구 및 작업대 재질이 유지된다. 한국어 인터넷 기사에 유일 모델의 해남 인근 실종과 굵은 ‘5억원 현상금 걸어’가 선명하다. 다만 지정된 회사명 ‘유빅사’ 대신 ‘유빙사’가 반복된다. 마츠다는 프레임 밖이며 찰리의 외형을 보여 주는 사진은 없다.",
        "hard_violations": [],
        "physics": "모니터 받침과 TV 아래 받침이 각각 무게를 지탱한다. 키보드와 마우스는 작업대에 닿아 있고, 공구는 용기에 담겨 있으며 뒤쪽 물건은 선반에 놓여 있다. 지지 없는 물체나 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면이 더 크게 잡히고 측면 각도와 배경 흐림이 살아 있어 지정된 근접 구도에 더 가깝지만, 화면 점유율과 ‘유빅사’의 정확한 표기는 미달한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "작업대와 현상금 기사는 충실하지만 참조 사진의 구도를 거의 유지해 화면이 더 작고 정면에 가까우며, ‘유빅사’도 ‘유빙사’로 표시된다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "모니터 표시 면은 작업대 앞 독자 쪽을 향하며, 카메라는 그 면을 얕은 측면 각도에서 본다. 상판과 받침 윗면이 보여 약한 내려다보기와 양립한다. 사람이나 시선, 무기, 이동체는 없다.",
        "built_space": "중앙 평판 모니터 1대, 왼쪽 본체 1대, 오른쪽 브라운관 TV 1대, 앞쪽 키보드 1개와 오른쪽 마우스 1개가 보인다. 뒤쪽 부품 서랍과 선반, 중앙 창, 벽 안내판은 참조 장소의 구성과 대응한다. 사람은 완전히 제외되어 있다. TV의 밝은 개구부 반사는 보이는 채광 환경에서 가능하다. 화면 자체는 전체 면적의 약 28%로, 요구된 3분의 1 초과에는 못 미치지만 B보다 크다.",
        "entities": "낡은 삼성 모니터와 본체, GoldStar TV, 공구통과 잡동사니가 참조와 대응한다. 실제 모니터 안에 한국어 실종 로봇 기사와 굵은 붉은색 ‘5억원 현상금 걸어’가 표시된다. 그러나 회사명은 요구된 ‘유빅사’가 아니라 ‘유빙사’로 읽히며, 유일 모델과 해남 인근 실종 설명에도 같은 차이가 있다. 마츠다는 보이지 않고 로봇 사진도 없어 인물 외모나 찰리의 의상은 검증 대상이 아니다.",
        "hard_violations": [],
        "physics": "모니터는 목과 넓은 받침으로 작업대에 지지되고, TV는 직사각형 받침 위에 놓여 있다. 본체·키보드·마우스·공구통은 작업대에 놓였고 공구는 통에 꽂혀 있다. 떠 있는 물체나 지지 없는 신체는 없다."
       },
       {
        "label": "A",
        "direction": "모니터는 작업대 앞을 향하고 카메라에 거의 정면으로 보이되 약한 측면 원근이 있다. 작업대와 받침 윗면이 보여 약한 내려다보기는 성립한다. 인물의 시선이나 이동 방향은 없으며, 표시 면이 뒤집히거나 독자 반대쪽을 향하지 않는다.",
        "built_space": "중앙 모니터 1대, 왼쪽 본체 1대, 오른쪽 브라운관 TV 1대, 전면 키보드 1개와 오른쪽 마우스 1개가 있다. 공구통, 부품 서랍, 창과 벽 안내판까지 참조 사진과 매우 가까운 배치다. 추가 인물이나 중복 설비는 없다. TV의 창 형태 반사는 이 공간에서 가능하다. 화면 면적은 전체의 약 26%로 요구보다 작고, 참조 사진의 시점과 구도를 거의 그대로 유지한다.",
        "entities": "참조의 낡은 컴퓨터와 TV, 공구 및 작업대 재질이 유지된다. 한국어 인터넷 기사에 유일 모델의 해남 인근 실종과 굵은 ‘5억원 현상금 걸어’가 선명하다. 다만 지정된 회사명 ‘유빅사’ 대신 ‘유빙사’가 반복된다. 마츠다는 프레임 밖이며 찰리의 외형을 보여 주는 사진은 없다.",
        "hard_violations": [],
        "physics": "모니터 받침과 TV 아래 받침이 각각 무게를 지탱한다. 키보드와 마우스는 작업대에 닿아 있고, 공구는 용기에 담겨 있으며 뒤쪽 물건은 선반에 놓여 있다. 지지 없는 물체나 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.8
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1800
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으나, B에 비해 주변 사물과 텍스트 디테일을 더 선명하게 유지했습니다."
   },
   {
    "label": "B",
    "score": 1800,
    "verdict_ko": "지시된 '유빅사' 텍스트를 '유빙사'로 잘못 출력했으며, 모니터 베젤 등 세부 묘사에서 미세한 열화가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L27B02.png",
    "asset_id": "298722bc-6576-4d5a-9f26-f871a3d03c86",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08b2-20f2-77aa-8b81-abc14802c0c7",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh15::signage": {
  "fp": "3c08dc561066c240",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh15": {
  "input_fingerprint": "c9425a423e2c9f1a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 멀어지는 이현우를 향해 다급하게 한 팔을 앞으로 뻗은 채 입을 크게 벌리고 있는 상체.\n\nLOCATION (lock): Inside the secondhand repair shop near its entrance, facing the departing visitors. Daylight enters from the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track from the entrance side of 마츠다 at upper-chest height, remaining oblique to his line of sight rather than standing between him and 이현우. Keep 마츠다's open mouth, forward-pitched torso and reaching arm together at center-right, with a cropped rear view of the departing 이현우 along the left foreground and the entrance beyond him. 마츠다 looks past the camera toward 이현우, who continues attending to the exit; emphasize the renewed human distance after the screen detail, with 찰리 already beyond this crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shop entrance (Being used by 이현우 to leave) — Seen obliquely from inside, beyond the left foreground figure; used as Provides the destination of 이현우's movement and the direction of 마츠다's appeal; Piled appliances and scrap (Crowding the shop interior); used as Frames 마츠다 from behind while leaving his reaching arm clear.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime shop ambience with readable facial detail and consistent contrast, allowing 마츠다's urgency to come from gesture rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The missing-robot article and its 500-million-won reward remain on the shop computer amid the piled appliances and scrap. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo while leaving. 마츠다: He remains inside the shop as the attempted purchase falls through. 이현우: He is leaving the shop, with his facial injuries and untreated leg bite unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 멀어지는 이현우를 향해 다급하게 한 팔을 앞으로 뻗은 채 입을 크게 벌리고 있는 상체.\n\nLOCATION (lock): Inside the secondhand repair shop near its entrance, facing the departing visitors. Daylight enters from the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track from the entrance side of 마츠다 at upper-chest height, remaining oblique to his line of sight rather than standing between him and 이현우. Keep 마츠다's open mouth, forward-pitched torso and reaching arm together at center-right, with a cropped rear view of the departing 이현우 along the left foreground and the entrance beyond him. 마츠다 looks past the camera toward 이현우, who continues attending to the exit; emphasize the renewed human distance after the screen detail, with 찰리 already beyond this crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shop entrance (Being used by 이현우 to leave) — Seen obliquely from inside, beyond the left foreground figure; used as Provides the destination of 이현우's movement and the direction of 마츠다's appeal; Piled appliances and scrap (Crowding the shop interior); used as Frames 마츠다 from behind while leaving his reaching arm clear.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime shop ambience with readable facial detail and consistent contrast, allowing 마츠다's urgency to come from gesture rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The missing-robot article and its 500-million-won reward remain on the shop computer amid the piled appliances and scrap. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo while leaving. 마츠다: He remains inside the shop as the attempted purchase falls through. 이현우: He is leaving the shop, with his facial injuries and untreated leg bite unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다가 멀어지는 이현우를 향해 다급하게 한 팔을 앞으로 뻗은 채 입을 크게 벌리고 있는 상체.\n\nLOCATION (lock): Inside the secondhand repair shop near its entrance, facing the departing visitors. Daylight enters from the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track from the entrance side of 마츠다 at upper-chest height, remaining oblique to his line of sight rather than standing between him and 이현우. Keep 마츠다's open mouth, forward-pitched torso and reaching arm together at center-right, with a cropped rear view of the departing 이현우 along the left foreground and the entrance beyond him. 마츠다 looks past the camera toward 이현우, who continues attending to the exit; emphasize the renewed human distance after the screen detail, with 찰리 already beyond this crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shop entrance (Being used by 이현우 to leave) — Seen obliquely from inside, beyond the left foreground figure; used as Provides the destination of 이현우's movement and the direction of 마츠다's appeal; Piled appliances and scrap (Crowding the shop interior); used as Frames 마츠다 from behind while leaving his reaching arm clear.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daytime shop ambience with readable facial detail and consistent contrast, allowing 마츠다's urgency to come from gesture rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The missing-robot article and its 500-million-won reward remain on the shop computer amid the piled appliances and scrap. Charlie retains the old coat and hat, blue-lit eyes and worn UBIK chest logo while leaving. 마츠다: He remains inside the shop as the attempted purchase falls through. 이현우: He is leaving the shop, with his facial injuries and untreated leg bite unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "마츠다의 시선과 앞으로 뻗은 오른팔이 왼쪽 전경에 뒷모습으로 위치한 이현우를 정확히 향하고 있음.",
    "built_space": "수리점 내부. 우측에 이전 샷과 동일한 기사가 띄워진 모니터가 있고, 뒤쪽 벽의 '중고 매입 판매' 간판이 레퍼런스와 일치함. 좌측 출입구로 채광이 들어옴.",
    "entities": "마츠다(60대, 오염된 조끼와 체크셔츠)와 이현우(남색 티셔츠 뒷모습)가 참조 이미지와 일치함. 모니터와 간판의 한글 텍스트 형상이 매우 정확함(배낭이 임의로 추가됨).",
    "hard_violations": [
     "[gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
    ],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 마츠다의 상체 기울기와 뻗은 팔이 다급한 동작을 물리적으로 잘 지지함."
   },
   {
    "label": "B",
    "direction": "마츠다의 시선과 뻗은 오른팔이 출입구 쪽으로 향하는 이현우를 향해 있음.",
    "built_space": "수리점 내부. 우측에 모니터가 배치되었으나 벽면 간판의 글씨가 뭉개져 있음. 좌측으로 출입구와 야외가 보임.",
    "entities": "마츠다와 이현우의 인상착의는 참조와 대체로 일치함. 모니터 화면의 기사 내용이 A에 비해 덜 선명하고 일부 텍스트가 왜곡됨(마찬가지로 배낭이 추가됨).",
    "hard_violations": [
     "[gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
    ],
    "physics": "바닥에 잘 지지되어 서 있으며 동작 자체는 성립하나, 뻗은 손의 형태와 관절 묘사가 다소 부자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 모니터 텍스트와 벽면 간판 등 고정된 디테일을 완벽히 유지하며, 지시된 카메라 구도와 다급한 동작을 가장 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 인물의 기본 배치는 지시를 따랐으나, 유지되어야 할 배경의 간판 텍스트가 왜곡되고 모니터 화면의 정교함이 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선과 앞으로 뻗은 오른팔이 왼쪽 전경에 뒷모습으로 위치한 이현우를 정확히 향하고 있음.",
        "built_space": "수리점 내부. 우측에 이전 샷과 동일한 기사가 띄워진 모니터가 있고, 뒤쪽 벽의 '중고 매입 판매' 간판이 레퍼런스와 일치함. 좌측 출입구로 채광이 들어옴.",
        "entities": "마츠다(60대, 오염된 조끼와 체크셔츠)와 이현우(남색 티셔츠 뒷모습)가 참조 이미지와 일치함. 모니터와 간판의 한글 텍스트 형상이 매우 정확함(배낭이 임의로 추가됨).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 마츠다의 상체 기울기와 뻗은 팔이 다급한 동작을 물리적으로 잘 지지함."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선과 뻗은 오른팔이 출입구 쪽으로 향하는 이현우를 향해 있음.",
        "built_space": "수리점 내부. 우측에 모니터가 배치되었으나 벽면 간판의 글씨가 뭉개져 있음. 좌측으로 출입구와 야외가 보임.",
        "entities": "마츠다와 이현우의 인상착의는 참조와 대체로 일치함. 모니터 화면의 기사 내용이 A에 비해 덜 선명하고 일부 텍스트가 왜곡됨(마찬가지로 배낭이 추가됨).",
        "hard_violations": [],
        "physics": "바닥에 잘 지지되어 서 있으며 동작 자체는 성립하나, 뻗은 손의 형태와 관절 묘사가 다소 부자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 모니터 텍스트와 벽면 간판 등 고정된 디테일을 완벽히 유지하며, 지시된 카메라 구도와 다급한 동작을 가장 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 인물의 기본 배치는 지시를 따랐으나, 유지되어야 할 배경의 간판 텍스트가 왜곡되고 모니터 화면의 정교함이 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 시선과 앞으로 뻗은 오른팔이 왼쪽 전경에 뒷모습으로 위치한 이현우를 정확히 향하고 있음.",
        "built_space": "수리점 내부. 우측에 이전 샷과 동일한 기사가 띄워진 모니터가 있고, 뒤쪽 벽의 '중고 매입 판매' 간판이 레퍼런스와 일치함. 좌측 출입구로 채광이 들어옴.",
        "entities": "마츠다(60대, 오염된 조끼와 체크셔츠)와 이현우(남색 티셔츠 뒷모습)가 참조 이미지와 일치함. 모니터와 간판의 한글 텍스트 형상이 매우 정확함(배낭이 임의로 추가됨).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 마츠다의 상체 기울기와 뻗은 팔이 다급한 동작을 물리적으로 잘 지지함."
       },
       {
        "label": "B",
        "direction": "마츠다의 시선과 뻗은 오른팔이 출입구 쪽으로 향하는 이현우를 향해 있음.",
        "built_space": "수리점 내부. 우측에 모니터가 배치되었으나 벽면 간판의 글씨가 뭉개져 있음. 좌측으로 출입구와 야외가 보임.",
        "entities": "마츠다와 이현우의 인상착의는 참조와 대체로 일치함. 모니터 화면의 기사 내용이 A에 비해 덜 선명하고 일부 텍스트가 왜곡됨(마찬가지로 배낭이 추가됨).",
        "hard_violations": [],
        "physics": "바닥에 잘 지지되어 서 있으며 동작 자체는 성립하나, 뻗은 손의 형태와 관절 묘사가 다소 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "크게 벌린 입과 앞으로 기운 상체, 이현우를 향한 팔의 방향이 더 충실하지만, 참조에 없는 배낭을 추가해 엄격한 소품 연속성을 위반합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 배치와 퇴장 방향은 맞지만 A보다 입 벌림과 상체의 다급한 전진이 약하며, 동일하게 근거 없는 배낭을 추가했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다의 눈과 얼굴은 왼쪽 전경의 이현우를 향하고, 펼친 손도 그쪽으로 뻗어 있습니다. 카메라는 두 사람의 시선축에서 비껴나 있습니다. 이현우는 뒤통수와 등을 보이며 왼쪽의 밝은 출입구를 향하고, 마츠다를 돌아보지 않습니다.",
        "built_space": "왼쪽에 유리 출입구 한 곳, 오른쪽에 컴퓨터 모니터 한 대와 키보드 한 개가 놓인 작업대가 보입니다. 뒤에는 여러 부품 서랍장과 가전 선반, 선풍기 한 대가 있습니다. 마츠다는 작업대 앞 실내에, 이현우는 출입구 가까운 전경에 있어 퇴장 동선이 성립하며 뻗은 팔도 가려지지 않습니다. 참조의 낡은 작업대, 부품 수납과 청색 한글 표지의 재질은 이어지지만 세부 배치가 정확히 일치하지는 않습니다. 불가능한 반사나 출입구 중복은 보이지 않습니다.",
        "entities": "두 사람만 보이고 찰리는 제외되어 있습니다. 마츠다는 참조와 유사한 노년 동아시아계 남성으로, 남색 모자와 얼룩진 청색 조끼, 체크 셔츠를 착용했습니다. 이현우는 짧은 검은 머리와 남색 티셔츠의 젊은 남성 뒷모습이며, 얼굴과 다리 부상은 이 구도에서 확인할 수 없습니다. 다만 참조에 없는 검은 배낭을 메고 있습니다. 오른쪽 모니터에는 로봇 실종 기사와 5억원 현상금 문구가 보입니다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
        ],
        "physics": "마츠다의 팔은 어깨와 팔꿈치에서 자연스럽게 이어져 앞으로 뻗고, 반대 팔은 아래로 내려와 있습니다. 상체의 기울기는 서 있는 사람이 외치는 동작으로 가능합니다. 두 사람의 발은 화면 밖이므로 바닥 접촉 자체는 확인할 수 없지만 공중에 떠 있는 징후는 없습니다. 배낭은 어깨끈으로 지지되고 모니터와 키보드는 작업대 위에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "마츠다는 왼쪽 전경의 이현우를 바라보며 한 손을 그를 향해 뻗습니다. 이현우의 얼굴과 몸은 왼쪽 출입구 쪽으로 돌아 있어 계속 떠나는 관계가 읽힙니다. 손이 약간 카메라 쪽으로 단축되어 보이지만 호소의 대상은 이현우로 식별됩니다.",
        "built_space": "왼쪽에 유리 출입구 한 곳, 오른쪽 작업대에 모니터 한 대와 키보드 한 개가 있습니다. 뒤쪽 선반에는 전자레인지 한 대와 여러 가전이 쌓여 있고, 부품 서랍장들과 오른쪽 위 철망 창 한 곳이 보입니다. 마츠다와 출입구 사이에 이현우가 있어 동선은 자연스럽습니다. 참조의 작업장 재료와 벽 표지는 이어지지만 수납물의 배치는 달라져 있습니다. 오른쪽 아래 작업대가 추가로 화면을 차지하며, 불가능한 반사는 보이지 않습니다.",
        "entities": "마츠다와 이현우 두 사람만 보입니다. 마츠다의 노년 얼굴, 모자, 청색 작업 조끼와 체크 셔츠는 참조와 대체로 맞습니다. 이현우는 헝클어진 검은 머리와 남색 티셔츠를 입은 젊은 남성의 뒷모습이며, 얼굴 정체성과 부상 상태는 가려져 확인할 수 없습니다. 이현우의 검은 배낭은 참조에 없는 소품입니다. 모니터에는 실종 기사와 5억원 현상금이 표시되어 있습니다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
        ],
        "physics": "뻗은 팔과 아래로 내린 반대 팔은 몸에 정상적으로 연결되어 있고, 상체는 약간 앞으로 기울어 있습니다. 발은 잘려 있으나 서 있는 자세와 모순되는 부유는 없습니다. 배낭은 어깨끈에 걸려 있고, 컴퓨터와 가전은 작업대 또는 선반에 지지되어 있습니다. 자세는 가능하지만 A보다 상체의 전진감이 약합니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "크게 벌린 입과 앞으로 기운 상체, 이현우를 향한 팔의 방향이 더 충실하지만, 참조에 없는 배낭을 추가해 엄격한 소품 연속성을 위반합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물 배치와 퇴장 방향은 맞지만 A보다 입 벌림과 상체의 다급한 전진이 약하며, 동일하게 근거 없는 배낭을 추가했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "마츠다의 눈과 얼굴은 왼쪽 전경의 이현우를 향하고, 펼친 손도 그쪽으로 뻗어 있습니다. 카메라는 두 사람의 시선축에서 비껴나 있습니다. 이현우는 뒤통수와 등을 보이며 왼쪽의 밝은 출입구를 향하고, 마츠다를 돌아보지 않습니다.",
        "built_space": "왼쪽에 유리 출입구 한 곳, 오른쪽에 컴퓨터 모니터 한 대와 키보드 한 개가 놓인 작업대가 보입니다. 뒤에는 여러 부품 서랍장과 가전 선반, 선풍기 한 대가 있습니다. 마츠다는 작업대 앞 실내에, 이현우는 출입구 가까운 전경에 있어 퇴장 동선이 성립하며 뻗은 팔도 가려지지 않습니다. 참조의 낡은 작업대, 부품 수납과 청색 한글 표지의 재질은 이어지지만 세부 배치가 정확히 일치하지는 않습니다. 불가능한 반사나 출입구 중복은 보이지 않습니다.",
        "entities": "두 사람만 보이고 찰리는 제외되어 있습니다. 마츠다는 참조와 유사한 노년 동아시아계 남성으로, 남색 모자와 얼룩진 청색 조끼, 체크 셔츠를 착용했습니다. 이현우는 짧은 검은 머리와 남색 티셔츠의 젊은 남성 뒷모습이며, 얼굴과 다리 부상은 이 구도에서 확인할 수 없습니다. 다만 참조에 없는 검은 배낭을 메고 있습니다. 오른쪽 모니터에는 로봇 실종 기사와 5억원 현상금 문구가 보입니다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
        ],
        "physics": "마츠다의 팔은 어깨와 팔꿈치에서 자연스럽게 이어져 앞으로 뻗고, 반대 팔은 아래로 내려와 있습니다. 상체의 기울기는 서 있는 사람이 외치는 동작으로 가능합니다. 두 사람의 발은 화면 밖이므로 바닥 접촉 자체는 확인할 수 없지만 공중에 떠 있는 징후는 없습니다. 배낭은 어깨끈으로 지지되고 모니터와 키보드는 작업대 위에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "마츠다는 왼쪽 전경의 이현우를 바라보며 한 손을 그를 향해 뻗습니다. 이현우의 얼굴과 몸은 왼쪽 출입구 쪽으로 돌아 있어 계속 떠나는 관계가 읽힙니다. 손이 약간 카메라 쪽으로 단축되어 보이지만 호소의 대상은 이현우로 식별됩니다.",
        "built_space": "왼쪽에 유리 출입구 한 곳, 오른쪽 작업대에 모니터 한 대와 키보드 한 개가 있습니다. 뒤쪽 선반에는 전자레인지 한 대와 여러 가전이 쌓여 있고, 부품 서랍장들과 오른쪽 위 철망 창 한 곳이 보입니다. 마츠다와 출입구 사이에 이현우가 있어 동선은 자연스럽습니다. 참조의 작업장 재료와 벽 표지는 이어지지만 수납물의 배치는 달라져 있습니다. 오른쪽 아래 작업대가 추가로 화면을 차지하며, 불가능한 반사는 보이지 않습니다.",
        "entities": "마츠다와 이현우 두 사람만 보입니다. 마츠다의 노년 얼굴, 모자, 청색 작업 조끼와 체크 셔츠는 참조와 대체로 맞습니다. 이현우는 헝클어진 검은 머리와 남색 티셔츠를 입은 젊은 남성의 뒷모습이며, 얼굴 정체성과 부상 상태는 가려져 확인할 수 없습니다. 이현우의 검은 배낭은 참조에 없는 소품입니다. 모니터에는 실종 기사와 5억원 현상금이 표시되어 있습니다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
        ],
        "physics": "뻗은 팔과 아래로 내린 반대 팔은 몸에 정상적으로 연결되어 있고, 상체는 약간 앞으로 기울어 있습니다. 발은 잘려 있으나 서 있는 자세와 모순되는 부유는 없습니다. 배낭은 어깨끈에 걸려 있고, 컴퓨터와 가전은 작업대 또는 선반에 지지되어 있습니다. 자세는 가능하지만 A보다 상체의 전진감이 약합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.464
   },
   "violations": {
    "B": [
     "[gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
    ],
    "A": [
     "[gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "이전 샷의 모니터 텍스트와 벽면 간판 등 고정된 디테일을 완벽히 유지하며, 지시된 카메라 구도와 다급한 동작을 가장 사실적으로 구현했습니다.  ★위반: [gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
   },
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "구도와 인물의 기본 배치는 지시를 따랐으나, 유지되어야 할 배경의 간판 텍스트가 왜곡되고 모니터 화면의 정교함이 떨어집니다.  ★위반: [gpt-high] 이현우에게 프롬프트와 인물 참조가 설정하지 않은 검은 배낭을 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S15sh9_sel.png",
    "asset_id": "b26871ad-1e85-4737-a7e3-7a5d17a0acb2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:889268>",
    "asset_id": "4300cf67-f254-4d25-bb5b-9b909d79bf38",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08b8-5265-7b20-9781-a3ec16720c63",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh9"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S15sh18::signage": {
  "fp": "0e8e8b3974a21059",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh18": {
  "input_fingerprint": "64d01abe49ff4223",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전화기 수화기를 귀에 바짝 댄 채 상점 밖 거리를 매섭게 경계하는 마츠다의 은밀한 눈빛.\n\nLOCATION (lock): At the telephone position inside the secondhand repair shop, with a sightline through the entrance to the daylight street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly-in inside the shop at 마츠다's eye level, retaining the established side three-quarter view and observing him directly rather than through the entrance. Place his face and the receiver in the left half, leaving space to the right toward the entrance as he presses the receiver against his ear and watches the street beyond the frame. Emphasize only the closing camera distance, keeping his sightline and the surrounding shop arrangement stable.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shop entrance (The opening connects the shop interior to the street) — Seen obliquely at the right edge, in the direction of Matsuda's outward sightline; used as Provides a narrow spatial reference for the street he is monitoring; Accumulated appliances and scrap (Piled inside the shop); used as Soft peripheral context behind Matsuda without competing with his eyes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast preserve the guarded expression without introducing a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's accumulated appliances, scrap, and daytime interior appearance from the reference. Exclude the departing teenager and robot; neither remains in the shop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer still displays the missing-robot article and reward, surrounded by the shop's piles of appliances and scrap. The departing Charlie retains the old coat and hat and worn UBIK chest logo. 마츠다: He remains in the shop using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전화기 수화기를 귀에 바짝 댄 채 상점 밖 거리를 매섭게 경계하는 마츠다의 은밀한 눈빛.\n\nLOCATION (lock): At the telephone position inside the secondhand repair shop, with a sightline through the entrance to the daylight street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly-in inside the shop at 마츠다's eye level, retaining the established side three-quarter view and observing him directly rather than through the entrance. Place his face and the receiver in the left half, leaving space to the right toward the entrance as he presses the receiver against his ear and watches the street beyond the frame. Emphasize only the closing camera distance, keeping his sightline and the surrounding shop arrangement stable.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shop entrance (The opening connects the shop interior to the street) — Seen obliquely at the right edge, in the direction of Matsuda's outward sightline; used as Provides a narrow spatial reference for the street he is monitoring; Accumulated appliances and scrap (Piled inside the shop); used as Soft peripheral context behind Matsuda without competing with his eyes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast preserve the guarded expression without introducing a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's accumulated appliances, scrap, and daytime interior appearance from the reference. Exclude the departing teenager and robot; neither remains in the shop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer still displays the missing-robot article and reward, surrounded by the shop's piles of appliances and scrap. The departing Charlie retains the old coat and hat and worn UBIK chest logo. 마츠다: He remains in the shop using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전화기 수화기를 귀에 바짝 댄 채 상점 밖 거리를 매섭게 경계하는 마츠다의 은밀한 눈빛.\n\nLOCATION (lock): At the telephone position inside the secondhand repair shop, with a sightline through the entrance to the daylight street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly-in inside the shop at 마츠다's eye level, retaining the established side three-quarter view and observing him directly rather than through the entrance. Place his face and the receiver in the left half, leaving space to the right toward the entrance as he presses the receiver against his ear and watches the street beyond the frame. Emphasize only the closing camera distance, keeping his sightline and the surrounding shop arrangement stable.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Shop entrance (The opening connects the shop interior to the street) — Seen obliquely at the right edge, in the direction of Matsuda's outward sightline; used as Provides a narrow spatial reference for the street he is monitoring; Accumulated appliances and scrap (Piled inside the shop); used as Soft peripheral context behind Matsuda without competing with his eyes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast preserve the guarded expression without introducing a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's accumulated appliances, scrap, and daytime interior appearance from the reference. Exclude the departing teenager and robot; neither remains in the shop.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer still displays the missing-robot article and reward, surrounded by the shop's piles of appliances and scrap. The departing Charlie retains the old coat and hat and worn UBIK chest logo. 마츠다: He remains in the shop using a telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 마츠다 right now, so 마츠다's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 마츠다: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 우측의 상점 입구 밖 거리를 향하고 있으며, 전화기 수화기는 왼쪽 귀에 바짝 대고 있음.",
    "built_space": "레퍼런스와 동일한 중고 수리점 내부. 우측에 거리가 보이는 입구가 위치하며, 배경에는 부품과 이전 샷과 레이아웃이 일치하는 컴퓨터 모니터가 배치됨.",
    "entities": "마츠다는 레퍼런스와 일치하는 인상착의(작업용 조끼, 체크 셔츠, 모자)를 갖추고 있음. 손에 검은색 전화기 수화기를 쥐고 있음.",
    "hard_violations": [],
    "physics": "수화기를 쥐고 있는 손의 형태가 자연스러우며, 전화선이 아래로 안정적으로 늘어져 중력을 따름."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 우측의 거리를 향하고 있으며, 수화기를 귀에 대고 있음.",
    "built_space": "수리점 내부 및 우측의 입구가 보이나, 배경의 컴퓨터 모니터 화면 구성이 이전 샷 레퍼런스와 다름(이전 샷에 없던 사진 썸네일이 나타남).",
    "entities": "마츠다의 복장과 외모는 일치하나, 수화기를 쥐고 있는 손의 구조가 비정상적임.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy: 수화기를 쥔 손의 손가락 개수와 관절 구조가 기형적임"
    ],
    "physics": "손이 수화기를 지탱하고 있으나 그 형태가 물리적/해부학적으로 불가능한 구조임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 화면 구성을 완벽히 유지하면서 자연스러운 손 모양과 배경 디테일을 잘 살려 프롬프트를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "마츠다의 시선과 구도는 좋으나, 배경 모니터의 디테일이 이전 샷과 일치하지 않고 수화기를 쥔 손의 해부학적 오류가 있어 실격 사유에 해당함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 우측의 상점 입구 밖 거리를 향하고 있으며, 전화기 수화기는 왼쪽 귀에 바짝 대고 있음.",
        "built_space": "레퍼런스와 동일한 중고 수리점 내부. 우측에 거리가 보이는 입구가 위치하며, 배경에는 부품과 이전 샷과 레이아웃이 일치하는 컴퓨터 모니터가 배치됨.",
        "entities": "마츠다는 레퍼런스와 일치하는 인상착의(작업용 조끼, 체크 셔츠, 모자)를 갖추고 있음. 손에 검은색 전화기 수화기를 쥐고 있음.",
        "hard_violations": [],
        "physics": "수화기를 쥐고 있는 손의 형태가 자연스러우며, 전화선이 아래로 안정적으로 늘어져 중력을 따름."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 우측의 거리를 향하고 있으며, 수화기를 귀에 대고 있음.",
        "built_space": "수리점 내부 및 우측의 입구가 보이나, 배경의 컴퓨터 모니터 화면 구성이 이전 샷 레퍼런스와 다름(이전 샷에 없던 사진 썸네일이 나타남).",
        "entities": "마츠다의 복장과 외모는 일치하나, 수화기를 쥐고 있는 손의 구조가 비정상적임.",
        "hard_violations": [
         "physically impossible anatomy: 수화기를 쥔 손의 손가락 개수와 관절 구조가 기형적임"
        ],
        "physics": "손이 수화기를 지탱하고 있으나 그 형태가 물리적/해부학적으로 불가능한 구조임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 화면 구성을 완벽히 유지하면서 자연스러운 손 모양과 배경 디테일을 잘 살려 프롬프트를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "마츠다의 시선과 구도는 좋으나, 배경 모니터의 디테일이 이전 샷과 일치하지 않고 수화기를 쥔 손의 해부학적 오류가 있어 실격 사유에 해당함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 우측의 상점 입구 밖 거리를 향하고 있으며, 전화기 수화기는 왼쪽 귀에 바짝 대고 있음.",
        "built_space": "레퍼런스와 동일한 중고 수리점 내부. 우측에 거리가 보이는 입구가 위치하며, 배경에는 부품과 이전 샷과 레이아웃이 일치하는 컴퓨터 모니터가 배치됨.",
        "entities": "마츠다는 레퍼런스와 일치하는 인상착의(작업용 조끼, 체크 셔츠, 모자)를 갖추고 있음. 손에 검은색 전화기 수화기를 쥐고 있음.",
        "hard_violations": [],
        "physics": "수화기를 쥐고 있는 손의 형태가 자연스러우며, 전화선이 아래로 안정적으로 늘어져 중력을 따름."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 우측의 거리를 향하고 있으며, 수화기를 귀에 대고 있음.",
        "built_space": "수리점 내부 및 우측의 입구가 보이나, 배경의 컴퓨터 모니터 화면 구성이 이전 샷 레퍼런스와 다름(이전 샷에 없던 사진 썸네일이 나타남).",
        "entities": "마츠다의 복장과 외모는 일치하나, 수화기를 쥐고 있는 손의 구조가 비정상적임.",
        "hard_violations": [
         "physically impossible anatomy: 수화기를 쥔 손의 손가락 개수와 관절 구조가 기형적임"
        ],
        "physics": "손이 수화기를 지탱하고 있으나 그 형태가 물리적/해부학적으로 불가능한 구조임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 얼굴·수화기 클로즈업과 오른쪽 거리 시선은 충실하지만, 벌어진 입과 비교적 풀린 표정이 은밀하고 매서운 경계심을 약화한다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "눈높이의 직접적인 클로즈업, 왼쪽 얼굴·수화기와 오른쪽 출입구 배치에 더해 다문 입과 날카로운 곁눈질이 요청한 경계의 순간을 가장 잘 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 오른쪽으로 조금 돌아가고 두 눈은 오른쪽 출입구 너머 거리를 향한다. 카메라를 응시하지 않는다. 수화기의 윗부분은 귀에 닿고 아랫부분은 입 아래에 놓여 통화 방향이 자연스럽다.",
        "built_space": "상점 안에서 인물을 직접 보는 눈높이 클로즈업이다. 얼굴과 수화기는 왼쪽 절반에 있고 오른쪽 가장자리에 출입구 하나와 낮의 거리가 보인다. 뒤에는 전자레인지 하나, 주요 부품 서랍장 두 개, 모니터 하나, 작업대와 한국어 안내판 하나가 보인다. 참고의 낡은 금속 선반과 수리점 재료는 이어지지만, 출입구와 후면 집기의 상대 배치가 같은 공간인지 확실히 대조하기는 어렵다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "참고와 닮은 60대 동아시아계 남성 한 명이며, 일본 국적 자체는 외형으로 확인할 수 없다. 짧은 회색 머리, 낡은 남색 모자, 얼룩진 청색 작업 조끼와 적갈색 체크 셔츠가 일치한다. 검은 유선 수화기와 이를 잡은 나이 든 남성의 손이 보인다. 모니터에는 기사 형태의 화면과 붉은 현상금 문구가 남아 있지만 세부 내용은 흐리다. 청소년이나 로봇은 없으며 바지는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "손가락으로 수화기 손잡이를 감싸 쥐고 귀에 눌러 지지한다. 손목과 팔뚝이 같은 인물의 소매로 자연스럽게 이어진다. 전화선은 수화기 아래로 늘어지고, 모니터와 집기는 작업대 및 선반에 놓여 있다. 하체의 지지는 클로즈업 밖이라 확인할 수 없지만, 보이는 부분에 부유하거나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "고개를 오른쪽으로 약간 돌린 채 눈동자를 오른쪽 출입구 너머 거리로 보낸다. 다문 입과 좁힌 눈매가 은밀한 감시로 읽히며 렌즈를 보지 않는다. 수화기는 귀에 밀착되고 송화부는 입 가까운 아래쪽을 향한다.",
        "built_space": "상점 내부에서 인물을 직접 관찰하는 눈높이 클로즈업이며 얼굴과 수화기는 왼쪽 절반에 놓인다. 오른쪽 가장자리에는 출입구 하나와 밝은 거리의 좁은 부분이 보인다. 배경에는 전자레인지 하나, 주요 부품 서랍장 두 개, 모니터 하나, 작업대와 한국어 안내판 하나가 있다. 낡은 선반과 쌓인 부품은 참고의 수리점 분위기를 유지한다. 다만 참고와 출입구·후면 집기의 상대 배치가 정확히 이어지는지는 확증하기 어렵다. 명백한 중복 설비나 잘못된 반사는 없다.",
        "entities": "등장인물은 참고와 닮은 60대 동아시아계 남성 한 명이다. 국적은 영상만으로 판별할 수 없지만 지정된 마츠다의 얼굴, 짧은 회색 머리, 남색 모자, 기름때 묻은 청색 조끼와 체크 셔츠에 부합한다. 수화기를 잡은 손도 피부와 나이가 얼굴에 맞는다. 검은 유선 전화와 배경의 기사·붉은 현상금 화면이 보이며, 기사의 상세 문장은 판독하기 어렵다. 청소년과 로봇은 없고 하의는 구도 밖이다.",
        "hard_violations": [],
        "physics": "손이 수화기를 확실하게 감싸 귀에 밀착시킨다. 굽힌 팔과 손목의 연결 및 손가락 접촉이 자연스럽고, 전화선은 아래로 처진다. 모니터는 받침대로 작업대에 지지되고 다른 장비도 선반 위에 놓여 있다. 발은 프레임 밖이지만 상체에서 불가능한 자세나 지지 없는 부유는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 얼굴·수화기 클로즈업과 오른쪽 거리 시선은 충실하지만, 벌어진 입과 비교적 풀린 표정이 은밀하고 매서운 경계심을 약화한다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "눈높이의 직접적인 클로즈업, 왼쪽 얼굴·수화기와 오른쪽 출입구 배치에 더해 다문 입과 날카로운 곁눈질이 요청한 경계의 순간을 가장 잘 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 오른쪽으로 조금 돌아가고 두 눈은 오른쪽 출입구 너머 거리를 향한다. 카메라를 응시하지 않는다. 수화기의 윗부분은 귀에 닿고 아랫부분은 입 아래에 놓여 통화 방향이 자연스럽다.",
        "built_space": "상점 안에서 인물을 직접 보는 눈높이 클로즈업이다. 얼굴과 수화기는 왼쪽 절반에 있고 오른쪽 가장자리에 출입구 하나와 낮의 거리가 보인다. 뒤에는 전자레인지 하나, 주요 부품 서랍장 두 개, 모니터 하나, 작업대와 한국어 안내판 하나가 보인다. 참고의 낡은 금속 선반과 수리점 재료는 이어지지만, 출입구와 후면 집기의 상대 배치가 같은 공간인지 확실히 대조하기는 어렵다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "참고와 닮은 60대 동아시아계 남성 한 명이며, 일본 국적 자체는 외형으로 확인할 수 없다. 짧은 회색 머리, 낡은 남색 모자, 얼룩진 청색 작업 조끼와 적갈색 체크 셔츠가 일치한다. 검은 유선 수화기와 이를 잡은 나이 든 남성의 손이 보인다. 모니터에는 기사 형태의 화면과 붉은 현상금 문구가 남아 있지만 세부 내용은 흐리다. 청소년이나 로봇은 없으며 바지는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "손가락으로 수화기 손잡이를 감싸 쥐고 귀에 눌러 지지한다. 손목과 팔뚝이 같은 인물의 소매로 자연스럽게 이어진다. 전화선은 수화기 아래로 늘어지고, 모니터와 집기는 작업대 및 선반에 놓여 있다. 하체의 지지는 클로즈업 밖이라 확인할 수 없지만, 보이는 부분에 부유하거나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "고개를 오른쪽으로 약간 돌린 채 눈동자를 오른쪽 출입구 너머 거리로 보낸다. 다문 입과 좁힌 눈매가 은밀한 감시로 읽히며 렌즈를 보지 않는다. 수화기는 귀에 밀착되고 송화부는 입 가까운 아래쪽을 향한다.",
        "built_space": "상점 내부에서 인물을 직접 관찰하는 눈높이 클로즈업이며 얼굴과 수화기는 왼쪽 절반에 놓인다. 오른쪽 가장자리에는 출입구 하나와 밝은 거리의 좁은 부분이 보인다. 배경에는 전자레인지 하나, 주요 부품 서랍장 두 개, 모니터 하나, 작업대와 한국어 안내판 하나가 있다. 낡은 선반과 쌓인 부품은 참고의 수리점 분위기를 유지한다. 다만 참고와 출입구·후면 집기의 상대 배치가 정확히 이어지는지는 확증하기 어렵다. 명백한 중복 설비나 잘못된 반사는 없다.",
        "entities": "등장인물은 참고와 닮은 60대 동아시아계 남성 한 명이다. 국적은 영상만으로 판별할 수 없지만 지정된 마츠다의 얼굴, 짧은 회색 머리, 남색 모자, 기름때 묻은 청색 조끼와 체크 셔츠에 부합한다. 수화기를 잡은 손도 피부와 나이가 얼굴에 맞는다. 검은 유선 전화와 배경의 기사·붉은 현상금 화면이 보이며, 기사의 상세 문장은 판독하기 어렵다. 청소년과 로봇은 없고 하의는 구도 밖이다.",
        "hard_violations": [],
        "physics": "손이 수화기를 확실하게 감싸 귀에 밀착시킨다. 굽힌 팔과 손목의 연결 및 손가락 접촉이 자연스럽고, 전화선은 아래로 처진다. 모니터는 받침대로 작업대에 지지되고 다른 장비도 선반 위에 놓여 있다. 발은 프레임 밖이지만 상체에서 불가능한 자세나 지지 없는 부유는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.289
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.039
   },
   "violations": {
    "B": [
     "[gemini-pro] physically impossible anatomy: 수화기를 쥔 손의 손가락 개수와 관절 구조가 기형적임"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1039
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스와 화면 구성을 완벽히 유지하면서 자연스러운 손 모양과 배경 디테일을 잘 살려 프롬프트를 충실히 구현함."
   },
   {
    "label": "B",
    "score": 1039,
    "verdict_ko": "마츠다의 시선과 구도는 좋으나, 배경 모니터의 디테일이 이전 샷과 일치하지 않고 수화기를 쥔 손의 해부학적 오류가 있어 실격 사유에 해당함.  ★위반: [gemini-pro] physically impossible anatomy: 수화기를 쥔 손의 손가락 개수와 관절 구조가 기형적임"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S15sh15_sel.png",
    "asset_id": "fcfdf058-d2f2-416a-ab05-18c97e55613a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:889268>",
    "asset_id": "4300cf67-f254-4d25-bb5b-9b909d79bf38",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08bd-0be4-7ef0-96f4-8d99714a257f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh15"
  }
 },
 "S16sh4::signage": {
  "fp": "b13c72ec91bcb1f4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::68b0e03d937a9abb": {
  "subjects": [],
  "subject_text": "라울의 컨테이너 앞마당\n낡은 철제 주거동 출입문 앞의 작은 마당. 주변으로 컨테이너 외벽이 이어지고, 물을 쓸 수 있는 호스가 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L28",
  "scope_role": "location_exterior",
  "scope_sha": "f73a8dec2043dccf"
 },
 "S16sh4::bgfirst_bg": {
  "input_fingerprint": "e96d90c19f6857d1",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨끗해진 자신의 몸을 내려다보며 양팔을 활짝 벌린 채 즐거운 표정을 짓는 찰리의 전신 구도.\n\nLOCATION (lock): In the outdoor washing area in front of a container home, on ground sprayed by a hose.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the diagonal retreat at 찰리's waist height with a slight upward tilt, retaining an oblique front view that shows his entire body and clear space beyond both extended hands and below his feet. Center him loosely as he tips his head down toward his newly cleaned torso, his open arms and delighted expression readable together; the water enters from the hose side, with 라울 and 앰버 outside the crop. Let increased camera distance be the sole emphasized change, without repositioning Charlie or changing the lighting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Raul's container home (The washing takes place outside the home) — An oblique portion of its exterior remains behind Charlie; used as Establishes the domestic setting and a stable scale reference; Hose water (Spraying Charlie and washing dirt from his body); used as A peripheral stream connects the off-screen hose side to Charlie's torso without obscuring his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination keeps the washing water and precise robot surfaces readable while allowing Charlie's delight to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨끗해진 자신의 몸을 내려다보며 양팔을 활짝 벌린 채 즐거운 표정을 짓는 찰리의 전신 구도.\n\nLOCATION (lock): In the outdoor washing area in front of a container home, on ground sprayed by a hose.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the diagonal retreat at 찰리's waist height with a slight upward tilt, retaining an oblique front view that shows his entire body and clear space beyond both extended hands and below his feet. Center him loosely as he tips his head down toward his newly cleaned torso, his open arms and delighted expression readable together; the water enters from the hose side, with 라울 and 앰버 outside the crop. Let increased camera distance be the sole emphasized change, without repositioning Charlie or changing the lighting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Raul's container home (The washing takes place outside the home) — An oblique portion of its exterior remains behind Charlie; used as Establishes the domestic setting and a stable scale reference; Hose water (Spraying Charlie and washing dirt from his body); used as A peripheral stream connects the off-screen hose side to Charlie's torso without obscuring his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination keeps the washing water and precise robot surfaces readable while allowing Charlie's delight to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S16sh4__bgfirst_bg.png",
  "asset_id": "e6bd925a-3eca-4f4c-b706-f4b94e618b81",
  "input_asset_ids": [
   "42e2691a-83b8-4772-b28b-520706f1436c",
   "91b36b81-eabb-4c94-984f-45b03850cbb1"
  ]
 },
 "S16sh4": {
  "input_fingerprint": "cb54cc920485b96b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨끗해진 자신의 몸을 내려다보며 양팔을 활짝 벌린 채 즐거운 표정을 짓는 찰리의 전신 구도.\n\nLOCATION (lock): In the outdoor washing area in front of a container home, on ground sprayed by a hose. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the diagonal retreat at 찰리's waist height with a slight upward tilt, retaining an oblique front view that shows his entire body and clear space beyond both extended hands and below his feet. Center him loosely as he tips his head down toward his newly cleaned torso, his open arms and delighted expression readable together; the water enters from the hose side, with 라울 and 앰버 outside the crop. Let increased camera distance be the sole emphasized change, without repositioning Charlie or changing the lighting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Raul's container home (The washing takes place outside the home) — An oblique portion of its exterior remains behind Charlie; used as Establishes the domestic setting and a stable scale reference; Hose water (Spraying Charlie and washing dirt from his body); used as A peripheral stream connects the off-screen hose side to Charlie's torso without obscuring his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination keeps the washing water and precise robot surfaces readable while allowing Charlie's delight to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A hose is spraying water outside Raul's container, and Charlie's accumulated dirt has been washed away. His old coat and hat, blue-lit eyes and worn UBIK chest logo remain established; washing does not restore the worn metal or logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨끗해진 자신의 몸을 내려다보며 양팔을 활짝 벌린 채 즐거운 표정을 짓는 찰리의 전신 구도.\n\nLOCATION (lock): In the outdoor washing area in front of a container home, on ground sprayed by a hose. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the diagonal retreat at 찰리's waist height with a slight upward tilt, retaining an oblique front view that shows his entire body and clear space beyond both extended hands and below his feet. Center him loosely as he tips his head down toward his newly cleaned torso, his open arms and delighted expression readable together; the water enters from the hose side, with 라울 and 앰버 outside the crop. Let increased camera distance be the sole emphasized change, without repositioning Charlie or changing the lighting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Raul's container home (The washing takes place outside the home) — An oblique portion of its exterior remains behind Charlie; used as Establishes the domestic setting and a stable scale reference; Hose water (Spraying Charlie and washing dirt from his body); used as A peripheral stream connects the off-screen hose side to Charlie's torso without obscuring his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination keeps the washing water and precise robot surfaces readable while allowing Charlie's delight to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A hose is spraying water outside Raul's container, and Charlie's accumulated dirt has been washed away. His old coat and hat, blue-lit eyes and worn UBIK chest logo remain established; washing does not restore the worn metal or logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨끗해진 자신의 몸을 내려다보며 양팔을 활짝 벌린 채 즐거운 표정을 짓는 찰리의 전신 구도.\n\nLOCATION (lock): In the outdoor washing area in front of a container home, on ground sprayed by a hose. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the diagonal retreat at 찰리's waist height with a slight upward tilt, retaining an oblique front view that shows his entire body and clear space beyond both extended hands and below his feet. Center him loosely as he tips his head down toward his newly cleaned torso, his open arms and delighted expression readable together; the water enters from the hose side, with 라울 and 앰버 outside the crop. Let increased camera distance be the sole emphasized change, without repositioning Charlie or changing the lighting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Raul's container home (The washing takes place outside the home) — An oblique portion of its exterior remains behind Charlie; used as Establishes the domestic setting and a stable scale reference; Hose water (Spraying Charlie and washing dirt from his body); used as A peripheral stream connects the off-screen hose side to Charlie's torso without obscuring his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination keeps the washing water and precise robot surfaces readable while allowing Charlie's delight to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A hose is spraying water outside Raul's container, and Charlie's accumulated dirt has been washed away. His old coat and hat, blue-lit eyes and worn UBIK chest logo remain established; washing does not restore the worn metal or logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S16sh4__bgfirst_bg.png",
     "asset_id": "e6bd925a-3eca-4f4c-b706-f4b94e618b81",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S16sh4.png",
     "asset_id": "42e2691a-83b8-4772-b28b-520706f1436c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L28B01.png",
     "asset_id": "91b36b81-eabb-4c94-984f-45b03850cbb1",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "고개를 약간 숙여 몸통을 바라보고 있으며, 우측 밖에서 뿜어진 물줄기가 가슴과 팔에 닿고 있음.",
    "built_space": "컨테이너 주택 앞 공터. 배경의 컨테이너와 집기들이 레퍼런스와 일치하게 배치되어 있음.",
    "entities": "찰리의 얼굴과 가슴 심볼은 비슷하나, 고릴라형 체형이 아닌 일반적인 인간형 비율(짧은 팔, 긴 다리)로 묘사되어 설정과 불일치함.",
    "hard_violations": [],
    "physics": "지면을 딛고 서 있으며 살포되는 물줄기와 신체의 상호작용에 무리가 없음."
   },
   {
    "label": "B",
    "direction": "찰리의 시선은 자신의 가슴을 향해 아래로 내려다보고 있으며, 화면 우측 밖에서 날아오는 물줄기가 몸통에 명중함.",
    "built_space": "컨테이너 주택 앞 공터. 레퍼런스와 동일한 위치에 창문, 문, 에어컨 실외기 및 배경의 호스 스탠드가 정확히 배치됨.",
    "entities": "찰리의 외형이 텍스트와 레퍼런스의 묘사(육중한 팔과 짧은 다리의 고릴라형 비율, 흰 마스크, 가슴의 원자로 심볼)와 일치함.",
    "hard_violations": [],
    "physics": "두 발로 젖은 지면을 안정적으로 딛고 서 있으며, 몸에 부딪혀 흩어지는 물줄기의 물리적 묘사가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 특수한 고릴라형 신체 비율을 완벽히 구현했으며, 지시된 전신 구도와 동작 지침을 충실히 따름."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 배경은 준수하나, 찰리의 신체 비율이 레퍼런스와 달리 인간형(긴 다리와 짧은 팔)으로 렌더링되어 캐릭터 묘사에서 크게 감점됨."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 시선은 자신의 가슴을 향해 아래로 내려다보고 있으며, 화면 우측 밖에서 날아오는 물줄기가 몸통에 명중함.",
        "built_space": "컨테이너 주택 앞 공터. 레퍼런스와 동일한 위치에 창문, 문, 에어컨 실외기 및 배경의 호스 스탠드가 정확히 배치됨.",
        "entities": "찰리의 외형이 텍스트와 레퍼런스의 묘사(육중한 팔과 짧은 다리의 고릴라형 비율, 흰 마스크, 가슴의 원자로 심볼)와 일치함.",
        "hard_violations": [],
        "physics": "두 발로 젖은 지면을 안정적으로 딛고 서 있으며, 몸에 부딪혀 흩어지는 물줄기의 물리적 묘사가 자연스러움."
       },
       {
        "label": "A",
        "direction": "고개를 약간 숙여 몸통을 바라보고 있으며, 우측 밖에서 뿜어진 물줄기가 가슴과 팔에 닿고 있음.",
        "built_space": "컨테이너 주택 앞 공터. 배경의 컨테이너와 집기들이 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "찰리의 얼굴과 가슴 심볼은 비슷하나, 고릴라형 체형이 아닌 일반적인 인간형 비율(짧은 팔, 긴 다리)로 묘사되어 설정과 불일치함.",
        "hard_violations": [],
        "physics": "지면을 딛고 서 있으며 살포되는 물줄기와 신체의 상호작용에 무리가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 특수한 고릴라형 신체 비율을 완벽히 구현했으며, 지시된 전신 구도와 동작 지침을 충실히 따름."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 배경은 준수하나, 찰리의 신체 비율이 레퍼런스와 달리 인간형(긴 다리와 짧은 팔)으로 렌더링되어 캐릭터 묘사에서 크게 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 시선은 자신의 가슴을 향해 아래로 내려다보고 있으며, 화면 우측 밖에서 날아오는 물줄기가 몸통에 명중함.",
        "built_space": "컨테이너 주택 앞 공터. 레퍼런스와 동일한 위치에 창문, 문, 에어컨 실외기 및 배경의 호스 스탠드가 정확히 배치됨.",
        "entities": "찰리의 외형이 텍스트와 레퍼런스의 묘사(육중한 팔과 짧은 다리의 고릴라형 비율, 흰 마스크, 가슴의 원자로 심볼)와 일치함.",
        "hard_violations": [],
        "physics": "두 발로 젖은 지면을 안정적으로 딛고 서 있으며, 몸에 부딪혀 흩어지는 물줄기의 물리적 묘사가 자연스러움."
       },
       {
        "label": "A",
        "direction": "고개를 약간 숙여 몸통을 바라보고 있으며, 우측 밖에서 뿜어진 물줄기가 가슴과 팔에 닿고 있음.",
        "built_space": "컨테이너 주택 앞 공터. 배경의 컨테이너와 집기들이 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "찰리의 얼굴과 가슴 심볼은 비슷하나, 고릴라형 체형이 아닌 일반적인 인간형 비율(짧은 팔, 긴 다리)로 묘사되어 설정과 불일치함.",
        "hard_violations": [],
        "physics": "지면을 딛고 서 있으며 살포되는 물줄기와 신체의 상호작용에 무리가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "양팔을 활짝 펼치고 깨끗해진 몸으로 고개를 숙인 전신 동작과 짧고 육중한 체형이 더 충실하지만, 푸른 눈·코트·모자·UBIK 로고와 차분한 낮 조명은 빠졌다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낮은 사선 전신 구도와 몸통으로 연결되는 물줄기는 맞지만, 팔이 더 처지고 다리가 길어져 핵심 동작과 고릴라형 체형이 A보다 약하며 지속 소품과 조명도 불일치한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 고개와 얼굴을 아래로 기울여 자신의 가슴과 배 쪽을 향한다. 양팔은 거의 어깨 높이로 넓게 벌리고 손바닥을 연다. 오른쪽 화면 밖에서 들어오는 물줄기는 오른팔과 어깨를 지나 상부 몸통에 닿으며 얼굴을 가리지 않는다.",
        "built_space": "회색 골판 컨테이너에 창문 두 개, 차양이 있는 출입문 한 개, 문 아래 계단 한 벌, 왼쪽 실외기 한 개가 보인다. 왼쪽 의자와 수납통 무리, 오른쪽 세척대와 수도·호스 자리도 참고 장소의 배치를 따른다. 찰리는 그 앞 젖은 마당에 서 있다. 배경은 약한 사선이며, 전신과 양손 밖 및 발 아래 여백이 확보되어 있다.",
        "entities": "등장 인물은 기계 찰리 한 개체뿐이며 라울과 앰버는 보이지 않는다. 샌드 베이지 장갑, 흰 마스크 얼굴, 점 형태의 눈과 선 형태의 미소, 푸른 원형 가슴 원자로, 마모된 금속은 참고와 부합한다. 육중한 팔과 비교적 짧은 다리도 유지된다. 눈은 참고처럼 주황색이지만 본문의 푸른 눈 지시와 다르며, 코트·모자·UBIK 로고는 보이지 않는다. 낮이지만 조명은 요구보다 직사광과 반짝임이 강하다.",
        "hard_violations": [],
        "physics": "벌린 두 발이 젖은 지면에 닿아 몸을 지지하고, 굽힌 무릎과 기계 관절이 펼친 팔을 지탱한다. 물은 화면 밖 호스 방향에서 분사되어 장갑에 부딪힌 뒤 아래로 떨어지고 바닥에 고인다. 배경 수도의 별도 낙수도 참고 사진에 있는 설비와 연결된다. 지지 없이 떠 있는 몸체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 머리를 아래와 화면 오른쪽으로 기울여 가슴 아래쪽을 향한다. 두 팔은 벌려져 있지만 어깨보다 낮고 특히 화면 왼쪽 손이 아래로 처져 있다. 오른쪽 아래 화면 밖에서 비스듬히 들어오는 물줄기는 팔 아래를 지나 몸통 옆과 가슴 부근에 닿는다. 얼굴은 물줄기에 가려지지 않는다.",
        "built_space": "컨테이너의 창문 두 개, 차양 달린 문 한 개, 문 앞 계단 한 벌, 왼쪽 실외기 한 개가 보인다. 왼쪽 수납통과 의자, 오른쪽 세척대·수도와 녹색 울타리도 같은 장소를 식별하게 한다. 찰리는 컨테이너 앞 젖은 마당에 서 있고 전신과 양손 바깥, 발 아래가 모두 들어온다. 배경 지붕선과 인물은 낮은 사선 시점을 비교적 분명하게 보여준다.",
        "entities": "기계 찰리만 등장한다. 흰 마스크, 주황색 점 눈, 선 형태의 미소, 베이지색 마모 장갑과 푸른 가슴 원자로는 참고의 주요 요소를 따른다. 다만 다리가 길고 몸통이 늘씬해 짧고 뚱뚱한 고릴라형 비율에서 더 멀어진다. 본문이 요구한 푸른 눈, 코트, 모자, 낡은 UBIK 로고는 보이지 않는다. 맑은 낮의 강한 햇빛으로 차분한 조명 요구와도 다르다.",
        "hard_violations": [],
        "physics": "두 발바닥이 지면에 접촉하고 벌린 다리가 몸의 무게를 받는다. 팔은 어깨와 팔꿈치 관절로 연결되어 펼쳐진 자세를 지탱한다. 분사수는 화면 밖 오른쪽에서 몸통까지 이어지며 충돌 후 물방울이 낙하한다. 바닥의 물웅덩이와 아래로 이어지는 반사는 가능한 배치이고, 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "양팔을 활짝 펼치고 깨끗해진 몸으로 고개를 숙인 전신 동작과 짧고 육중한 체형이 더 충실하지만, 푸른 눈·코트·모자·UBIK 로고와 차분한 낮 조명은 빠졌다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낮은 사선 전신 구도와 몸통으로 연결되는 물줄기는 맞지만, 팔이 더 처지고 다리가 길어져 핵심 동작과 고릴라형 체형이 A보다 약하며 지속 소품과 조명도 불일치한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 고개와 얼굴을 아래로 기울여 자신의 가슴과 배 쪽을 향한다. 양팔은 거의 어깨 높이로 넓게 벌리고 손바닥을 연다. 오른쪽 화면 밖에서 들어오는 물줄기는 오른팔과 어깨를 지나 상부 몸통에 닿으며 얼굴을 가리지 않는다.",
        "built_space": "회색 골판 컨테이너에 창문 두 개, 차양이 있는 출입문 한 개, 문 아래 계단 한 벌, 왼쪽 실외기 한 개가 보인다. 왼쪽 의자와 수납통 무리, 오른쪽 세척대와 수도·호스 자리도 참고 장소의 배치를 따른다. 찰리는 그 앞 젖은 마당에 서 있다. 배경은 약한 사선이며, 전신과 양손 밖 및 발 아래 여백이 확보되어 있다.",
        "entities": "등장 인물은 기계 찰리 한 개체뿐이며 라울과 앰버는 보이지 않는다. 샌드 베이지 장갑, 흰 마스크 얼굴, 점 형태의 눈과 선 형태의 미소, 푸른 원형 가슴 원자로, 마모된 금속은 참고와 부합한다. 육중한 팔과 비교적 짧은 다리도 유지된다. 눈은 참고처럼 주황색이지만 본문의 푸른 눈 지시와 다르며, 코트·모자·UBIK 로고는 보이지 않는다. 낮이지만 조명은 요구보다 직사광과 반짝임이 강하다.",
        "hard_violations": [],
        "physics": "벌린 두 발이 젖은 지면에 닿아 몸을 지지하고, 굽힌 무릎과 기계 관절이 펼친 팔을 지탱한다. 물은 화면 밖 호스 방향에서 분사되어 장갑에 부딪힌 뒤 아래로 떨어지고 바닥에 고인다. 배경 수도의 별도 낙수도 참고 사진에 있는 설비와 연결된다. 지지 없이 떠 있는 몸체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 머리를 아래와 화면 오른쪽으로 기울여 가슴 아래쪽을 향한다. 두 팔은 벌려져 있지만 어깨보다 낮고 특히 화면 왼쪽 손이 아래로 처져 있다. 오른쪽 아래 화면 밖에서 비스듬히 들어오는 물줄기는 팔 아래를 지나 몸통 옆과 가슴 부근에 닿는다. 얼굴은 물줄기에 가려지지 않는다.",
        "built_space": "컨테이너의 창문 두 개, 차양 달린 문 한 개, 문 앞 계단 한 벌, 왼쪽 실외기 한 개가 보인다. 왼쪽 수납통과 의자, 오른쪽 세척대·수도와 녹색 울타리도 같은 장소를 식별하게 한다. 찰리는 컨테이너 앞 젖은 마당에 서 있고 전신과 양손 바깥, 발 아래가 모두 들어온다. 배경 지붕선과 인물은 낮은 사선 시점을 비교적 분명하게 보여준다.",
        "entities": "기계 찰리만 등장한다. 흰 마스크, 주황색 점 눈, 선 형태의 미소, 베이지색 마모 장갑과 푸른 가슴 원자로는 참고의 주요 요소를 따른다. 다만 다리가 길고 몸통이 늘씬해 짧고 뚱뚱한 고릴라형 비율에서 더 멀어진다. 본문이 요구한 푸른 눈, 코트, 모자, 낡은 UBIK 로고는 보이지 않는다. 맑은 낮의 강한 햇빛으로 차분한 조명 요구와도 다르다.",
        "hard_violations": [],
        "physics": "두 발바닥이 지면에 접촉하고 벌린 다리가 몸의 무게를 받는다. 팔은 어깨와 팔꿈치 관절로 연결되어 펼쳐진 자세를 지탱한다. 분사수는 화면 밖 오른쪽에서 몸통까지 이어지며 충돌 후 물방울이 낙하한다. 바닥의 물웅덩이와 아래로 이어지는 반사는 가능한 배치이고, 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.482,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.482,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1482
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "찰리의 특수한 고릴라형 신체 비율을 완벽히 구현했으며, 지시된 전신 구도와 동작 지침을 충실히 따름."
   },
   {
    "label": "A",
    "score": 1482,
    "verdict_ko": "구도와 배경은 준수하나, 찰리의 신체 비율이 레퍼런스와 달리 인간형(긴 다리와 짧은 팔)으로 렌더링되어 캐릭터 묘사에서 크게 감점됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L28B01.png",
    "asset_id": "91b36b81-eabb-4c94-984f-45b03850cbb1",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08c1-6c01-78a4-985a-70a85b0a039f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S16sh4__bgfirst_bg.png",
   "bg_asset_id": "e6bd925a-3eca-4f4c-b706-f4b94e618b81",
   "bg_record_key": "S16sh4::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S16sh8::signage": {
  "fp": "fa4abcc72b2dc83d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S16sh8": {
  "input_fingerprint": "68a931e037554500",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 팔짱 낀 이현우와 그 옆에서 눈을 반짝이는 페드로의 시선이 동시에 찰리 쪽을 향한 구도.\n\nLOCATION (lock): At the edge of the container home's outdoor forecourt, a short distance from the hose-washing area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the shallow arc on the Charlie-facing side of 이현우 and 페드로 at shoulder height, forming an oblique waist-up two-shot with 이현우 left and 페드로 right, neither face overlapping the other. 이현우 settles his weight behind folded arms while 페드로 inclines slightly forward, both looking past the same frame edge toward 찰리, who remains off camera as required by this flow stage. Hold distance and exposure steady through the endpoint so their shared gaze, rather than a fresh camera approach, becomes the emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container-home exterior (The observers remain outside Raul's home, apart from the washing) — A peripheral exterior portion sits behind the two observers; used as Maintains location continuity while leaving their faces and folded arms unobstructed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daytime ambient light and gentle facial separation, with Pedro's bright-eyed interest providing the expressive lift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's body is now washed clean of the accumulated dirt. His worn metal, faded chest logo, blue-lit eyes and established coat-and-hat disguise remain. 이현우: He remains at a distance from the washing, with facial injuries and the untreated leg bite. 페드로: He remains at the distant observation spot.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 팔짱 낀 이현우와 그 옆에서 눈을 반짝이는 페드로의 시선이 동시에 찰리 쪽을 향한 구도.\n\nLOCATION (lock): At the edge of the container home's outdoor forecourt, a short distance from the hose-washing area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the shallow arc on the Charlie-facing side of 이현우 and 페드로 at shoulder height, forming an oblique waist-up two-shot with 이현우 left and 페드로 right, neither face overlapping the other. 이현우 settles his weight behind folded arms while 페드로 inclines slightly forward, both looking past the same frame edge toward 찰리, who remains off camera as required by this flow stage. Hold distance and exposure steady through the endpoint so their shared gaze, rather than a fresh camera approach, becomes the emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container-home exterior (The observers remain outside Raul's home, apart from the washing) — A peripheral exterior portion sits behind the two observers; used as Maintains location continuity while leaving their faces and folded arms unobstructed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daytime ambient light and gentle facial separation, with Pedro's bright-eyed interest providing the expressive lift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's body is now washed clean of the accumulated dirt. His worn metal, faded chest logo, blue-lit eyes and established coat-and-hat disguise remain. 이현우: He remains at a distance from the washing, with facial injuries and the untreated leg bite. 페드로: He remains at the distant observation spot.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 팔짱 낀 이현우와 그 옆에서 눈을 반짝이는 페드로의 시선이 동시에 찰리 쪽을 향한 구도.\n\nLOCATION (lock): At the edge of the container home's outdoor forecourt, a short distance from the hose-washing area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the shallow arc on the Charlie-facing side of 이현우 and 페드로 at shoulder height, forming an oblique waist-up two-shot with 이현우 left and 페드로 right, neither face overlapping the other. 이현우 settles his weight behind folded arms while 페드로 inclines slightly forward, both looking past the same frame edge toward 찰리, who remains off camera as required by this flow stage. Hold distance and exposure steady through the endpoint so their shared gaze, rather than a fresh camera approach, becomes the emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container-home exterior (The observers remain outside Raul's home, apart from the washing) — A peripheral exterior portion sits behind the two observers; used as Maintains location continuity while leaving their faces and folded arms unobstructed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daytime ambient light and gentle facial separation, with Pedro's bright-eyed interest providing the expressive lift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hose-washing setup remains outside the container, and Charlie's body is now washed clean of the accumulated dirt. His worn metal, faded chest logo, blue-lit eyes and established coat-and-hat disguise remain. 이현우: He remains at a distance from the washing, with facial injuries and the untreated leg bite. 페드로: He remains at the distant observation spot.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
    "built_space": "컨테이너의 창문과 문이 보이나, 에어컨 실외기가 창문 아래가 아닌 좌측으로 이동해 레퍼런스의 배치와 다름.",
    "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
    "hard_violations": [
     "[gemini-pro] 우측 배경에 지지대나 연결된 호스 없이 허공에서 생성되어 뿜어지는 물줄기"
    ],
    "physics": "인물의 자세는 자연스러우나, 우측 배경의 물줄기가 물리적 원천 없이 공중에 떠 있음."
   },
   {
    "label": "B",
    "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
    "built_space": "창문 아래의 에어컨 실외기, 좌측의 선반과 노란 상자, 우측의 수도꼭지 등 레퍼런스의 공간 구조와 정확히 일치함.",
    "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
    "hard_violations": [],
    "physics": "팔짱을 낀 이현우와 앞으로 살짝 숙인 페드로의 자세가 자연스럽게 지지되고 있으며, 우측 수도꼭지의 물줄기도 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 인물의 특징은 잘 살렸으나, 우측 배경에 허공에서 뿜어지는 물줄기가 발생하고 구조물 위치가 어긋남."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물의 구도와 시선을 정확히 구현했으며, 락(lock)이 걸린 배경의 디테일까지 완벽하게 재현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
        "built_space": "컨테이너의 창문과 문이 보이나, 에어컨 실외기가 창문 아래가 아닌 좌측으로 이동해 레퍼런스의 배치와 다름.",
        "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
        "hard_violations": [
         "우측 배경에 지지대나 연결된 호스 없이 허공에서 생성되어 뿜어지는 물줄기"
        ],
        "physics": "인물의 자세는 자연스러우나, 우측 배경의 물줄기가 물리적 원천 없이 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
        "built_space": "창문 아래의 에어컨 실외기, 좌측의 선반과 노란 상자, 우측의 수도꼭지 등 레퍼런스의 공간 구조와 정확히 일치함.",
        "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
        "hard_violations": [],
        "physics": "팔짱을 낀 이현우와 앞으로 살짝 숙인 페드로의 자세가 자연스럽게 지지되고 있으며, 우측 수도꼭지의 물줄기도 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 인물의 특징은 잘 살렸으나, 우측 배경에 허공에서 뿜어지는 물줄기가 발생하고 구조물 위치가 어긋남."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인물의 구도와 시선을 정확히 구현했으며, 락(lock)이 걸린 배경의 디테일까지 완벽하게 재현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
        "built_space": "컨테이너의 창문과 문이 보이나, 에어컨 실외기가 창문 아래가 아닌 좌측으로 이동해 레퍼런스의 배치와 다름.",
        "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
        "hard_violations": [
         "우측 배경에 지지대나 연결된 호스 없이 허공에서 생성되어 뿜어지는 물줄기"
        ],
        "physics": "인물의 자세는 자연스러우나, 우측 배경의 물줄기가 물리적 원천 없이 공중에 떠 있음."
       },
       {
        "label": "B",
        "direction": "이현우와 페드로의 시선이 화면 우측 프레임 밖의 찰리를 향해 나란히 고정되어 있음.",
        "built_space": "창문 아래의 에어컨 실외기, 좌측의 선반과 노란 상자, 우측의 수도꼭지 등 레퍼런스의 공간 구조와 정확히 일치함.",
        "entities": "이현우(상처 난 얼굴, 인이어, 어두운 셔츠)와 페드로(비니, 집업 자켓) 모두 지정된 외형 및 의상과 일치함.",
        "hard_violations": [],
        "physics": "팔짱을 낀 이현우와 앞으로 살짝 숙인 페드로의 자세가 자연스럽게 지지되고 있으며, 우측 수도꼭지의 물줄기도 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 이현우의 팔짱과 오른쪽 페드로의 전경사, 화면 밖 오른쪽으로 모이는 시선을 정확히 구현하며 인물 외형과 컨테이너 설비 배치의 연속성도 더 충실하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "허리 위 투숏과 공동 시선은 충실하지만, 이현우의 뒤로 실린 자세가 덜 뚜렷하고 실외기의 창문 대비 위치가 이전 장소와 다소 달라 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로 모두 얼굴과 눈동자를 화면 오른쪽 바깥으로 돌리고 있다. 두 시선은 지시된 화면 밖 찰리의 위치와 양립하며, 서로 또는 카메라를 보는 모습은 아니다. 찰리는 보이지 않는다. 오른쪽 배경의 호스 물줄기는 아래쪽 마당으로 떨어진다.",
        "built_space": "녹슨 회색 골판 컨테이너에 창문 두 개, 중앙 출입문 한 개와 작은 차양, 문 왼쪽 벽등 한 개가 보인다. 왼쪽 창문 아래에는 실외기 한 대와 수납통들이 있고, 오른쪽에는 병들이 놓인 작업대와 호스 급수대 한 곳, 울타리와 기대 세운 팔레트가 보인다. 두 사람은 이 설비들보다 앞선 야외에 서 있으며 세척 설비와 떨어져 있다. 배경이 얼굴이나 팔짱을 가리지 않고, 설비가 비정상적으로 확대되지 않았다.",
        "entities": "인물은 두 명뿐이다. 왼쪽 이현우는 참조와 가까운 앳된 동아시아계 남성 얼굴, 헝클어진 검은 머리, 마른 체격이며 얼굴과 목의 상처, 귀의 소형 인이어, 피와 먼지가 묻은 어두운 셔츠가 보인다. 오른쪽 페드로는 참조와 가까운 젊은 라틴계 혼혈 남성 외형이며 검은 비니와 낡은 올리브색 집업을 착용했다. 눈은 정상적인 형태이고 살짝 벌린 입과 집중한 눈빛으로 흥미를 표현한다. 바지와 다리의 물린 상처는 지정된 상반신 구도 밖이므로 확인 대상이 아니다.",
        "hard_violations": [],
        "physics": "이현우는 두 팔을 몸 앞에서 자연스럽게 포개고 손을 반대쪽 팔에 붙이고 있다. 몸통은 곧게 서면서 페드로보다 뒤로 안정되게 놓인다. 페드로는 허리와 상체를 조금 앞으로 기울인 가능한 기립 자세다. 발은 화면 밖이지만 공중에 뜬 정황은 없다. 호스는 급수대와 연결되어 바닥에 놓이고, 물줄기는 아래로 휘어 젖은 지면에 닿는다."
       },
       {
        "label": "B",
        "direction": "두 사람 모두 화면 오른쪽 밖을 바라보며, 페드로의 얼굴도 같은 쪽으로 돌아가 있다. 화면 밖 찰리를 함께 관찰한다는 설정에 맞고 카메라를 직접 보지 않는다. 찰리나 다른 인물은 들어오지 않았다. 배경의 물줄기는 오른쪽 급수대에서 아래쪽 지면으로 향한다.",
        "built_space": "회색 컨테이너의 창문 두 개, 중앙 문 한 개와 차양, 벽등 한 개가 보인다. 오른쪽에는 병들이 놓인 작업대, 호스 급수대 한 곳, 울타리와 팔레트가 유지된다. 실외기는 한 대지만 왼쪽 창문 아래보다 화면 중앙의 두 인물 사이에 드러나 참조의 창문 대비 위치와 다소 다르게 읽힌다. 인물들은 마당 앞쪽에 서 있고 배경 설비와 겹쳐 서 있지는 않다. 어깨 높이의 사선 허리 위 투숏이며 얼굴끼리 겹치지 않는다.",
        "entities": "왼쪽의 젊은 동아시아계 남성과 오른쪽의 젊은 라틴계 혼혈 남성으로 두 명만 보인다. 이현우의 검은 헝클어진 머리, 상처 난 얼굴과 목, 인이어, 피와 흙먼지가 묻은 셔츠가 유지된다. 페드로의 비니, 얼굴의 옅은 수염과 낡은 올리브색 집업도 참조에 가깝다. 페드로는 자연스러운 눈동자와 열린 표정으로 관심을 드러낸다. 하체 복장과 다리 상처는 프레임 밖이며, 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "이현우의 팔짱은 한 손이 반대쪽 위팔에 얹히고 다른 손이 팔 아래로 들어가는 자연스러운 구조다. 다만 머리와 어깨가 조금 앞으로 기울어 뒤로 체중을 싣는 지시는 A보다 덜 명확하다. 페드로는 허리에서 살짝 앞으로 숙이며 양팔을 몸 옆으로 내린 가능한 자세다. 발이 잘려 있지만 부유하거나 지지 없이 매달린 모습은 아니다. 배경 설비는 지면이나 작업대에 지지되고 물은 급수대에서 바닥으로 떨어진다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 이현우의 팔짱과 오른쪽 페드로의 전경사, 화면 밖 오른쪽으로 모이는 시선을 정확히 구현하며 인물 외형과 컨테이너 설비 배치의 연속성도 더 충실하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "허리 위 투숏과 공동 시선은 충실하지만, 이현우의 뒤로 실린 자세가 덜 뚜렷하고 실외기의 창문 대비 위치가 이전 장소와 다소 달라 보인다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우와 페드로 모두 얼굴과 눈동자를 화면 오른쪽 바깥으로 돌리고 있다. 두 시선은 지시된 화면 밖 찰리의 위치와 양립하며, 서로 또는 카메라를 보는 모습은 아니다. 찰리는 보이지 않는다. 오른쪽 배경의 호스 물줄기는 아래쪽 마당으로 떨어진다.",
        "built_space": "녹슨 회색 골판 컨테이너에 창문 두 개, 중앙 출입문 한 개와 작은 차양, 문 왼쪽 벽등 한 개가 보인다. 왼쪽 창문 아래에는 실외기 한 대와 수납통들이 있고, 오른쪽에는 병들이 놓인 작업대와 호스 급수대 한 곳, 울타리와 기대 세운 팔레트가 보인다. 두 사람은 이 설비들보다 앞선 야외에 서 있으며 세척 설비와 떨어져 있다. 배경이 얼굴이나 팔짱을 가리지 않고, 설비가 비정상적으로 확대되지 않았다.",
        "entities": "인물은 두 명뿐이다. 왼쪽 이현우는 참조와 가까운 앳된 동아시아계 남성 얼굴, 헝클어진 검은 머리, 마른 체격이며 얼굴과 목의 상처, 귀의 소형 인이어, 피와 먼지가 묻은 어두운 셔츠가 보인다. 오른쪽 페드로는 참조와 가까운 젊은 라틴계 혼혈 남성 외형이며 검은 비니와 낡은 올리브색 집업을 착용했다. 눈은 정상적인 형태이고 살짝 벌린 입과 집중한 눈빛으로 흥미를 표현한다. 바지와 다리의 물린 상처는 지정된 상반신 구도 밖이므로 확인 대상이 아니다.",
        "hard_violations": [],
        "physics": "이현우는 두 팔을 몸 앞에서 자연스럽게 포개고 손을 반대쪽 팔에 붙이고 있다. 몸통은 곧게 서면서 페드로보다 뒤로 안정되게 놓인다. 페드로는 허리와 상체를 조금 앞으로 기울인 가능한 기립 자세다. 발은 화면 밖이지만 공중에 뜬 정황은 없다. 호스는 급수대와 연결되어 바닥에 놓이고, 물줄기는 아래로 휘어 젖은 지면에 닿는다."
       },
       {
        "label": "A",
        "direction": "두 사람 모두 화면 오른쪽 밖을 바라보며, 페드로의 얼굴도 같은 쪽으로 돌아가 있다. 화면 밖 찰리를 함께 관찰한다는 설정에 맞고 카메라를 직접 보지 않는다. 찰리나 다른 인물은 들어오지 않았다. 배경의 물줄기는 오른쪽 급수대에서 아래쪽 지면으로 향한다.",
        "built_space": "회색 컨테이너의 창문 두 개, 중앙 문 한 개와 차양, 벽등 한 개가 보인다. 오른쪽에는 병들이 놓인 작업대, 호스 급수대 한 곳, 울타리와 팔레트가 유지된다. 실외기는 한 대지만 왼쪽 창문 아래보다 화면 중앙의 두 인물 사이에 드러나 참조의 창문 대비 위치와 다소 다르게 읽힌다. 인물들은 마당 앞쪽에 서 있고 배경 설비와 겹쳐 서 있지는 않다. 어깨 높이의 사선 허리 위 투숏이며 얼굴끼리 겹치지 않는다.",
        "entities": "왼쪽의 젊은 동아시아계 남성과 오른쪽의 젊은 라틴계 혼혈 남성으로 두 명만 보인다. 이현우의 검은 헝클어진 머리, 상처 난 얼굴과 목, 인이어, 피와 흙먼지가 묻은 셔츠가 유지된다. 페드로의 비니, 얼굴의 옅은 수염과 낡은 올리브색 집업도 참조에 가깝다. 페드로는 자연스러운 눈동자와 열린 표정으로 관심을 드러낸다. 하체 복장과 다리 상처는 프레임 밖이며, 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "이현우의 팔짱은 한 손이 반대쪽 위팔에 얹히고 다른 손이 팔 아래로 들어가는 자연스러운 구조다. 다만 머리와 어깨가 조금 앞으로 기울어 뒤로 체중을 싣는 지시는 A보다 덜 명확하다. 페드로는 허리에서 살짝 앞으로 숙이며 양팔을 몸 옆으로 내린 가능한 자세다. 발이 잘려 있지만 부유하거나 지지 없이 매달린 모습은 아니다. 배경 설비는 지면이나 작업대에 지지되고 물은 급수대에서 바닥으로 떨어진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.46,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.21,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 우측 배경에 지지대나 연결된 호스 없이 허공에서 생성되어 뿜어지는 물줄기"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1210,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1210,
    "verdict_ko": "구도와 인물의 특징은 잘 살렸으나, 우측 배경에 허공에서 뿜어지는 물줄기가 발생하고 구조물 위치가 어긋남.  ★위반: [gemini-pro] 우측 배경에 지지대나 연결된 호스 없이 허공에서 생성되어 뿜어지는 물줄기"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 인물의 구도와 시선을 정확히 구현했으며, 락(lock)이 걸린 배경의 디테일까지 완벽하게 재현함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S16sh4_sel.png",
    "asset_id": "c4c97849-8f3f-4674-a147-2392a77b8a06",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08c8-52e8-7954-b9f2-0600e3ec4c91",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh4"
  }
 },
 "S17sh5::signage": {
  "fp": "fc9e19cdc300d349",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh5": {
  "input_fingerprint": "7f955314ca45d94c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손가락 세 개를 펴 보인 채 '박철진의 수하 1'을 응시하는 마츠다의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the cluttered secondhand shop, with daylight entering from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage beside 마츠다 at shoulder height, nearly perpendicular to his facing direction and remaining on the established side of the conversational axis. Place his upper body left of center and his three extended fingers in the central gap, with only 박철진의 수하 1's near shoulder and torso at the right edge; 마츠다 studies the henchman's face just beyond that edge while the henchman's head remains outside the crop. Keep the entry moment settled before the lateral track begins, making Matsuda's bargaining hand the principal positional emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Television under repair (Matsuda's repair has been interrupted by the visitors) — Only an oblique portion of the housing is visible; screen content is not featured; used as A small lower-background reference to the interrupted work; Desk (Present in Matsuda's work area) — Its near edge runs beneath the bargaining gesture; used as Grounds the hand gesture within the workspace without obstructing it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime interior illumination with controlled contrast keeps the three-finger demand and Matsuda's watchful profile equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop interior, piled electrical appliances, scrap, and daytime lighting from the reference. Exclude the teenage customer and the robot, who have already left.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains crowded with appliances and scrap, with a television at the repair station. 마츠다: He is still upright inside the shop, having interrupted his television repair. 박철진의 수하 1: He is inside the shop wearing the militia's self-styled uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손가락 세 개를 펴 보인 채 '박철진의 수하 1'을 응시하는 마츠다의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the cluttered secondhand shop, with daylight entering from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage beside 마츠다 at shoulder height, nearly perpendicular to his facing direction and remaining on the established side of the conversational axis. Place his upper body left of center and his three extended fingers in the central gap, with only 박철진의 수하 1's near shoulder and torso at the right edge; 마츠다 studies the henchman's face just beyond that edge while the henchman's head remains outside the crop. Keep the entry moment settled before the lateral track begins, making Matsuda's bargaining hand the principal positional emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Television under repair (Matsuda's repair has been interrupted by the visitors) — Only an oblique portion of the housing is visible; screen content is not featured; used as A small lower-background reference to the interrupted work; Desk (Present in Matsuda's work area) — Its near edge runs beneath the bargaining gesture; used as Grounds the hand gesture within the workspace without obstructing it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime interior illumination with controlled contrast keeps the three-finger demand and Matsuda's watchful profile equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop interior, piled electrical appliances, scrap, and daytime lighting from the reference. Exclude the teenage customer and the robot, who have already left.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains crowded with appliances and scrap, with a television at the repair station. 마츠다: He is still upright inside the shop, having interrupted his television repair. 박철진의 수하 1: He is inside the shop wearing the militia's self-styled uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손가락 세 개를 펴 보인 채 '박철진의 수하 1'을 응시하는 마츠다의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the cluttered secondhand shop, with daylight entering from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage beside 마츠다 at shoulder height, nearly perpendicular to his facing direction and remaining on the established side of the conversational axis. Place his upper body left of center and his three extended fingers in the central gap, with only 박철진의 수하 1's near shoulder and torso at the right edge; 마츠다 studies the henchman's face just beyond that edge while the henchman's head remains outside the crop. Keep the entry moment settled before the lateral track begins, making Matsuda's bargaining hand the principal positional emphasis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Television under repair (Matsuda's repair has been interrupted by the visitors) — Only an oblique portion of the housing is visible; screen content is not featured; used as A small lower-background reference to the interrupted work; Desk (Present in Matsuda's work area) — Its near edge runs beneath the bargaining gesture; used as Grounds the hand gesture within the workspace without obstructing it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime interior illumination with controlled contrast keeps the three-finger demand and Matsuda's watchful profile equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop interior, piled electrical appliances, scrap, and daytime lighting from the reference. Exclude the teenage customer and the robot, who have already left.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop remains crowded with appliances and scrap, with a television at the repair station. 마츠다: He is still upright inside the shop, having interrupted his television repair. 박철진의 수하 1: He is inside the shop wearing the militia's self-styled uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — wearing: 기름과 얼룩이 진 빛바랜 청색 작업 조끼와 체크 무늬 셔츠, 낡은 바지.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "마츠다의 시선이 우측의 수하를 자연스럽게 향함.",
    "built_space": "지정된 전파사 내부 환경과 책상, 수리 중인 TV가 올바르게 표현됨.",
    "entities": "마츠다의 인상착의는 훌륭하나 수하의 머리 일부가 화면에 노출되었고, 3개의 손가락을 편 손의 형태가 어색함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (마츠다의 오른팔에 엄지손가락 방향이 반대인 왼손이 연결되어 있음)"
    ],
    "physics": "손을 들어올린 동작 자체는 신체에 의해 지탱되나 해부학적 구조가 붕괴됨."
   },
   {
    "label": "B",
    "direction": "마츠다의 시선이 화면 우측 가장자리에 위치한 수하의 얼굴 방향을 정확히 향하고 있음.",
    "built_space": "레퍼런스의 전파사 내부 공간과 일치하며, 하단 배경에 책상과 수리 중인 TV 하우징이 올바르게 배치됨.",
    "entities": "마츠다의 얼굴과 작업복이 레퍼런스와 완벽히 일치하며, 손가락 3개를 편 동작이 잘 묘사됨. 우측 수하도 규정된 복장을 착용함.",
    "hard_violations": [],
    "physics": "마츠다가 선 상태로 팔을 들어올린 자세가 신체에 의해 안정적으로 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도와 마츠다의 손동작을 훌륭하게 구현했으나, 수하의 머리가 프레임 밖으로 완전히 배제되지 않고 일부 노출된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "마츠다의 오른팔에 왼손이 달려 있는 치명적인 해부학적 오류가 발생하여 우선적으로 실격 처리됩니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "마츠다의 시선이 화면 우측 가장자리에 위치한 수하의 얼굴 방향을 정확히 향하고 있음.",
        "built_space": "레퍼런스의 전파사 내부 공간과 일치하며, 하단 배경에 책상과 수리 중인 TV 하우징이 올바르게 배치됨.",
        "entities": "마츠다의 얼굴과 작업복이 레퍼런스와 완벽히 일치하며, 손가락 3개를 편 동작이 잘 묘사됨. 우측 수하도 규정된 복장을 착용함.",
        "hard_violations": [],
        "physics": "마츠다가 선 상태로 팔을 들어올린 자세가 신체에 의해 안정적으로 지탱되고 있음."
       },
       {
        "label": "A",
        "direction": "마츠다의 시선이 우측의 수하를 자연스럽게 향함.",
        "built_space": "지정된 전파사 내부 환경과 책상, 수리 중인 TV가 올바르게 표현됨.",
        "entities": "마츠다의 인상착의는 훌륭하나 수하의 머리 일부가 화면에 노출되었고, 3개의 손가락을 편 손의 형태가 어색함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (마츠다의 오른팔에 엄지손가락 방향이 반대인 왼손이 연결되어 있음)"
        ],
        "physics": "손을 들어올린 동작 자체는 신체에 의해 지탱되나 해부학적 구조가 붕괴됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도와 마츠다의 손동작을 훌륭하게 구현했으나, 수하의 머리가 프레임 밖으로 완전히 배제되지 않고 일부 노출된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "마츠다의 오른팔에 왼손이 달려 있는 치명적인 해부학적 오류가 발생하여 우선적으로 실격 처리됩니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "마츠다의 시선이 화면 우측 가장자리에 위치한 수하의 얼굴 방향을 정확히 향하고 있음.",
        "built_space": "레퍼런스의 전파사 내부 공간과 일치하며, 하단 배경에 책상과 수리 중인 TV 하우징이 올바르게 배치됨.",
        "entities": "마츠다의 얼굴과 작업복이 레퍼런스와 완벽히 일치하며, 손가락 3개를 편 동작이 잘 묘사됨. 우측 수하도 규정된 복장을 착용함.",
        "hard_violations": [],
        "physics": "마츠다가 선 상태로 팔을 들어올린 자세가 신체에 의해 안정적으로 지탱되고 있음."
       },
       {
        "label": "A",
        "direction": "마츠다의 시선이 우측의 수하를 자연스럽게 향함.",
        "built_space": "지정된 전파사 내부 환경과 책상, 수리 중인 TV가 올바르게 표현됨.",
        "entities": "마츠다의 인상착의는 훌륭하나 수하의 머리 일부가 화면에 노출되었고, 3개의 손가락을 편 손의 형태가 어색함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (마츠다의 오른팔에 엄지손가락 방향이 반대인 왼손이 연결되어 있음)"
        ],
        "physics": "손을 들어올린 동작 자체는 신체에 의해 지탱되나 해부학적 구조가 붕괴됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "상대 쪽으로 내민 세 손가락과 비스듬한 몸 방향은 더 적합하지만, 요구한 측면 시점이 아니며 수하의 머리를 화면 밖으로 잘라내지 못했다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 작업실의 연속성은 유지하지만, 마츠다의 몸과 손바닥을 카메라에 정면으로 보여 측면 협상 구도에서 더 멀어지고 수하의 머리도 노출한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마츠다는 오른쪽 수하의 얼굴을 바라본다. 세 손가락은 중앙에서 위쪽으로 펴져 있고, 팔은 수하 쪽으로 뻗어 손등이 카메라에 보인다. 수하도 마츠다 쪽으로 몸과 머리를 돌렸지만 눈은 보이지 않는다. 응시 대상은 맞으나 마츠다의 얼굴은 정확한 측면보다 정면에 가까운 사선이다.",
        "built_space": "오른쪽에 채광되는 출입구 하나, 뒤쪽에 한글 매입·판매 표지 하나와 부품 서랍장들, 왼쪽 선반에 여러 중고 화면 기기가 보인다. 수리 책상 하나의 가장자리가 손짓 아래를 지나며 오른쪽 하단에 열린 전자기기 외함 하나가 놓여 있다. 외함은 비스듬하고 화면 내용은 노출하지 않지만 텔레비전인지는 명확하지 않다. 두 사람은 책상 앞에서 마주하며, 수하의 어깨와 몸통뿐 아니라 귀·턱·뒤통수 일부까지 화면에 들어온다.",
        "entities": "노년 동아시아계 남성인 마츠다는 참고의 얼굴, 짧은 회색 머리, 낡은 남색 모자, 얼룩진 청색 조끼와 체크 셔츠에 가깝다. 일본 국적 자체는 외관으로 확인할 수 없다. 수하는 짧은 검은 머리의 성인 남성으로 보이며 검은 제복풍 상의와 붉은 완장을 착용한다. 얼굴 대부분이 가려져 참고 인물과의 정확한 일치는 판단하기 어렵다. 세 손가락이 분명히 펴져 있으며, 추가 인물·청소년·로봇은 없다.",
        "hard_violations": [],
        "physics": "올린 손은 손목과 굽힌 팔을 통해 마츠다의 몸에 자연스럽게 이어지고, 나머지 손가락을 접은 자세도 가능하다. 두 사람의 하체와 발은 잘려 있지만 상체는 정상적으로 서 있는 자세다. 전자기기 외함과 공구는 책상에, 뒤쪽 기기들은 선반에 놓여 있어 지지 없는 부유물은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "마츠다의 눈은 오른쪽 수하의 얼굴을 향하지만 몸통은 거의 카메라 정면을 향한다. 중앙의 세 손가락은 위로 펴져 있고 손바닥은 수하보다 카메라 쪽을 향한다. 수하는 왼쪽 마츠다를 향해 돌아서 있으나 눈은 보이지 않는다. 상대를 응시하는 행동은 맞지만 요구한 옆모습 관찰 시점은 아니다.",
        "built_space": "오른쪽 출입구 하나에서 낮빛이 들어오며 뒤쪽 표지 하나, 부품 서랍장들과 왼쪽 중고 기기 선반이 보인다. 수리 책상 하나가 손 아래에 있고, 오른쪽 하단에는 회로가 드러난 외함 하나가 놓여 있다. 외함의 화면은 보이지 않지만 텔레비전임을 확정하기 어렵고 작은 배경 단서보다는 비교적 눈에 띈다. 수하는 오른쪽 전경을 차지하며 어깨·몸통 외에 턱·귀·머리 일부까지 노출된다.",
        "entities": "마츠다는 참고와 가까운 노년 동아시아계 남성으로, 회색 머리와 남색 모자, 기름 얼룩이 있는 청색 조끼, 체크 셔츠를 유지한다. 수하는 성인 남성의 목과 턱, 짧은 검은 머리 일부가 보이고 검은 제복풍 옷과 붉은 완장을 착용한다. 국적과 수하의 정확한 얼굴 일치는 보이는 부분만으로 확인할 수 없다. 펼친 손가락은 세 개이며 다른 사람이나 로봇은 없다.",
        "hard_violations": [],
        "physics": "마츠다는 팔꿈치를 굽혀 손을 들어 올렸으며 손·손목·팔의 연결과 세 손가락 자세는 물리적으로 가능하다. 발은 프레임 밖이지만 두 상체에 부유하거나 비정상적으로 기울어진 징후는 없다. 열린 외함과 작업 물품은 책상 위에 놓이고 배경 기기들은 선반이 받친다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "상대 쪽으로 내민 세 손가락과 비스듬한 몸 방향은 더 적합하지만, 요구한 측면 시점이 아니며 수하의 머리를 화면 밖으로 잘라내지 못했다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 작업실의 연속성은 유지하지만, 마츠다의 몸과 손바닥을 카메라에 정면으로 보여 측면 협상 구도에서 더 멀어지고 수하의 머리도 노출한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "마츠다는 오른쪽 수하의 얼굴을 바라본다. 세 손가락은 중앙에서 위쪽으로 펴져 있고, 팔은 수하 쪽으로 뻗어 손등이 카메라에 보인다. 수하도 마츠다 쪽으로 몸과 머리를 돌렸지만 눈은 보이지 않는다. 응시 대상은 맞으나 마츠다의 얼굴은 정확한 측면보다 정면에 가까운 사선이다.",
        "built_space": "오른쪽에 채광되는 출입구 하나, 뒤쪽에 한글 매입·판매 표지 하나와 부품 서랍장들, 왼쪽 선반에 여러 중고 화면 기기가 보인다. 수리 책상 하나의 가장자리가 손짓 아래를 지나며 오른쪽 하단에 열린 전자기기 외함 하나가 놓여 있다. 외함은 비스듬하고 화면 내용은 노출하지 않지만 텔레비전인지는 명확하지 않다. 두 사람은 책상 앞에서 마주하며, 수하의 어깨와 몸통뿐 아니라 귀·턱·뒤통수 일부까지 화면에 들어온다.",
        "entities": "노년 동아시아계 남성인 마츠다는 참고의 얼굴, 짧은 회색 머리, 낡은 남색 모자, 얼룩진 청색 조끼와 체크 셔츠에 가깝다. 일본 국적 자체는 외관으로 확인할 수 없다. 수하는 짧은 검은 머리의 성인 남성으로 보이며 검은 제복풍 상의와 붉은 완장을 착용한다. 얼굴 대부분이 가려져 참고 인물과의 정확한 일치는 판단하기 어렵다. 세 손가락이 분명히 펴져 있으며, 추가 인물·청소년·로봇은 없다.",
        "hard_violations": [],
        "physics": "올린 손은 손목과 굽힌 팔을 통해 마츠다의 몸에 자연스럽게 이어지고, 나머지 손가락을 접은 자세도 가능하다. 두 사람의 하체와 발은 잘려 있지만 상체는 정상적으로 서 있는 자세다. 전자기기 외함과 공구는 책상에, 뒤쪽 기기들은 선반에 놓여 있어 지지 없는 부유물은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "마츠다의 눈은 오른쪽 수하의 얼굴을 향하지만 몸통은 거의 카메라 정면을 향한다. 중앙의 세 손가락은 위로 펴져 있고 손바닥은 수하보다 카메라 쪽을 향한다. 수하는 왼쪽 마츠다를 향해 돌아서 있으나 눈은 보이지 않는다. 상대를 응시하는 행동은 맞지만 요구한 옆모습 관찰 시점은 아니다.",
        "built_space": "오른쪽 출입구 하나에서 낮빛이 들어오며 뒤쪽 표지 하나, 부품 서랍장들과 왼쪽 중고 기기 선반이 보인다. 수리 책상 하나가 손 아래에 있고, 오른쪽 하단에는 회로가 드러난 외함 하나가 놓여 있다. 외함의 화면은 보이지 않지만 텔레비전임을 확정하기 어렵고 작은 배경 단서보다는 비교적 눈에 띈다. 수하는 오른쪽 전경을 차지하며 어깨·몸통 외에 턱·귀·머리 일부까지 노출된다.",
        "entities": "마츠다는 참고와 가까운 노년 동아시아계 남성으로, 회색 머리와 남색 모자, 기름 얼룩이 있는 청색 조끼, 체크 셔츠를 유지한다. 수하는 성인 남성의 목과 턱, 짧은 검은 머리 일부가 보이고 검은 제복풍 옷과 붉은 완장을 착용한다. 국적과 수하의 정확한 얼굴 일치는 보이는 부분만으로 확인할 수 없다. 펼친 손가락은 세 개이며 다른 사람이나 로봇은 없다.",
        "hard_violations": [],
        "physics": "마츠다는 팔꿈치를 굽혀 손을 들어 올렸으며 손·손목·팔의 연결과 세 손가락 자세는 물리적으로 가능하다. 발은 프레임 밖이지만 두 상체에 부유하거나 비정상적으로 기울어진 징후는 없다. 열린 외함과 작업 물품은 책상 위에 놓이고 배경 기기들은 선반이 받친다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.262,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.012,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (마츠다의 오른팔에 엄지손가락 방향이 반대인 왼손이 연결되어 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1012
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 구도와 마츠다의 손동작을 훌륭하게 구현했으나, 수하의 머리가 프레임 밖으로 완전히 배제되지 않고 일부 노출된 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 1012,
    "verdict_ko": "마츠다의 오른팔에 왼손이 달려 있는 치명적인 해부학적 오류가 발생하여 우선적으로 실격 처리됩니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (마츠다의 오른팔에 엄지손가락 방향이 반대인 왼손이 연결되어 있음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S15sh18_sel.png",
    "asset_id": "e10ee162-e2fd-4a5d-89ea-36d61454ea34",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:889268>",
    "asset_id": "4300cf67-f254-4d25-bb5b-9b909d79bf38",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1245048>",
    "asset_id": "ee70391d-c03c-42da-a6c4-ad74f0a7d728",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08cd-1330-7e7f-b7fe-a155079e3c14",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh18"
  },
  "staged_characters_added": [
   "C16"
  ]
 },
 "S17sh9::signage": {
  "fp": "d926f6b948616902",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh9": {
  "input_fingerprint": "230dc15078c52f35",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다를 향해 소음기가 달린 권총의 총구를 겨눈 '박철진의 수하 1'의 팔 근접 구도.\n\nLOCATION (lock): In the customer-facing space immediately beside the secondhand shop's repair desk, lit by daytime light from the entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the same side of the axis to elbow height beside 박철진의 수하 1, viewing his extended forearm and suppressed rifle obliquely from slightly below rather than looking down the barrel. The forearm crosses the lower-right portion toward 마츠다's soft, partial torso at left, with the weapon occupying less than two-fifths of the frame and its scale checked by both bodies; the henchman's attention follows his aim toward Matsuda's face beyond the upper crop. Emphasize the reduced camera distance, preserve the existing light, and capture only the held aim before the camera turns away from any discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Suppressed rifle (Aimed toward Matsuda before the unseen discharge) — Seen obliquely from the side, with the muzzle pointing left toward Matsuda rather than toward the lens; used as Connects the foreground forearm to the threatened figure while remaining subordinate in frame area; Desk (Still beneath Matsuda's work position) — A partial edge remains behind the arm; used as Maintains spatial continuity and a practical scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the shop's subdued daytime illumination without muzzle flash or any heightened lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding shop clutter, work furnishings, and daytime lighting from the reference. Exclude the earlier three-finger bargaining gesture as a repeated event.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The appliance-and-scrap shop and the television repair station remain unchanged. 마츠다: He is still upright in the shop immediately before the shot and collapse. 박철진의 수하 1: He remains inside the shop in a militia uniform and armband.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다를 향해 소음기가 달린 권총의 총구를 겨눈 '박철진의 수하 1'의 팔 근접 구도.\n\nLOCATION (lock): In the customer-facing space immediately beside the secondhand shop's repair desk, lit by daytime light from the entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the same side of the axis to elbow height beside 박철진의 수하 1, viewing his extended forearm and suppressed rifle obliquely from slightly below rather than looking down the barrel. The forearm crosses the lower-right portion toward 마츠다's soft, partial torso at left, with the weapon occupying less than two-fifths of the frame and its scale checked by both bodies; the henchman's attention follows his aim toward Matsuda's face beyond the upper crop. Emphasize the reduced camera distance, preserve the existing light, and capture only the held aim before the camera turns away from any discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Suppressed rifle (Aimed toward Matsuda before the unseen discharge) — Seen obliquely from the side, with the muzzle pointing left toward Matsuda rather than toward the lens; used as Connects the foreground forearm to the threatened figure while remaining subordinate in frame area; Desk (Still beneath Matsuda's work position) — A partial edge remains behind the arm; used as Maintains spatial continuity and a practical scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the shop's subdued daytime illumination without muzzle flash or any heightened lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding shop clutter, work furnishings, and daytime lighting from the reference. Exclude the earlier three-finger bargaining gesture as a repeated event.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The appliance-and-scrap shop and the television repair station remain unchanged. 마츠다: He is still upright in the shop immediately before the shot and collapse. 박철진의 수하 1: He remains inside the shop in a militia uniform and armband.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마츠다를 향해 소음기가 달린 권총의 총구를 겨눈 '박철진의 수하 1'의 팔 근접 구도.\n\nLOCATION (lock): In the customer-facing space immediately beside the secondhand shop's repair desk, lit by daytime light from the entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the same side of the axis to elbow height beside 박철진의 수하 1, viewing his extended forearm and suppressed rifle obliquely from slightly below rather than looking down the barrel. The forearm crosses the lower-right portion toward 마츠다's soft, partial torso at left, with the weapon occupying less than two-fifths of the frame and its scale checked by both bodies; the henchman's attention follows his aim toward Matsuda's face beyond the upper crop. Emphasize the reduced camera distance, preserve the existing light, and capture only the held aim before the camera turns away from any discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Suppressed rifle (Aimed toward Matsuda before the unseen discharge) — Seen obliquely from the side, with the muzzle pointing left toward Matsuda rather than toward the lens; used as Connects the foreground forearm to the threatened figure while remaining subordinate in frame area; Desk (Still beneath Matsuda's work position) — A partial edge remains behind the arm; used as Maintains spatial continuity and a practical scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the shop's subdued daytime illumination without muzzle flash or any heightened lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the surrounding shop clutter, work furnishings, and daytime lighting from the reference. Exclude the earlier three-finger bargaining gesture as a repeated event.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The appliance-and-scrap shop and the television repair station remain unchanged. 마츠다: He is still upright in the shop immediately before the shot and collapse. 박철진의 수하 1: He remains inside the shop in a militia uniform and armband.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수하의 총구가 마츠다의 가슴과 어깨 부위를 향해 수평으로 겨눠짐.",
    "built_space": "중고 매장 내부. 작업대 모서리와 배경의 전자제품 선반, 간판 배치가 레퍼런스와 일치함.",
    "entities": "마츠다(레퍼런스 복장), 수하 1(남색 전투복, 붉은 완장), 소음기가 달린 권총이 모두 올바르게 묘사됨.",
    "hard_violations": [],
    "physics": "오른손이 권총을 쥐고 지지하며, 인물들은 바닥에 서서 자세를 유지함."
   },
   {
    "label": "B",
    "direction": "수하의 총구가 마츠다의 얼굴 부위를 향해 약간 위쪽으로 겨눠짐.",
    "built_space": "중고 매장 내부. 작업대, 배경 선반 및 한글 간판 요소가 레퍼런스 공간과 정확히 일치함.",
    "entities": "마츠다(60대 남성, 레퍼런스 복장), 수하 1(전투복, 붉은 완장), 소음기가 달린 권총이 정확히 식별됨.",
    "hard_violations": [],
    "physics": "오른손이 자연스럽게 권총을 쥐고 있으며, 두 인물 모두 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "총구가 마츠다의 얼굴이 아닌 가슴을 향하고 있으며, 마츠다의 얼굴이 상단으로 크롭되지 않고 프레임에 모두 포함되어 샷 구도 지시를 온전히 따르지 못했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "마츠다의 얼굴이 프레임 안에 온전히 포함된 점은 동일한 오류이나, 총구가 지시된 대로 마츠다의 얼굴을 향해 위로 겨눠져 있어 방향성과 구도 의도를 더 잘 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하의 총구가 마츠다의 가슴과 어깨 부위를 향해 수평으로 겨눠짐.",
        "built_space": "중고 매장 내부. 작업대 모서리와 배경의 전자제품 선반, 간판 배치가 레퍼런스와 일치함.",
        "entities": "마츠다(레퍼런스 복장), 수하 1(남색 전투복, 붉은 완장), 소음기가 달린 권총이 모두 올바르게 묘사됨.",
        "hard_violations": [],
        "physics": "오른손이 권총을 쥐고 지지하며, 인물들은 바닥에 서서 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "수하의 총구가 마츠다의 얼굴 부위를 향해 약간 위쪽으로 겨눠짐.",
        "built_space": "중고 매장 내부. 작업대, 배경 선반 및 한글 간판 요소가 레퍼런스 공간과 정확히 일치함.",
        "entities": "마츠다(60대 남성, 레퍼런스 복장), 수하 1(전투복, 붉은 완장), 소음기가 달린 권총이 정확히 식별됨.",
        "hard_violations": [],
        "physics": "오른손이 자연스럽게 권총을 쥐고 있으며, 두 인물 모두 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "총구가 마츠다의 얼굴이 아닌 가슴을 향하고 있으며, 마츠다의 얼굴이 상단으로 크롭되지 않고 프레임에 모두 포함되어 샷 구도 지시를 온전히 따르지 못했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "마츠다의 얼굴이 프레임 안에 온전히 포함된 점은 동일한 오류이나, 총구가 지시된 대로 마츠다의 얼굴을 향해 위로 겨눠져 있어 방향성과 구도 의도를 더 잘 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수하의 총구가 마츠다의 가슴과 어깨 부위를 향해 수평으로 겨눠짐.",
        "built_space": "중고 매장 내부. 작업대 모서리와 배경의 전자제품 선반, 간판 배치가 레퍼런스와 일치함.",
        "entities": "마츠다(레퍼런스 복장), 수하 1(남색 전투복, 붉은 완장), 소음기가 달린 권총이 모두 올바르게 묘사됨.",
        "hard_violations": [],
        "physics": "오른손이 권총을 쥐고 지지하며, 인물들은 바닥에 서서 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "수하의 총구가 마츠다의 얼굴 부위를 향해 약간 위쪽으로 겨눠짐.",
        "built_space": "중고 매장 내부. 작업대, 배경 선반 및 한글 간판 요소가 레퍼런스 공간과 정확히 일치함.",
        "entities": "마츠다(60대 남성, 레퍼런스 복장), 수하 1(전투복, 붉은 완장), 소음기가 달린 권총이 정확히 식별됨.",
        "hard_violations": [],
        "physics": "오른손이 자연스럽게 권총을 쥐고 있으며, 두 인물 모두 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "마츠다를 향한 소음 권총과 장소 연속성은 맞지만, 얼굴과 상체를 크게 보여주는 어깨너머 구도라 지시된 팔 근접 구도에서 멀다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "팔과 권총에 더 가까이 접근하고 왼쪽 마츠다를 부분적으로 잘라 A보다 충실하지만, 얼굴을 상단 밖으로 제외한 낮은 팔 근접 구도까지는 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "소음기 총구는 렌즈가 아니라 화면 왼쪽 마츠다의 턱·목 높이를 향한다. 마츠다는 오른쪽 수하의 얼굴 쪽을 보고 있다. 수하의 머리도 마츠다 쪽으로 돌아가 있으나 눈은 보이지 않아 정확한 시선은 확인할 수 없다.",
        "built_space": "하단 중앙부터 오른쪽에 수리 작업대 하나, 그 위에 외장이 열린 전자기기 하나와 공구·천·부품 상자가 보인다. 뒤에는 전자기기가 쌓인 금속 선반들이 있고 오른쪽에 주광이 들어오는 출입구 하나, 위쪽에 기존 문구의 간판 하나가 있다. 마츠다는 작업대 왼쪽 앞, 수하는 오른쪽 앞에 있어 이전 장면의 공간 관계가 유지된다. 다만 작업대와 마츠다의 상체를 넓게 보여주며, 팔꿈치 높이에서 살짝 올려다보는 근접 시점으로 읽히지는 않는다.",
        "entities": "인물은 두 명뿐이다. 마츠다는 고령 동아시아계 남성으로 이전 장면의 낡은 모자, 체크 셔츠, 청조끼와 얼굴 특징을 유지한다. 수하는 짧은 검은 머리의 성인 남성으로 어두운 전투복과 붉은 완장을 착용하며, 보이는 손도 성인 남성의 손으로 자연스럽다. 국적은 외모만으로 확인할 수 없다. 총은 소총이 아니라 소음기가 장착된 권총으로, 최우선 한국어 샷 텍스트에는 맞지만 영문 세부 지시와는 다르다. 총의 화면 점유 면적은 5분의 2 미만이다. 마츠다의 얼굴 전체가 선명하게 보여 흐릿한 부분 몸통과 상단 밖 얼굴이라는 지시는 충족하지 않는다.",
        "hard_violations": [],
        "physics": "권총 손잡이는 수하의 손에 잡혀 있고 손목과 전완은 오른쪽 어깨로 자연스럽게 이어진다. 한 손으로 정지 조준할 수 있는 자세이며 공중에 지지 없이 떠 있는 물체는 없다. 두 사람은 곧게 선 상체로 보이고 발은 화면 밖이다. 작업대 위 기기와 공구는 상판에 놓여 있다. 발사광이나 반동은 없어 발사 전 순간에 맞는다."
       },
       {
        "label": "B",
        "direction": "권총과 소음기는 비스듬한 측면으로 보이며 총구는 왼쪽 마츠다의 목과 윗가슴 방향을 향한다. 렌즈를 향한 조준은 아니다. 마츠다의 눈은 오른쪽 수하를 향하며 수하의 머리도 마츠다 쪽을 향하지만, 수하의 눈은 잘려 정확한 시선은 확인할 수 없다.",
        "built_space": "하단에 수리 작업대 하나가 있고, 오른쪽에는 외장이 열린 전자기기 하나, 중앙에는 천·부품 상자·작은 기기, 앞쪽에는 공구가 놓여 있다. 배경의 금속 선반들과 낡은 전자기기, 오른쪽 출입구 하나, 위쪽 기존 간판 하나가 이전 장소와 연결된다. 마츠다는 왼쪽, 수하는 오른쪽에서 작업대 앞 공간을 사이에 두고 마주한다. A보다 가까운 시점에서 팔 뒤로 작업대 가장자리가 남지만, 명확하게 팔꿈치 높이에서 올려다보는 각도는 아니다.",
        "entities": "추가 인물 없이 마츠다와 수하의 일부만 보인다. 마츠다의 고령 남성 얼굴, 모자, 체크 셔츠와 청조끼가 이전 장면과 맞고, 수하의 보이는 턱·목·손 및 어두운 소매와 붉은 완장도 인물 설정과 양립한다. 잘린 얼굴만으로 수하의 정확한 신원이나 국적을 확정할 수는 없다. 소음 권총은 최우선 한국어 샷 텍스트와 일치하며 소총이라는 영문 항목과는 다르다. 총은 화면 면적의 5분의 2보다 작고 전완이 오른쪽 아래에서 중앙 왼쪽으로 뻗는다. 다만 마츠다의 얼굴 대부분이 여전히 화면에 남고 몸통도 충분히 흐려지지 않았다.",
        "hard_violations": [],
        "physics": "수하의 손가락이 권총 손잡이를 감싸고 손목·전완·상완이 연결되어 총을 지탱한다. 정지 상태의 한 손 조준으로 가능한 자세이며 지지 없는 부유는 없다. 두 사람의 하체는 화면 밖이지만 보이는 상체에 붕괴나 비현실적 기울기는 없다. 수리 기기와 공구는 작업대가 받치고 있다. 총구 화염이나 탄피, 반동 없이 조준만 유지한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "마츠다를 향한 소음 권총과 장소 연속성은 맞지만, 얼굴과 상체를 크게 보여주는 어깨너머 구도라 지시된 팔 근접 구도에서 멀다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "팔과 권총에 더 가까이 접근하고 왼쪽 마츠다를 부분적으로 잘라 A보다 충실하지만, 얼굴을 상단 밖으로 제외한 낮은 팔 근접 구도까지는 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "소음기 총구는 렌즈가 아니라 화면 왼쪽 마츠다의 턱·목 높이를 향한다. 마츠다는 오른쪽 수하의 얼굴 쪽을 보고 있다. 수하의 머리도 마츠다 쪽으로 돌아가 있으나 눈은 보이지 않아 정확한 시선은 확인할 수 없다.",
        "built_space": "하단 중앙부터 오른쪽에 수리 작업대 하나, 그 위에 외장이 열린 전자기기 하나와 공구·천·부품 상자가 보인다. 뒤에는 전자기기가 쌓인 금속 선반들이 있고 오른쪽에 주광이 들어오는 출입구 하나, 위쪽에 기존 문구의 간판 하나가 있다. 마츠다는 작업대 왼쪽 앞, 수하는 오른쪽 앞에 있어 이전 장면의 공간 관계가 유지된다. 다만 작업대와 마츠다의 상체를 넓게 보여주며, 팔꿈치 높이에서 살짝 올려다보는 근접 시점으로 읽히지는 않는다.",
        "entities": "인물은 두 명뿐이다. 마츠다는 고령 동아시아계 남성으로 이전 장면의 낡은 모자, 체크 셔츠, 청조끼와 얼굴 특징을 유지한다. 수하는 짧은 검은 머리의 성인 남성으로 어두운 전투복과 붉은 완장을 착용하며, 보이는 손도 성인 남성의 손으로 자연스럽다. 국적은 외모만으로 확인할 수 없다. 총은 소총이 아니라 소음기가 장착된 권총으로, 최우선 한국어 샷 텍스트에는 맞지만 영문 세부 지시와는 다르다. 총의 화면 점유 면적은 5분의 2 미만이다. 마츠다의 얼굴 전체가 선명하게 보여 흐릿한 부분 몸통과 상단 밖 얼굴이라는 지시는 충족하지 않는다.",
        "hard_violations": [],
        "physics": "권총 손잡이는 수하의 손에 잡혀 있고 손목과 전완은 오른쪽 어깨로 자연스럽게 이어진다. 한 손으로 정지 조준할 수 있는 자세이며 공중에 지지 없이 떠 있는 물체는 없다. 두 사람은 곧게 선 상체로 보이고 발은 화면 밖이다. 작업대 위 기기와 공구는 상판에 놓여 있다. 발사광이나 반동은 없어 발사 전 순간에 맞는다."
       },
       {
        "label": "A",
        "direction": "권총과 소음기는 비스듬한 측면으로 보이며 총구는 왼쪽 마츠다의 목과 윗가슴 방향을 향한다. 렌즈를 향한 조준은 아니다. 마츠다의 눈은 오른쪽 수하를 향하며 수하의 머리도 마츠다 쪽을 향하지만, 수하의 눈은 잘려 정확한 시선은 확인할 수 없다.",
        "built_space": "하단에 수리 작업대 하나가 있고, 오른쪽에는 외장이 열린 전자기기 하나, 중앙에는 천·부품 상자·작은 기기, 앞쪽에는 공구가 놓여 있다. 배경의 금속 선반들과 낡은 전자기기, 오른쪽 출입구 하나, 위쪽 기존 간판 하나가 이전 장소와 연결된다. 마츠다는 왼쪽, 수하는 오른쪽에서 작업대 앞 공간을 사이에 두고 마주한다. A보다 가까운 시점에서 팔 뒤로 작업대 가장자리가 남지만, 명확하게 팔꿈치 높이에서 올려다보는 각도는 아니다.",
        "entities": "추가 인물 없이 마츠다와 수하의 일부만 보인다. 마츠다의 고령 남성 얼굴, 모자, 체크 셔츠와 청조끼가 이전 장면과 맞고, 수하의 보이는 턱·목·손 및 어두운 소매와 붉은 완장도 인물 설정과 양립한다. 잘린 얼굴만으로 수하의 정확한 신원이나 국적을 확정할 수는 없다. 소음 권총은 최우선 한국어 샷 텍스트와 일치하며 소총이라는 영문 항목과는 다르다. 총은 화면 면적의 5분의 2보다 작고 전완이 오른쪽 아래에서 중앙 왼쪽으로 뻗는다. 다만 마츠다의 얼굴 대부분이 여전히 화면에 남고 몸통도 충분히 흐려지지 않았다.",
        "hard_violations": [],
        "physics": "수하의 손가락이 권총 손잡이를 감싸고 손목·전완·상완이 연결되어 총을 지탱한다. 정지 상태의 한 손 조준으로 가능한 자세이며 지지 없는 부유는 없다. 두 사람의 하체는 화면 밖이지만 보이는 상체에 붕괴나 비현실적 기울기는 없다. 수리 기기와 공구는 작업대가 받치고 있다. 총구 화염이나 탄피, 반동 없이 조준만 유지한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "총구가 마츠다의 얼굴이 아닌 가슴을 향하고 있으며, 마츠다의 얼굴이 상단으로 크롭되지 않고 프레임에 모두 포함되어 샷 구도 지시를 온전히 따르지 못했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "마츠다의 얼굴이 프레임 안에 온전히 포함된 점은 동일한 오류이나, 총구가 지시된 대로 마츠다의 얼굴을 향해 위로 겨눠져 있어 방향성과 구도 의도를 더 잘 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S17sh5_sel.png",
    "asset_id": "24bc04d9-10a3-483b-80a2-75e08b6da96c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:728720>",
    "asset_id": "445f3c52-b87c-457e-ae88-317bd9f1202e",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1351055>",
    "asset_id": "48373447-e257-4c63-b5b6-c48339a86042",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08d2-2c53-7660-9a70-95fdceaf9782",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh5"
  },
  "staged_characters_added": [
   "C14"
  ]
 },
 "S17sh12::signage": {
  "fp": "4cb250a1ec0e2ddf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh12": {
  "input_fingerprint": "26070054b278f1c3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 마츠다의 시신을 등진 채 무전기를 입가에 댄 '박철진의 수하 1'의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the secondhand shop, with the entrance admitting daylight behind the shopfront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the outer-side track at 박철진의 수하 1's shoulder height, perpendicular to his new facing direction, with a slight downward tilt connecting his right-foreground profile to the desk at left rear. He holds the radio at his mouth and concentrates on the report while looking away into the off-screen shop interior; behind his turned back, 마츠다 lies face-down across the desk with no readable eye contact. Make the henchman's turn away from the body the dominant positional change, retaining the established subdued exposure and avoiding any view of the firing.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (Matsuda's collapsed body rests across it) — Its top is visible obliquely behind the henchman on the left; used as Anchors the body behind the reporting figure within one continuous shop space; Radio (Held at the henchman's mouth during his report) — Seen from the side beside his lower face; used as A small foreground action detail, clearly separated from the body behind him.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained interior illumination gives the report and the motionless body an unsentimental, controlled clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's worktable, accumulated appliances, and daylight appearance from the reference. Exclude the earlier bargaining moment and the departing customer's robot.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Matsuda is slumped forward onto the desk after being shot, with his upper body collapsed against the desktop. The precise turn of his head, the placement of his arms and legs, and whether his lower body remains seated are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television repair station and piles of shop appliances remain in place. 마츠다: He is slumped forward onto the desk after being shot. 박철진의 수하 1: He remains in his militia uniform and armband, holding a radio near his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리); 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 마츠다의 시신을 등진 채 무전기를 입가에 댄 '박철진의 수하 1'의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the secondhand shop, with the entrance admitting daylight behind the shopfront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the outer-side track at 박철진의 수하 1's shoulder height, perpendicular to his new facing direction, with a slight downward tilt connecting his right-foreground profile to the desk at left rear. He holds the radio at his mouth and concentrates on the report while looking away into the off-screen shop interior; behind his turned back, 마츠다 lies face-down across the desk with no readable eye contact. Make the henchman's turn away from the body the dominant positional change, retaining the established subdued exposure and avoiding any view of the firing.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (Matsuda's collapsed body rests across it) — Its top is visible obliquely behind the henchman on the left; used as Anchors the body behind the reporting figure within one continuous shop space; Radio (Held at the henchman's mouth during his report) — Seen from the side beside his lower face; used as A small foreground action detail, clearly separated from the body behind him.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained interior illumination gives the report and the motionless body an unsentimental, controlled clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's worktable, accumulated appliances, and daylight appearance from the reference. Exclude the earlier bargaining moment and the departing customer's robot.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Matsuda is slumped forward onto the desk after being shot, with his upper body collapsed against the desktop. The precise turn of his head, the placement of his arms and legs, and whether his lower body remains seated are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television repair station and piles of shop appliances remain in place. 마츠다: He is slumped forward onto the desk after being shot. 박철진의 수하 1: He remains in his militia uniform and armband, holding a radio near his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리); 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 마츠다의 시신을 등진 채 무전기를 입가에 댄 '박철진의 수하 1'의 측면 구도.\n\nLOCATION (lock): Beside the repair desk inside the secondhand shop, with the entrance admitting daylight behind the shopfront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the outer-side track at 박철진의 수하 1's shoulder height, perpendicular to his new facing direction, with a slight downward tilt connecting his right-foreground profile to the desk at left rear. He holds the radio at his mouth and concentrates on the report while looking away into the off-screen shop interior; behind his turned back, 마츠다 lies face-down across the desk with no readable eye contact. Make the henchman's turn away from the body the dominant positional change, retaining the established subdued exposure and avoiding any view of the firing.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Desk (Matsuda's collapsed body rests across it) — Its top is visible obliquely behind the henchman on the left; used as Anchors the body behind the reporting figure within one continuous shop space; Radio (Held at the henchman's mouth during his report) — Seen from the side beside his lower face; used as A small foreground action detail, clearly separated from the body behind him.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained interior illumination gives the report and the motionless body an unsentimental, controlled clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shop's worktable, accumulated appliances, and daylight appearance from the reference. Exclude the earlier bargaining moment and the departing customer's robot.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Matsuda is slumped forward onto the desk after being shot, with his upper body collapsed against the desktop. The precise turn of his head, the placement of his arms and legs, and whether his lower body remains seated are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The television repair station and piles of shop appliances remain in place. 마츠다: He is slumped forward onto the desk after being shot. 박철진의 수하 1: He remains in his militia uniform and armband, holding a radio near his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진의 수하 1 right now, so 박철진의 수하 1's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진의 수하 1: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마츠다 (일본인 남성, 60대의 얼굴, 짧은 머리); 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수하 1의 시선은 화면 우측 오프스크린을 향하고 있으며, 무전기는 입가에 정확히 위치함.",
    "built_space": "좌측 후경에 작업대와 전자기기, '중고 매입 판매' 간판이 위치함. 우측 배경으로 외부의 밝은 채광이 들어오며 공간의 배치가 지시문과 일치함.",
    "entities": "수하 1은 남색 전투복과 붉은 완장(캐릭터 레퍼런스 이미지와 일치하는 형태)을 착용함. 마츠다는 체크 셔츠와 데님 조끼를 입고 엎어져 있으며, 야구 모자는 책상 위에 놓임. 무전기는 프롭 레퍼런스와 형태가 완벽히 일치함.",
    "hard_violations": [],
    "physics": "마츠다의 시신은 작업대에 온전히 무게를 기댄 채 중력에 순응하여 늘어져 있음. 수하 1의 손은 허공에 뜨거나 왜곡되지 않고 무전기를 안정적으로 지탱함."
   },
   {
    "label": "B",
    "direction": "수하 1의 시선은 우측 오프스크린을 향하며, 무전기는 입가 근처에 위치함.",
    "built_space": "좌측에 작업대와 간판이 있으며, 배경 밖으로 채광이 들어옴. 전반적인 공간 배치는 기준을 충족함.",
    "entities": "수하 1은 전투복과 로고가 포함된 완장을 착용함. 그러나 마츠다가 레퍼런스와 전혀 다른 형태의 챙 모자(뉴스보이 캡)를 쓰고 있음. 무전기는 다이얼과 안테나 형태가 레퍼런스와 다르게 일그러짐.",
    "hard_violations": [
     "[gemini-pro] 마츠다의 모자가 레퍼런스에 없는 완전히 다른 형태(뉴스보이 캡)로 발명됨",
     "[gemini-pro] 무전기를 쥔 손가락의 해부학적 오류 및 무전기 형태 왜곡"
    ],
    "physics": "마츠다의 몸은 책상에 엎어져 지탱되고 있으나, 수하 1이 무전기를 쥔 손가락이 부자연스럽게 뭉개져 기기를 제대로 쥐고 있는 것으로 보이지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 측면 구도와 엎어진 시신의 상태를 정확히 묘사하였으며, 특히 무전기 프롭의 세부 디테일과 쥐고 있는 손의 형태가 매우 뛰어남."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이전 컷의 완장 로고는 유지했으나, 마츠다가 엉뚱한 종류의 모자를 쓰고 있으며 무전기를 쥔 손과 프롭 자체가 심하게 왜곡됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하 1의 시선은 화면 우측 오프스크린을 향하고 있으며, 무전기는 입가에 정확히 위치함.",
        "built_space": "좌측 후경에 작업대와 전자기기, '중고 매입 판매' 간판이 위치함. 우측 배경으로 외부의 밝은 채광이 들어오며 공간의 배치가 지시문과 일치함.",
        "entities": "수하 1은 남색 전투복과 붉은 완장(캐릭터 레퍼런스 이미지와 일치하는 형태)을 착용함. 마츠다는 체크 셔츠와 데님 조끼를 입고 엎어져 있으며, 야구 모자는 책상 위에 놓임. 무전기는 프롭 레퍼런스와 형태가 완벽히 일치함.",
        "hard_violations": [],
        "physics": "마츠다의 시신은 작업대에 온전히 무게를 기댄 채 중력에 순응하여 늘어져 있음. 수하 1의 손은 허공에 뜨거나 왜곡되지 않고 무전기를 안정적으로 지탱함."
       },
       {
        "label": "B",
        "direction": "수하 1의 시선은 우측 오프스크린을 향하며, 무전기는 입가 근처에 위치함.",
        "built_space": "좌측에 작업대와 간판이 있으며, 배경 밖으로 채광이 들어옴. 전반적인 공간 배치는 기준을 충족함.",
        "entities": "수하 1은 전투복과 로고가 포함된 완장을 착용함. 그러나 마츠다가 레퍼런스와 전혀 다른 형태의 챙 모자(뉴스보이 캡)를 쓰고 있음. 무전기는 다이얼과 안테나 형태가 레퍼런스와 다르게 일그러짐.",
        "hard_violations": [
         "마츠다의 모자가 레퍼런스에 없는 완전히 다른 형태(뉴스보이 캡)로 발명됨",
         "무전기를 쥔 손가락의 해부학적 오류 및 무전기 형태 왜곡"
        ],
        "physics": "마츠다의 몸은 책상에 엎어져 지탱되고 있으나, 수하 1이 무전기를 쥔 손가락이 부자연스럽게 뭉개져 기기를 제대로 쥐고 있는 것으로 보이지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 측면 구도와 엎어진 시신의 상태를 정확히 묘사하였으며, 특히 무전기 프롭의 세부 디테일과 쥐고 있는 손의 형태가 매우 뛰어남."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이전 컷의 완장 로고는 유지했으나, 마츠다가 엉뚱한 종류의 모자를 쓰고 있으며 무전기를 쥔 손과 프롭 자체가 심하게 왜곡됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수하 1의 시선은 화면 우측 오프스크린을 향하고 있으며, 무전기는 입가에 정확히 위치함.",
        "built_space": "좌측 후경에 작업대와 전자기기, '중고 매입 판매' 간판이 위치함. 우측 배경으로 외부의 밝은 채광이 들어오며 공간의 배치가 지시문과 일치함.",
        "entities": "수하 1은 남색 전투복과 붉은 완장(캐릭터 레퍼런스 이미지와 일치하는 형태)을 착용함. 마츠다는 체크 셔츠와 데님 조끼를 입고 엎어져 있으며, 야구 모자는 책상 위에 놓임. 무전기는 프롭 레퍼런스와 형태가 완벽히 일치함.",
        "hard_violations": [],
        "physics": "마츠다의 시신은 작업대에 온전히 무게를 기댄 채 중력에 순응하여 늘어져 있음. 수하 1의 손은 허공에 뜨거나 왜곡되지 않고 무전기를 안정적으로 지탱함."
       },
       {
        "label": "B",
        "direction": "수하 1의 시선은 우측 오프스크린을 향하며, 무전기는 입가 근처에 위치함.",
        "built_space": "좌측에 작업대와 간판이 있으며, 배경 밖으로 채광이 들어옴. 전반적인 공간 배치는 기준을 충족함.",
        "entities": "수하 1은 전투복과 로고가 포함된 완장을 착용함. 그러나 마츠다가 레퍼런스와 전혀 다른 형태의 챙 모자(뉴스보이 캡)를 쓰고 있음. 무전기는 다이얼과 안테나 형태가 레퍼런스와 다르게 일그러짐.",
        "hard_violations": [
         "마츠다의 모자가 레퍼런스에 없는 완전히 다른 형태(뉴스보이 캡)로 발명됨",
         "무전기를 쥔 손가락의 해부학적 오류 및 무전기 형태 왜곡"
        ],
        "physics": "마츠다의 몸은 책상에 엎어져 지탱되고 있으나, 수하 1이 무전기를 쥔 손가락이 부자연스럽게 뭉개져 기기를 제대로 쥐고 있는 것으로 보이지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 전경의 수하가 시신에 등을 돌린 측면과 입가의 무전기가 더 정확하며, 이전 장면의 문양 있는 붉은 완장도 유지한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "보고 행동과 왼쪽 뒤 책상에 엎어진 시신은 맞지만, 수하의 가슴이 더 드러나는 사선 구도로 돌아서기의 강조가 약하고 완장 문양이 보이지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하는 오른쪽 화면 밖을 응시하며 왼쪽 뒤 마츠다에게 등과 뒤쪽 어깨를 향한다. 얼굴은 거의 옆모습이고 무전기는 입 바로 앞에 있다. 마츠다의 얼굴은 책상 쪽으로 내려가 눈맞춤이 없다. 무전기 안테나는 위로 향하고, 기기의 측면과 후면 위주로 보여 사용 방향에 뚜렷한 모순은 없다. 발사나 조준은 보이지 않는다.",
        "built_space": "왼쪽에 연속된 수리 작업대 한 곳, 오른쪽 뒤에 밝은 출입구 한 곳, 뒤 벽에 기존 문구의 간판 한 개가 보인다. 왼쪽과 중앙 선반에는 여러 폐가전과 노출 회로 장치가 쌓여 있다. 수하는 작업대 오른쪽 앞에 있고 마츠다는 그 뒤 왼쪽 작업대에 엎어져 있어 요구된 공간 관계가 성립한다. 비스듬히 보이는 상판, 낡은 재료와 낮의 유입광은 참고 장소와 부합하지만 개별 가전의 배열은 달라졌다. 반사에 의존하는 배치는 없다.",
        "entities": "보이는 사람은 성인 동아시아계 남성 수하와 회색 머리의 고령 동아시아계 남성 마츠다 두 명뿐이다. 국적은 외형만으로 확인할 수 없다. 수하의 얼굴과 체격은 인물 참고와 대체로 맞고 검은 비니, 어두운 남색 상의, 붉은 완장을 착용한다. 완장의 흰 원형 문양은 이전 장면과 가깝다. 마츠다는 이전 장면의 모자, 체크 셔츠, 낡은 청조끼를 유지하며 얼굴은 가려져 세부 동일성은 확인하기 어렵다. 무전기는 검은 휴대형 장비이나 참고 소품의 은색 전면 격자와 화면은 이 각도에서 확인되지 않고 안테나 비례도 다르다. 작업대에는 수리 장비와 공구가 있으며 로봇이나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "수하의 손이 무전기 몸체를 감싸 쥐고 손목과 소매가 자연스럽게 이어진다. 마츠다의 머리와 앞으로 접힌 상체는 작업대에 기대고, 가까운 팔은 작업대 가장자리와 아래 장비 위에 걸친 뒤 손목 쪽으로 처진다. 시신이 손을 능동적으로 들고 있는 모습은 없다. 하체 지지는 대부분 가려져 있으나 보이는 부분에서 공중 부양이나 불가능한 관절은 발견되지 않는다. 공구와 장비는 상판 또는 선반 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "수하는 오른쪽 화면 밖을 보며 입 앞에 무전기를 들고 있고 마츠다는 왼쪽 뒤에 남아 있다. 다만 얼굴과 가슴이 카메라 쪽으로 더 열려 있어 정확한 측면보다는 사선 측면이다. 마츠다의 머리는 상판을 향해 숙여져 눈맞춤이 없다. 무전기 안테나는 위로 향하며 기기의 측면이 주로 보인다. 무기 조준이나 발사는 없다.",
        "built_space": "왼쪽 수리 작업대 한 곳, 중앙 오른쪽의 밝은 출입구 한 곳, 뒤 벽의 기존 간판 한 개가 보인다. 왼쪽과 중앙의 다층 선반에는 모니터와 분해된 가전이 쌓여 있다. 수하는 작업대 오른쪽 전경에 서 있고 마츠다는 왼쪽 후경 상판에 상체를 걸친다. 한 공간 안의 앞뒤 관계와 낮의 채광은 맞지만 참고 사진과 개별 가전 배치는 다르다. 작업대 상판은 비스듬히 보이며 중복된 주요 고정 설비나 불가능한 반사는 없다.",
        "entities": "성인 동아시아계 남성 수하와 고령 동아시아계 남성 마츠다 두 명만 보인다. 수하의 얼굴, 검은 비니와 어두운 남색 작업복형 전투복은 인물 참고에 가깝다. 붉은 완장은 있으나 이전 장면의 흰 문양 대신 매듭 쪽이 보인다. 마츠다는 회색 짧은 머리, 체크 셔츠와 낡은 청조끼를 유지하며 모자는 옆 상판에 놓여 있다. 얼굴이 숨겨져 정확한 얼굴 일치는 판단할 수 없다. 무전기는 검은 휴대형이지만 참고 소품보다 안테나가 가늘고 길어 보이며 은색 전면부의 일치는 확인되지 않는다. 수리 장비와 공구는 있고 추가 인물이나 로봇은 없다.",
        "hard_violations": [],
        "physics": "수하의 손가락이 무전기를 확실히 잡고 있어 기기는 손으로 지지된다. 마츠다의 머리와 상체는 상판에 기대고 가까운 팔은 책상 가장자리 아래로 축 처진다. 팔을 공중에 들고 버티는 자세는 아니다. 골반 일부는 보이지만 좌석과 발은 화면 밖이어서 하체 지지 방식은 확정할 수 없으며, 상판에 기댄 붕괴 자세 자체는 가능하다. 모자와 공구는 책상 위에 놓여 있고 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 전경의 수하가 시신에 등을 돌린 측면과 입가의 무전기가 더 정확하며, 이전 장면의 문양 있는 붉은 완장도 유지한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "보고 행동과 왼쪽 뒤 책상에 엎어진 시신은 맞지만, 수하의 가슴이 더 드러나는 사선 구도로 돌아서기의 강조가 약하고 완장 문양이 보이지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "수하는 오른쪽 화면 밖을 응시하며 왼쪽 뒤 마츠다에게 등과 뒤쪽 어깨를 향한다. 얼굴은 거의 옆모습이고 무전기는 입 바로 앞에 있다. 마츠다의 얼굴은 책상 쪽으로 내려가 눈맞춤이 없다. 무전기 안테나는 위로 향하고, 기기의 측면과 후면 위주로 보여 사용 방향에 뚜렷한 모순은 없다. 발사나 조준은 보이지 않는다.",
        "built_space": "왼쪽에 연속된 수리 작업대 한 곳, 오른쪽 뒤에 밝은 출입구 한 곳, 뒤 벽에 기존 문구의 간판 한 개가 보인다. 왼쪽과 중앙 선반에는 여러 폐가전과 노출 회로 장치가 쌓여 있다. 수하는 작업대 오른쪽 앞에 있고 마츠다는 그 뒤 왼쪽 작업대에 엎어져 있어 요구된 공간 관계가 성립한다. 비스듬히 보이는 상판, 낡은 재료와 낮의 유입광은 참고 장소와 부합하지만 개별 가전의 배열은 달라졌다. 반사에 의존하는 배치는 없다.",
        "entities": "보이는 사람은 성인 동아시아계 남성 수하와 회색 머리의 고령 동아시아계 남성 마츠다 두 명뿐이다. 국적은 외형만으로 확인할 수 없다. 수하의 얼굴과 체격은 인물 참고와 대체로 맞고 검은 비니, 어두운 남색 상의, 붉은 완장을 착용한다. 완장의 흰 원형 문양은 이전 장면과 가깝다. 마츠다는 이전 장면의 모자, 체크 셔츠, 낡은 청조끼를 유지하며 얼굴은 가려져 세부 동일성은 확인하기 어렵다. 무전기는 검은 휴대형 장비이나 참고 소품의 은색 전면 격자와 화면은 이 각도에서 확인되지 않고 안테나 비례도 다르다. 작업대에는 수리 장비와 공구가 있으며 로봇이나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "수하의 손이 무전기 몸체를 감싸 쥐고 손목과 소매가 자연스럽게 이어진다. 마츠다의 머리와 앞으로 접힌 상체는 작업대에 기대고, 가까운 팔은 작업대 가장자리와 아래 장비 위에 걸친 뒤 손목 쪽으로 처진다. 시신이 손을 능동적으로 들고 있는 모습은 없다. 하체 지지는 대부분 가려져 있으나 보이는 부분에서 공중 부양이나 불가능한 관절은 발견되지 않는다. 공구와 장비는 상판 또는 선반 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "수하는 오른쪽 화면 밖을 보며 입 앞에 무전기를 들고 있고 마츠다는 왼쪽 뒤에 남아 있다. 다만 얼굴과 가슴이 카메라 쪽으로 더 열려 있어 정확한 측면보다는 사선 측면이다. 마츠다의 머리는 상판을 향해 숙여져 눈맞춤이 없다. 무전기 안테나는 위로 향하며 기기의 측면이 주로 보인다. 무기 조준이나 발사는 없다.",
        "built_space": "왼쪽 수리 작업대 한 곳, 중앙 오른쪽의 밝은 출입구 한 곳, 뒤 벽의 기존 간판 한 개가 보인다. 왼쪽과 중앙의 다층 선반에는 모니터와 분해된 가전이 쌓여 있다. 수하는 작업대 오른쪽 전경에 서 있고 마츠다는 왼쪽 후경 상판에 상체를 걸친다. 한 공간 안의 앞뒤 관계와 낮의 채광은 맞지만 참고 사진과 개별 가전 배치는 다르다. 작업대 상판은 비스듬히 보이며 중복된 주요 고정 설비나 불가능한 반사는 없다.",
        "entities": "성인 동아시아계 남성 수하와 고령 동아시아계 남성 마츠다 두 명만 보인다. 수하의 얼굴, 검은 비니와 어두운 남색 작업복형 전투복은 인물 참고에 가깝다. 붉은 완장은 있으나 이전 장면의 흰 문양 대신 매듭 쪽이 보인다. 마츠다는 회색 짧은 머리, 체크 셔츠와 낡은 청조끼를 유지하며 모자는 옆 상판에 놓여 있다. 얼굴이 숨겨져 정확한 얼굴 일치는 판단할 수 없다. 무전기는 검은 휴대형이지만 참고 소품보다 안테나가 가늘고 길어 보이며 은색 전면부의 일치는 확인되지 않는다. 수리 장비와 공구는 있고 추가 인물이나 로봇은 없다.",
        "hard_violations": [],
        "physics": "수하의 손가락이 무전기를 확실히 잡고 있어 기기는 손으로 지지된다. 마츠다의 머리와 상체는 상판에 기대고 가까운 팔은 책상 가장자리 아래로 축 처진다. 팔을 공중에 들고 버티는 자세는 아니다. 골반 일부는 보이지만 좌석과 발은 화면 밖이어서 하체 지지 방식은 확정할 수 없으며, 상판에 기댄 붕괴 자세 자체는 가능하다. 모자와 공구는 책상 위에 놓여 있고 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 마츠다의 모자가 레퍼런스에 없는 완전히 다른 형태(뉴스보이 캡)로 발명됨",
     "[gemini-pro] 무전기를 쥔 손가락의 해부학적 오류 및 무전기 형태 왜곡"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지시된 측면 구도와 엎어진 시신의 상태를 정확히 묘사하였으며, 특히 무전기 프롭의 세부 디테일과 쥐고 있는 손의 형태가 매우 뛰어남."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "이전 컷의 완장 로고는 유지했으나, 마츠다가 엉뚱한 종류의 모자를 쓰고 있으며 무전기를 쥔 손과 프롭 자체가 심하게 왜곡됨.  ★위반: [gemini-pro] 마츠다의 모자가 레퍼런스에 없는 완전히 다른 형태(뉴스보이 캡)로 발명됨 / [gemini-pro] 무전기를 쥔 손가락의 해부학적 오류 및 무전기 형태 왜곡"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 마츠다, 박철진의 수하 1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S17sh5_sel.png",
    "asset_id": "24bc04d9-10a3-483b-80a2-75e08b6da96c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 마츠다: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1351458>",
    "asset_id": "57dec84c-f070-40df-9f3b-d8604bf5d6af",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:728720>",
    "asset_id": "445f3c52-b87c-457e-ae88-317bd9f1202e",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 무전기: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1027091>",
    "asset_id": "68be9f54-1b7a-4963-ad00-d3aea327c7a6",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08d8-961d-7e26-b88d-a1ca7b809cb3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh5"
  }
 },
 "S18sh2::signage": {
  "fp": "8353b982feed854b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::6f4cd3dd9a9eed07": {
  "subjects": [],
  "subject_text": "수도권 도심 민병대 사무실·박철진의 사무실\n책상과 맞은편 빈 벽이 있는 사무실. 책상 위에는 펼쳐진 지도와 스피커폰이 놓여 있고, 벽에는 다트 표적이 마련되어 있다.",
  "identity": "canonical",
  "scope_id": "L30",
  "scope_role": "location_interior",
  "scope_sha": "9a33cdf17fec567a"
 },
 "S18sh2::bgfirst_bg": {
  "input_fingerprint": "02501c14dac62b02",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 스피커폰 맞은편에 서서 하얀 손수건을 코끝에 댄 윤성찬의 근접 구도.\n\nLOCATION (lock): At the speakerphone conversation area inside a militia commander's office. The office is illuminated for daytime use.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 윤성찬's eye height on the established side of his axis with 박철진, retaining a three-quarter close view rather than moving onto his frontal line. Place his face just right of center and the white handkerchief beneath his nose in the lower middle, leaving look room left toward the off-screen 박철진 as he listens to the speakerphone conversation. Emphasize only the narrowing distance, preserving his listening posture and the existing illumination.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White handkerchief (Held against Yoon's nose as he listens); used as A small action detail below his eyes, leaving his expression readable; Office wall behind Yoon (The knife has not yet been thrown into it) — A limited oblique section appears behind his shoulder; used as Establishes the background plane that will carry the subsequent threat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior light maintains facial detail and the white handkerchief without turning either into a harsh highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 스피커폰 맞은편에 서서 하얀 손수건을 코끝에 댄 윤성찬의 근접 구도.\n\nLOCATION (lock): At the speakerphone conversation area inside a militia commander's office. The office is illuminated for daytime use.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 윤성찬's eye height on the established side of his axis with 박철진, retaining a three-quarter close view rather than moving onto his frontal line. Place his face just right of center and the white handkerchief beneath his nose in the lower middle, leaving look room left toward the off-screen 박철진 as he listens to the speakerphone conversation. Emphasize only the narrowing distance, preserving his listening posture and the existing illumination.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White handkerchief (Held against Yoon's nose as he listens); used as A small action detail below his eyes, leaving his expression readable; Office wall behind Yoon (The knife has not yet been thrown into it) — A limited oblique section appears behind his shoulder; used as Establishes the background plane that will carry the subsequent threat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior light maintains facial detail and the white handkerchief without turning either into a harsh highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh2__bgfirst_bg.png",
  "asset_id": "07e30494-bd9f-4f2b-9fd3-fa1665889592",
  "input_asset_ids": [
   "b77b5c1d-f086-4988-a584-51f3d06b92a2",
   "4974b9e4-5b59-4fa8-b4ea-8a3c9daa3ec6"
  ]
 },
 "S18sh2": {
  "input_fingerprint": "ba4cd1aa18d50ae5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스피커폰 맞은편에 서서 하얀 손수건을 코끝에 댄 윤성찬의 근접 구도.\n\nLOCATION (lock): At the speakerphone conversation area inside a militia commander's office. The office is illuminated for daytime use. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 윤성찬's eye height on the established side of his axis with 박철진, retaining a three-quarter close view rather than moving onto his frontal line. Place his face just right of center and the white handkerchief beneath his nose in the lower middle, leaving look room left toward the off-screen 박철진 as he listens to the speakerphone conversation. Emphasize only the narrowing distance, preserving his listening posture and the existing illumination.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White handkerchief (Held against Yoon's nose as he listens); used as A small action detail below his eyes, leaving his expression readable; Office wall behind Yoon (The knife has not yet been thrown into it) — A limited oblique section appears behind his shoulder; used as Establishes the background plane that will carry the subsequent threat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior light maintains facial detail and the white handkerchief without turning either into a harsh highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office speakerphone is in use. 윤성찬: He stands in the office holding a handkerchief to his nose and repeatedly wiping it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스피커폰 맞은편에 서서 하얀 손수건을 코끝에 댄 윤성찬의 근접 구도.\n\nLOCATION (lock): At the speakerphone conversation area inside a militia commander's office. The office is illuminated for daytime use. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 윤성찬's eye height on the established side of his axis with 박철진, retaining a three-quarter close view rather than moving onto his frontal line. Place his face just right of center and the white handkerchief beneath his nose in the lower middle, leaving look room left toward the off-screen 박철진 as he listens to the speakerphone conversation. Emphasize only the narrowing distance, preserving his listening posture and the existing illumination.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White handkerchief (Held against Yoon's nose as he listens); used as A small action detail below his eyes, leaving his expression readable; Office wall behind Yoon (The knife has not yet been thrown into it) — A limited oblique section appears behind his shoulder; used as Establishes the background plane that will carry the subsequent threat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior light maintains facial detail and the white handkerchief without turning either into a harsh highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office speakerphone is in use. 윤성찬: He stands in the office holding a handkerchief to his nose and repeatedly wiping it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스피커폰 맞은편에 서서 하얀 손수건을 코끝에 댄 윤성찬의 근접 구도.\n\nLOCATION (lock): At the speakerphone conversation area inside a militia commander's office. The office is illuminated for daytime use. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 윤성찬's eye height on the established side of his axis with 박철진, retaining a three-quarter close view rather than moving onto his frontal line. Place his face just right of center and the white handkerchief beneath his nose in the lower middle, leaving look room left toward the off-screen 박철진 as he listens to the speakerphone conversation. Emphasize only the narrowing distance, preserving his listening posture and the existing illumination.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: White handkerchief (Held against Yoon's nose as he listens); used as A small action detail below his eyes, leaving his expression readable; Office wall behind Yoon (The knife has not yet been thrown into it) — A limited oblique section appears behind his shoulder; used as Establishes the background plane that will carry the subsequent threat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior light maintains facial detail and the white handkerchief without turning either into a harsh highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office speakerphone is in use. 윤성찬: He stands in the office holding a handkerchief to his nose and repeatedly wiping it.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 윤성찬 right now, so 윤성찬's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 윤성찬: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh2__bgfirst_bg.png",
     "asset_id": "07e30494-bd9f-4f2b-9fd3-fa1665889592",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S18sh2.png",
     "asset_id": "b77b5c1d-f086-4988-a584-51f3d06b92a2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L30B01.png",
     "asset_id": "4974b9e4-5b59-4fa8-b4ea-8a3c9daa3ec6",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 왼쪽 여백을 향하고, 손에 든 손수건이 코끝에 위치합니다.",
    "built_space": "사무실 공간으로 책상과 스피커폰, 책장, 지도가 보이나, 오른쪽 배경 벽에 지시와 달리 단검이 꽂혀 있습니다.",
    "entities": "차콜 정장과 흰 손수건은 존재하나, 레퍼런스 인물의 주요 특징인 콧수염이 누락되었습니다.",
    "hard_violations": [
     "[gpt-high] 아직 칼이 던져져 있지 않아야 하는 사무실 벽에 칼이 이미 붙어 있어, 후속 위협 이전이라는 명시적 장면 상태를 위반합니다."
    ],
    "physics": "손이 손수건을 쥐고 얼굴에 밀착하여 지지하고 있습니다."
   },
   {
    "label": "B",
    "direction": "시선은 화면 밖 왼쪽을 향하고 있으며, 오른손은 손수건을 코끝으로 향하게 하고 있습니다.",
    "built_space": "사무실 내부로, 인물 뒤로 책장, 서류함, 한반도 지도, 화이트보드가 보이며 지시된 대로 벽에 칼은 보이지 않습니다.",
    "entities": "캐릭터 레퍼런스와 일치하는 콧수염과 백발의 남성, 차콜 그레이 정장과 넥타이, 흰 손수건이 확인됩니다.",
    "hard_violations": [],
    "physics": "오른손이 손수건을 쥐고 코 부분에 안정적으로 대고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스 이미지의 인물 특징(콧수염 등)을 정확히 재현했으며, 칼이 보이지 않아야 한다는 배경 조건과 근접 구도를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터의 콧수염이 누락되었으며, 배경 벽에 아직 없어야 할 단검이 꽂혀 있어 프롬프트의 세부 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "시선은 화면 밖 왼쪽을 향하고 있으며, 오른손은 손수건을 코끝으로 향하게 하고 있습니다.",
        "built_space": "사무실 내부로, 인물 뒤로 책장, 서류함, 한반도 지도, 화이트보드가 보이며 지시된 대로 벽에 칼은 보이지 않습니다.",
        "entities": "캐릭터 레퍼런스와 일치하는 콧수염과 백발의 남성, 차콜 그레이 정장과 넥타이, 흰 손수건이 확인됩니다.",
        "hard_violations": [],
        "physics": "오른손이 손수건을 쥐고 코 부분에 안정적으로 대고 있습니다."
       },
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽 여백을 향하고, 손에 든 손수건이 코끝에 위치합니다.",
        "built_space": "사무실 공간으로 책상과 스피커폰, 책장, 지도가 보이나, 오른쪽 배경 벽에 지시와 달리 단검이 꽂혀 있습니다.",
        "entities": "차콜 정장과 흰 손수건은 존재하나, 레퍼런스 인물의 주요 특징인 콧수염이 누락되었습니다.",
        "hard_violations": [],
        "physics": "손이 손수건을 쥐고 얼굴에 밀착하여 지지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스 이미지의 인물 특징(콧수염 등)을 정확히 재현했으며, 칼이 보이지 않아야 한다는 배경 조건과 근접 구도를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터의 콧수염이 누락되었으며, 배경 벽에 아직 없어야 할 단검이 꽂혀 있어 프롬프트의 세부 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선은 화면 밖 왼쪽을 향하고 있으며, 오른손은 손수건을 코끝으로 향하게 하고 있습니다.",
        "built_space": "사무실 내부로, 인물 뒤로 책장, 서류함, 한반도 지도, 화이트보드가 보이며 지시된 대로 벽에 칼은 보이지 않습니다.",
        "entities": "캐릭터 레퍼런스와 일치하는 콧수염과 백발의 남성, 차콜 그레이 정장과 넥타이, 흰 손수건이 확인됩니다.",
        "hard_violations": [],
        "physics": "오른손이 손수건을 쥐고 코 부분에 안정적으로 대고 있습니다."
       },
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽 여백을 향하고, 손에 든 손수건이 코끝에 위치합니다.",
        "built_space": "사무실 공간으로 책상과 스피커폰, 책장, 지도가 보이나, 오른쪽 배경 벽에 지시와 달리 단검이 꽂혀 있습니다.",
        "entities": "차콜 정장과 흰 손수건은 존재하나, 레퍼런스 인물의 주요 특징인 콧수염이 누락되었습니다.",
        "hard_violations": [],
        "physics": "손이 손수건을 쥐고 얼굴에 밀착하여 지지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽을 향한 경청 시선, 눈높이의 사선 클로즈업, 코끝에 댄 손수건과 칼 없는 배경이 지시에 충실하며, 배경 노출은 요구보다 조금 넓습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "손수건 동작과 왼쪽 시선은 맞지만, 아직 칼이 없어야 할 벽에 칼이 있고 정면에 가까운 시점과 넓은 사무실 배경이 지정 구도를 약화합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 얼굴과 두 눈은 화면 왼쪽의 화면 밖 상대를 향합니다. 박철진 자체는 보이지 않지만 그를 향한 경청 방향과 일치하며, 카메라를 응시하지 않습니다. 손수건 윗부분은 코끝과 콧구멍 아래에 닿아 있습니다.",
        "built_space": "왼쪽 창 하나, 그 옆 금속 선반 하나와 부분적으로 가려진 서류함 하나, 뒤쪽 한반도 지도 하나, 오른쪽 화이트보드 일부, 왼쪽 아래 의자 등받이 하나가 보입니다. 참조 사무실의 재료와 주요 배치 관계를 유지합니다. 어깨 뒤 벽에 칼은 보이지 않습니다. 얼굴은 중앙보다 오른쪽에 있고 왼쪽에 시선 여백이 있으나, 요구한 제한적 벽면 외에 선반과 창도 상당히 드러납니다. 스피커폰과 발은 클로즈업 밖이므로 정확한 대면 위치는 확인할 수 없습니다.",
        "entities": "인물은 고령의 한국인 남성으로 제시된 참조와 부합하는 외형이며, 은백색의 정돈된 머리, 콧수염, 얼굴 주름과 체형이 대체로 일치합니다. 머리의 윗부분은 참조보다 약간 풍성합니다. 차콜 그레이 정장, 흰 셔츠, 어두운 넥타이와 흰 손수건이 보입니다. 손도 주름과 피부 질감이 있는 노인의 손으로 얼굴 및 소매와 일관됩니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "윤성찬의 오른손이 손수건을 직접 쥐고 코에 대며, 손목과 전완은 정장 소매로 자연스럽게 이어집니다. 천은 손가락에 눌리고 아래로 접혀 늘어져 실제 직물의 지지와 무게가 읽힙니다. 상체는 수직이며 떠 있는 자세가 아닙니다. 하체가 잘려 발의 지지는 확인할 수 없지만 물리적 모순은 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "두 눈은 화면 왼쪽의 화면 밖 상대를 향해 있어 박철진 쪽을 보며 듣는 방향과 맞습니다. 몸통은 카메라를 거의 정면으로 향하고 얼굴만 왼쪽으로 조금 돌려, 요구한 사선 시점은 A보다 약합니다. 손수건은 코 아래에 밀착되어 있습니다.",
        "built_space": "왼쪽 창 하나, 금속 선반 하나, 무전기들이 놓인 서류함 하나, 책상 하나와 그 뒤 의자 하나, 책상 왼쪽 스피커폰 하나가 보입니다. 벽에는 왼쪽 가장자리 지도와 중앙 지도 일부가 있고 오른쪽에는 문 하나와 수납 선반 일부가 있습니다. 참조의 사무실 구조는 잘 식별되지만 책상과 방 전체가 넓게 노출되어 제한된 배경 벽면이라는 요구와 거리가 있습니다. 특히 인물 오른쪽 뒤 벽에 칼 하나가 이미 붙어 있습니다.",
        "entities": "인물은 참조와 유사한 고령의 한국인 남성 외형으로, 은백색 머리와 깊은 주름, 차콜 정장, 흰 셔츠, 어두운 넥타이를 갖췄습니다. 흰 손수건과 이를 쥔 노년의 손이 보이며, 입과 콧수염 부분은 천에 가려 직접 비교하기 어렵습니다. 스피커폰의 형태는 참조와 일치하지만 통화 중인지는 정지 화면만으로 확인할 수 없습니다. 추가 인물은 없습니다.",
        "hard_violations": [
         "아직 칼이 던져져 있지 않아야 하는 사무실 벽에 칼이 이미 붙어 있어, 후속 위협 이전이라는 명시적 장면 상태를 위반합니다."
        ],
        "physics": "오른손이 손수건을 쥐고 코에 누르고 있으며 손목과 팔의 연결은 자연스럽습니다. 천은 손의 압력에 따라 접혀 아래로 드리워집니다. 상체는 수직으로 유지되고, 화면 밖 하체의 지지는 확인할 수 없으나 부유 징후는 없습니다. 스피커폰과 지도는 책상 위에 놓여 있고, 배경 칼은 벽에 고정된 상태로 읽히므로 부유 문제가 아니라 시점상 존재해서는 안 되는 소품 문제입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽을 향한 경청 시선, 눈높이의 사선 클로즈업, 코끝에 댄 손수건과 칼 없는 배경이 지시에 충실하며, 배경 노출은 요구보다 조금 넓습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "손수건 동작과 왼쪽 시선은 맞지만, 아직 칼이 없어야 할 벽에 칼이 있고 정면에 가까운 시점과 넓은 사무실 배경이 지정 구도를 약화합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬의 얼굴과 두 눈은 화면 왼쪽의 화면 밖 상대를 향합니다. 박철진 자체는 보이지 않지만 그를 향한 경청 방향과 일치하며, 카메라를 응시하지 않습니다. 손수건 윗부분은 코끝과 콧구멍 아래에 닿아 있습니다.",
        "built_space": "왼쪽 창 하나, 그 옆 금속 선반 하나와 부분적으로 가려진 서류함 하나, 뒤쪽 한반도 지도 하나, 오른쪽 화이트보드 일부, 왼쪽 아래 의자 등받이 하나가 보입니다. 참조 사무실의 재료와 주요 배치 관계를 유지합니다. 어깨 뒤 벽에 칼은 보이지 않습니다. 얼굴은 중앙보다 오른쪽에 있고 왼쪽에 시선 여백이 있으나, 요구한 제한적 벽면 외에 선반과 창도 상당히 드러납니다. 스피커폰과 발은 클로즈업 밖이므로 정확한 대면 위치는 확인할 수 없습니다.",
        "entities": "인물은 고령의 한국인 남성으로 제시된 참조와 부합하는 외형이며, 은백색의 정돈된 머리, 콧수염, 얼굴 주름과 체형이 대체로 일치합니다. 머리의 윗부분은 참조보다 약간 풍성합니다. 차콜 그레이 정장, 흰 셔츠, 어두운 넥타이와 흰 손수건이 보입니다. 손도 주름과 피부 질감이 있는 노인의 손으로 얼굴 및 소매와 일관됩니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "윤성찬의 오른손이 손수건을 직접 쥐고 코에 대며, 손목과 전완은 정장 소매로 자연스럽게 이어집니다. 천은 손가락에 눌리고 아래로 접혀 늘어져 실제 직물의 지지와 무게가 읽힙니다. 상체는 수직이며 떠 있는 자세가 아닙니다. 하체가 잘려 발의 지지는 확인할 수 없지만 물리적 모순은 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "두 눈은 화면 왼쪽의 화면 밖 상대를 향해 있어 박철진 쪽을 보며 듣는 방향과 맞습니다. 몸통은 카메라를 거의 정면으로 향하고 얼굴만 왼쪽으로 조금 돌려, 요구한 사선 시점은 A보다 약합니다. 손수건은 코 아래에 밀착되어 있습니다.",
        "built_space": "왼쪽 창 하나, 금속 선반 하나, 무전기들이 놓인 서류함 하나, 책상 하나와 그 뒤 의자 하나, 책상 왼쪽 스피커폰 하나가 보입니다. 벽에는 왼쪽 가장자리 지도와 중앙 지도 일부가 있고 오른쪽에는 문 하나와 수납 선반 일부가 있습니다. 참조의 사무실 구조는 잘 식별되지만 책상과 방 전체가 넓게 노출되어 제한된 배경 벽면이라는 요구와 거리가 있습니다. 특히 인물 오른쪽 뒤 벽에 칼 하나가 이미 붙어 있습니다.",
        "entities": "인물은 참조와 유사한 고령의 한국인 남성 외형으로, 은백색 머리와 깊은 주름, 차콜 정장, 흰 셔츠, 어두운 넥타이를 갖췄습니다. 흰 손수건과 이를 쥔 노년의 손이 보이며, 입과 콧수염 부분은 천에 가려 직접 비교하기 어렵습니다. 스피커폰의 형태는 참조와 일치하지만 통화 중인지는 정지 화면만으로 확인할 수 없습니다. 추가 인물은 없습니다.",
        "hard_violations": [
         "아직 칼이 던져져 있지 않아야 하는 사무실 벽에 칼이 이미 붙어 있어, 후속 위협 이전이라는 명시적 장면 상태를 위반합니다."
        ],
        "physics": "오른손이 손수건을 쥐고 코에 누르고 있으며 손목과 팔의 연결은 자연스럽습니다. 천은 손의 압력에 따라 접혀 아래로 드리워집니다. 상체는 수직으로 유지되고, 화면 밖 하체의 지지는 확인할 수 없으나 부유 징후는 없습니다. 스피커폰과 지도는 책상 위에 놓여 있고, 배경 칼은 벽에 고정된 상태로 읽히므로 부유 문제가 아니라 시점상 존재해서는 안 되는 소품 문제입니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.016,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.766,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 아직 칼이 던져져 있지 않아야 하는 사무실 벽에 칼이 이미 붙어 있어, 후속 위협 이전이라는 명시적 장면 상태를 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 766
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스 이미지의 인물 특징(콧수염 등)을 정확히 재현했으며, 칼이 보이지 않아야 한다는 배경 조건과 근접 구도를 충실히 따랐습니다."
   },
   {
    "label": "A",
    "score": 766,
    "verdict_ko": "캐릭터의 콧수염이 누락되었으며, 배경 벽에 아직 없어야 할 단검이 꽂혀 있어 프롬프트의 세부 지시를 위반했습니다.  ★위반: [gpt-high] 아직 칼이 던져져 있지 않아야 하는 사무실 벽에 칼이 이미 붙어 있어, 후속 위협 이전이라는 명시적 장면 상태를 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L30B01.png",
    "asset_id": "4974b9e4-5b59-4fa8-b4ea-8a3c9daa3ec6",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08dd-ba3c-70d2-90df-b672b0491d3d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh2__bgfirst_bg.png",
   "bg_asset_id": "07e30494-bd9f-4f2b-9fd3-fa1665889592",
   "bg_record_key": "S18sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S18sh7::signage": {
  "fp": "4ac0a759fad70dd5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S18sh7": {
  "input_fingerprint": "4411ec68a15b6216",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽에 박힌 단검 옆에서 두 눈을 부릅뜬 채 붉게 달아오른 윤성찬의 굳은 얼굴.\n\nLOCATION (lock): Immediately in front of the wall behind the visitor's position in the militia commander's office, under daytime office illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly to the established side of 윤성찬 at close distance, with the lens slightly above his eyes and tilted down toward his oblique face and the neighboring wall. Hold his widened eyes and flushed, rigid face left of center while the embedded knife remains a smaller, separate shape to his right in the background; his stare stays fixed on 박철진 off-screen left, not on the knife or lens. Keep this stage-entry frame static, making the newly shortened distance the emphasis while retaining the preceding light.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wall area with embedded knife, clear of Yoon's facial outline in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Knife (Embedded in the wall behind Yoon) — The projecting handle and visible blade portion are viewed obliquely beside, not across, his facial outline; used as A compact background threat that shares the reaction frame without exaggerated scale; Office wall (Now holds the thrown knife); used as Provides the continuous background plane separating the knife from Yoon's head.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral, controlled interior illumination so Yoon's flushed skin reads as a frightened physical reaction rather than a change in light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is lodged in the office wall; it is no longer being held. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽에 박힌 단검 옆에서 두 눈을 부릅뜬 채 붉게 달아오른 윤성찬의 굳은 얼굴.\n\nLOCATION (lock): Immediately in front of the wall behind the visitor's position in the militia commander's office, under daytime office illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly to the established side of 윤성찬 at close distance, with the lens slightly above his eyes and tilted down toward his oblique face and the neighboring wall. Hold his widened eyes and flushed, rigid face left of center while the embedded knife remains a smaller, separate shape to his right in the background; his stare stays fixed on 박철진 off-screen left, not on the knife or lens. Keep this stage-entry frame static, making the newly shortened distance the emphasis while retaining the preceding light.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wall area with embedded knife, clear of Yoon's facial outline in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Knife (Embedded in the wall behind Yoon) — The projecting handle and visible blade portion are viewed obliquely beside, not across, his facial outline; used as A compact background threat that shares the reaction frame without exaggerated scale; Office wall (Now holds the thrown knife); used as Provides the continuous background plane separating the knife from Yoon's head.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral, controlled interior illumination so Yoon's flushed skin reads as a frightened physical reaction rather than a change in light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is lodged in the office wall; it is no longer being held. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽에 박힌 단검 옆에서 두 눈을 부릅뜬 채 붉게 달아오른 윤성찬의 굳은 얼굴.\n\nLOCATION (lock): Immediately in front of the wall behind the visitor's position in the militia commander's office, under daytime office illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly to the established side of 윤성찬 at close distance, with the lens slightly above his eyes and tilted down toward his oblique face and the neighboring wall. Hold his widened eyes and flushed, rigid face left of center while the embedded knife remains a smaller, separate shape to his right in the background; his stare stays fixed on 박철진 off-screen left, not on the knife or lens. Keep this stage-entry frame static, making the newly shortened distance the emphasis while retaining the preceding light.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Wall area with embedded knife, clear of Yoon's facial outline in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Knife (Embedded in the wall behind Yoon) — The projecting handle and visible blade portion are viewed obliquely beside, not across, his facial outline; used as A compact background threat that shares the reaction frame without exaggerated scale; Office wall (Now holds the thrown knife); used as Provides the continuous background plane separating the knife from Yoon's head.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral, controlled interior illumination so Yoon's flushed skin reads as a frightened physical reaction rather than a change in light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thrown knife is lodged in the office wall; it is no longer being held. The speakerphone remains in the office. 윤성찬: He remains standing with his handkerchief, his face flushed from the sudden fright.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬의 두 눈은 화면 왼쪽 바깥을 뚜렷하게 응시하고 있습니다.",
    "built_space": "배경으로 한반도 지도와 종이, 창문 일부 및 캐비닛이 보이며 인물 우측 벽면에 단검이 박혀 있습니다.",
    "entities": "인물의 외모, 정장, 손수건, 단검은 지정된 형태와 일치하나, 붉게 달아오른 얼굴이라는 조건에 비해 안색의 변화가 미미합니다.",
    "hard_violations": [],
    "physics": "단검은 벽면에 단단히 꽂혀 있고, 오른손이 손수건을 제대로 파지하고 있습니다."
   },
   {
    "label": "B",
    "direction": "윤성찬의 시선은 화면 밖 왼쪽을 향하고 있으며 단검이나 카메라를 보지 않습니다.",
    "built_space": "사무실 내부로, 인물 뒤편 벽면에는 한반도 지도와 종이가 붙어 있고 인물 우측으로 벽에 꽂힌 단검이 위치합니다. 우측 끝에는 이전 샷에 있던 화이트보드의 일부가 보입니다.",
    "entities": "윤성찬(70대 후반 남성, 정장 차림)과 손수건, 단검 모두 레퍼런스 및 프롬프트 묘사와 일치합니다. 피부의 붉은 홍조가 명확히 확인됩니다.",
    "hard_violations": [],
    "physics": "벽에 박힌 단검은 안정적으로 고정되어 있고, 윤성찬의 오른손이 손수건을 자연스럽게 쥐고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 '붉게 달아오른' 얼굴 표현이 매우 사실적으로 구현되었으며, 인물의 경직된 표정과 시선 처리, 배경의 단검 배치 등 모든 요소가 훌륭하게 반영되었습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 인물의 기본 인상 및 디테일은 잘 유지되었으나, 프롬프트에서 핵심적으로 요구한 '붉게 달아오른' 피부 묘사가 부족하여 다소 평이한 안색으로 보입니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬의 시선은 화면 밖 왼쪽을 향하고 있으며 단검이나 카메라를 보지 않습니다.",
        "built_space": "사무실 내부로, 인물 뒤편 벽면에는 한반도 지도와 종이가 붙어 있고 인물 우측으로 벽에 꽂힌 단검이 위치합니다. 우측 끝에는 이전 샷에 있던 화이트보드의 일부가 보입니다.",
        "entities": "윤성찬(70대 후반 남성, 정장 차림)과 손수건, 단검 모두 레퍼런스 및 프롬프트 묘사와 일치합니다. 피부의 붉은 홍조가 명확히 확인됩니다.",
        "hard_violations": [],
        "physics": "벽에 박힌 단검은 안정적으로 고정되어 있고, 윤성찬의 오른손이 손수건을 자연스럽게 쥐고 있습니다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 두 눈은 화면 왼쪽 바깥을 뚜렷하게 응시하고 있습니다.",
        "built_space": "배경으로 한반도 지도와 종이, 창문 일부 및 캐비닛이 보이며 인물 우측 벽면에 단검이 박혀 있습니다.",
        "entities": "인물의 외모, 정장, 손수건, 단검은 지정된 형태와 일치하나, 붉게 달아오른 얼굴이라는 조건에 비해 안색의 변화가 미미합니다.",
        "hard_violations": [],
        "physics": "단검은 벽면에 단단히 꽂혀 있고, 오른손이 손수건을 제대로 파지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 '붉게 달아오른' 얼굴 표현이 매우 사실적으로 구현되었으며, 인물의 경직된 표정과 시선 처리, 배경의 단검 배치 등 모든 요소가 훌륭하게 반영되었습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 인물의 기본 인상 및 디테일은 잘 유지되었으나, 프롬프트에서 핵심적으로 요구한 '붉게 달아오른' 피부 묘사가 부족하여 다소 평이한 안색으로 보입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬의 시선은 화면 밖 왼쪽을 향하고 있으며 단검이나 카메라를 보지 않습니다.",
        "built_space": "사무실 내부로, 인물 뒤편 벽면에는 한반도 지도와 종이가 붙어 있고 인물 우측으로 벽에 꽂힌 단검이 위치합니다. 우측 끝에는 이전 샷에 있던 화이트보드의 일부가 보입니다.",
        "entities": "윤성찬(70대 후반 남성, 정장 차림)과 손수건, 단검 모두 레퍼런스 및 프롬프트 묘사와 일치합니다. 피부의 붉은 홍조가 명확히 확인됩니다.",
        "hard_violations": [],
        "physics": "벽에 박힌 단검은 안정적으로 고정되어 있고, 윤성찬의 오른손이 손수건을 자연스럽게 쥐고 있습니다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 두 눈은 화면 왼쪽 바깥을 뚜렷하게 응시하고 있습니다.",
        "built_space": "배경으로 한반도 지도와 종이, 창문 일부 및 캐비닛이 보이며 인물 우측 벽면에 단검이 박혀 있습니다.",
        "entities": "인물의 외모, 정장, 손수건, 단검은 지정된 형태와 일치하나, 붉게 달아오른 얼굴이라는 조건에 비해 안색의 변화가 미미합니다.",
        "hard_violations": [],
        "physics": "단검은 벽면에 단단히 꽂혀 있고, 오른손이 손수건을 제대로 파지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽을 응시하는 붉어진 얼굴과 벽에 박힌 작은 단검은 충실하지만, 손수건이 입가 일부를 가리고 내려다보는 사선 시점이 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "약간 높은 사선 클로즈업에서 굳고 붉어진 얼굴을 가림 없이 보여주며, 화면 밖 왼쪽 응시와 얼굴에서 분리된 배경 단검을 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈의 시선은 화면 밖 왼쪽을 향해 박철진의 지정 위치와 맞으며, 렌즈나 칼을 보지 않는다. 단검은 오른쪽 벽에 칼끝이 들어가 있고 손잡이는 왼쪽 앞쪽으로 돌출되어 얼굴 윤곽과 겹치지 않는다.",
        "built_space": "왼쪽 창 한 개, 그 옆 선반 한 개, 뒤쪽 벽의 지도 한 장, 오른쪽 가장자리의 화이트보드 일부와 부착 종이가 보인다. 이전 장면의 회색 벽과 사무실 구성이 이어지며 중복 설비나 불가능한 반사는 없다. 얼굴은 중앙 왼쪽에 크게 놓이고 단검 한 자루는 오른쪽 배경에 작게 분리되지만, 시점은 거의 눈높이에 가깝다.",
        "entities": "윤성찬에 해당하는 노년의 동아시아계 남성 한 명만 보인다. 짧게 정돈한 은회색 머리, 콧수염, 주름과 검버섯, 차콜 정장과 어두운 넥타이가 참조와 잘 맞는다. 두 눈은 정상적인 해부학적 형태로 크게 떠 있으며 볼과 이마가 붉다. 흰 손수건과 금속 날·검은 손잡이의 단검이 보인다. 스피커폰은 클로즈업 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "손수건은 손가락으로 잡혀 입가에 닿아 있고, 손과 팔의 연결도 자연스럽다. 단검은 벽의 파손된 삽입부에 날이 고정되어 지지된다. 상체는 곧게 서 있는 모습이며 발과 바닥은 프레임 밖이다. 떠 있거나 지지 없이 놓인 물체는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴과 두 눈이 화면 밖 왼쪽을 향해 있어 박철진을 응시한다는 지시에 부합한다. 단검이나 카메라를 향한 시선은 아니다. 오른쪽 배경 단검의 날은 벽으로 들어가고 손잡이는 왼쪽 앞쪽으로 비스듬히 돌출되어 얼굴과 확실히 분리된다.",
        "built_space": "왼쪽 창 한 개와 선반 한 개, 뒤쪽 지도 한 장, 오른쪽 벽의 부착 종이 한 장이 보인다. 회색 벽과 낮의 창광이 이전 사무실의 재질 및 조명을 이어간다. 화이트보드는 오른쪽 화면 밖으로 빠져 있으며 설비 중복이나 반사는 없다. 눈보다 약간 높은 사선 시점에서 얼굴을 중앙 왼쪽에 크게 배치한다. 단검은 작은 배경 요소로 유지되지만 지정한 중간 오른쪽보다는 가장자리 쪽에 가깝다.",
        "entities": "참조와 일치하는 은회색 머리와 콧수염, 고령의 얼굴 특징을 가진 동아시아계 남성 한 명이 보인다. 차콜 정장, 흰 셔츠, 어두운 넥타이와 흰 포켓스퀘어가 유지된다. 붉어진 볼, 부릅뜬 정상적인 눈, 굳은 입매가 드러난다. 흰 손수건은 턱 아래에 내려와 얼굴을 가리지 않는다. 단검 한 자루가 벽에 박혀 있으며 스피커폰은 촬영 범위 밖이다.",
        "hard_violations": [],
        "physics": "흰 손수건은 가슴 앞에서 손가락으로 확실히 쥐고 있어 중력에 따라 접혀 내려온다. 단검은 벽의 균열과 삽입 지점에 날이 고정되어 지지된다. 목과 어깨, 손의 자세는 서서 놀라 굳은 순간으로 가능한 형태다. 하체는 프레임 밖이며 부유하거나 지지 없는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽을 응시하는 붉어진 얼굴과 벽에 박힌 작은 단검은 충실하지만, 손수건이 입가 일부를 가리고 내려다보는 사선 시점이 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "약간 높은 사선 클로즈업에서 굳고 붉어진 얼굴을 가림 없이 보여주며, 화면 밖 왼쪽 응시와 얼굴에서 분리된 배경 단검을 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈의 시선은 화면 밖 왼쪽을 향해 박철진의 지정 위치와 맞으며, 렌즈나 칼을 보지 않는다. 단검은 오른쪽 벽에 칼끝이 들어가 있고 손잡이는 왼쪽 앞쪽으로 돌출되어 얼굴 윤곽과 겹치지 않는다.",
        "built_space": "왼쪽 창 한 개, 그 옆 선반 한 개, 뒤쪽 벽의 지도 한 장, 오른쪽 가장자리의 화이트보드 일부와 부착 종이가 보인다. 이전 장면의 회색 벽과 사무실 구성이 이어지며 중복 설비나 불가능한 반사는 없다. 얼굴은 중앙 왼쪽에 크게 놓이고 단검 한 자루는 오른쪽 배경에 작게 분리되지만, 시점은 거의 눈높이에 가깝다.",
        "entities": "윤성찬에 해당하는 노년의 동아시아계 남성 한 명만 보인다. 짧게 정돈한 은회색 머리, 콧수염, 주름과 검버섯, 차콜 정장과 어두운 넥타이가 참조와 잘 맞는다. 두 눈은 정상적인 해부학적 형태로 크게 떠 있으며 볼과 이마가 붉다. 흰 손수건과 금속 날·검은 손잡이의 단검이 보인다. 스피커폰은 클로즈업 밖이라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "손수건은 손가락으로 잡혀 입가에 닿아 있고, 손과 팔의 연결도 자연스럽다. 단검은 벽의 파손된 삽입부에 날이 고정되어 지지된다. 상체는 곧게 서 있는 모습이며 발과 바닥은 프레임 밖이다. 떠 있거나 지지 없이 놓인 물체는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴과 두 눈이 화면 밖 왼쪽을 향해 있어 박철진을 응시한다는 지시에 부합한다. 단검이나 카메라를 향한 시선은 아니다. 오른쪽 배경 단검의 날은 벽으로 들어가고 손잡이는 왼쪽 앞쪽으로 비스듬히 돌출되어 얼굴과 확실히 분리된다.",
        "built_space": "왼쪽 창 한 개와 선반 한 개, 뒤쪽 지도 한 장, 오른쪽 벽의 부착 종이 한 장이 보인다. 회색 벽과 낮의 창광이 이전 사무실의 재질 및 조명을 이어간다. 화이트보드는 오른쪽 화면 밖으로 빠져 있으며 설비 중복이나 반사는 없다. 눈보다 약간 높은 사선 시점에서 얼굴을 중앙 왼쪽에 크게 배치한다. 단검은 작은 배경 요소로 유지되지만 지정한 중간 오른쪽보다는 가장자리 쪽에 가깝다.",
        "entities": "참조와 일치하는 은회색 머리와 콧수염, 고령의 얼굴 특징을 가진 동아시아계 남성 한 명이 보인다. 차콜 정장, 흰 셔츠, 어두운 넥타이와 흰 포켓스퀘어가 유지된다. 붉어진 볼, 부릅뜬 정상적인 눈, 굳은 입매가 드러난다. 흰 손수건은 턱 아래에 내려와 얼굴을 가리지 않는다. 단검 한 자루가 벽에 박혀 있으며 스피커폰은 촬영 범위 밖이다.",
        "hard_violations": [],
        "physics": "흰 손수건은 가슴 앞에서 손가락으로 확실히 쥐고 있어 중력에 따라 접혀 내려온다. 단검은 벽의 균열과 삽입 지점에 날이 고정되어 지지된다. 목과 어깨, 손의 자세는 서서 놀라 굳은 순간으로 가능한 형태다. 하체는 프레임 밖이며 부유하거나 지지 없는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.889
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.889
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1889,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "프롬프트가 요구한 '붉게 달아오른' 얼굴 표현이 매우 사실적으로 구현되었으며, 인물의 경직된 표정과 시선 처리, 배경의 단검 배치 등 모든 요소가 훌륭하게 반영되었습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "구도와 인물의 기본 인상 및 디테일은 잘 유지되었으나, 프롬프트에서 핵심적으로 요구한 '붉게 달아오른' 피부 묘사가 부족하여 다소 평이한 안색으로 보입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh2_sel.png",
    "asset_id": "54a5ba05-0239-46f1-825d-90946c0e1f31",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08e3-d29a-737e-b7a0-f85666ef2931",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh2"
  }
 },
 "S18sh8::signage": {
  "fp": "7f7c000622906a92",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S18sh8": {
  "input_fingerprint": "f4242d86de01358d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 겁에 질린 윤성찬을 향해 허리를 90도로 굽힌 채 조롱하듯 인사 자세를 취한 박철진의 전신.\n\nLOCATION (lock): In the open floor area between the militia commander's position and his visitor, inside the daylit office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the diagonal dolly-out on the same side of the conversation axis, lowering to 박철진's standing chest height and retaining a lateral three-quarter view as his full figure enters the composition. Place 박철진 left of center at the bottom of his ninety-degree bow toward 윤성찬 on the right by the wall, leaving his head, bent torso, legs, and feet unobstructed; his lowered eyes face the floor beneath the bow while 윤성찬 looks down at him with a guarded backward lean. Emphasize the widening camera distance rather than a new lighting or eyeline setup, allowing the mocking courtesy and Yoon's fear to coexist in one direct view.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office wall behind Yoon (The thrown knife remains embedded) — The wall is viewed obliquely behind Yoon on the right; used as Keeps the earlier threat spatially attached to Yoon while the bow occupies the open foreground; Embedded knife (Remains fixed in the wall after the throw) — Its projecting handle is visible as a small background detail beside Yoon; used as Maintains threat continuity without competing with the full-body gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued interior illumination, with sufficient separation to read the complete bow and Yoon's tense reaction without theatrical relighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains lodged in the office wall, and the speakerphone remains in place. 윤성찬: He remains standing with the handkerchief, his face still flushed from the fright. 박철진: He bends into a ninety-degree bow, no longer holding the thrown knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 겁에 질린 윤성찬을 향해 허리를 90도로 굽힌 채 조롱하듯 인사 자세를 취한 박철진의 전신.\n\nLOCATION (lock): In the open floor area between the militia commander's position and his visitor, inside the daylit office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the diagonal dolly-out on the same side of the conversation axis, lowering to 박철진's standing chest height and retaining a lateral three-quarter view as his full figure enters the composition. Place 박철진 left of center at the bottom of his ninety-degree bow toward 윤성찬 on the right by the wall, leaving his head, bent torso, legs, and feet unobstructed; his lowered eyes face the floor beneath the bow while 윤성찬 looks down at him with a guarded backward lean. Emphasize the widening camera distance rather than a new lighting or eyeline setup, allowing the mocking courtesy and Yoon's fear to coexist in one direct view.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office wall behind Yoon (The thrown knife remains embedded) — The wall is viewed obliquely behind Yoon on the right; used as Keeps the earlier threat spatially attached to Yoon while the bow occupies the open foreground; Embedded knife (Remains fixed in the wall after the throw) — Its projecting handle is visible as a small background detail beside Yoon; used as Maintains threat continuity without competing with the full-body gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued interior illumination, with sufficient separation to read the complete bow and Yoon's tense reaction without theatrical relighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains lodged in the office wall, and the speakerphone remains in place. 윤성찬: He remains standing with the handkerchief, his face still flushed from the fright. 박철진: He bends into a ninety-degree bow, no longer holding the thrown knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 겁에 질린 윤성찬을 향해 허리를 90도로 굽힌 채 조롱하듯 인사 자세를 취한 박철진의 전신.\n\nLOCATION (lock): In the open floor area between the militia commander's position and his visitor, inside the daylit office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the diagonal dolly-out on the same side of the conversation axis, lowering to 박철진's standing chest height and retaining a lateral three-quarter view as his full figure enters the composition. Place 박철진 left of center at the bottom of his ninety-degree bow toward 윤성찬 on the right by the wall, leaving his head, bent torso, legs, and feet unobstructed; his lowered eyes face the floor beneath the bow while 윤성찬 looks down at him with a guarded backward lean. Emphasize the widening camera distance rather than a new lighting or eyeline setup, allowing the mocking courtesy and Yoon's fear to coexist in one direct view.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office wall behind Yoon (The thrown knife remains embedded) — The wall is viewed obliquely behind Yoon on the right; used as Keeps the earlier threat spatially attached to Yoon while the bow occupies the open foreground; Embedded knife (Remains fixed in the wall after the throw) — Its projecting handle is visible as a small background detail beside Yoon; used as Maintains threat continuity without competing with the full-body gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same subdued interior illumination, with sufficient separation to read the complete bow and Yoon's tense reaction without theatrical relighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife remains lodged in the office wall, and the speakerphone remains in place. 윤성찬: He remains standing with the handkerchief, his face still flushed from the fright. 박철진: He bends into a ninety-degree bow, no longer holding the thrown knife.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진은 윤성찬을 향해 몸을 굽히고 있으며, 윤성찬은 아래로 그를 주시함.",
    "built_space": "사무실 내부. 좌측에 책상과 창문, 우측 벽면에 지도, 화이트보드, 박힌 칼이 위치함.",
    "entities": "박철진(모자, 남색 전투복, 붉은 완장), 윤성찬(정장, 손수건), 박힌 칼 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 안정적으로 딛고 서 있으며, 굽힌 허리와 손수건을 쥔 손의 묘사가 자연스러움."
   },
   {
    "label": "B",
    "direction": "박철진은 윤성찬을 향해 고개를 숙이고, 윤성찬은 그를 내려다봄.",
    "built_space": "사무실 내부. 좌측의 책상, 중앙 테이블의 스피커폰, 우측 벽의 칼 등이 배치됨.",
    "entities": "윤성찬과 벽의 칼은 잘 묘사되었으나, 박철진의 왼팔에 있어야 할 붉은 완장이 없음.",
    "hard_violations": [],
    "physics": "두 사람 모두 지면에 서서 중력에 맞게 자세를 유지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "붉은 완장, 90도 인사 자세, 윤성찬의 경계하는 태도 등 프롬프트와 레퍼런스를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 배경 디테일은 우수하나, 박철진의 필수 요소인 붉은 완장이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 윤성찬을 향해 몸을 굽히고 있으며, 윤성찬은 아래로 그를 주시함.",
        "built_space": "사무실 내부. 좌측에 책상과 창문, 우측 벽면에 지도, 화이트보드, 박힌 칼이 위치함.",
        "entities": "박철진(모자, 남색 전투복, 붉은 완장), 윤성찬(정장, 손수건), 박힌 칼 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 딛고 서 있으며, 굽힌 허리와 손수건을 쥔 손의 묘사가 자연스러움."
       },
       {
        "label": "B",
        "direction": "박철진은 윤성찬을 향해 고개를 숙이고, 윤성찬은 그를 내려다봄.",
        "built_space": "사무실 내부. 좌측의 책상, 중앙 테이블의 스피커폰, 우측 벽의 칼 등이 배치됨.",
        "entities": "윤성찬과 벽의 칼은 잘 묘사되었으나, 박철진의 왼팔에 있어야 할 붉은 완장이 없음.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 서서 중력에 맞게 자세를 유지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "붉은 완장, 90도 인사 자세, 윤성찬의 경계하는 태도 등 프롬프트와 레퍼런스를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 구도와 배경 디테일은 우수하나, 박철진의 필수 요소인 붉은 완장이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 윤성찬을 향해 몸을 굽히고 있으며, 윤성찬은 아래로 그를 주시함.",
        "built_space": "사무실 내부. 좌측에 책상과 창문, 우측 벽면에 지도, 화이트보드, 박힌 칼이 위치함.",
        "entities": "박철진(모자, 남색 전투복, 붉은 완장), 윤성찬(정장, 손수건), 박힌 칼 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 딛고 서 있으며, 굽힌 허리와 손수건을 쥔 손의 묘사가 자연스러움."
       },
       {
        "label": "B",
        "direction": "박철진은 윤성찬을 향해 고개를 숙이고, 윤성찬은 그를 내려다봄.",
        "built_space": "사무실 내부. 좌측의 책상, 중앙 테이블의 스피커폰, 우측 벽의 칼 등이 배치됨.",
        "entities": "윤성찬과 벽의 칼은 잘 묘사되었으나, 박철진의 왼팔에 있어야 할 붉은 완장이 없음.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 서서 중력에 맞게 자세를 유지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전신 와이드 구도와 두 사람의 방향은 맞지만, 박철진의 상체가 크게 기울어진 채 남아 있어 핵심인 90도 인사의 최저점과 거리가 멀다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박철진의 가려지지 않은 전신, 더 깊은 인사, 오른쪽 윤성찬과 벽의 칼을 충실하게 배치했으나 정확한 90도 굽힘과 윤성찬의 뒤로 물러선 반응은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 화면 왼쪽에서 오른쪽 윤성찬을 향해 몸을 숙이고, 얼굴과 눈은 두 사람 사이의 바닥을 향한다. 윤성찬은 왼쪽 아래 박철진을 바라본다. 칼은 날이 오른쪽 벽에 박히고 손잡이가 왼쪽으로 돌출되어 있어 투척 후 상태에 맞는다.",
        "built_space": "뒤쪽에 큰 창면 하나와 책상 하나, 그 뒤 업무용 의자 하나가 있다. 왼쪽에는 소파 하나와 방문객용 의자 두 개가 보이고, 중앙 낮은 탁자 하나 위에 스피커폰 하나가 놓였다. 뒤쪽 오른편에는 책장 하나와 금속 수납장 하나, 오른쪽 벽에는 지도 하나·화이트보드 하나·박힌 칼 하나가 보인다. 두 사람은 가구 사이 빈 바닥에 서며 박철진의 머리부터 신발까지 가리지 않는다. 참조의 창틀, 서류 책장, 밝은 벽과 지도는 이어지지만, 참조의 좁은 화면만으로 새로 드러난 가구의 원래 위치까지 확인할 수는 없다.",
        "entities": "인물은 중년 동아시아계 남성 박철진과 고령 동아시아계 남성 윤성찬 두 명이다. 박철진의 남색 전투복, 챙 있는 모자와 허리 주머니는 인물 참조와 부합하며 얼굴은 숙인 측면만 보여 정확한 동일성 확인이 제한된다. 붉은 완장은 보이지 않아 확인할 수 없다. 윤성찬은 은회색 머리와 콧수염, 짙은 정장·흰 셔츠·어두운 넥타이로 이전 장면의 외양을 유지하고, 붉어진 얼굴 가까이에 흰 손수건을 들고 있다. 벽의 칼과 탁자의 스피커폰이 각각 하나씩 보이고 박철진은 칼을 들지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 신발이 바닥에 닿아 체중을 지지한다. 박철진은 골반에서 상체를 숙이고 손을 허벅지 옆에 내려 둔, 가능한 인사 자세다. 다만 상체가 수평보다 상당히 높아 허리 굽힘은 대략 50도 안팎으로 보이며 90도 인사의 최저점은 아니다. 윤성찬의 손수건은 손에 잡혀 있고, 칼은 벽에 박힌 날로, 스피커폰은 탁자 상판으로 지지된다. 윤성찬은 거의 곧게 서 있어 경계하며 뒤로 기대는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "박철진은 왼쪽에서 오른쪽 벽 앞의 윤성찬에게 인사하며 눈을 자신의 숙인 머리 아래 바닥으로 내린다. 윤성찬은 고개와 시선을 왼쪽 아래 박철진에게 향한다. 벽의 칼은 날이 벽 안쪽으로 들어가고 검은 손잡이가 왼쪽으로 돌출되어 있다.",
        "built_space": "왼쪽 뒤에 큰 창면과 책상 하나, 업무용 의자 하나, 깃대 하나가 보인다. 책상 위에는 스피커폰 하나와 조명 하나가 있다. 후면에는 서류 선반 두 열과 그 아래·옆의 낮은 수납장들, 창 아래 서랍장 두 개가 보인다. 오른쪽 벽에는 지도 하나·화이트보드 하나·박힌 칼 하나가 있다. 박철진은 중앙 왼쪽 빈 바닥에, 윤성찬은 오른쪽 벽 가까이에 서 있다. 전경 가구는 박철진의 전신을 가리지 않으며 칼은 작은 배경 소품으로 유지된다. 창, 서류 수납부, 지도와 손상된 벽의 재질은 참조와 연결되지만 화면 밖이었던 수납 가구의 정확한 배치는 검증하기 어렵다.",
        "entities": "인물은 지정된 두 남성뿐이다. 박철진은 중년 동아시아계 남성으로 보이며 남색 전투복과 모자, 허리 주머니, 흰 문양이 있는 붉은 완장이 참조와 잘 맞는다. 숙인 얼굴 때문에 정면 얼굴의 동일성은 제한적으로만 확인된다. 윤성찬은 고령의 얼굴, 정돈된 은회색 머리와 콧수염, 짙은 정장과 흰 셔츠를 유지한다. 흰 손수건을 입 가까이 쥐고 긴장한 표정을 보인다. 칼은 벽에 남아 있고 박철진의 보이는 손은 비어 있으며, 스피커폰도 책상 위에 보인다.",
        "hard_violations": [],
        "physics": "박철진은 두 부츠를 바닥에 대고 다리와 골반으로 숙인 몸을 지탱한다. 상체는 A보다 수평에 가까워 더 깊은 인사지만, 어깨가 골반보다 높아 정확한 90도 굽힘에는 못 미친다. 손은 허벅지 옆에 자연스럽게 내려와 있다. 윤성찬도 두 발로 서 있고 손으로 손수건을 잡는다. 벽이 칼을, 책상 상판이 스피커폰을 지지하므로 떠 있는 물체는 없다. 윤성찬의 자세는 긴장되어 있으나 명확한 뒤쪽 체중 이동은 약하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전신 와이드 구도와 두 사람의 방향은 맞지만, 박철진의 상체가 크게 기울어진 채 남아 있어 핵심인 90도 인사의 최저점과 거리가 멀다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박철진의 가려지지 않은 전신, 더 깊은 인사, 오른쪽 윤성찬과 벽의 칼을 충실하게 배치했으나 정확한 90도 굽힘과 윤성찬의 뒤로 물러선 반응은 부족하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진은 화면 왼쪽에서 오른쪽 윤성찬을 향해 몸을 숙이고, 얼굴과 눈은 두 사람 사이의 바닥을 향한다. 윤성찬은 왼쪽 아래 박철진을 바라본다. 칼은 날이 오른쪽 벽에 박히고 손잡이가 왼쪽으로 돌출되어 있어 투척 후 상태에 맞는다.",
        "built_space": "뒤쪽에 큰 창면 하나와 책상 하나, 그 뒤 업무용 의자 하나가 있다. 왼쪽에는 소파 하나와 방문객용 의자 두 개가 보이고, 중앙 낮은 탁자 하나 위에 스피커폰 하나가 놓였다. 뒤쪽 오른편에는 책장 하나와 금속 수납장 하나, 오른쪽 벽에는 지도 하나·화이트보드 하나·박힌 칼 하나가 보인다. 두 사람은 가구 사이 빈 바닥에 서며 박철진의 머리부터 신발까지 가리지 않는다. 참조의 창틀, 서류 책장, 밝은 벽과 지도는 이어지지만, 참조의 좁은 화면만으로 새로 드러난 가구의 원래 위치까지 확인할 수는 없다.",
        "entities": "인물은 중년 동아시아계 남성 박철진과 고령 동아시아계 남성 윤성찬 두 명이다. 박철진의 남색 전투복, 챙 있는 모자와 허리 주머니는 인물 참조와 부합하며 얼굴은 숙인 측면만 보여 정확한 동일성 확인이 제한된다. 붉은 완장은 보이지 않아 확인할 수 없다. 윤성찬은 은회색 머리와 콧수염, 짙은 정장·흰 셔츠·어두운 넥타이로 이전 장면의 외양을 유지하고, 붉어진 얼굴 가까이에 흰 손수건을 들고 있다. 벽의 칼과 탁자의 스피커폰이 각각 하나씩 보이고 박철진은 칼을 들지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 신발이 바닥에 닿아 체중을 지지한다. 박철진은 골반에서 상체를 숙이고 손을 허벅지 옆에 내려 둔, 가능한 인사 자세다. 다만 상체가 수평보다 상당히 높아 허리 굽힘은 대략 50도 안팎으로 보이며 90도 인사의 최저점은 아니다. 윤성찬의 손수건은 손에 잡혀 있고, 칼은 벽에 박힌 날로, 스피커폰은 탁자 상판으로 지지된다. 윤성찬은 거의 곧게 서 있어 경계하며 뒤로 기대는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "박철진은 왼쪽에서 오른쪽 벽 앞의 윤성찬에게 인사하며 눈을 자신의 숙인 머리 아래 바닥으로 내린다. 윤성찬은 고개와 시선을 왼쪽 아래 박철진에게 향한다. 벽의 칼은 날이 벽 안쪽으로 들어가고 검은 손잡이가 왼쪽으로 돌출되어 있다.",
        "built_space": "왼쪽 뒤에 큰 창면과 책상 하나, 업무용 의자 하나, 깃대 하나가 보인다. 책상 위에는 스피커폰 하나와 조명 하나가 있다. 후면에는 서류 선반 두 열과 그 아래·옆의 낮은 수납장들, 창 아래 서랍장 두 개가 보인다. 오른쪽 벽에는 지도 하나·화이트보드 하나·박힌 칼 하나가 있다. 박철진은 중앙 왼쪽 빈 바닥에, 윤성찬은 오른쪽 벽 가까이에 서 있다. 전경 가구는 박철진의 전신을 가리지 않으며 칼은 작은 배경 소품으로 유지된다. 창, 서류 수납부, 지도와 손상된 벽의 재질은 참조와 연결되지만 화면 밖이었던 수납 가구의 정확한 배치는 검증하기 어렵다.",
        "entities": "인물은 지정된 두 남성뿐이다. 박철진은 중년 동아시아계 남성으로 보이며 남색 전투복과 모자, 허리 주머니, 흰 문양이 있는 붉은 완장이 참조와 잘 맞는다. 숙인 얼굴 때문에 정면 얼굴의 동일성은 제한적으로만 확인된다. 윤성찬은 고령의 얼굴, 정돈된 은회색 머리와 콧수염, 짙은 정장과 흰 셔츠를 유지한다. 흰 손수건을 입 가까이 쥐고 긴장한 표정을 보인다. 칼은 벽에 남아 있고 박철진의 보이는 손은 비어 있으며, 스피커폰도 책상 위에 보인다.",
        "hard_violations": [],
        "physics": "박철진은 두 부츠를 바닥에 대고 다리와 골반으로 숙인 몸을 지탱한다. 상체는 A보다 수평에 가까워 더 깊은 인사지만, 어깨가 골반보다 높아 정확한 90도 굽힘에는 못 미친다. 손은 허벅지 옆에 자연스럽게 내려와 있다. 윤성찬도 두 발로 서 있고 손으로 손수건을 잡는다. 벽이 칼을, 책상 상판이 스피커폰을 지지하므로 떠 있는 물체는 없다. 윤성찬의 자세는 긴장되어 있으나 명확한 뒤쪽 체중 이동은 약하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.464
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.464
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "붉은 완장, 90도 인사 자세, 윤성찬의 경계하는 태도 등 프롬프트와 레퍼런스를 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "전반적인 구도와 배경 디테일은 우수하나, 박철진의 필수 요소인 붉은 완장이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh7_sel.png",
    "asset_id": "be6063cb-385b-4fae-8aa5-fd219183f545",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1456670>",
    "asset_id": "cb968520-53a5-4b55-8260-e357fc83e9f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08e7-b2f9-790a-aa03-b469c425b200",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh7"
  },
  "staged_characters_added": [
   "C17"
  ]
 },
 "S19sh3::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 샷은 고급 세단의 내부를 배경으로 하며, 뒷좌석에 앉은 인물이 앞좌석의 운전자를 향해 지시를 내리는 장면입니다. 인물들이 차량 내부의 어느 좌석(앞좌석과 뒷좌석)에 위치해 있는지와 그들 간의 상대적인 방향 및 시선 처리가 매우 중요하므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "43247fc1e24a73fa"
 },
 "S19sh3::signage": {
  "fp": "0d28b1228c570156",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::ced2602cdeba": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_ced2602cdeba.png",
  "place_text": "Inside the rear passenger compartment of a luxury sedan, facing the chauffeur's front seat. Daylight enters through the side windows.",
  "input_fingerprint": "e98bb6b84c76dde1"
 },
 "S19sh3::confined_fp": {
  "reads": {
   "controls": "Steering wheel is attached to the right side of the front seat.",
   "mirrors": "No mirrors are indicated in the diagram.",
   "camera": "Positioned at the rear right passenger location, pointing left across the back row toward the rear left seat.",
   "occupants": "윤성찬의 비서 is in the front seat; 윤성찬 is in the rear left passenger seat."
  },
  "mismatches": [],
  "scene_description_en": "Shot from the rear right passenger area looking toward the rear left, the interior of a luxury sedan is visible. On the left side of the screen, Yoon Seong-chan (elderly man in a charcoal suit) is seen in profile, leaning forward with an open mouth. He points emphatically toward the right side of the frame. On the screen right, the rear of the front seat is visible in the background, showing only the back shoulder and edge of the head of the secretary, who faces forward. The bottom of the frame is grounded by the rear seating area, all lit by subdued ambient daylight.",
  "fixed": true,
  "input_fingerprint": "45e05a559b315557"
 },
 "era_assess::34e67fa5db2041ac": {
  "subjects": [],
  "subject_text": "윤성찬의 고급 세단 내부\n고급스럽게 마감된 밀폐형 객실. 앞쪽 운전석과 뒤쪽 좌석이 구분되며, 옆문 창과 전면 유리를 통해 외부가 보인다.",
  "identity": "canonical",
  "scope_id": "L31",
  "scope_role": "location_interior",
  "scope_sha": "9802432ba344c7c4"
 },
 "S19sh3": {
  "input_fingerprint": "d84fe07bd74eeee1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞좌석의 '윤성찬의 비서'를 향해 손가락을 뻗은 채 입을 열고 매섭게 지시하는 윤성찬의 측면.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury sedan, facing the chauffeur's front seat. Daylight enters through the side windows. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from the opposite rear-seat position at 윤성찬's seated upper-chest height, looking across his profile with a slight upward angle. Keep him left of center, leaning forward with his mouth open and pointing toward the front seat at screen right, where only 윤성찬의 비서's rear shoulder and head edge are visible; Yoon addresses the secretary's face beyond the crop while the secretary remains oriented toward the car's front. Emphasize the forward placement of Yoon's pointing hand without enlarging it through exaggerated perspective, and hold here before the camera slides between the seats.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Front seat (Occupied by the secretary) — Its rear-facing portion appears at the right edge from the opposite rear-seat camera position; used as Establishes the destination of Yoon's pointing gesture without blocking his hand; Rear seating area (Occupied by Yoon) — Seen laterally beneath and behind his seated body; used as Grounds his forward lean and makes the rear-to-front relationship legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination inside the car preserves Yoon's severe expression without adding an unsupported directional light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated inside the car after leaving the office and still carries his handkerchief. 윤성찬의 비서: He is seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nShot from the rear right passenger area looking toward the rear left, the interior of a luxury sedan is visible. On the left side of the screen, Yoon Seong-chan (elderly man in a charcoal suit) is seen in profile, leaning forward with an open mouth. He points emphatically toward the right side of the frame. On the screen right, the rear of the front seat is visible in the background, showing only the back shoulder and edge of the head of the secretary, who faces forward. The bottom of the frame is grounded by the rear seating area, all lit by subdued ambient daylight.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞좌석의 '윤성찬의 비서'를 향해 손가락을 뻗은 채 입을 열고 매섭게 지시하는 윤성찬의 측면.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury sedan, facing the chauffeur's front seat. Daylight enters through the side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination inside the car preserves Yoon's severe expression without adding an unsupported directional light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated inside the car after leaving the office and still carries his handkerchief. 윤성찬의 비서: He is seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nShot from the rear right passenger area looking toward the rear left, the interior of a luxury sedan is visible. On the left side of the screen, Yoon Seong-chan (elderly man in a charcoal suit) is seen in profile, leaning forward with an open mouth. He points emphatically toward the right side of the frame. On the screen right, the rear of the front seat is visible in the background, showing only the back shoulder and edge of the head of the secretary, who faces forward. The bottom of the frame is grounded by the rear seating area, all lit by subdued ambient daylight.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앞좌석의 '윤성찬의 비서'를 향해 손가락을 뻗은 채 입을 열고 매섭게 지시하는 윤성찬의 측면.\n\nLOCATION (lock): Inside the rear passenger compartment of a luxury sedan, facing the chauffeur's front seat. Daylight enters through the side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination inside the car preserves Yoon's severe expression without adding an unsupported directional light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬: He is seated inside the car after leaving the office and still carries his handkerchief. 윤성찬의 비서: He is seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    },
    {
     "label": "윤성찬의 비서",
     "path": "<bytes:1155240>",
     "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh3_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    },
    {
     "label": "윤성찬의 비서",
     "path": "<bytes:1155240>",
     "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽의 윤성찬이 고개를 돌려 화면 오른쪽 앞좌석(운전석)에 앉은 비서를 향해 시선을 두고 있으며, 뻗은 오른손의 검지손가락 역시 비서의 머리와 어깨 쪽을 정확히 가리키고 있음.",
    "built_space": "고급 세단의 뒷좌석 내부. 평면도와 지시문에 따라 카메라가 우측 뒷좌석에 위치하여 좌측 뒷좌석의 윤성찬의 측면을 담고 있으며, 우측 전방에는 운전석의 뒷면과 비서의 뒷모습이 올바르게 배치됨. 중앙 콘솔과 창문 등의 형태와 위치가 자연스러움.",
    "entities": "윤성찬은 참조 이미지에 맞는 노인의 얼굴과 콧수염을 지니고 있으며, 차콜 그레이 정장과 흰색 손수건 등 의상과 소품 설정이 정확히 일치함. 비서 역시 검은 머리의 남성으로 올바르게 묘사됨.",
    "hard_violations": [],
    "physics": "윤성찬은 엉덩이와 등을 시트에 대고 안정적으로 앉아 있으며, 손수건을 쥔 왼손은 무릎 위에 올려져 체중과 중력을 자연스럽게 받고 있음. 가리키는 오른팔 역시 어깨를 축으로 물리적으로 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "화면 왼쪽의 윤성찬이 화면 오른쪽 전방의 비서를 향해 시선을 두고 오른손을 뻗어 가리키고 있음.",
    "built_space": "고급 차량의 내부 공간으로, 우측 뒷좌석에서 좌측 뒷좌석 인물과 전방 운전석을 바라보는 카메라 구도와 공간 배치는 프롬프트의 지시와 평면도를 잘 따르고 있음.",
    "entities": "윤성찬과 비서의 전반적인 외형과 위치는 프롬프트와 일치하며 손수건 소품도 존재함. 그러나 윤성찬의 가슴팍(넥타이 옆)에 프롬프트에 없는 얇은 천 가닥이 추가로 생성되어 의상 구조가 훼손됨.",
    "hard_violations": [
     "[gemini-pro] 윤성찬의 넥타이 옆에 생성된 정체불명의 얇은 천 (발명된 사물)",
     "[gemini-pro] 가리키는 오른손 검지손가락의 비정상적인 길이와 뭉개진 나머지 손가락들, 그리고 손수건을 쥔 왼손의 기형적인 관절 (물리적으로 불가능한 해부학)"
    ],
    "physics": "인물이 좌석에 앉아 있는 전반적인 무게 중심은 유지되고 있으나, 오른손과 왼손의 손가락 구조가 정상적인 인체의 물리적 뼈대와 관절 형태를 벗어나 심하게 왜곡되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 우측 뒷좌석에서의 카메라 구도, 인물의 공간적 위치 관계, 역동적으로 지시하는 동작과 디테일(손수건)을 해부학적 오류 없이 완벽하게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "공간 배치와 카메라 구도는 프롬프트를 잘 따랐으나, 손가락의 심각한 해부학적 왜곡과 넥타이 부근의 존재하지 않는 의상 요소가 추가되는 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽의 윤성찬이 고개를 돌려 화면 오른쪽 앞좌석(운전석)에 앉은 비서를 향해 시선을 두고 있으며, 뻗은 오른손의 검지손가락 역시 비서의 머리와 어깨 쪽을 정확히 가리키고 있음.",
        "built_space": "고급 세단의 뒷좌석 내부. 평면도와 지시문에 따라 카메라가 우측 뒷좌석에 위치하여 좌측 뒷좌석의 윤성찬의 측면을 담고 있으며, 우측 전방에는 운전석의 뒷면과 비서의 뒷모습이 올바르게 배치됨. 중앙 콘솔과 창문 등의 형태와 위치가 자연스러움.",
        "entities": "윤성찬은 참조 이미지에 맞는 노인의 얼굴과 콧수염을 지니고 있으며, 차콜 그레이 정장과 흰색 손수건 등 의상과 소품 설정이 정확히 일치함. 비서 역시 검은 머리의 남성으로 올바르게 묘사됨.",
        "hard_violations": [],
        "physics": "윤성찬은 엉덩이와 등을 시트에 대고 안정적으로 앉아 있으며, 손수건을 쥔 왼손은 무릎 위에 올려져 체중과 중력을 자연스럽게 받고 있음. 가리키는 오른팔 역시 어깨를 축으로 물리적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽의 윤성찬이 화면 오른쪽 전방의 비서를 향해 시선을 두고 오른손을 뻗어 가리키고 있음.",
        "built_space": "고급 차량의 내부 공간으로, 우측 뒷좌석에서 좌측 뒷좌석 인물과 전방 운전석을 바라보는 카메라 구도와 공간 배치는 프롬프트의 지시와 평면도를 잘 따르고 있음.",
        "entities": "윤성찬과 비서의 전반적인 외형과 위치는 프롬프트와 일치하며 손수건 소품도 존재함. 그러나 윤성찬의 가슴팍(넥타이 옆)에 프롬프트에 없는 얇은 천 가닥이 추가로 생성되어 의상 구조가 훼손됨.",
        "hard_violations": [
         "윤성찬의 넥타이 옆에 생성된 정체불명의 얇은 천 (발명된 사물)",
         "가리키는 오른손 검지손가락의 비정상적인 길이와 뭉개진 나머지 손가락들, 그리고 손수건을 쥔 왼손의 기형적인 관절 (물리적으로 불가능한 해부학)"
        ],
        "physics": "인물이 좌석에 앉아 있는 전반적인 무게 중심은 유지되고 있으나, 오른손과 왼손의 손가락 구조가 정상적인 인체의 물리적 뼈대와 관절 형태를 벗어나 심하게 왜곡되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 우측 뒷좌석에서의 카메라 구도, 인물의 공간적 위치 관계, 역동적으로 지시하는 동작과 디테일(손수건)을 해부학적 오류 없이 완벽하게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "공간 배치와 카메라 구도는 프롬프트를 잘 따랐으나, 손가락의 심각한 해부학적 왜곡과 넥타이 부근의 존재하지 않는 의상 요소가 추가되는 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽의 윤성찬이 고개를 돌려 화면 오른쪽 앞좌석(운전석)에 앉은 비서를 향해 시선을 두고 있으며, 뻗은 오른손의 검지손가락 역시 비서의 머리와 어깨 쪽을 정확히 가리키고 있음.",
        "built_space": "고급 세단의 뒷좌석 내부. 평면도와 지시문에 따라 카메라가 우측 뒷좌석에 위치하여 좌측 뒷좌석의 윤성찬의 측면을 담고 있으며, 우측 전방에는 운전석의 뒷면과 비서의 뒷모습이 올바르게 배치됨. 중앙 콘솔과 창문 등의 형태와 위치가 자연스러움.",
        "entities": "윤성찬은 참조 이미지에 맞는 노인의 얼굴과 콧수염을 지니고 있으며, 차콜 그레이 정장과 흰색 손수건 등 의상과 소품 설정이 정확히 일치함. 비서 역시 검은 머리의 남성으로 올바르게 묘사됨.",
        "hard_violations": [],
        "physics": "윤성찬은 엉덩이와 등을 시트에 대고 안정적으로 앉아 있으며, 손수건을 쥔 왼손은 무릎 위에 올려져 체중과 중력을 자연스럽게 받고 있음. 가리키는 오른팔 역시 어깨를 축으로 물리적으로 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽의 윤성찬이 화면 오른쪽 전방의 비서를 향해 시선을 두고 오른손을 뻗어 가리키고 있음.",
        "built_space": "고급 차량의 내부 공간으로, 우측 뒷좌석에서 좌측 뒷좌석 인물과 전방 운전석을 바라보는 카메라 구도와 공간 배치는 프롬프트의 지시와 평면도를 잘 따르고 있음.",
        "entities": "윤성찬과 비서의 전반적인 외형과 위치는 프롬프트와 일치하며 손수건 소품도 존재함. 그러나 윤성찬의 가슴팍(넥타이 옆)에 프롬프트에 없는 얇은 천 가닥이 추가로 생성되어 의상 구조가 훼손됨.",
        "hard_violations": [
         "윤성찬의 넥타이 옆에 생성된 정체불명의 얇은 천 (발명된 사물)",
         "가리키는 오른손 검지손가락의 비정상적인 길이와 뭉개진 나머지 손가락들, 그리고 손수건을 쥔 왼손의 기형적인 관절 (물리적으로 불가능한 해부학)"
        ],
        "physics": "인물이 좌석에 앉아 있는 전반적인 무게 중심은 유지되고 있으나, 오른손과 왼손의 손가락 구조가 정상적인 인체의 물리적 뼈대와 관절 형태를 벗어나 심하게 왜곡되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "B보다 전방으로 숙인 자세와 측면성이 강하고 손가락도 비서의 목·머리 방향에 가깝지만, 요구한 미디엄 숏보다 넓고 앞좌석이 오른쪽 가장자리를 넘어 크게 들어온다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "입을 열고 지시하는 행동은 맞지만, 손가락이 비서의 얼굴보다 낮은 등받이·어깨 쪽을 향하며 몸의 전방 기울기와 측면 구도가 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 오른쪽 앞좌석의 비서를 바라보며 입을 열고 있다. 뻗은 검지는 오른쪽으로 약간 올라가며, 연장하면 비서의 목 아래·어깨 윗부분 쪽에 닿는다. 얼굴 자체를 정확히 겨냥하지는 않지만 지시 대상은 비서로 읽힌다. 비서는 뒤돌아보지 않고 차량 전방을 향한다.",
        "built_space": "왼쪽에 윤성찬을 받치는 뒷좌석과 머리받침 하나, 오른쪽 중경에 앞좌석 등받이와 머리받침 하나, 오른쪽 전경에 다른 앞좌석의 일부가 보인다. 측면 창과 문 손잡이 하나, 스피커 하나, 천장 손잡이 두 개, 하단 송풍구가 보이며 앞뒤 좌석 관계는 성립한다. 다만 앞좌석 등받이가 가장자리에만 걸리지 않고 화면 안쪽을 상당히 차지한다. 카메라는 반대편 뒷좌석에서 가로질러 보는 위치에 가깝지만 명확한 약한 올려다보기는 아니며, 허벅지와 무릎까지 포함해 지정 미디엄 숏보다 넓다. 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 윤성찬과 비서 두 명뿐이다. 윤성찬은 참고와 유사한 노년 동아시아 남성의 얼굴, 짧은 흰머리와 콧수염, 차콜 정장과 넥타이를 갖췄고 왼손에 흰 손수건을 쥐고 있다. 비서는 짧은 검은 머리의 성인 남성이지만 얼굴은 가려져 신원 세부를 확인할 수 없다. 비서의 검은 정장과 흰 셔츠 깃은 참고의 남색 셔츠와 다르다. 가죽과 목재 장식의 세단 실내, 낮의 창밖 빛은 조건에 맞으며 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "윤성찬의 골반과 허벅지는 뒷좌석 쿠션에 놓여 있고 상체만 앞으로 기울어 있다. 지시하는 팔은 어깨와 팔꿈치로 자연스럽게 연결되며 손수건은 반대 손에 잡혀 있다. 비서의 하체는 앞좌석에 가려져 있지만 상체 위치는 착석으로 설명된다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 눈과 열린 입은 오른쪽 비서를 향한다. 검지는 거의 수평으로 뻗어 있어 연장선이 앞좌석 등받이 위쪽과 비서의 어깨·등 높이를 지나며, 화면 밖 얼굴을 향하는 방향은 A보다 약하다. 비서는 차량 전방을 향하고 있다.",
        "built_space": "왼쪽에 뒷좌석과 일부 머리받침 하나, 오른쪽 중경에 앞좌석 등받이와 머리받침 하나, 맨 오른쪽 전경에 다른 앞좌석의 머리받침과 등받이 일부가 있다. 두 앞좌석 자체는 정상적인 세단 구성이다. 측면 문 손잡이 하나와 스피커 하나, 창문, 하단 송풍구가 보인다. 윤성찬은 뒷좌석에, 비서는 앞좌석에 놓여 있으나 비서의 등과 어깨가 요구한 가장자리 노출보다 넓게 보인다. 윤성찬의 무릎과 넓은 좌석 전경까지 담아 지정 미디엄 숏보다 넓고, 약한 올려다보기보다는 수평에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "윤성찬과 비서 외의 인물은 없다. 윤성찬의 노년 얼굴, 짧게 정돈한 흰머리와 콧수염, 차콜 정장, 넥타이는 참고와 대체로 일치하며 흰 손수건도 왼손에 있다. 비서는 짧은 검은 머리의 성인 남성으로 보이나 뒷모습만으로 얼굴 일치는 확인할 수 없다. 비서의 정장과 흰 셔츠는 참고의 남색 셔츠와 다르다. 실내 재질과 주간 자연광은 요청에 맞으며 문자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "윤성찬의 골반과 다리는 뒷좌석에 지지되며, 상체를 약간 기울여 팔을 뻗는 동작은 가능하다. 손수건은 왼손이 직접 잡고 있다. 비서는 앞좌석 등받이 앞에 앉아 있고 어깨 뒤로 안전벨트 일부가 보인다. 인체 연결이나 물체 지지에서 명백한 물리적 불가능은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "B보다 전방으로 숙인 자세와 측면성이 강하고 손가락도 비서의 목·머리 방향에 가깝지만, 요구한 미디엄 숏보다 넓고 앞좌석이 오른쪽 가장자리를 넘어 크게 들어온다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "입을 열고 지시하는 행동은 맞지만, 손가락이 비서의 얼굴보다 낮은 등받이·어깨 쪽을 향하며 몸의 전방 기울기와 측면 구도가 A보다 약하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬은 오른쪽 앞좌석의 비서를 바라보며 입을 열고 있다. 뻗은 검지는 오른쪽으로 약간 올라가며, 연장하면 비서의 목 아래·어깨 윗부분 쪽에 닿는다. 얼굴 자체를 정확히 겨냥하지는 않지만 지시 대상은 비서로 읽힌다. 비서는 뒤돌아보지 않고 차량 전방을 향한다.",
        "built_space": "왼쪽에 윤성찬을 받치는 뒷좌석과 머리받침 하나, 오른쪽 중경에 앞좌석 등받이와 머리받침 하나, 오른쪽 전경에 다른 앞좌석의 일부가 보인다. 측면 창과 문 손잡이 하나, 스피커 하나, 천장 손잡이 두 개, 하단 송풍구가 보이며 앞뒤 좌석 관계는 성립한다. 다만 앞좌석 등받이가 가장자리에만 걸리지 않고 화면 안쪽을 상당히 차지한다. 카메라는 반대편 뒷좌석에서 가로질러 보는 위치에 가깝지만 명확한 약한 올려다보기는 아니며, 허벅지와 무릎까지 포함해 지정 미디엄 숏보다 넓다. 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 윤성찬과 비서 두 명뿐이다. 윤성찬은 참고와 유사한 노년 동아시아 남성의 얼굴, 짧은 흰머리와 콧수염, 차콜 정장과 넥타이를 갖췄고 왼손에 흰 손수건을 쥐고 있다. 비서는 짧은 검은 머리의 성인 남성이지만 얼굴은 가려져 신원 세부를 확인할 수 없다. 비서의 검은 정장과 흰 셔츠 깃은 참고의 남색 셔츠와 다르다. 가죽과 목재 장식의 세단 실내, 낮의 창밖 빛은 조건에 맞으며 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "윤성찬의 골반과 허벅지는 뒷좌석 쿠션에 놓여 있고 상체만 앞으로 기울어 있다. 지시하는 팔은 어깨와 팔꿈치로 자연스럽게 연결되며 손수건은 반대 손에 잡혀 있다. 비서의 하체는 앞좌석에 가려져 있지만 상체 위치는 착석으로 설명된다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 눈과 열린 입은 오른쪽 비서를 향한다. 검지는 거의 수평으로 뻗어 있어 연장선이 앞좌석 등받이 위쪽과 비서의 어깨·등 높이를 지나며, 화면 밖 얼굴을 향하는 방향은 A보다 약하다. 비서는 차량 전방을 향하고 있다.",
        "built_space": "왼쪽에 뒷좌석과 일부 머리받침 하나, 오른쪽 중경에 앞좌석 등받이와 머리받침 하나, 맨 오른쪽 전경에 다른 앞좌석의 머리받침과 등받이 일부가 있다. 두 앞좌석 자체는 정상적인 세단 구성이다. 측면 문 손잡이 하나와 스피커 하나, 창문, 하단 송풍구가 보인다. 윤성찬은 뒷좌석에, 비서는 앞좌석에 놓여 있으나 비서의 등과 어깨가 요구한 가장자리 노출보다 넓게 보인다. 윤성찬의 무릎과 넓은 좌석 전경까지 담아 지정 미디엄 숏보다 넓고, 약한 올려다보기보다는 수평에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "윤성찬과 비서 외의 인물은 없다. 윤성찬의 노년 얼굴, 짧게 정돈한 흰머리와 콧수염, 차콜 정장, 넥타이는 참고와 대체로 일치하며 흰 손수건도 왼손에 있다. 비서는 짧은 검은 머리의 성인 남성으로 보이나 뒷모습만으로 얼굴 일치는 확인할 수 없다. 비서의 정장과 흰 셔츠는 참고의 남색 셔츠와 다르다. 실내 재질과 주간 자연광은 요청에 맞으며 문자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "윤성찬의 골반과 다리는 뒷좌석에 지지되며, 상체를 약간 기울여 팔을 뻗는 동작은 가능하다. 손수건은 왼손이 직접 잡고 있다. 비서는 앞좌석 등받이 앞에 앉아 있고 어깨 뒤로 안전벨트 일부가 보인다. 인체 연결이나 물체 지지에서 명백한 물리적 불가능은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 윤성찬의 넥타이 옆에 생성된 정체불명의 얇은 천 (발명된 사물)",
     "[gemini-pro] 가리키는 오른손 검지손가락의 비정상적인 길이와 뭉개진 나머지 손가락들, 그리고 손수건을 쥔 왼손의 기형적인 관절 (물리적으로 불가능한 해부학)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "프롬프트가 요구한 우측 뒷좌석에서의 카메라 구도, 인물의 공간적 위치 관계, 역동적으로 지시하는 동작과 디테일(손수건)을 해부학적 오류 없이 완벽하게 구현한 훌륭한 결과물입니다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "공간 배치와 카메라 구도는 프롬프트를 잘 따랐으나, 손가락의 심각한 해부학적 왜곡과 넥타이 부근의 존재하지 않는 의상 요소가 추가되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 윤성찬의 넥타이 옆에 생성된 정체불명의 얇은 천 (발명된 사물) / [gemini-pro] 가리키는 오른손 검지손가락의 비정상적인 길이와 뭉개진 나머지 손가락들, 그리고 손수건을 쥔 왼손의 기형적인 관절 (물리적으로 불가능한 해부학)"
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh3_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "윤성찬",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   },
   {
    "label": "윤성찬의 비서",
    "path": "<bytes:1155240>",
    "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08ec-c0b3-714c-81ab-98a7346d0408",
  "confined_fp": {
   "base_key": "confinedfp::ced2602cdeba",
   "apt_reason": "이 샷은 고급 세단의 내부를 배경으로 하며, 뒷좌석에 앉은 인물이 앞좌석의 운전자를 향해 지시를 내리는 장면입니다. 인물들이 차량 내부의 어느 좌석(앞좌석과 뒷좌석)에 위치해 있는지와 그들 간의 상대적인 방향 및 시선 처리가 매우 중요하므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C18"
  ]
 },
 "S19sh4::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 숏은 차량 내부 운전석에서 진행되며, 운전자인 비서가 뒷좌석의 승객을 돌아보는 상황이므로 차량 내 인물들의 좌석 배치와 시선 방향이 정확하게 표현되어야 합니다.",
  "input_fingerprint": "3e44b5e85d7ffea0"
 },
 "S19sh4::signage": {
  "fp": "9eedebc9f4056749",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::557950ae2d7a": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_557950ae2d7a.png",
  "place_text": "At the chauffeur's position inside the luxury sedan, directly ahead of the rear passenger compartment. Daylight enters through the windshield and side windows.",
  "input_fingerprint": "cd1379b6de306db0"
 },
 "S19sh4::confined_fp": {
  "reads": {
   "controls": "A steering wheel is attached to the front-left driver's seat.",
   "mirrors": "No mirrors or reflective surfaces are depicted in the diagram.",
   "camera": "The camera is positioned in the center of the rear passenger seating area, pointing diagonally forward and left toward the driver's seat.",
   "occupants": "The front-left driver's seat is occupied by '윤성찬의 비서'."
  },
  "mismatches": [],
  "scene_description_en": "The camera sits in the center rear of the car, looking diagonally forward and left directly at the driver's seat. The driver, '윤성찬의 비서', occupies this front-left position and appears in the near-midground on the right half of the screen, his head turned sharply backward to face the left edge. The steering wheel is attached to the dashboard in front of him, appearing in the background on the left side of the screen and facing rearward toward the seat. The windshield spans the far background across the upper left, facing inward to let in daylight. No mirrors or reflective surfaces are present in this view.",
  "fixed": false,
  "input_fingerprint": "2e3eb75c086b7ad0"
 },
 "S19sh4": {
  "input_fingerprint": "5dbe9f0f416f11c4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 지시에 깜짝 놀라 고개를 뒤로 돌린 채 굳어 있는 '윤성찬의 비서'의 당혹스러운 얼굴.\n\nLOCATION (lock): At the chauffeur's position inside the luxury sedan, directly ahead of the rear passenger compartment. Daylight enters through the windshield and side windows. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a direct observational close view from near the center of the rear seating area, diagonally behind the secretary and slightly above seated eye level, looking gently downward. 윤성찬의 비서's turned face occupies the right half in three-quarter view, with a narrow seat edge establishing the forward-facing body beneath the backward turn; their startled gaze stays on 윤성찬 outside the left edge. Keep the camera motionless at this reaction-stage midpoint, allowing the arrested neck turn and parted expression to carry the shock.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Front seat (Occupied by the secretary) — A narrow portion of the rear-facing seat edge appears beneath the turned face; used as Establishes the mismatch between the forward-facing seated body and the backward reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination inside the car remains subdued, with controlled facial contrast and no stylized perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬의 비서 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬의 비서: He remains seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera sits in the center rear of the car, looking diagonally forward and left directly at the driver's seat. The driver, '윤성찬의 비서', occupies this front-left position and appears in the near-midground on the right half of the screen, his head turned sharply backward to face the left edge. The steering wheel is attached to the dashboard in front of him, appearing in the background on the left side of the screen and facing rearward toward the seat. The windshield spans the far background across the upper left, facing inward to let in daylight. No mirrors or reflective surfaces are present in this view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 지시에 깜짝 놀라 고개를 뒤로 돌린 채 굳어 있는 '윤성찬의 비서'의 당혹스러운 얼굴.\n\nLOCATION (lock): At the chauffeur's position inside the luxury sedan, directly ahead of the rear passenger compartment. Daylight enters through the windshield and side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination inside the car remains subdued, with controlled facial contrast and no stylized perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬의 비서: He remains seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera sits in the center rear of the car, looking diagonally forward and left directly at the driver's seat. The driver, '윤성찬의 비서', occupies this front-left position and appears in the near-midground on the right half of the screen, his head turned sharply backward to face the left edge. The steering wheel is attached to the dashboard in front of him, appearing in the background on the left side of the screen and facing rearward toward the seat. The windshield spans the far background across the upper left, facing inward to let in daylight. No mirrors or reflective surfaces are present in this view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 지시에 깜짝 놀라 고개를 뒤로 돌린 채 굳어 있는 '윤성찬의 비서'의 당혹스러운 얼굴.\n\nLOCATION (lock): At the chauffeur's position inside the luxury sedan, directly ahead of the rear passenger compartment. Daylight enters through the windshield and side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination inside the car remains subdued, with controlled facial contrast and no stylized perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 윤성찬의 비서: He remains seated in the front of the car as its driver.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬의 비서 (한국인 남성, 성인의 얼굴, 정돈된 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh4_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬의 비서",
     "path": "<bytes:1155240>",
     "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh4_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "윤성찬의 비서",
     "path": "<bytes:1155240>",
     "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "운전석에 앉은 윤성찬의 비서의 시선이 화면 왼쪽 가장자리 밖(윤성찬이 있는 타깃 위치)을 정확히 향하고 있다.",
    "built_space": "뒷좌석 중앙에서 좌측 전방을 바라보는 앵글로, 화면 왼쪽에 운전대와 대시보드, 중앙에 운전석의 인물, 오른쪽 전경에 조수석 등받이가 좌핸들(LHD) 규격에 맞게 올바른 구조로 배치되어 있다.",
    "entities": "윤성찬의 비서(짧고 정돈된 검은 머리의 한국인 남성, 네이비 버튼업 셔츠)가 레퍼런스와 동일한 외모와 복장으로 정확히 묘사되었다.",
    "hard_violations": [
     "[gpt-high] 비서가 우측 앞좌석의 운전대를 사용하는 구조로 묘사되어, 좌측 운전석을 고정한 도면 및 후석 중앙에서 보는 카메라 관계와 반대입니다."
    ],
    "physics": "운전자의 몸통은 앞을 향한 상태(셔츠의 등판이 보임)에서 고개를 왼쪽 어깨 너머로 돌린 자세가 해부학적으로 자연스럽게 구현되었고, 보이지 않는 왼팔이 운전대에 닿아 지지되는 구조가 사실적이다."
   },
   {
    "label": "B",
    "direction": "운전석에 앉은 인물의 시선이 화면 왼쪽 가장자리를 향하고 있다.",
    "built_space": "차량 내부 구조가 왜곡되어 운전대 왼쪽에 내비게이션 화면이 위치하는 등 좌핸들 차량 구조와 평면도 레퍼런스를 위반한 물리적으로 불가능한 배치를 보여준다.",
    "entities": "윤성찬의 비서의 얼굴, 헤어스타일, 네이비 셔츠 의상이 레퍼런스와 잘 일치한다.",
    "hard_violations": [
     "[gemini-pro] 전방을 향한 몸(forward-facing body)이어야 한다는 지문을 어기고 상체 전체가 뒤(카메라 쪽)를 돌아보아 셔츠 앞면 단추가 보이는 잘못된 자세",
     "[gemini-pro] 차량 내부 중앙에 있어야 할 센터 콘솔 스크린이 운전대 왼쪽에 위치하는 물리적으로 불가능한 구조",
     "[gpt-high] 좌측 운전석 도면과 달리 중앙 장치 왼쪽·운전자 측 창문 오른쪽의 우측 운전석 구조로 배치하여, 비서의 착석 위치와 후석 중앙 카메라의 관찰 관계가 좌우 반전되었습니다."
    ],
    "physics": "몸통이 카메라(차량 뒤쪽)를 향하고 있음에도 고개를 부자연스럽게 비틀어 뒤를 돌아보는 척하는 해부학적 오류가 있으며, 운전대를 잡은 손의 구조도 모순된다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에서 요구한 전방을 향한 몸통과 왼쪽 어깨 너머로 뒤를 돌아보는 고개의 자세, 좌핸들 차량의 내부 구조와 카메라 앵글을 완벽하게 구현한 탁월한 샷입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "몸통이 앞을 향해야 한다는 지시를 어기고 셔츠 앞면이 카메라를 향한 해부학적 오류를 범했으며, 운전대 왼쪽에 센터 콘솔이 배치되는 불가능한 차량 구조를 그렸습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "운전석에 앉은 윤성찬의 비서의 시선이 화면 왼쪽 가장자리 밖(윤성찬이 있는 타깃 위치)을 정확히 향하고 있다.",
        "built_space": "뒷좌석 중앙에서 좌측 전방을 바라보는 앵글로, 화면 왼쪽에 운전대와 대시보드, 중앙에 운전석의 인물, 오른쪽 전경에 조수석 등받이가 좌핸들(LHD) 규격에 맞게 올바른 구조로 배치되어 있다.",
        "entities": "윤성찬의 비서(짧고 정돈된 검은 머리의 한국인 남성, 네이비 버튼업 셔츠)가 레퍼런스와 동일한 외모와 복장으로 정확히 묘사되었다.",
        "hard_violations": [],
        "physics": "운전자의 몸통은 앞을 향한 상태(셔츠의 등판이 보임)에서 고개를 왼쪽 어깨 너머로 돌린 자세가 해부학적으로 자연스럽게 구현되었고, 보이지 않는 왼팔이 운전대에 닿아 지지되는 구조가 사실적이다."
       },
       {
        "label": "B",
        "direction": "운전석에 앉은 인물의 시선이 화면 왼쪽 가장자리를 향하고 있다.",
        "built_space": "차량 내부 구조가 왜곡되어 운전대 왼쪽에 내비게이션 화면이 위치하는 등 좌핸들 차량 구조와 평면도 레퍼런스를 위반한 물리적으로 불가능한 배치를 보여준다.",
        "entities": "윤성찬의 비서의 얼굴, 헤어스타일, 네이비 셔츠 의상이 레퍼런스와 잘 일치한다.",
        "hard_violations": [
         "전방을 향한 몸(forward-facing body)이어야 한다는 지문을 어기고 상체 전체가 뒤(카메라 쪽)를 돌아보아 셔츠 앞면 단추가 보이는 잘못된 자세",
         "차량 내부 중앙에 있어야 할 센터 콘솔 스크린이 운전대 왼쪽에 위치하는 물리적으로 불가능한 구조"
        ],
        "physics": "몸통이 카메라(차량 뒤쪽)를 향하고 있음에도 고개를 부자연스럽게 비틀어 뒤를 돌아보는 척하는 해부학적 오류가 있으며, 운전대를 잡은 손의 구조도 모순된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에서 요구한 전방을 향한 몸통과 왼쪽 어깨 너머로 뒤를 돌아보는 고개의 자세, 좌핸들 차량의 내부 구조와 카메라 앵글을 완벽하게 구현한 탁월한 샷입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "몸통이 앞을 향해야 한다는 지시를 어기고 셔츠 앞면이 카메라를 향한 해부학적 오류를 범했으며, 운전대 왼쪽에 센터 콘솔이 배치되는 불가능한 차량 구조를 그렸습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "운전석에 앉은 윤성찬의 비서의 시선이 화면 왼쪽 가장자리 밖(윤성찬이 있는 타깃 위치)을 정확히 향하고 있다.",
        "built_space": "뒷좌석 중앙에서 좌측 전방을 바라보는 앵글로, 화면 왼쪽에 운전대와 대시보드, 중앙에 운전석의 인물, 오른쪽 전경에 조수석 등받이가 좌핸들(LHD) 규격에 맞게 올바른 구조로 배치되어 있다.",
        "entities": "윤성찬의 비서(짧고 정돈된 검은 머리의 한국인 남성, 네이비 버튼업 셔츠)가 레퍼런스와 동일한 외모와 복장으로 정확히 묘사되었다.",
        "hard_violations": [],
        "physics": "운전자의 몸통은 앞을 향한 상태(셔츠의 등판이 보임)에서 고개를 왼쪽 어깨 너머로 돌린 자세가 해부학적으로 자연스럽게 구현되었고, 보이지 않는 왼팔이 운전대에 닿아 지지되는 구조가 사실적이다."
       },
       {
        "label": "B",
        "direction": "운전석에 앉은 인물의 시선이 화면 왼쪽 가장자리를 향하고 있다.",
        "built_space": "차량 내부 구조가 왜곡되어 운전대 왼쪽에 내비게이션 화면이 위치하는 등 좌핸들 차량 구조와 평면도 레퍼런스를 위반한 물리적으로 불가능한 배치를 보여준다.",
        "entities": "윤성찬의 비서의 얼굴, 헤어스타일, 네이비 셔츠 의상이 레퍼런스와 잘 일치한다.",
        "hard_violations": [
         "전방을 향한 몸(forward-facing body)이어야 한다는 지문을 어기고 상체 전체가 뒤(카메라 쪽)를 돌아보아 셔츠 앞면 단추가 보이는 잘못된 자세",
         "차량 내부 중앙에 있어야 할 센터 콘솔 스크린이 운전대 왼쪽에 위치하는 물리적으로 불가능한 구조"
        ],
        "physics": "몸통이 카메라(차량 뒤쪽)를 향하고 있음에도 고개를 부자연스럽게 비틀어 뒤를 돌아보는 척하는 해부학적 오류가 있으며, 운전대를 잡은 손의 구조도 모순된다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "놀라 뒤돌아본 표정과 인물 외형은 맞지만, 시선이 화면 오른쪽으로 향하고 운전석 배치가 도면과 반대이며 상반신·계기판까지 넓게 담아 지정 클로즈업에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "A보다 얼굴을 크게 담아 반응 클로즈업에 가깝지만, 화면 오른쪽을 보는 시선과 좌우가 뒤집힌 운전석 배치는 동일하게 실패하며 좌석도 좁은 가장자리 이상으로 드러납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "비서는 몸을 운전대 쪽에 둔 채 고개를 뒤로 돌렸으나, 두 눈은 화면 오른쪽 바깥을 향합니다. 지정된 화면 왼쪽 밖의 윤성찬을 보는 시선이 아닙니다. 벌어진 입과 올라간 눈썹은 놀란 반응을 나타냅니다.",
        "built_space": "운전대 1개, 계기판 1개, 왼쪽 중앙 디스플레이 1개, 룸미러 1개, 큰 운전석 머리받침 1개와 등받이가 보입니다. 중앙 디스플레이가 운전대 왼쪽에 있고 운전자 옆 창문이 오른쪽에 있어 우측 운전석 배치로 읽히며, 좌측 운전석을 지정한 도면과 반대입니다. 카메라도 운전자의 왼쪽 뒤에서 보는 관계입니다. 좌석은 좁은 가장자리가 아니라 화면 오른쪽과 하단을 크게 차지합니다. 룸미러에는 식별 가능한 인물 반사가 없습니다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 한국인 남성이라는 설정과 충돌하는 외형은 없습니다. 짧게 정돈한 검은 머리, 얼굴 생김새와 남색 셔츠는 인물 참조에 대체로 부합합니다. 윤성찬이나 다른 사람은 등장하지 않습니다. 가죽 좌석과 목재 장식은 고급 세단으로 읽힙니다. 별도의 이전 장면 사진은 제공되지 않아 정확한 내장재 연속성은 확인할 수 없습니다.",
        "hard_violations": [
         "좌측 운전석 도면과 달리 중앙 장치 왼쪽·운전자 측 창문 오른쪽의 우측 운전석 구조로 배치하여, 비서의 착석 위치와 후석 중앙 카메라의 관찰 관계가 좌우 반전되었습니다."
        ],
        "physics": "보이는 등과 어깨는 좌석에 앉은 몸으로 자연스럽게 이어지고, 한 손은 운전대 아래쪽 테두리를 감싸 쥐고 있습니다. 몸통을 일부 틀며 뒤돌아보는 목의 자세는 가능한 동작입니다. 엉덩이와 발은 프레임 밖이지만 공중에 뜬 몸이나 지지 없는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "비서는 운전대를 향한 몸 위로 고개를 뒤돌리고 눈을 화면 오른쪽 바깥으로 보냅니다. 화면 왼쪽 밖에 있어야 하는 윤성찬에게 시선이 향하지 않습니다. 입술이 벌어지고 눈썹이 들려 당혹스러운 정지 순간은 드러납니다.",
        "built_space": "운전대 1개, 계기판 1개, 오른쪽의 운전석 머리받침 1개와 등받이, 왼쪽 아래의 다른 앞좌석 일부가 보입니다. 운전자 바로 오른쪽에 측면 창문이 있고 다른 앞좌석이 왼쪽에 있어 도면의 좌측 운전석과 반대입니다. 얼굴은 A보다 크고 오른쪽에 놓였지만, 큰 머리받침과 넓은 등받이 및 운전대까지 보여 지정된 좁은 좌석 가장자리 구성에는 못 미칩니다. 불가능한 반사는 보이지 않습니다.",
        "entities": "참조와 유사한 성인 동아시아계 남성 한 명이 짧은 검은 머리와 남색 셔츠 차림으로 등장합니다. 얼굴과 머리 모양은 비서의 인물 참조에 대체로 맞고 다른 사람이나 추가 신체는 보이지 않습니다. 가죽 좌석과 차량 장치는 실물 세단 내부로 읽히며 낮의 외광도 확인됩니다. 제공된 도면만으로 이전 장면의 정확한 재질과 마모까지 비교할 수는 없습니다.",
        "hard_violations": [
         "비서가 우측 앞좌석의 운전대를 사용하는 구조로 묘사되어, 좌측 운전석을 고정한 도면 및 후석 중앙에서 보는 카메라 관계와 반대입니다."
        ],
        "physics": "몸은 앞좌석 등받이 앞에 자리하며 어깨와 목의 뒤돌림이 해부학적으로 가능합니다. 운전대 오른쪽 위에 손이 걸쳐 있고 아래쪽에도 다른 손 일부가 보여 운전 동작을 멈춘 상태로 읽힙니다. 운전대는 조향축으로, 머리받침은 금속 지지대로 고정되어 있으며 지지 없이 떠 있는 대상은 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "놀라 뒤돌아본 표정과 인물 외형은 맞지만, 시선이 화면 오른쪽으로 향하고 운전석 배치가 도면과 반대이며 상반신·계기판까지 넓게 담아 지정 클로즈업에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "A보다 얼굴을 크게 담아 반응 클로즈업에 가깝지만, 화면 오른쪽을 보는 시선과 좌우가 뒤집힌 운전석 배치는 동일하게 실패하며 좌석도 좁은 가장자리 이상으로 드러납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "비서는 몸을 운전대 쪽에 둔 채 고개를 뒤로 돌렸으나, 두 눈은 화면 오른쪽 바깥을 향합니다. 지정된 화면 왼쪽 밖의 윤성찬을 보는 시선이 아닙니다. 벌어진 입과 올라간 눈썹은 놀란 반응을 나타냅니다.",
        "built_space": "운전대 1개, 계기판 1개, 왼쪽 중앙 디스플레이 1개, 룸미러 1개, 큰 운전석 머리받침 1개와 등받이가 보입니다. 중앙 디스플레이가 운전대 왼쪽에 있고 운전자 옆 창문이 오른쪽에 있어 우측 운전석 배치로 읽히며, 좌측 운전석을 지정한 도면과 반대입니다. 카메라도 운전자의 왼쪽 뒤에서 보는 관계입니다. 좌석은 좁은 가장자리가 아니라 화면 오른쪽과 하단을 크게 차지합니다. 룸미러에는 식별 가능한 인물 반사가 없습니다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 한국인 남성이라는 설정과 충돌하는 외형은 없습니다. 짧게 정돈한 검은 머리, 얼굴 생김새와 남색 셔츠는 인물 참조에 대체로 부합합니다. 윤성찬이나 다른 사람은 등장하지 않습니다. 가죽 좌석과 목재 장식은 고급 세단으로 읽힙니다. 별도의 이전 장면 사진은 제공되지 않아 정확한 내장재 연속성은 확인할 수 없습니다.",
        "hard_violations": [
         "좌측 운전석 도면과 달리 중앙 장치 왼쪽·운전자 측 창문 오른쪽의 우측 운전석 구조로 배치하여, 비서의 착석 위치와 후석 중앙 카메라의 관찰 관계가 좌우 반전되었습니다."
        ],
        "physics": "보이는 등과 어깨는 좌석에 앉은 몸으로 자연스럽게 이어지고, 한 손은 운전대 아래쪽 테두리를 감싸 쥐고 있습니다. 몸통을 일부 틀며 뒤돌아보는 목의 자세는 가능한 동작입니다. 엉덩이와 발은 프레임 밖이지만 공중에 뜬 몸이나 지지 없는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "비서는 운전대를 향한 몸 위로 고개를 뒤돌리고 눈을 화면 오른쪽 바깥으로 보냅니다. 화면 왼쪽 밖에 있어야 하는 윤성찬에게 시선이 향하지 않습니다. 입술이 벌어지고 눈썹이 들려 당혹스러운 정지 순간은 드러납니다.",
        "built_space": "운전대 1개, 계기판 1개, 오른쪽의 운전석 머리받침 1개와 등받이, 왼쪽 아래의 다른 앞좌석 일부가 보입니다. 운전자 바로 오른쪽에 측면 창문이 있고 다른 앞좌석이 왼쪽에 있어 도면의 좌측 운전석과 반대입니다. 얼굴은 A보다 크고 오른쪽에 놓였지만, 큰 머리받침과 넓은 등받이 및 운전대까지 보여 지정된 좁은 좌석 가장자리 구성에는 못 미칩니다. 불가능한 반사는 보이지 않습니다.",
        "entities": "참조와 유사한 성인 동아시아계 남성 한 명이 짧은 검은 머리와 남색 셔츠 차림으로 등장합니다. 얼굴과 머리 모양은 비서의 인물 참조에 대체로 맞고 다른 사람이나 추가 신체는 보이지 않습니다. 가죽 좌석과 차량 장치는 실물 세단 내부로 읽히며 낮의 외광도 확인됩니다. 제공된 도면만으로 이전 장면의 정확한 재질과 마모까지 비교할 수는 없습니다.",
        "hard_violations": [
         "비서가 우측 앞좌석의 운전대를 사용하는 구조로 묘사되어, 좌측 운전석을 고정한 도면 및 후석 중앙에서 보는 카메라 관계와 반대입니다."
        ],
        "physics": "몸은 앞좌석 등받이 앞에 자리하며 어깨와 목의 뒤돌림이 해부학적으로 가능합니다. 운전대 오른쪽 위에 손이 걸쳐 있고 아래쪽에도 다른 손 일부가 보여 운전 동작을 멈춘 상태로 읽힙니다. 운전대는 조향축으로, 머리받침은 금속 지지대로 고정되어 있으며 지지 없이 떠 있는 대상은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.95
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.7
   },
   "violations": {
    "B": [
     "[gemini-pro] 전방을 향한 몸(forward-facing body)이어야 한다는 지문을 어기고 상체 전체가 뒤(카메라 쪽)를 돌아보아 셔츠 앞면 단추가 보이는 잘못된 자세",
     "[gemini-pro] 차량 내부 중앙에 있어야 할 센터 콘솔 스크린이 운전대 왼쪽에 위치하는 물리적으로 불가능한 구조",
     "[gpt-high] 좌측 운전석 도면과 달리 중앙 장치 왼쪽·운전자 측 창문 오른쪽의 우측 운전석 구조로 배치하여, 비서의 착석 위치와 후석 중앙 카메라의 관찰 관계가 좌우 반전되었습니다."
    ],
    "A": [
     "[gpt-high] 비서가 우측 앞좌석의 운전대를 사용하는 구조로 묘사되어, 좌측 운전석을 고정한 도면 및 후석 중앙에서 보는 카메라 관계와 반대입니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 700
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지문에서 요구한 전방을 향한 몸통과 왼쪽 어깨 너머로 뒤를 돌아보는 고개의 자세, 좌핸들 차량의 내부 구조와 카메라 앵글을 완벽하게 구현한 탁월한 샷입니다.  ★위반: [gpt-high] 비서가 우측 앞좌석의 운전대를 사용하는 구조로 묘사되어, 좌측 운전석을 고정한 도면 및 후석 중앙에서 보는 카메라 관계와 반대입니다."
   },
   {
    "label": "B",
    "score": 700,
    "verdict_ko": "몸통이 앞을 향해야 한다는 지시를 어기고 셔츠 앞면이 카메라를 향한 해부학적 오류를 범했으며, 운전대 왼쪽에 센터 콘솔이 배치되는 불가능한 차량 구조를 그렸습니다.  ★위반: [gemini-pro] 전방을 향한 몸(forward-facing body)이어야 한다는 지문을 어기고 상체 전체가 뒤(카메라 쪽)를 돌아보아 셔츠 앞면 단추가 보이는 잘못된 자세 / [gemini-pro] 차량 내부 중앙에 있어야 할 센터 콘솔 스크린이 운전대 왼쪽에 위치하는 물리적으로 불가능한 구조 / [gpt-high] 좌측 운전석 도면과 달리 중앙 장치 왼쪽·운전자 측 창문 오른쪽의 우측 운전석 구조로 배치하여, 비서의 착석 위치와 후석 중앙 카메라의 관찰 관계가 좌우 반전되었습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S19sh4_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "윤성찬의 비서",
    "path": "<bytes:1155240>",
    "asset_id": "736a43b8-dbe6-41d5-86af-0d923c2084a9",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab08f9-42a1-732b-9ae8-25b67256697b",
  "confined_fp": {
   "base_key": "confinedfp::557950ae2d7a",
   "apt_reason": "이 숏은 차량 내부 운전석에서 진행되며, 운전자인 비서가 뒷좌석의 승객을 돌아보는 상황이므로 차량 내 인물들의 좌석 배치와 시선 방향이 정확하게 표현되어야 합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S19sh3"
  }
 },
 "S20sh2::signage": {
  "fp": "57c106ffa309ad3f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S20sh2::bgfirst_bg": {
  "input_fingerprint": "a023bbb782d32cd2",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 컨테이너 안을 난장판으로 뒤지고 있는 민병대원들을 향해 입을 벌린 채 굳어있는 이현우의 상체.\n\nLOCATION (lock): In the settlement alley outside the family's container home, with a view into the dwelling being searched.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach-stage dolly on 이현우's front three-quarter side at approximately his eye level, keeping his open-mouthed upper body in the left-center and 페드로 partially visible beside him. Both look beyond the right edge toward the container interior, their interrupted forward movement held in unequal weight shifts rather than matching poses. Keep the directly observed militia activity as a narrow contextual layer at the right edge: one searcher bends farther into the container while another turns within it, preserving the searching action without duplicating their silhouettes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 이현우's container interior (Being searched and disrupted by militia members) — Only an oblique slice of the accessible interior appears at the right edge; used as Provides immediate evidence for the reaction without competing with the upper-body framing; Militia uniforms and armbands (Worn by the searchers) — Partial sleeve and torso views reveal the armbands on differently angled bodies; used as Identifies the searching group while staggered bends and shoulder turns prevent a cloned formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the stunned faces readable without separating the container into a different lighting scheme.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 컨테이너 안을 난장판으로 뒤지고 있는 민병대원들을 향해 입을 벌린 채 굳어있는 이현우의 상체.\n\nLOCATION (lock): In the settlement alley outside the family's container home, with a view into the dwelling being searched.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach-stage dolly on 이현우's front three-quarter side at approximately his eye level, keeping his open-mouthed upper body in the left-center and 페드로 partially visible beside him. Both look beyond the right edge toward the container interior, their interrupted forward movement held in unequal weight shifts rather than matching poses. Keep the directly observed militia activity as a narrow contextual layer at the right edge: one searcher bends farther into the container while another turns within it, preserving the searching action without duplicating their silhouettes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 이현우's container interior (Being searched and disrupted by militia members) — Only an oblique slice of the accessible interior appears at the right edge; used as Provides immediate evidence for the reaction without competing with the upper-body framing; Militia uniforms and armbands (Worn by the searchers) — Partial sleeve and torso views reveal the armbands on differently angled bodies; used as Identifies the searching group while staggered bends and shoulder turns prevent a cloned formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the stunned faces readable without separating the container into a different lighting scheme.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh2__bgfirst_bg.png",
  "asset_id": "d3c4f5b3-7201-4533-a614-4b374d18ec93",
  "input_asset_ids": [
   "9be84bf2-5f5f-453b-860d-e8e8fda8712e",
   "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b"
  ]
 },
 "S20sh2": {
  "input_fingerprint": "7ac46fccbd0aa659",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 컨테이너 안을 난장판으로 뒤지고 있는 민병대원들을 향해 입을 벌린 채 굳어있는 이현우의 상체.\n\nLOCATION (lock): In the settlement alley outside the family's container home, with a view into the dwelling being searched. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach-stage dolly on 이현우's front three-quarter side at approximately his eye level, keeping his open-mouthed upper body in the left-center and 페드로 partially visible beside him. Both look beyond the right edge toward the container interior, their interrupted forward movement held in unequal weight shifts rather than matching poses. Keep the directly observed militia activity as a narrow contextual layer at the right edge: one searcher bends farther into the container while another turns within it, preserving the searching action without duplicating their silhouettes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 이현우's container interior (Being searched and disrupted by militia members) — Only an oblique slice of the accessible interior appears at the right edge; used as Provides immediate evidence for the reaction without competing with the upper-body framing; Militia uniforms and armbands (Worn by the searchers) — Partial sleeve and torso views reveal the armbands on differently angled bodies; used as Identifies the searching group while staggered bends and shoulder turns prevent a cloned formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the stunned faces readable without separating the container into a different lighting scheme.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 is being searched, leaving its interior disturbed. 이현우: He is hiding near the container, still bearing his earlier facial injuries and untreated leg bite. 페드로: He is hiding near the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 컨테이너 안을 난장판으로 뒤지고 있는 민병대원들을 향해 입을 벌린 채 굳어있는 이현우의 상체.\n\nLOCATION (lock): In the settlement alley outside the family's container home, with a view into the dwelling being searched. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach-stage dolly on 이현우's front three-quarter side at approximately his eye level, keeping his open-mouthed upper body in the left-center and 페드로 partially visible beside him. Both look beyond the right edge toward the container interior, their interrupted forward movement held in unequal weight shifts rather than matching poses. Keep the directly observed militia activity as a narrow contextual layer at the right edge: one searcher bends farther into the container while another turns within it, preserving the searching action without duplicating their silhouettes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 이현우's container interior (Being searched and disrupted by militia members) — Only an oblique slice of the accessible interior appears at the right edge; used as Provides immediate evidence for the reaction without competing with the upper-body framing; Militia uniforms and armbands (Worn by the searchers) — Partial sleeve and torso views reveal the armbands on differently angled bodies; used as Identifies the searching group while staggered bends and shoulder turns prevent a cloned formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the stunned faces readable without separating the container into a different lighting scheme.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 is being searched, leaving its interior disturbed. 이현우: He is hiding near the container, still bearing his earlier facial injuries and untreated leg bite. 페드로: He is hiding near the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 컨테이너 안을 난장판으로 뒤지고 있는 민병대원들을 향해 입을 벌린 채 굳어있는 이현우의 상체.\n\nLOCATION (lock): In the settlement alley outside the family's container home, with a view into the dwelling being searched. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach-stage dolly on 이현우's front three-quarter side at approximately his eye level, keeping his open-mouthed upper body in the left-center and 페드로 partially visible beside him. Both look beyond the right edge toward the container interior, their interrupted forward movement held in unequal weight shifts rather than matching poses. Keep the directly observed militia activity as a narrow contextual layer at the right edge: one searcher bends farther into the container while another turns within it, preserving the searching action without duplicating their silhouettes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 이현우's container interior (Being searched and disrupted by militia members) — Only an oblique slice of the accessible interior appears at the right edge; used as Provides immediate evidence for the reaction without competing with the upper-body framing; Militia uniforms and armbands (Worn by the searchers) — Partial sleeve and torso views reveal the armbands on differently angled bodies; used as Identifies the searching group while staggered bends and shoulder turns prevent a cloned formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the stunned faces readable without separating the container into a different lighting scheme.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 is being searched, leaving its interior disturbed. 이현우: He is hiding near the container, still bearing his earlier facial injuries and untreated leg bite. 페드로: He is hiding near the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh2__bgfirst_bg.png",
     "asset_id": "d3c4f5b3-7201-4533-a614-4b374d18ec93",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S20sh2.png",
     "asset_id": "9be84bf2-5f5f-453b-860d-e8e8fda8712e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1278860>",
     "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_front_d0287f.png",
     "asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1278860>",
     "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 페드로의 시선이 모두 화면 우측의 컨테이너 내부를 향하고 있음.",
    "built_space": "레퍼런스와 유사한 골목길 배경이나, 컨테이너 입구에 명시된 7-31 표지판이 누락됨.",
    "entities": "이현우는 오염된 셔츠와 인이어를 착용했으나, 페드로는 명시되지 않은 모자를 쓰고 있어 짧은 머리 레퍼런스를 가림.",
    "hard_violations": [
     "[gemini-pro] 페드로에게 프롬프트 및 레퍼런스에 없는 모자가 추가됨 (발명된 사물)"
    ],
    "physics": "두 인물 모두 몸을 약간 숙이고 체중을 앞으로 실은 상태로 바닥을 자연스럽게 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "이현우가 입을 벌린 채 페드로와 함께 우측의 컨테이너 내부를 굳은 표정으로 응시함.",
    "built_space": "골목길 및 컨테이너 외벽이 레퍼런스와 일치하며 7-31 표지판이 올바른 위치에 있음.",
    "entities": "이현우와 페드로의 얼굴 및 헤어스타일이 레퍼런스와 완벽히 일치하며, 우측에 완장을 찬 민병대원들이 식별됨.",
    "hard_violations": [
     "[gpt-high] 두 청년 뒤 골목에 지정되지 않은 추가 인물이 최소 두 명 등장한다."
    ],
    "physics": "두 인물이 무게 중심을 이동한 채 구조물과 바닥에 안정적으로 의지하여 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물들의 외모와 헤어스타일을 레퍼런스와 정확히 일치시켰고, 7-31 표지판과 컨테이너 내부를 뒤지는 민병대까지 프롬프트의 지시사항을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "페드로에게 지시되지 않은 모자를 씌워 레퍼런스 외형을 위반했으며, 컨테이너의 핵심 요소인 7-31 표지판이 누락됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로의 시선이 모두 화면 우측의 컨테이너 내부를 향하고 있음.",
        "built_space": "레퍼런스와 유사한 골목길 배경이나, 컨테이너 입구에 명시된 7-31 표지판이 누락됨.",
        "entities": "이현우는 오염된 셔츠와 인이어를 착용했으나, 페드로는 명시되지 않은 모자를 쓰고 있어 짧은 머리 레퍼런스를 가림.",
        "hard_violations": [
         "페드로에게 프롬프트 및 레퍼런스에 없는 모자가 추가됨 (발명된 사물)"
        ],
        "physics": "두 인물 모두 몸을 약간 숙이고 체중을 앞으로 실은 상태로 바닥을 자연스럽게 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 입을 벌린 채 페드로와 함께 우측의 컨테이너 내부를 굳은 표정으로 응시함.",
        "built_space": "골목길 및 컨테이너 외벽이 레퍼런스와 일치하며 7-31 표지판이 올바른 위치에 있음.",
        "entities": "이현우와 페드로의 얼굴 및 헤어스타일이 레퍼런스와 완벽히 일치하며, 우측에 완장을 찬 민병대원들이 식별됨.",
        "hard_violations": [],
        "physics": "두 인물이 무게 중심을 이동한 채 구조물과 바닥에 안정적으로 의지하여 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물들의 외모와 헤어스타일을 레퍼런스와 정확히 일치시켰고, 7-31 표지판과 컨테이너 내부를 뒤지는 민병대까지 프롬프트의 지시사항을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "페드로에게 지시되지 않은 모자를 씌워 레퍼런스 외형을 위반했으며, 컨테이너의 핵심 요소인 7-31 표지판이 누락됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로의 시선이 모두 화면 우측의 컨테이너 내부를 향하고 있음.",
        "built_space": "레퍼런스와 유사한 골목길 배경이나, 컨테이너 입구에 명시된 7-31 표지판이 누락됨.",
        "entities": "이현우는 오염된 셔츠와 인이어를 착용했으나, 페드로는 명시되지 않은 모자를 쓰고 있어 짧은 머리 레퍼런스를 가림.",
        "hard_violations": [
         "페드로에게 프롬프트 및 레퍼런스에 없는 모자가 추가됨 (발명된 사물)"
        ],
        "physics": "두 인물 모두 몸을 약간 숙이고 체중을 앞으로 실은 상태로 바닥을 자연스럽게 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "이현우가 입을 벌린 채 페드로와 함께 우측의 컨테이너 내부를 굳은 표정으로 응시함.",
        "built_space": "골목길 및 컨테이너 외벽이 레퍼런스와 일치하며 7-31 표지판이 올바른 위치에 있음.",
        "entities": "이현우와 페드로의 얼굴 및 헤어스타일이 레퍼런스와 완벽히 일치하며, 우측에 완장을 찬 민병대원들이 식별됨.",
        "hard_violations": [],
        "physics": "두 인물이 무게 중심을 이동한 채 구조물과 바닥에 안정적으로 의지하여 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 열린 입과 좌중앙 상체, 정면 사선 구도는 충실하지만, 골목에 지정되지 않은 인물들이 추가되어 실격이다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "불필요한 인물 없이 오른쪽 실내 수색과 반응을 연결하지만, 이현우가 측면에 가깝고 두 청년의 기울기가 비슷하며 페드로의 외모·복장이 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로 모두 카메라가 아니라 화면 오른쪽 컨테이너 내부를 바라본다. 앞쪽 수색자는 허리를 굽혀 아래쪽 물건을 뒤지고, 뒤쪽 수색자는 몸을 오른쪽으로 돌려 실내 작업면을 향한다. 앞 수색자의 몸 옆에 보이는 총기는 아래쪽으로 늘어져 있으며 청년들을 겨누지 않는다.",
        "built_space": "좌측 골목과 우측 주거 출입구가 이어진다. 출입구 하나, 7-31 표찰 하나, 전기함 하나와 수직 배관이 보이고, 실내에는 선반과 흐트러진 생활물품이 있다. 청년들은 문밖, 수색자 둘은 문 안에 있어 위치 관계가 성립한다. 실내가 화면 오른쪽 약 4분의 1을 차지해 보조 맥락으로 유지된다. 참조의 낡은 금속 외벽과 젖은 골목을 재현하며 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 참조와 가까운 얼굴·체격, 피와 먼지가 묻은 어두운 셔츠, 얼굴 상처와 인이어 장치를 갖췄다. 입도 명확히 벌어져 있다. 페드로는 짧은 검은 머리와 옅은 콧수염, 어두운 티셔츠로 참조에 비교적 가깝다. 군복과 완장을 착용한 수색자는 두 명이다. 그러나 두 청년 뒤 골목에 별도의 행인이 최소 두 명 보인다. 다리 상처는 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [
         "두 청년 뒤 골목에 지정되지 않은 추가 인물이 최소 두 명 등장한다."
        ],
        "physics": "두 청년은 하체가 잘린 서 있는 자세이며, 페드로가 더 숙이고 이현우는 비교적 세워져 있어 서로 다른 정지 순간으로 읽힌다. 발은 보이지 않지만 공중에 떠 있다는 단서는 없다. 수색자들은 허리와 무릎을 사용해 몸을 굽히거나 돌리고 있으며, 장비는 몸의 끈과 벨트에 연결되어 있다. 실내 물품은 선반이나 바닥에 놓여 있어 지지 없는 부유물은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 청년의 얼굴과 시선은 오른쪽 열린 출입구와 그 안의 수색자들을 향한다. 이현우의 얼굴은 요구한 정면 사선보다 옆모습에 가깝다. 앞 수색자는 오른쪽 아래 서랍과 물건을 향해 숙이고, 뒤 수색자는 몸을 틀어 손에 든 종이 쪽을 내려다본다.",
        "built_space": "좌측 골목, 우측 컨테이너 벽과 출입구 하나가 보인다. 외벽에는 창 하나, 전기함 하나, 수직 배관이 있으며 문밖에는 의자 두 개가 놓여 있다. 내부에는 천장등 하나와 선반·수납장, 열린 서랍들이 보인다. 청년들은 바깥, 수색자 둘은 안쪽에 있어 공간 관계가 자연스럽다. 참조의 창·전기함·배관과 낡은 외벽을 잘 유지하지만, 실내를 비교적 정면으로 넓게 보여 좁은 사선 조각이라는 요구는 약해진다. 7-31 표찰은 보이지 않는다.",
        "entities": "이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 마른 체격, 오염된 어두운 셔츠, 얼굴 상처와 인이어 장치를 갖췄고 입을 벌리고 있다. 옆 청년은 페드로 참조와 얼굴 인상이 상당히 다르며, 참조의 티셔츠 대신 모자와 겹쳐 입은 셔츠를 착용한다. 군복과 완장을 착용한 수색자 두 명은 서로 다른 동작으로 구별된다. 지정 인원 외에 명확한 추가 인물은 보이지 않는다. 다리 상처는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "두 청년은 골반에서 상체를 앞으로 기울인 서 있는 자세로, 중단된 접근 동작으로 가능한 모습이다. 다만 두 사람의 기울기가 비슷해 서로 다른 체중 이동이라는 지시는 약하다. 앞 수색자는 바닥에 놓인 다리로 몸을 지탱하며 허리를 굽히고, 뒤 수색자는 서서 손으로 종이를 잡는다. 의자와 흩어진 물건은 바닥에, 수납 물품은 선반과 서랍에 지지되어 있으며 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 열린 입과 좌중앙 상체, 정면 사선 구도는 충실하지만, 골목에 지정되지 않은 인물들이 추가되어 실격이다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "불필요한 인물 없이 오른쪽 실내 수색과 반응을 연결하지만, 이현우가 측면에 가깝고 두 청년의 기울기가 비슷하며 페드로의 외모·복장이 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우와 페드로 모두 카메라가 아니라 화면 오른쪽 컨테이너 내부를 바라본다. 앞쪽 수색자는 허리를 굽혀 아래쪽 물건을 뒤지고, 뒤쪽 수색자는 몸을 오른쪽으로 돌려 실내 작업면을 향한다. 앞 수색자의 몸 옆에 보이는 총기는 아래쪽으로 늘어져 있으며 청년들을 겨누지 않는다.",
        "built_space": "좌측 골목과 우측 주거 출입구가 이어진다. 출입구 하나, 7-31 표찰 하나, 전기함 하나와 수직 배관이 보이고, 실내에는 선반과 흐트러진 생활물품이 있다. 청년들은 문밖, 수색자 둘은 문 안에 있어 위치 관계가 성립한다. 실내가 화면 오른쪽 약 4분의 1을 차지해 보조 맥락으로 유지된다. 참조의 낡은 금속 외벽과 젖은 골목을 재현하며 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 참조와 가까운 얼굴·체격, 피와 먼지가 묻은 어두운 셔츠, 얼굴 상처와 인이어 장치를 갖췄다. 입도 명확히 벌어져 있다. 페드로는 짧은 검은 머리와 옅은 콧수염, 어두운 티셔츠로 참조에 비교적 가깝다. 군복과 완장을 착용한 수색자는 두 명이다. 그러나 두 청년 뒤 골목에 별도의 행인이 최소 두 명 보인다. 다리 상처는 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [
         "두 청년 뒤 골목에 지정되지 않은 추가 인물이 최소 두 명 등장한다."
        ],
        "physics": "두 청년은 하체가 잘린 서 있는 자세이며, 페드로가 더 숙이고 이현우는 비교적 세워져 있어 서로 다른 정지 순간으로 읽힌다. 발은 보이지 않지만 공중에 떠 있다는 단서는 없다. 수색자들은 허리와 무릎을 사용해 몸을 굽히거나 돌리고 있으며, 장비는 몸의 끈과 벨트에 연결되어 있다. 실내 물품은 선반이나 바닥에 놓여 있어 지지 없는 부유물은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 청년의 얼굴과 시선은 오른쪽 열린 출입구와 그 안의 수색자들을 향한다. 이현우의 얼굴은 요구한 정면 사선보다 옆모습에 가깝다. 앞 수색자는 오른쪽 아래 서랍과 물건을 향해 숙이고, 뒤 수색자는 몸을 틀어 손에 든 종이 쪽을 내려다본다.",
        "built_space": "좌측 골목, 우측 컨테이너 벽과 출입구 하나가 보인다. 외벽에는 창 하나, 전기함 하나, 수직 배관이 있으며 문밖에는 의자 두 개가 놓여 있다. 내부에는 천장등 하나와 선반·수납장, 열린 서랍들이 보인다. 청년들은 바깥, 수색자 둘은 안쪽에 있어 공간 관계가 자연스럽다. 참조의 창·전기함·배관과 낡은 외벽을 잘 유지하지만, 실내를 비교적 정면으로 넓게 보여 좁은 사선 조각이라는 요구는 약해진다. 7-31 표찰은 보이지 않는다.",
        "entities": "이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 검은 머리, 마른 체격, 오염된 어두운 셔츠, 얼굴 상처와 인이어 장치를 갖췄고 입을 벌리고 있다. 옆 청년은 페드로 참조와 얼굴 인상이 상당히 다르며, 참조의 티셔츠 대신 모자와 겹쳐 입은 셔츠를 착용한다. 군복과 완장을 착용한 수색자 두 명은 서로 다른 동작으로 구별된다. 지정 인원 외에 명확한 추가 인물은 보이지 않는다. 다리 상처는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "두 청년은 골반에서 상체를 앞으로 기울인 서 있는 자세로, 중단된 접근 동작으로 가능한 모습이다. 다만 두 사람의 기울기가 비슷해 서로 다른 체중 이동이라는 지시는 약하다. 앞 수색자는 바닥에 놓인 다리로 몸을 지탱하며 허리를 굽히고, 뒤 수색자는 서서 손으로 종이를 잡는다. 의자와 흩어진 물건은 바닥에, 수납 물품은 선반과 서랍에 지지되어 있으며 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.464,
    "B": 1.179
   },
   "violations": {
    "A": [
     "[gemini-pro] 페드로에게 프롬프트 및 레퍼런스에 없는 모자가 추가됨 (발명된 사물)"
    ],
    "B": [
     "[gpt-high] 두 청년 뒤 골목에 지정되지 않은 추가 인물이 최소 두 명 등장한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1179,
   "A": 1464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "인물들의 외모와 헤어스타일을 레퍼런스와 정확히 일치시켰고, 7-31 표지판과 컨테이너 내부를 뒤지는 민병대까지 프롬프트의 지시사항을 충실히 구현함.  ★위반: [gpt-high] 두 청년 뒤 골목에 지정되지 않은 추가 인물이 최소 두 명 등장한다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "페드로에게 지시되지 않은 모자를 씌워 레퍼런스 외형을 위반했으며, 컨테이너의 핵심 요소인 7-31 표지판이 누락됨.  ★위반: [gemini-pro] 페드로에게 프롬프트 및 레퍼런스에 없는 모자가 추가됨 (발명된 사물)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_front_d0287f.png",
    "asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1278860>",
    "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0907-ac28-722f-af1e-8b4573597990",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh2__bgfirst_bg.png",
   "bg_asset_id": "d3c4f5b3-7201-4533-a614-4b374d18ec93",
   "bg_record_key": "S20sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "family_container_front",
   "groupbg_asset_id": "55bc1a18-3b38-4dc6-9fe5-319a6e9a3c1b"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C07"
  ]
 },
 "S20sh7::signage": {
  "fp": "abd4cd77458e34d8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S20sh7": {
  "input_fingerprint": "7206d8484c4f7474",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손이 눈앞의 사복 경찰의 가슴팍에 닿아 힘껏 밀어내고 있는 능동적 신체 접촉의 찰나.\n\nLOCATION (lock): On the refugee-settlement lane near the searched container home, where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track from the established side at chest height, retaining a slight upward view but easing the tilt down enough to show both palms contacting the officer's chest. Place 페드로 on the left in profile and the 사복 경찰 on the right in oblique view, with their upper bodies and the compressed interval between them sharing the frame; 페드로's attention angles downward along his extended forearms while the distracted officer's face remains turned toward off-screen space. Capture the direct, unmediated instant of 페드로 transferring his weight forward and the officer beginning to yield backward, making body displacement the dominant change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp passage (The retreat route is blocked by the officer) — The passage continues behind the officer's right-side silhouette; used as A limited strip of visible passage makes the obstruction and backward displacement legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding subdued daylight and controlled contrast so the contact reads through posture rather than a lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 페드로 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is at the interception point with his arms extended in a forceful pushing action.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손이 눈앞의 사복 경찰의 가슴팍에 닿아 힘껏 밀어내고 있는 능동적 신체 접촉의 찰나.\n\nLOCATION (lock): On the refugee-settlement lane near the searched container home, where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track from the established side at chest height, retaining a slight upward view but easing the tilt down enough to show both palms contacting the officer's chest. Place 페드로 on the left in profile and the 사복 경찰 on the right in oblique view, with their upper bodies and the compressed interval between them sharing the frame; 페드로's attention angles downward along his extended forearms while the distracted officer's face remains turned toward off-screen space. Capture the direct, unmediated instant of 페드로 transferring his weight forward and the officer beginning to yield backward, making body displacement the dominant change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp passage (The retreat route is blocked by the officer) — The passage continues behind the officer's right-side silhouette; used as A limited strip of visible passage makes the obstruction and backward displacement legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding subdued daylight and controlled contrast so the contact reads through posture rather than a lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 페드로 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is at the interception point with his arms extended in a forceful pushing action.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 페드로의 두 손이 눈앞의 사복 경찰의 가슴팍에 닿아 힘껏 밀어내고 있는 능동적 신체 접촉의 찰나.\n\nLOCATION (lock): On the refugee-settlement lane near the searched container home, where plainclothes police block the escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track from the established side at chest height, retaining a slight upward view but easing the tilt down enough to show both palms contacting the officer's chest. Place 페드로 on the left in profile and the 사복 경찰 on the right in oblique view, with their upper bodies and the compressed interval between them sharing the frame; 페드로's attention angles downward along his extended forearms while the distracted officer's face remains turned toward off-screen space. Capture the direct, unmediated instant of 페드로 transferring his weight forward and the officer beginning to yield backward, making body displacement the dominant change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp passage (The retreat route is blocked by the officer) — The passage continues behind the officer's right-side silhouette; used as A limited strip of visible passage makes the obstruction and backward displacement legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding subdued daylight and controlled contrast so the contact reads through posture rather than a lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 페드로 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the ongoing search. 페드로: He is at the interception point with his arms extended in a forceful pushing action.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "페드로는 경찰의 가슴을 향해 팔을 뻗고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
    "built_space": "우측에 이전 샷과 동일한 컨테이너(창문, 배전반)가 위치하며 좌측에 골목이 보임.",
    "entities": "페드로는 이전 샷의 외양과 일치하며, 경찰은 캐릭터 레퍼런스의 비니와 재킷을 착용함.",
    "hard_violations": [],
    "physics": "페드로가 앞으로 체중을 싣고 경찰이 뒤로 밀리는 자세가 자연스럽게 지면에 지탱됨."
   },
   {
    "label": "B",
    "direction": "페드로는 양손으로 경찰의 가슴을 밀고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
    "built_space": "우측에 컨테이너가 위치하고 좌측에 골목길이 보임.",
    "entities": "페드로는 이전 샷과 일치하나, 경찰은 이전 샷에 있던 동료의 얼굴과 모자를 그대로 사용함.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 등장한 동료의 얼굴과 모자를 이 샷의 경찰에게 적용하여 다른 인물 복사 금지 지침을 명백히 위반함",
     "[gpt-high] 왼쪽 배경 골목에 장면이 허용하지 않은 제삼자의 몸이 보인다."
    ],
    "physics": "두 인물 모두 지면에 서서 밀고 밀리는 힘의 작용이 올바르게 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 인물과 배경을 정확히 유지하고 물리적 충돌을 잘 표현했으나, 프레이밍 지시와 달리 양손 접촉이 모두 보이지는 않음."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이전 샷에 등장한 동료의 얼굴과 모자를 사복 경찰에게 그대로 적용하여 절대 금지 지침을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 경찰의 가슴을 향해 팔을 뻗고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
        "built_space": "우측에 이전 샷과 동일한 컨테이너(창문, 배전반)가 위치하며 좌측에 골목이 보임.",
        "entities": "페드로는 이전 샷의 외양과 일치하며, 경찰은 캐릭터 레퍼런스의 비니와 재킷을 착용함.",
        "hard_violations": [],
        "physics": "페드로가 앞으로 체중을 싣고 경찰이 뒤로 밀리는 자세가 자연스럽게 지면에 지탱됨."
       },
       {
        "label": "B",
        "direction": "페드로는 양손으로 경찰의 가슴을 밀고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
        "built_space": "우측에 컨테이너가 위치하고 좌측에 골목길이 보임.",
        "entities": "페드로는 이전 샷과 일치하나, 경찰은 이전 샷에 있던 동료의 얼굴과 모자를 그대로 사용함.",
        "hard_violations": [
         "이전 샷에 등장한 동료의 얼굴과 모자를 이 샷의 경찰에게 적용하여 다른 인물 복사 금지 지침을 명백히 위반함"
        ],
        "physics": "두 인물 모두 지면에 서서 밀고 밀리는 힘의 작용이 올바르게 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 인물과 배경을 정확히 유지하고 물리적 충돌을 잘 표현했으나, 프레이밍 지시와 달리 양손 접촉이 모두 보이지는 않음."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이전 샷에 등장한 동료의 얼굴과 모자를 사복 경찰에게 그대로 적용하여 절대 금지 지침을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "페드로는 경찰의 가슴을 향해 팔을 뻗고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
        "built_space": "우측에 이전 샷과 동일한 컨테이너(창문, 배전반)가 위치하며 좌측에 골목이 보임.",
        "entities": "페드로는 이전 샷의 외양과 일치하며, 경찰은 캐릭터 레퍼런스의 비니와 재킷을 착용함.",
        "hard_violations": [],
        "physics": "페드로가 앞으로 체중을 싣고 경찰이 뒤로 밀리는 자세가 자연스럽게 지면에 지탱됨."
       },
       {
        "label": "B",
        "direction": "페드로는 양손으로 경찰의 가슴을 밀고 시선을 아래로 향하며, 경찰은 화면 밖 우측을 응시함.",
        "built_space": "우측에 컨테이너가 위치하고 좌측에 골목길이 보임.",
        "entities": "페드로는 이전 샷과 일치하나, 경찰은 이전 샷에 있던 동료의 얼굴과 모자를 그대로 사용함.",
        "hard_violations": [
         "이전 샷에 등장한 동료의 얼굴과 모자를 이 샷의 경찰에게 적용하여 다른 인물 복사 금지 지침을 명백히 위반함"
        ],
        "physics": "두 인물 모두 지면에 서서 밀고 밀리는 힘의 작용이 올바르게 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "양손의 가슴 접촉과 경찰의 화면 밖 시선은 명확하지만, 골목에 허용되지 않은 제삼자가 보이며 경찰 뒤의 퇴로 대신 컨테이너 출입구가 배치되어 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "추가 인물 없이 중간 크기 측면 구도와 전진·후퇴의 밀기 동작을 구현했지만, 양손 접촉이 겹쳐 보이고 경찰 뒤로 이어져야 할 골목 및 페드로의 인물 참조는 충분히 맞지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 청년은 고개와 눈을 뻗은 팔을 따라 오른쪽 아래의 접촉 부위로 향한다. 두 손바닥은 오른쪽 남자의 가슴 양쪽에 각각 닿는다. 경찰은 얼굴과 시선을 화면 오른쪽 밖으로 돌려 청년을 보지 않는다. 청년의 힘은 오른쪽으로 전달되고 경찰의 상체는 뒤로 젖혀진다.",
        "built_space": "녹슨 청회색 컨테이너 벽, 창 하나, 외벽 전기함 하나와 세로 배관, 열린 출입구 하나, 실내 형광등 하나와 선반이 보인다. 낮은 금속 가구와 콘크리트 받침도 있어 이전 장소의 재료와 설비는 대체로 이어진다. 다만 골목은 청년 뒤의 화면 왼쪽으로 뻗고 경찰 뒤에는 실내 출입구가 있어, 경찰 오른쪽 윤곽 뒤로 좁게 이어지는 퇴로를 막는 배치는 구현되지 않았다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "전경에는 젊은 남성과 성인 남성 경찰이 있고, 왼쪽 먼 골목에 별도의 사람 한 명이 보인다. 청년의 헝클어진 검은 머리, 볼 상처, 이어피스와 낡은 갈색 단추 셔츠는 이전 장면의 중앙 인물과 일치한다. 그러나 별도 페드로 참조의 얼굴, 검은 비니, 녹색 집업과는 다르며 요청한 라틴계 혼혈 정체성을 외형만으로 확인할 수 없다. 경찰은 챙 모자와 녹색 사복 재킷을 입은 성인 남성으로 표현된다. 추가 문구나 그래픽 표시는 없다.",
        "hard_violations": [
         "왼쪽 배경 골목에 장면이 허용하지 않은 제삼자의 몸이 보인다."
        ],
        "physics": "청년의 두 손은 경찰의 가슴에 실제로 접촉하며 손가락과 손목이 각각 구별된다. 앞으로 숙인 청년의 몸통과 뒤로 물러나는 경찰의 몸통은 밀기 동작으로 가능한 자세다. 두 사람의 발은 화면 밖이므로 지면 지지점을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다. 경찰의 허리 주머니는 벨트에 고정되어 있고 배경 가구는 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 청년은 팔을 따라 가슴 접촉 지점 쪽으로 시선을 낮춘다. 두 팔은 오른쪽 경찰의 가슴으로 향하지만 손들이 가까이 겹쳐 각각의 손바닥 접촉은 A보다 덜 분명하다. 경찰은 청년의 얼굴보다 높은 화면 왼쪽 바깥 방향을 바라본다. 청년은 오른쪽으로 체중을 보내고 경찰의 상체는 오른쪽 뒤로 기울어진다.",
        "built_space": "녹슨 컨테이너 외벽, 창 하나, 전기함 하나와 배관, 열린 출입구 하나, 실내 형광등 하나, 선반과 열린 서랍이 보인다. 외벽 아래 콘크리트 받침과 젖은 골목 바닥도 이전 장소와 연결된다. 그러나 경찰은 골목 횡단 지점보다 출입구 앞에 놓여 있고, 통로는 화면 왼쪽으로 이어진다. 따라서 경찰의 오른쪽 윤곽 뒤로 퇴로가 남는 지정 배치는 아니다. 명백히 중복된 고정 설비나 불가능한 반사는 없다.",
        "entities": "보이는 사람은 밀고 있는 청년과 밀리는 성인 남성 두 명뿐이다. 청년의 머리, 볼 상처, 이어피스, 갈색 단추 셔츠는 이전 장면 중앙 인물의 모습을 유지하지만 별도 페드로 인물 참조의 얼굴과 녹색 집업은 재현하지 않는다. 경찰은 검은 비니와 녹색 집업을 입어 오히려 페드로 인물 참조의 복장에 가깝다. 청년은 십대 후반으로 읽힐 수 있으나 지정된 혼혈 배경은 외형만으로 확정할 수 없다. 읽을 수 있는 새 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "청년의 두 전완과 손목은 경찰 가슴의 겹친 손 접촉부로 이어지며, 명백한 추가 손이나 불가능한 관절은 보이지 않는다. 청년의 전방 기울기와 경찰의 골반에 대한 상체 후방 기울기는 밀려 균형을 잃기 시작하는 순간으로 가능하다. 하체와 발은 잘려 있어 지면 접촉을 직접 볼 수 없지만, 몸이 무지지 상태로 떠 있는 장면은 아니다. 경찰의 들어 올린 손은 팔에 자연스럽게 연결되고 허리 주머니는 벨트가 지지한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "양손의 가슴 접촉과 경찰의 화면 밖 시선은 명확하지만, 골목에 허용되지 않은 제삼자가 보이며 경찰 뒤의 퇴로 대신 컨테이너 출입구가 배치되어 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "추가 인물 없이 중간 크기 측면 구도와 전진·후퇴의 밀기 동작을 구현했지만, 양손 접촉이 겹쳐 보이고 경찰 뒤로 이어져야 할 골목 및 페드로의 인물 참조는 충분히 맞지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 청년은 고개와 눈을 뻗은 팔을 따라 오른쪽 아래의 접촉 부위로 향한다. 두 손바닥은 오른쪽 남자의 가슴 양쪽에 각각 닿는다. 경찰은 얼굴과 시선을 화면 오른쪽 밖으로 돌려 청년을 보지 않는다. 청년의 힘은 오른쪽으로 전달되고 경찰의 상체는 뒤로 젖혀진다.",
        "built_space": "녹슨 청회색 컨테이너 벽, 창 하나, 외벽 전기함 하나와 세로 배관, 열린 출입구 하나, 실내 형광등 하나와 선반이 보인다. 낮은 금속 가구와 콘크리트 받침도 있어 이전 장소의 재료와 설비는 대체로 이어진다. 다만 골목은 청년 뒤의 화면 왼쪽으로 뻗고 경찰 뒤에는 실내 출입구가 있어, 경찰 오른쪽 윤곽 뒤로 좁게 이어지는 퇴로를 막는 배치는 구현되지 않았다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "전경에는 젊은 남성과 성인 남성 경찰이 있고, 왼쪽 먼 골목에 별도의 사람 한 명이 보인다. 청년의 헝클어진 검은 머리, 볼 상처, 이어피스와 낡은 갈색 단추 셔츠는 이전 장면의 중앙 인물과 일치한다. 그러나 별도 페드로 참조의 얼굴, 검은 비니, 녹색 집업과는 다르며 요청한 라틴계 혼혈 정체성을 외형만으로 확인할 수 없다. 경찰은 챙 모자와 녹색 사복 재킷을 입은 성인 남성으로 표현된다. 추가 문구나 그래픽 표시는 없다.",
        "hard_violations": [
         "왼쪽 배경 골목에 장면이 허용하지 않은 제삼자의 몸이 보인다."
        ],
        "physics": "청년의 두 손은 경찰의 가슴에 실제로 접촉하며 손가락과 손목이 각각 구별된다. 앞으로 숙인 청년의 몸통과 뒤로 물러나는 경찰의 몸통은 밀기 동작으로 가능한 자세다. 두 사람의 발은 화면 밖이므로 지면 지지점을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다. 경찰의 허리 주머니는 벨트에 고정되어 있고 배경 가구는 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 청년은 팔을 따라 가슴 접촉 지점 쪽으로 시선을 낮춘다. 두 팔은 오른쪽 경찰의 가슴으로 향하지만 손들이 가까이 겹쳐 각각의 손바닥 접촉은 A보다 덜 분명하다. 경찰은 청년의 얼굴보다 높은 화면 왼쪽 바깥 방향을 바라본다. 청년은 오른쪽으로 체중을 보내고 경찰의 상체는 오른쪽 뒤로 기울어진다.",
        "built_space": "녹슨 컨테이너 외벽, 창 하나, 전기함 하나와 배관, 열린 출입구 하나, 실내 형광등 하나, 선반과 열린 서랍이 보인다. 외벽 아래 콘크리트 받침과 젖은 골목 바닥도 이전 장소와 연결된다. 그러나 경찰은 골목 횡단 지점보다 출입구 앞에 놓여 있고, 통로는 화면 왼쪽으로 이어진다. 따라서 경찰의 오른쪽 윤곽 뒤로 퇴로가 남는 지정 배치는 아니다. 명백히 중복된 고정 설비나 불가능한 반사는 없다.",
        "entities": "보이는 사람은 밀고 있는 청년과 밀리는 성인 남성 두 명뿐이다. 청년의 머리, 볼 상처, 이어피스, 갈색 단추 셔츠는 이전 장면 중앙 인물의 모습을 유지하지만 별도 페드로 인물 참조의 얼굴과 녹색 집업은 재현하지 않는다. 경찰은 검은 비니와 녹색 집업을 입어 오히려 페드로 인물 참조의 복장에 가깝다. 청년은 십대 후반으로 읽힐 수 있으나 지정된 혼혈 배경은 외형만으로 확정할 수 없다. 읽을 수 있는 새 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "청년의 두 전완과 손목은 경찰 가슴의 겹친 손 접촉부로 이어지며, 명백한 추가 손이나 불가능한 관절은 보이지 않는다. 청년의 전방 기울기와 경찰의 골반에 대한 상체 후방 기울기는 밀려 균형을 잃기 시작하는 순간으로 가능하다. 하체와 발은 잘려 있어 지면 접촉을 직접 볼 수 없지만, 몸이 무지지 상태로 떠 있는 장면은 아니다. 경찰의 들어 올린 손은 팔에 자연스럽게 연결되고 허리 주머니는 벨트가 지지한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷에 등장한 동료의 얼굴과 모자를 이 샷의 경찰에게 적용하여 다른 인물 복사 금지 지침을 명백히 위반함",
     "[gpt-high] 왼쪽 배경 골목에 장면이 허용하지 않은 제삼자의 몸이 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이전 샷의 인물과 배경을 정확히 유지하고 물리적 충돌을 잘 표현했으나, 프레이밍 지시와 달리 양손 접촉이 모두 보이지는 않음."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "이전 샷에 등장한 동료의 얼굴과 모자를 사복 경찰에게 그대로 적용하여 절대 금지 지침을 위반함.  ★위반: [gemini-pro] 이전 샷에 등장한 동료의 얼굴과 모자를 이 샷의 경찰에게 적용하여 다른 인물 복사 금지 지침을 명백히 위반함 / [gpt-high] 왼쪽 배경 골목에 장면이 허용하지 않은 제삼자의 몸이 보인다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 페드로 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh2_sel.png",
    "asset_id": "9f99fd35-ca47-4354-8483-d666b954509d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab090e-142b-73cc-9af5-fcefe713a3f2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh2"
  }
 },
 "S20sh13::signage": {
  "fp": "545a9a2714a6ac00",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::market_narrow_alley": {
  "input_fingerprint": "7d8b2470727b2174",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_narrow_alley",
    "tags": [
     "S20sh13"
    ]
   },
   "context_sig": "c0a00fcbe0bce3a1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 더욱 좁은 골목길로 도망가려던 찰나-!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 더욱 좁은 골목길로 도망가려던 찰나-!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_market_narrow_alley_c1d5a1.png",
  "asset_id": "570e3429-0e1d-4937-98b2-40b1870f158a",
  "input_asset_ids": [
   "3574d533-8ad5-4acf-b7f4-1cb64f757d44"
  ],
  "origin_tag": "S20sh13",
  "place_text": "At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.",
  "origin_inputs": {
   "place_text": "At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.",
   "time_of_day_en": "day",
   "conti_asset_id": "3574d533-8ad5-4acf-b7f4-1cb64f757d44"
  }
 },
 "S20sh13::bgfirst_bg": {
  "input_fingerprint": "fa338e40204da9aa",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사복 경찰이 휘두른 소총 개머리판이 이현우의 머리를 향해 강하게 뻗어져 닿기 직전인 결정적 찰나.\n\nLOCATION (lock): At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the narrow alley entrance, hold the handheld camera beside 이현우's flank at upper-torso height, looking slightly upward and obliquely across his route while remaining outside the rifle butt's trajectory. Frame 이현우 on the right with his running torso pitched forward and his attention still directed into the alley beyond the right edge; the 사복 경찰 enters from the left, looking downward along the swing toward the lowered head. Keep both upper bodies visible and preserve an unmistakable open interval between the butt and 이현우's head, with the weapon occupying only a small central portion of this directly observed pre-contact frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Narrow alley entrance (이현우 is about to enter when intercepted) — The opening recedes behind and to the right of 이현우; used as Establishes the interrupted escape route and supplies negative space behind the pre-contact gap; Rifle and butt (Being swung, not yet touching 이현우) — The butt approaches the side of the head across the image plane rather than pointing into the lens; used as Makes the imminent impact readable without foreshortening or concealing the separation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue restrained daylight with enough local contrast to distinguish the weapon, the face, and the remaining gap without an artificial impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사복 경찰이 휘두른 소총 개머리판이 이현우의 머리를 향해 강하게 뻗어져 닿기 직전인 결정적 찰나.\n\nLOCATION (lock): At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the narrow alley entrance, hold the handheld camera beside 이현우's flank at upper-torso height, looking slightly upward and obliquely across his route while remaining outside the rifle butt's trajectory. Frame 이현우 on the right with his running torso pitched forward and his attention still directed into the alley beyond the right edge; the 사복 경찰 enters from the left, looking downward along the swing toward the lowered head. Keep both upper bodies visible and preserve an unmistakable open interval between the butt and 이현우's head, with the weapon occupying only a small central portion of this directly observed pre-contact frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Narrow alley entrance (이현우 is about to enter when intercepted) — The opening recedes behind and to the right of 이현우; used as Establishes the interrupted escape route and supplies negative space behind the pre-contact gap; Rifle and butt (Being swung, not yet touching 이현우) — The butt approaches the side of the head across the image plane rather than pointing into the lens; used as Makes the imminent impact readable without foreshortening or concealing the separation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue restrained daylight with enough local contrast to distinguish the weapon, the face, and the remaining gap without an artificial impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh13__bgfirst_bg.png",
  "asset_id": "fe73bf94-aa04-4021-9e36-51a9ab3d1bd1",
  "input_asset_ids": [
   "3574d533-8ad5-4acf-b7f4-1cb64f757d44",
   "570e3429-0e1d-4937-98b2-40b1870f158a"
  ]
 },
 "S20sh13": {
  "input_fingerprint": "0ccb0e94766c985d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사복 경찰이 휘두른 소총 개머리판이 이현우의 머리를 향해 강하게 뻗어져 닿기 직전인 결정적 찰나.\n\nLOCATION (lock): At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the narrow alley entrance, hold the handheld camera beside 이현우's flank at upper-torso height, looking slightly upward and obliquely across his route while remaining outside the rifle butt's trajectory. Frame 이현우 on the right with his running torso pitched forward and his attention still directed into the alley beyond the right edge; the 사복 경찰 enters from the left, looking downward along the swing toward the lowered head. Keep both upper bodies visible and preserve an unmistakable open interval between the butt and 이현우's head, with the weapon occupying only a small central portion of this directly observed pre-contact frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Narrow alley entrance (이현우 is about to enter when intercepted) — The opening recedes behind and to the right of 이현우; used as Establishes the interrupted escape route and supplies negative space behind the pre-contact gap; Rifle and butt (Being swung, not yet touching 이현우) — The butt approaches the side of the head across the image plane rather than pointing into the lens; used as Makes the imminent impact readable without foreshortening or concealing the separation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue restrained daylight with enough local contrast to distinguish the weapon, the face, and the remaining gap without an artificial impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes knocked over during the escape remain scattered in the market alley. The earlier search has left container 7-31 disturbed. 이현우: He has reached the entrance to a narrower alley while fleeing. His earlier facial injuries and untreated leg bite remain; the impending rifle-butt strike has not yet landed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사복 경찰 right now, so 사복 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사복 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사복 경찰이 휘두른 소총 개머리판이 이현우의 머리를 향해 강하게 뻗어져 닿기 직전인 결정적 찰나.\n\nLOCATION (lock): At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the narrow alley entrance, hold the handheld camera beside 이현우's flank at upper-torso height, looking slightly upward and obliquely across his route while remaining outside the rifle butt's trajectory. Frame 이현우 on the right with his running torso pitched forward and his attention still directed into the alley beyond the right edge; the 사복 경찰 enters from the left, looking downward along the swing toward the lowered head. Keep both upper bodies visible and preserve an unmistakable open interval between the butt and 이현우's head, with the weapon occupying only a small central portion of this directly observed pre-contact frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Narrow alley entrance (이현우 is about to enter when intercepted) — The opening recedes behind and to the right of 이현우; used as Establishes the interrupted escape route and supplies negative space behind the pre-contact gap; Rifle and butt (Being swung, not yet touching 이현우) — The butt approaches the side of the head across the image plane rather than pointing into the lens; used as Makes the imminent impact readable without foreshortening or concealing the separation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue restrained daylight with enough local contrast to distinguish the weapon, the face, and the remaining gap without an artificial impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes knocked over during the escape remain scattered in the market alley. The earlier search has left container 7-31 disturbed. 이현우: He has reached the entrance to a narrower alley while fleeing. His earlier facial injuries and untreated leg bite remain; the impending rifle-butt strike has not yet landed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사복 경찰 right now, so 사복 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사복 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사복 경찰이 휘두른 소총 개머리판이 이현우의 머리를 향해 강하게 뻗어져 닿기 직전인 결정적 찰나.\n\nLOCATION (lock): At the mouth of a particularly narrow alley branching off the refugee settlement's market lanes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the narrow alley entrance, hold the handheld camera beside 이현우's flank at upper-torso height, looking slightly upward and obliquely across his route while remaining outside the rifle butt's trajectory. Frame 이현우 on the right with his running torso pitched forward and his attention still directed into the alley beyond the right edge; the 사복 경찰 enters from the left, looking downward along the swing toward the lowered head. Keep both upper bodies visible and preserve an unmistakable open interval between the butt and 이현우's head, with the weapon occupying only a small central portion of this directly observed pre-contact frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Narrow alley entrance (이현우 is about to enter when intercepted) — The opening recedes behind and to the right of 이현우; used as Establishes the interrupted escape route and supplies negative space behind the pre-contact gap; Rifle and butt (Being swung, not yet touching 이현우) — The butt approaches the side of the head across the image plane rather than pointing into the lens; used as Makes the imminent impact readable without foreshortening or concealing the separation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue restrained daylight with enough local contrast to distinguish the weapon, the face, and the remaining gap without an artificial impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boxes knocked over during the escape remain scattered in the market alley. The earlier search has left container 7-31 disturbed. 이현우: He has reached the entrance to a narrower alley while fleeing. His earlier facial injuries and untreated leg bite remain; the impending rifle-butt strike has not yet landed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사복 경찰 right now, so 사복 경찰's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사복 경찰: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh13__bgfirst_bg.png",
     "asset_id": "fe73bf94-aa04-4021-9e36-51a9ab3d1bd1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S20sh13.png",
     "asset_id": "3574d533-8ad5-4acf-b7f4-1cb64f757d44",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_market_narrow_alley_c1d5a1.png",
     "asset_id": "570e3429-0e1d-4937-98b2-40b1870f158a",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "사복 경찰이 소총 개머리판을 이현우의 머리를 향해 겨냥하고 있으며, 이현우의 시선은 오른쪽 골목 안을 향함.",
    "built_space": "기준 이미지와 동일한 좁은 컨테이너 골목 입구이며, 바닥에 상자와 과일들이 흩어져 있음.",
    "entities": "이현우의 측면 모습과 의상이 기준과 일치하며 인이어 무전기를 착용함. 사복 경찰이 소총을 들고 등장함.",
    "hard_violations": [],
    "physics": "경찰이 한 손으로 소총을 쥐고 앞으로 뻗는 자세이며, 이현우는 몸을 숙이고 이동하려는 안정적인 자세를 유지함."
   },
   {
    "label": "B",
    "direction": "경찰이 무기의 끝을 이현우의 머리로 향하게 휘두르고 있으며, 이현우는 전방 골목을 주시함.",
    "built_space": "기준 이미지와 일치하는 시장 골목 배경이며, 엎어진 상자 등 환경 디테일이 반영됨.",
    "entities": "이현우의 정면 얼굴과 상처가 기준과 잘 일치하나, 소총의 방아쇠 울이 개머리판 중간에 달린 기형적인 형태임. 배경 골목에 프롬프트에 없는 인물이 존재함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 허가되지 않은 인물이 배경 골목에 추가됨.",
     "[gemini-pro] 방아쇠 울이 개머리판 중앙에 달린 물리적으로 불가능한 구조의 소총 형태.",
     "[gpt-high] 숏에 지정되지 않은 배경 인물 두 명이 골목 안에 추가되어 있다."
    ],
    "physics": "이현우가 한 발을 들고 달리는 역동적인 자세이며, 경찰이 두 손으로 무기를 휘두르는 동작을 취함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 미디엄 샷 프레이밍과 개머리판이 머리로 향하는 궤적을 정확히 구현하였으며, 불필요한 인물 추가 없이 지시사항을 충실히 따름."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이현우의 외모와 상처 묘사는 우수하나, 배경에 프롬프트가 허용하지 않은 인물이 등장하고 소총 구조가 왜곡되어 규정을 크게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사복 경찰이 소총 개머리판을 이현우의 머리를 향해 겨냥하고 있으며, 이현우의 시선은 오른쪽 골목 안을 향함.",
        "built_space": "기준 이미지와 동일한 좁은 컨테이너 골목 입구이며, 바닥에 상자와 과일들이 흩어져 있음.",
        "entities": "이현우의 측면 모습과 의상이 기준과 일치하며 인이어 무전기를 착용함. 사복 경찰이 소총을 들고 등장함.",
        "hard_violations": [],
        "physics": "경찰이 한 손으로 소총을 쥐고 앞으로 뻗는 자세이며, 이현우는 몸을 숙이고 이동하려는 안정적인 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "경찰이 무기의 끝을 이현우의 머리로 향하게 휘두르고 있으며, 이현우는 전방 골목을 주시함.",
        "built_space": "기준 이미지와 일치하는 시장 골목 배경이며, 엎어진 상자 등 환경 디테일이 반영됨.",
        "entities": "이현우의 정면 얼굴과 상처가 기준과 잘 일치하나, 소총의 방아쇠 울이 개머리판 중간에 달린 기형적인 형태임. 배경 골목에 프롬프트에 없는 인물이 존재함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 허가되지 않은 인물이 배경 골목에 추가됨.",
         "방아쇠 울이 개머리판 중앙에 달린 물리적으로 불가능한 구조의 소총 형태."
        ],
        "physics": "이현우가 한 발을 들고 달리는 역동적인 자세이며, 경찰이 두 손으로 무기를 휘두르는 동작을 취함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 미디엄 샷 프레이밍과 개머리판이 머리로 향하는 궤적을 정확히 구현하였으며, 불필요한 인물 추가 없이 지시사항을 충실히 따름."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이현우의 외모와 상처 묘사는 우수하나, 배경에 프롬프트가 허용하지 않은 인물이 등장하고 소총 구조가 왜곡되어 규정을 크게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "사복 경찰이 소총 개머리판을 이현우의 머리를 향해 겨냥하고 있으며, 이현우의 시선은 오른쪽 골목 안을 향함.",
        "built_space": "기준 이미지와 동일한 좁은 컨테이너 골목 입구이며, 바닥에 상자와 과일들이 흩어져 있음.",
        "entities": "이현우의 측면 모습과 의상이 기준과 일치하며 인이어 무전기를 착용함. 사복 경찰이 소총을 들고 등장함.",
        "hard_violations": [],
        "physics": "경찰이 한 손으로 소총을 쥐고 앞으로 뻗는 자세이며, 이현우는 몸을 숙이고 이동하려는 안정적인 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "경찰이 무기의 끝을 이현우의 머리로 향하게 휘두르고 있으며, 이현우는 전방 골목을 주시함.",
        "built_space": "기준 이미지와 일치하는 시장 골목 배경이며, 엎어진 상자 등 환경 디테일이 반영됨.",
        "entities": "이현우의 정면 얼굴과 상처가 기준과 잘 일치하나, 소총의 방아쇠 울이 개머리판 중간에 달린 기형적인 형태임. 배경 골목에 프롬프트에 없는 인물이 존재함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 허가되지 않은 인물이 배경 골목에 추가됨.",
         "방아쇠 울이 개머리판 중앙에 달린 물리적으로 불가능한 구조의 소총 형태."
        ],
        "physics": "이현우가 한 발을 들고 달리는 역동적인 자세이며, 경찰이 두 손으로 무기를 휘두르는 동작을 취함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "골목 뒤에 불필요한 인물이 추가됐고 머리를 향하는 부분도 개머리판이 아니며, 무릎까지 넓힌 구도가 지정된 상반신 중심의 미디엄 숏에서 벗어난다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 경찰이 오른쪽으로 내민 개머리판과 낮아진 머리 사이의 빈틈, 골목으로 향하는 이현우의 시선과 상반신 구도를 가장 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽으로 달리며 화면 오른쪽 밖을 보고, 경찰은 무기를 따라 이현우의 머리 쪽을 내려다본다. 그러나 소총의 넓은 개머리판은 경찰 쪽 왼쪽 아래에 있고, 머리에 접근하는 오른쪽 끝은 총열과 앞몸통 쪽으로 보인다. 머리와 무기 사이 간격은 있지만 지정된 개머리판 타격 방향이 아니다.",
        "built_space": "왼쪽 시장 천막과 상자 더미, 중앙의 흰색·청색 컨테이너 모서리와 수직 배관, 오른쪽으로 깊어지는 좁은 골목이 보인다. 모서리에는 큰 전기함 하나와 일부 가려진 작은 함 하나가 식별되고, 골목 뒤에는 쓰레기통 하나가 있다. 장소의 주요 재료와 배치는 참조와 유사하지만, 경찰과 이현우의 다리까지 크게 포함해 요구된 상반신 중심 구도보다 넓다.",
        "entities": "전경에는 사복 차림의 성인 남성 경찰과 짧고 헝클어진 검은 머리의 동아시아계 청소년 남성이 있다. 이현우의 마른 체격, 어두운 낡은 셔츠, 얼굴 상처, 인이어는 대체로 맞으며 국적 자체는 외형으로 확인할 수 없다. 소총은 보이지만 타격하는 끝이 개머리판으로 읽히지 않는다. 골목 뒤에는 추가 인물 두 명으로 읽히는 형상이 있다. 바닥에는 흩어진 상자와 농산물이 있으며, 컨테이너 7-31의 상태는 식별할 수 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 배경 인물 두 명이 골목 안에 추가되어 있다."
        ],
        "physics": "소총은 경찰의 손이 몸통 부근을 잡아 지지하며 손과 소매의 연결도 보인다. 이현우의 뒤쪽 다리는 무릎을 접어 들린 달리기 자세이고, 반대쪽 다리는 화면 아래로 이어져 발의 접지는 확인되지 않는다. 상체와 다리의 관계는 가능한 달리기 동작이며 근거 없이 떠 있는 몸으로 볼 이유는 없다. 상자와 농산물은 바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "경찰은 왼쪽에서 이현우의 낮아진 머리를 향해 시선을 내리고, 오른쪽으로 향한 소총 개머리판을 머리 뒤쪽 측면 높이로 내민다. 총열 쪽은 반대인 왼쪽으로 이어지며 렌즈를 겨누지 않는다. 개머리판과 머리 사이에는 분명한 열린 간격이 있다. 이현우는 몸을 앞으로 기울이고 얼굴과 시선을 오른쪽 골목 안으로 돌려 도주 방향을 유지한다.",
        "built_space": "왼쪽 천막 시장, 중앙의 흰색·청색 컨테이너 모서리와 배관, 오른쪽의 좁은 골목이 참조 장소의 관계를 유지한다. 위쪽의 금속 함 하나와 청색 모서리의 작은 전기함 하나, 골목 바닥의 배수구 두 곳, 뒤쪽 쓰레기통 하나가 식별된다. 골목은 이현우 뒤와 오른쪽으로 이어지며 두 인물은 입구에서 서로 가로막는 위치에 있다. 두 상반신이 중심인 미디엄 구도이며, 약한 올려다보기는 두드러지지 않는다.",
        "entities": "사복 경찰과 이현우 두 사람만 보인다. 이현우는 참조와 부합하는 짧은 검은 머리, 젊은 동아시아계 남성의 외형, 마른 체격, 흙먼지와 핏자국이 묻은 어두운 셔츠, 작은 인이어와 얼굴 상처를 갖췄다. 얼굴이 옆뒤로 돌아 정확한 정면 인상 비교는 제한된다. 소총은 탄창과 목제 개머리판이 구별된다. 시장 쪽 상자와 흩어진 농산물은 보이지만 컨테이너 7-31의 교란 상태는 확인되지 않는다. 다리 물린 상처는 이 구도 밖이다.",
        "hard_violations": [],
        "physics": "경찰의 손이 소총 손잡이 부근을 확실히 쥐고 있으며 손목과 팔뚝이 같은 인물의 소매로 이어진다. 개머리판을 앞으로 밀어 휘두르는 동작은 팔과 기울어진 몸통으로 지지된다. 멜빵은 총에 연결되어 아래로 처진다. 이현우는 몸통을 숙이고 팔을 굽힌 달리기 중간 자세이며 하체가 프레임 밖으로 이어져 접지 여부는 확인되지 않지만 부유를 나타내는 모순은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "골목 뒤에 불필요한 인물이 추가됐고 머리를 향하는 부분도 개머리판이 아니며, 무릎까지 넓힌 구도가 지정된 상반신 중심의 미디엄 숏에서 벗어난다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 경찰이 오른쪽으로 내민 개머리판과 낮아진 머리 사이의 빈틈, 골목으로 향하는 이현우의 시선과 상반신 구도를 가장 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽으로 달리며 화면 오른쪽 밖을 보고, 경찰은 무기를 따라 이현우의 머리 쪽을 내려다본다. 그러나 소총의 넓은 개머리판은 경찰 쪽 왼쪽 아래에 있고, 머리에 접근하는 오른쪽 끝은 총열과 앞몸통 쪽으로 보인다. 머리와 무기 사이 간격은 있지만 지정된 개머리판 타격 방향이 아니다.",
        "built_space": "왼쪽 시장 천막과 상자 더미, 중앙의 흰색·청색 컨테이너 모서리와 수직 배관, 오른쪽으로 깊어지는 좁은 골목이 보인다. 모서리에는 큰 전기함 하나와 일부 가려진 작은 함 하나가 식별되고, 골목 뒤에는 쓰레기통 하나가 있다. 장소의 주요 재료와 배치는 참조와 유사하지만, 경찰과 이현우의 다리까지 크게 포함해 요구된 상반신 중심 구도보다 넓다.",
        "entities": "전경에는 사복 차림의 성인 남성 경찰과 짧고 헝클어진 검은 머리의 동아시아계 청소년 남성이 있다. 이현우의 마른 체격, 어두운 낡은 셔츠, 얼굴 상처, 인이어는 대체로 맞으며 국적 자체는 외형으로 확인할 수 없다. 소총은 보이지만 타격하는 끝이 개머리판으로 읽히지 않는다. 골목 뒤에는 추가 인물 두 명으로 읽히는 형상이 있다. 바닥에는 흩어진 상자와 농산물이 있으며, 컨테이너 7-31의 상태는 식별할 수 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 배경 인물 두 명이 골목 안에 추가되어 있다."
        ],
        "physics": "소총은 경찰의 손이 몸통 부근을 잡아 지지하며 손과 소매의 연결도 보인다. 이현우의 뒤쪽 다리는 무릎을 접어 들린 달리기 자세이고, 반대쪽 다리는 화면 아래로 이어져 발의 접지는 확인되지 않는다. 상체와 다리의 관계는 가능한 달리기 동작이며 근거 없이 떠 있는 몸으로 볼 이유는 없다. 상자와 농산물은 바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "경찰은 왼쪽에서 이현우의 낮아진 머리를 향해 시선을 내리고, 오른쪽으로 향한 소총 개머리판을 머리 뒤쪽 측면 높이로 내민다. 총열 쪽은 반대인 왼쪽으로 이어지며 렌즈를 겨누지 않는다. 개머리판과 머리 사이에는 분명한 열린 간격이 있다. 이현우는 몸을 앞으로 기울이고 얼굴과 시선을 오른쪽 골목 안으로 돌려 도주 방향을 유지한다.",
        "built_space": "왼쪽 천막 시장, 중앙의 흰색·청색 컨테이너 모서리와 배관, 오른쪽의 좁은 골목이 참조 장소의 관계를 유지한다. 위쪽의 금속 함 하나와 청색 모서리의 작은 전기함 하나, 골목 바닥의 배수구 두 곳, 뒤쪽 쓰레기통 하나가 식별된다. 골목은 이현우 뒤와 오른쪽으로 이어지며 두 인물은 입구에서 서로 가로막는 위치에 있다. 두 상반신이 중심인 미디엄 구도이며, 약한 올려다보기는 두드러지지 않는다.",
        "entities": "사복 경찰과 이현우 두 사람만 보인다. 이현우는 참조와 부합하는 짧은 검은 머리, 젊은 동아시아계 남성의 외형, 마른 체격, 흙먼지와 핏자국이 묻은 어두운 셔츠, 작은 인이어와 얼굴 상처를 갖췄다. 얼굴이 옆뒤로 돌아 정확한 정면 인상 비교는 제한된다. 소총은 탄창과 목제 개머리판이 구별된다. 시장 쪽 상자와 흩어진 농산물은 보이지만 컨테이너 7-31의 교란 상태는 확인되지 않는다. 다리 물린 상처는 이 구도 밖이다.",
        "hard_violations": [],
        "physics": "경찰의 손이 소총 손잡이 부근을 확실히 쥐고 있으며 손목과 팔뚝이 같은 인물의 소매로 이어진다. 개머리판을 앞으로 밀어 휘두르는 동작은 팔과 기울어진 몸통으로 지지된다. 멜빵은 총에 연결되어 아래로 처진다. 이현우는 몸통을 숙이고 팔을 굽힌 달리기 중간 자세이며 하체가 프레임 밖으로 이어져 접지 여부는 확인되지 않지만 부유를 나타내는 모순은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.619
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.369
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 허가되지 않은 인물이 배경 골목에 추가됨.",
     "[gemini-pro] 방아쇠 울이 개머리판 중앙에 달린 물리적으로 불가능한 구조의 소총 형태.",
     "[gpt-high] 숏에 지정되지 않은 배경 인물 두 명이 골목 안에 추가되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 369
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 미디엄 샷 프레이밍과 개머리판이 머리로 향하는 궤적을 정확히 구현하였으며, 불필요한 인물 추가 없이 지시사항을 충실히 따름."
   },
   {
    "label": "B",
    "score": 369,
    "verdict_ko": "이현우의 외모와 상처 묘사는 우수하나, 배경에 프롬프트가 허용하지 않은 인물이 등장하고 소총 구조가 왜곡되어 규정을 크게 위반함.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 허가되지 않은 인물이 배경 골목에 추가됨. / [gemini-pro] 방아쇠 울이 개머리판 중앙에 달린 물리적으로 불가능한 구조의 소총 형태. / [gpt-high] 숏에 지정되지 않은 배경 인물 두 명이 골목 안에 추가되어 있다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_market_narrow_alley_c1d5a1.png",
    "asset_id": "570e3429-0e1d-4937-98b2-40b1870f158a",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0914-4f3b-7a66-9462-7e64bfa3264e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S20sh13__bgfirst_bg.png",
   "bg_asset_id": "fe73bf94-aa04-4021-9e36-51a9ab3d1bd1",
   "bg_record_key": "S20sh13::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "market_narrow_alley",
   "groupbg_asset_id": "570e3429-0e1d-4937-98b2-40b1870f158a"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S21sh2::signage": {
  "fp": "81a87c59f4e54180",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::36aee137da1872d9": {
  "subjects": [],
  "subject_text": "라울의 컨테이너 내부\n낡은 철제 주거동 내부의 소박한 방. 출입문과 잠자리가 있으며, 좁은 실내에 기본 생활용품이 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L29",
  "scope_role": "location_interior",
  "scope_sha": "68237f3bfc9a7857"
 },
 "S21sh2::bgfirst_bg": {
  "input_fingerprint": "749443cffb837e7b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 컨테이너 철문 사이로 땀범벅이 된 라울이 뒷발로 문턱을 차고 상체를 앞으로 크게 기울인 mid-stride 자세.\n\nLOCATION (lock): Just inside the open front threshold of a container home, with daylight entering through the doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the low, offset position inside the container, pan with 라울's entry and look upward toward his three-quarter profile without moving into his path. Show his full forward-driving figure across the center, keeping the trailing foot against the threshold in the lower-left portion and the leaning torso advancing toward the right; his gaze reaches toward 앰버 outside that edge. The direct observational frame catches the urgent crossing rather than a completed entrance, with the open doorway providing a stable reference for his momentum.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Doorway threshold beneath the trailing foot in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Container iron door (Thrown open for 라울's entrance) — Seen obliquely from inside, with the open leaf remaining near the frame edge; used as Frames the entrance without obscuring the forward-leaning torso; Doorway threshold (라울's trailing foot is pushing away from it) — Runs diagonally through the lower-left part of the frame; used as Anchors the physical instant of departure into the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the container interior preserves the sweat on 라울's face with subdued brightness and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 컨테이너 철문 사이로 땀범벅이 된 라울이 뒷발로 문턱을 차고 상체를 앞으로 크게 기울인 mid-stride 자세.\n\nLOCATION (lock): Just inside the open front threshold of a container home, with daylight entering through the doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the low, offset position inside the container, pan with 라울's entry and look upward toward his three-quarter profile without moving into his path. Show his full forward-driving figure across the center, keeping the trailing foot against the threshold in the lower-left portion and the leaning torso advancing toward the right; his gaze reaches toward 앰버 outside that edge. The direct observational frame catches the urgent crossing rather than a completed entrance, with the open doorway providing a stable reference for his momentum.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Doorway threshold beneath the trailing foot in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Container iron door (Thrown open for 라울's entrance) — Seen obliquely from inside, with the open leaf remaining near the frame edge; used as Frames the entrance without obscuring the forward-leaning torso; Doorway threshold (라울's trailing foot is pushing away from it) — Runs diagonally through the lower-left part of the frame; used as Anchors the physical instant of departure into the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the container interior preserves the sweat on 라울's face with subdued brightness and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S21sh2__bgfirst_bg.png",
  "asset_id": "5177e3d9-8c8d-4af2-8128-284537731e8c",
  "input_asset_ids": [
   "9779949b-627e-4745-9cf6-6ea75a73ca87",
   "d35a2e6a-cbe6-4fa1-8c57-2961d294afc0"
  ]
 },
 "S21sh2": {
  "input_fingerprint": "8db2f7806b53aba4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 철문 사이로 땀범벅이 된 라울이 뒷발로 문턱을 차고 상체를 앞으로 크게 기울인 mid-stride 자세.\n\nLOCATION (lock): Just inside the open front threshold of a container home, with daylight entering through the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the low, offset position inside the container, pan with 라울's entry and look upward toward his three-quarter profile without moving into his path. Show his full forward-driving figure across the center, keeping the trailing foot against the threshold in the lower-left portion and the leaning torso advancing toward the right; his gaze reaches toward 앰버 outside that edge. The direct observational frame catches the urgent crossing rather than a completed entrance, with the open doorway providing a stable reference for his momentum.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Doorway threshold beneath the trailing foot in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Container iron door (Thrown open for 라울's entrance) — Seen obliquely from inside, with the open leaf remaining near the frame edge; used as Frames the entrance without obscuring the forward-leaning torso; Doorway threshold (라울's trailing foot is pushing away from it) — Runs diagonally through the lower-left part of the frame; used as Anchors the physical instant of departure into the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the container interior preserves the sweat on 라울's face with subdued brightness and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container entrance door has been thrown open. 라울: He is entering through the newly opened doorway.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 철문 사이로 땀범벅이 된 라울이 뒷발로 문턱을 차고 상체를 앞으로 크게 기울인 mid-stride 자세.\n\nLOCATION (lock): Just inside the open front threshold of a container home, with daylight entering through the doorway. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the low, offset position inside the container, pan with 라울's entry and look upward toward his three-quarter profile without moving into his path. Show his full forward-driving figure across the center, keeping the trailing foot against the threshold in the lower-left portion and the leaning torso advancing toward the right; his gaze reaches toward 앰버 outside that edge. The direct observational frame catches the urgent crossing rather than a completed entrance, with the open doorway providing a stable reference for his momentum.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Doorway threshold beneath the trailing foot in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Container iron door (Thrown open for 라울's entrance) — Seen obliquely from inside, with the open leaf remaining near the frame edge; used as Frames the entrance without obscuring the forward-leaning torso; Doorway threshold (라울's trailing foot is pushing away from it) — Runs diagonally through the lower-left part of the frame; used as Anchors the physical instant of departure into the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the container interior preserves the sweat on 라울's face with subdued brightness and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container entrance door has been thrown open. 라울: He is entering through the newly opened doorway.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 철문 사이로 땀범벅이 된 라울이 뒷발로 문턱을 차고 상체를 앞으로 크게 기울인 mid-stride 자세.\n\nLOCATION (lock): Just inside the open front threshold of a container home, with daylight entering through the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the low, offset position inside the container, pan with 라울's entry and look upward toward his three-quarter profile without moving into his path. Show his full forward-driving figure across the center, keeping the trailing foot against the threshold in the lower-left portion and the leaning torso advancing toward the right; his gaze reaches toward 앰버 outside that edge. The direct observational frame catches the urgent crossing rather than a completed entrance, with the open doorway providing a stable reference for his momentum.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Doorway threshold beneath the trailing foot in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Container iron door (Thrown open for 라울's entrance) — Seen obliquely from inside, with the open leaf remaining near the frame edge; used as Frames the entrance without obscuring the forward-leaning torso; Doorway threshold (라울's trailing foot is pushing away from it) — Runs diagonally through the lower-left part of the frame; used as Anchors the physical instant of departure into the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the container interior preserves the sweat on 라울's face with subdued brightness and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The container entrance door has been thrown open. 라울: He is entering through the newly opened doorway.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S21sh2__bgfirst_bg.png",
     "asset_id": "5177e3d9-8c8d-4af2-8128-284537731e8c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S21sh2.png",
     "asset_id": "9779949b-627e-4745-9cf6-6ea75a73ca87",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L29B01.png",
     "asset_id": "d35a2e6a-cbe6-4fa1-8c57-2961d294afc0",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선과 몸의 방향이 프레임 우측 밖을 향하고 있음.",
    "built_space": "컨테이너 내부 뷰. 좌측 하단 문턱부터 배경의 침구와 벽면 구조까지 기준 사진의 공간적 특징을 잘 반영함.",
    "entities": "라울의 인물 특징(혼혈, 꽁지머리, 의상 등)과 땀 흘리는 묘사가 기준과 잘 부합함.",
    "hard_violations": [],
    "physics": "왼발은 문턱에 위치하나 오른발이 바닥에 평평하게 닿아 있어 체중이 양발에 분산된 형태임."
   },
   {
    "label": "B",
    "direction": "인물의 시선과 진행 방향 모두 프레임 우측 바깥(앰버가 있는 방향)을 향함.",
    "built_space": "컨테이너 내부 로우 앵글. 좌측 하단의 열린 철문과 문턱, 배경의 창문 및 침구 위치가 기준 이미지의 공간 구조와 일치함.",
    "entities": "라울(10세 라틴/흑인 혼혈, 꽁지머리, 땀범벅, 낡은 녹색 티셔츠와 반바지)의 외형이 기준과 정확히 일치함.",
    "hard_violations": [],
    "physics": "왼발로 문턱을 힘차게 딛고 있으며, 오른발은 공중에 떠 있어 달리는 도중의 체중 이동과 도약이 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 프레이밍을 정확히 따랐으며, 공중에 뜬 앞발을 통해 문턱을 차고 전진하는 역동적인 mid-stride 자세를 훌륭하게 구현함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프레이밍과 캐릭터의 디테일은 우수하나, 앞발이 바닥에 완전히 닿아 있어 텍스트가 요구한 강한 전진 모멘텀이 다소 반감됨."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "인물의 시선과 진행 방향 모두 프레임 우측 바깥(앰버가 있는 방향)을 향함.",
        "built_space": "컨테이너 내부 로우 앵글. 좌측 하단의 열린 철문과 문턱, 배경의 창문 및 침구 위치가 기준 이미지의 공간 구조와 일치함.",
        "entities": "라울(10세 라틴/흑인 혼혈, 꽁지머리, 땀범벅, 낡은 녹색 티셔츠와 반바지)의 외형이 기준과 정확히 일치함.",
        "hard_violations": [],
        "physics": "왼발로 문턱을 힘차게 딛고 있으며, 오른발은 공중에 떠 있어 달리는 도중의 체중 이동과 도약이 물리적으로 자연스러움."
       },
       {
        "label": "A",
        "direction": "인물의 시선과 몸의 방향이 프레임 우측 밖을 향하고 있음.",
        "built_space": "컨테이너 내부 뷰. 좌측 하단 문턱부터 배경의 침구와 벽면 구조까지 기준 사진의 공간적 특징을 잘 반영함.",
        "entities": "라울의 인물 특징(혼혈, 꽁지머리, 의상 등)과 땀 흘리는 묘사가 기준과 잘 부합함.",
        "hard_violations": [],
        "physics": "왼발은 문턱에 위치하나 오른발이 바닥에 평평하게 닿아 있어 체중이 양발에 분산된 형태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 프레이밍을 정확히 따랐으며, 공중에 뜬 앞발을 통해 문턱을 차고 전진하는 역동적인 mid-stride 자세를 훌륭하게 구현함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프레이밍과 캐릭터의 디테일은 우수하나, 앞발이 바닥에 완전히 닿아 있어 텍스트가 요구한 강한 전진 모멘텀이 다소 반감됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 시선과 진행 방향 모두 프레임 우측 바깥(앰버가 있는 방향)을 향함.",
        "built_space": "컨테이너 내부 로우 앵글. 좌측 하단의 열린 철문과 문턱, 배경의 창문 및 침구 위치가 기준 이미지의 공간 구조와 일치함.",
        "entities": "라울(10세 라틴/흑인 혼혈, 꽁지머리, 땀범벅, 낡은 녹색 티셔츠와 반바지)의 외형이 기준과 정확히 일치함.",
        "hard_violations": [],
        "physics": "왼발로 문턱을 힘차게 딛고 있으며, 오른발은 공중에 떠 있어 달리는 도중의 체중 이동과 도약이 물리적으로 자연스러움."
       },
       {
        "label": "A",
        "direction": "인물의 시선과 몸의 방향이 프레임 우측 밖을 향하고 있음.",
        "built_space": "컨테이너 내부 뷰. 좌측 하단 문턱부터 배경의 침구와 벽면 구조까지 기준 사진의 공간적 특징을 잘 반영함.",
        "entities": "라울의 인물 특징(혼혈, 꽁지머리, 의상 등)과 땀 흘리는 묘사가 기준과 잘 부합함.",
        "hard_violations": [],
        "physics": "왼발은 문턱에 위치하나 오른발이 바닥에 평평하게 닿아 있어 체중이 양발에 분산된 형태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 실내 시점과 오른쪽을 향한 삼사분면 얼굴은 잘 맞지만, 머리가 상단에 거의 닿고 팔과 보폭의 동작이 B보다 약해 문턱을 강하게 차며 진입하는 순간이 덜 선명하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 아래 문턱을 미는 뒷발부터 오른쪽으로 기울인 상체까지 전신의 추진 동작이 명확해 지정된 진입 순간과 와이드 구도를 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울의 몸과 얼굴은 오른쪽으로 향하고 시선도 화면 오른쪽 밖을 향한다. 화면 밖 앰버를 바라보라는 방향과 맞으며, 앰버 자체는 보이지 않는다. 뒷다리는 왼쪽 문턱에 남고 앞다리는 실내 오른쪽으로 나아간다.",
        "built_space": "카메라는 실내 바닥 가까이에서 출입구를 비스듬히 본다. 왼쪽 가장자리에 열린 철문 한 짝, 왼쪽 아래에 대각선 문턱이 있다. 왼쪽 벽 창문 하나와 어두운 커튼, 바닥 침구 한 벌과 베개 하나, 뒤쪽 수납장과 선반, 천장 등 하나, 오른쪽 돌출 칸막이와 그 옆 통로가 보여 장소 참조의 주요 구성을 유지한다. 라울은 침구 앞의 빈 진입 공간에 있으며 시설과 충돌하지 않는다.",
        "entities": "인물은 라울로 보이는 어린 남자아이 한 명뿐이다. 갈색 피부, 어린 얼굴과 가는 체격, 뒤로 묶은 곱슬머리는 혼혈 아동으로 설정된 참조 외형에 부합한다. 얼룩진 올리브색 티셔츠, 베이지색 카고 반바지, 벨트, 회색 양말과 낡은 운동화도 일치한다. 얼굴과 목의 땀, 젖은 티셔츠가 분명하다. 별도 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "뒷신발 앞부분이 문턱 위에 닿아 지지점이 되고, 앞신발은 실내 바닥 바로 위에서 착지를 앞둔 모습이다. 굽힌 앞무릎과 앞으로 이동한 무게중심은 진입 보폭으로 가능하다. 팔은 다소 아래로 내려가 있지만 불가능한 자세는 아니며, 몸이 근거 없이 공중에 떠 있지는 않다."
       },
       {
        "label": "B",
        "direction": "라울은 왼쪽 출입구에서 실내 오른쪽으로 전진한다. 상체와 앞무릎이 오른쪽으로 향하고, 눈은 오른쪽 화면 밖을 주시해 앰버의 지정된 화면 밖 위치와 맞는다. 다만 얼굴은 요청한 삼사분면보다 측면에 가깝다.",
        "built_space": "실내의 낮고 비껴난 시점에서 전신을 담는다. 왼쪽 가장자리의 열린 철문 한 짝과 왼쪽 아래로 비스듬히 이어지는 문턱이 진입 기준점으로 읽힌다. 왼쪽 창문 하나와 커튼, 침구 한 벌과 베개 하나, 천장 등 하나, 뒤쪽 벽과 부분적으로 가려진 수납 시설, 오른쪽 돌출 칸막이와 좁은 통로가 참조 장소와 대응한다. 라울은 문턱과 빈 바닥 사이에 있어 동선이 자연스럽다.",
        "entities": "라울에 해당하는 어린 남자아이 한 명만 있다. 참조와 유사한 갈색 피부와 얼굴, 가는 아동 체격, 뒤로 묶은 곱슬머리가 보인다. 바랜 올리브색 티셔츠와 얼룩진 베이지색 반바지, 회색 양말과 낡은 운동화가 참조 의상에 부합한다. 얼굴과 목에는 땀의 광택이 있다. 앰버나 다른 사람은 추가되지 않았고 그래픽 문구도 없다.",
        "hard_violations": [],
        "physics": "뒷발의 앞부분이 문턱 모서리에 닿고 뒤꿈치는 들려 있어 문턱을 밀어내는 지지가 읽힌다. 앞발은 실내 바닥에 닿아 착지 하중을 받을 수 있다. 뻗은 뒷다리, 굽힌 앞무릎, 앞으로 기울인 몸통과 굽힌 팔이 달려 들어오는 보폭으로 연결되며, 지지 없는 부유나 해부학적으로 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 실내 시점과 오른쪽을 향한 삼사분면 얼굴은 잘 맞지만, 머리가 상단에 거의 닿고 팔과 보폭의 동작이 B보다 약해 문턱을 강하게 차며 진입하는 순간이 덜 선명하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 아래 문턱을 미는 뒷발부터 오른쪽으로 기울인 상체까지 전신의 추진 동작이 명확해 지정된 진입 순간과 와이드 구도를 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "라울의 몸과 얼굴은 오른쪽으로 향하고 시선도 화면 오른쪽 밖을 향한다. 화면 밖 앰버를 바라보라는 방향과 맞으며, 앰버 자체는 보이지 않는다. 뒷다리는 왼쪽 문턱에 남고 앞다리는 실내 오른쪽으로 나아간다.",
        "built_space": "카메라는 실내 바닥 가까이에서 출입구를 비스듬히 본다. 왼쪽 가장자리에 열린 철문 한 짝, 왼쪽 아래에 대각선 문턱이 있다. 왼쪽 벽 창문 하나와 어두운 커튼, 바닥 침구 한 벌과 베개 하나, 뒤쪽 수납장과 선반, 천장 등 하나, 오른쪽 돌출 칸막이와 그 옆 통로가 보여 장소 참조의 주요 구성을 유지한다. 라울은 침구 앞의 빈 진입 공간에 있으며 시설과 충돌하지 않는다.",
        "entities": "인물은 라울로 보이는 어린 남자아이 한 명뿐이다. 갈색 피부, 어린 얼굴과 가는 체격, 뒤로 묶은 곱슬머리는 혼혈 아동으로 설정된 참조 외형에 부합한다. 얼룩진 올리브색 티셔츠, 베이지색 카고 반바지, 벨트, 회색 양말과 낡은 운동화도 일치한다. 얼굴과 목의 땀, 젖은 티셔츠가 분명하다. 별도 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "뒷신발 앞부분이 문턱 위에 닿아 지지점이 되고, 앞신발은 실내 바닥 바로 위에서 착지를 앞둔 모습이다. 굽힌 앞무릎과 앞으로 이동한 무게중심은 진입 보폭으로 가능하다. 팔은 다소 아래로 내려가 있지만 불가능한 자세는 아니며, 몸이 근거 없이 공중에 떠 있지는 않다."
       },
       {
        "label": "A",
        "direction": "라울은 왼쪽 출입구에서 실내 오른쪽으로 전진한다. 상체와 앞무릎이 오른쪽으로 향하고, 눈은 오른쪽 화면 밖을 주시해 앰버의 지정된 화면 밖 위치와 맞는다. 다만 얼굴은 요청한 삼사분면보다 측면에 가깝다.",
        "built_space": "실내의 낮고 비껴난 시점에서 전신을 담는다. 왼쪽 가장자리의 열린 철문 한 짝과 왼쪽 아래로 비스듬히 이어지는 문턱이 진입 기준점으로 읽힌다. 왼쪽 창문 하나와 커튼, 침구 한 벌과 베개 하나, 천장 등 하나, 뒤쪽 벽과 부분적으로 가려진 수납 시설, 오른쪽 돌출 칸막이와 좁은 통로가 참조 장소와 대응한다. 라울은 문턱과 빈 바닥 사이에 있어 동선이 자연스럽다.",
        "entities": "라울에 해당하는 어린 남자아이 한 명만 있다. 참조와 유사한 갈색 피부와 얼굴, 가는 아동 체격, 뒤로 묶은 곱슬머리가 보인다. 바랜 올리브색 티셔츠와 얼룩진 베이지색 반바지, 회색 양말과 낡은 운동화가 참조 의상에 부합한다. 얼굴과 목에는 땀의 광택이 있다. 앰버나 다른 사람은 추가되지 않았고 그래픽 문구도 없다.",
        "hard_violations": [],
        "physics": "뒷발의 앞부분이 문턱 모서리에 닿고 뒤꿈치는 들려 있어 문턱을 밀어내는 지지가 읽힌다. 앞발은 실내 바닥에 닿아 착지 하중을 받을 수 있다. 뻗은 뒷다리, 굽힌 앞무릎, 앞으로 기울인 몸통과 굽힌 팔이 달려 들어오는 보폭으로 연결되며, 지지 없는 부유나 해부학적으로 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.889
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.889
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1889,
   "A": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "요구된 프레이밍을 정확히 따랐으며, 공중에 뜬 앞발을 통해 문턱을 차고 전진하는 역동적인 mid-stride 자세를 훌륭하게 구현함."
   },
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "프레이밍과 캐릭터의 디테일은 우수하나, 앞발이 바닥에 완전히 닿아 있어 텍스트가 요구한 강한 전진 모멘텀이 다소 반감됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L29B01.png",
    "asset_id": "d35a2e6a-cbe6-4fa1-8c57-2961d294afc0",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0925-2a18-7fc5-aaf9-0556096b84a9",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S21sh2__bgfirst_bg.png",
   "bg_asset_id": "5177e3d9-8c8d-4af2-8128-284537731e8c",
   "bg_record_key": "S21sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S21sh6::signage": {
  "fp": "c9a102e4558f5bd3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S21sh6": {
  "input_fingerprint": "84c24abef6678635",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 텅 빈 컨테이너 안의 구석자리를 향해 시선을 황급히 돌린 측면 구도.\n\nLOCATION (lock): In the sleeping area of the container home, facing an empty interior corner. Daylight enters from the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral move beside 앰버 at her seated eye level, perpendicular to her line of sight, and settle with her upper body in profile along the left third. Her shoulders remain caught in the act of rising while her head turns sharply toward the empty corner on the right, making the redirected attention—not a new lighting or distance effect—the expressive change. Observe both face and vacant space directly in the same frame, leaving the corner readable without inserting an imagined 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Empty corner at the end of 앰버's rightward gaze in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Empty container corner (Unoccupied where 앰버 searches for 찰리) — The meeting interior planes are visible beyond 앰버's profile on the right; used as Holds the destination of her gaze and gives the missing presence a concrete spatial location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime interior illumination and consistent contrast across 앰버 and the empty corner, treating the absence as physical rather than spectral.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance door remains open, with no subsequent closing described. 앰버: She has woken and sat up inside the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 텅 빈 컨테이너 안의 구석자리를 향해 시선을 황급히 돌린 측면 구도.\n\nLOCATION (lock): In the sleeping area of the container home, facing an empty interior corner. Daylight enters from the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral move beside 앰버 at her seated eye level, perpendicular to her line of sight, and settle with her upper body in profile along the left third. Her shoulders remain caught in the act of rising while her head turns sharply toward the empty corner on the right, making the redirected attention—not a new lighting or distance effect—the expressive change. Observe both face and vacant space directly in the same frame, leaving the corner readable without inserting an imagined 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Empty corner at the end of 앰버's rightward gaze in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Empty container corner (Unoccupied where 앰버 searches for 찰리) — The meeting interior planes are visible beyond 앰버's profile on the right; used as Holds the destination of her gaze and gives the missing presence a concrete spatial location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime interior illumination and consistent contrast across 앰버 and the empty corner, treating the absence as physical rather than spectral.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance door remains open, with no subsequent closing described. 앰버: She has woken and sat up inside the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 앰버가 텅 빈 컨테이너 안의 구석자리를 향해 시선을 황급히 돌린 측면 구도.\n\nLOCATION (lock): In the sleeping area of the container home, facing an empty interior corner. Daylight enters from the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral move beside 앰버 at her seated eye level, perpendicular to her line of sight, and settle with her upper body in profile along the left third. Her shoulders remain caught in the act of rising while her head turns sharply toward the empty corner on the right, making the redirected attention—not a new lighting or distance effect—the expressive change. Observe both face and vacant space directly in the same frame, leaving the corner readable without inserting an imagined 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Empty corner at the end of 앰버's rightward gaze in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Empty container corner (Unoccupied where 앰버 searches for 찰리) — The meeting interior planes are visible beyond 앰버's profile on the right; used as Holds the destination of her gaze and gives the missing presence a concrete spatial location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime interior illumination and consistent contrast across 앰버 and the empty corner, treating the absence as physical rather than spectral.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance door remains open, with no subsequent closing described. 앰버: She has woken and sat up inside the container.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선이 화면 오른쪽의 비어 있는 컨테이너 구석을 명확히 향하고 있음.",
    "built_space": "레퍼런스와 일치하는 컨테이너 내부 구조로, 열린 문, 바닥의 매트리스, 선반, 벽면의 디테일이 정확한 위치에 있음.",
    "entities": "앰버의 외모(금발, 어린 얼굴)와 복장(작업복, 방진 마스크, 공구 벨트)이 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "매트리스 위에 앉아 왼팔로 바닥을 짚어 몸의 무게를 자연스럽게 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "앰버가 화면 오른쪽을 향해 시선을 돌리고 있으나 타겟 지점이 다소 모호함.",
    "built_space": "컨테이너 내부는 레퍼런스를 따르고 있으나, 인물이 잠자리(매트리스)가 아닌 문가에 위치해 공간 사용이 어색함.",
    "entities": "앰버의 외모와 의상은 캐릭터 레퍼런스의 요건을 잘 충족하고 있음.",
    "hard_violations": [
     "[gemini-pro] 지정된 스테이징(매트리스에서 일어나 앉은 상태)을 무시하고 문가에 엉거주춤하게 서 있는 자세로 배치됨."
    ],
    "physics": "어정쩡하게 몸을 굽힌 상태로 지지점(손의 위치)이 불분명하여 물리적 안정감이 떨어짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 '매트리스 위에 일어나 앉은' 자세와 시선 방향, 측면 구도를 매우 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 위치와 자세가 '앉아있는 상태(seated)'라는 프롬프트의 명확한 연출 지시를 어기고 어정쩡하게 서 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선이 화면 오른쪽의 비어 있는 컨테이너 구석을 명확히 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 컨테이너 내부 구조로, 열린 문, 바닥의 매트리스, 선반, 벽면의 디테일이 정확한 위치에 있음.",
        "entities": "앰버의 외모(금발, 어린 얼굴)와 복장(작업복, 방진 마스크, 공구 벨트)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "매트리스 위에 앉아 왼팔로 바닥을 짚어 몸의 무게를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "앰버가 화면 오른쪽을 향해 시선을 돌리고 있으나 타겟 지점이 다소 모호함.",
        "built_space": "컨테이너 내부는 레퍼런스를 따르고 있으나, 인물이 잠자리(매트리스)가 아닌 문가에 위치해 공간 사용이 어색함.",
        "entities": "앰버의 외모와 의상은 캐릭터 레퍼런스의 요건을 잘 충족하고 있음.",
        "hard_violations": [
         "지정된 스테이징(매트리스에서 일어나 앉은 상태)을 무시하고 문가에 엉거주춤하게 서 있는 자세로 배치됨."
        ],
        "physics": "어정쩡하게 몸을 굽힌 상태로 지지점(손의 위치)이 불분명하여 물리적 안정감이 떨어짐."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 '매트리스 위에 일어나 앉은' 자세와 시선 방향, 측면 구도를 매우 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 위치와 자세가 '앉아있는 상태(seated)'라는 프롬프트의 명확한 연출 지시를 어기고 어정쩡하게 서 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선이 화면 오른쪽의 비어 있는 컨테이너 구석을 명확히 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 컨테이너 내부 구조로, 열린 문, 바닥의 매트리스, 선반, 벽면의 디테일이 정확한 위치에 있음.",
        "entities": "앰버의 외모(금발, 어린 얼굴)와 복장(작업복, 방진 마스크, 공구 벨트)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "매트리스 위에 앉아 왼팔로 바닥을 짚어 몸의 무게를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "앰버가 화면 오른쪽을 향해 시선을 돌리고 있으나 타겟 지점이 다소 모호함.",
        "built_space": "컨테이너 내부는 레퍼런스를 따르고 있으나, 인물이 잠자리(매트리스)가 아닌 문가에 위치해 공간 사용이 어색함.",
        "entities": "앰버의 외모와 의상은 캐릭터 레퍼런스의 요건을 잘 충족하고 있음.",
        "hard_violations": [
         "지정된 스테이징(매트리스에서 일어나 앉은 상태)을 무시하고 문가에 엉거주춤하게 서 있는 자세로 배치됨."
        ],
        "physics": "어정쩡하게 몸을 굽힌 상태로 지지점(손의 위치)이 불분명하여 물리적 안정감이 떨어짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽 빈 구석을 보는 측면과 장소는 맞지만, 허벅지까지 넓어진 구도와 이미 일어선 듯한 자세가 앉은 눈높이의 미디엄 숏 및 일어나려는 순간에서 벗어난다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "침구에 앉아 손으로 몸을 밀어 올리며 오른쪽 빈 구석을 돌아보는 순간을 보여 주어, 왼쪽 측면 배치와 앉은 눈높이의 미디엄 숏 지시를 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 얼굴과 눈은 화면 오른쪽을 향하며, 시선 앞에는 칸막이와 오른쪽 벽이 만나는 비어 있는 구석이 보인다. 카메라를 보지 않으며 다른 사람이나 시선을 가로막는 인물도 없다. 얼굴은 측면이지만 몸통은 카메라 쪽으로 더 열려 있다.",
        "built_space": "왼쪽에 열린 출입구 하나와 커튼이 있는 창, 바닥에 침구 하나와 베개 하나가 보인다. 뒤쪽에는 수납장 하나, 검은 선반 하나, 낮은 상자 하나가 있고 오른쪽에는 칸막이와 내부 문구멍 하나, 걸린 천과 벽 부착물이 있다. 낡은 밝은 금속 벽과 마모된 바닥은 장소 참고와 대체로 이어진다. 앰버는 침구 앞쪽 가장자리 부근에서 상당히 높이 올라온 상태라 앉은 눈높이의 측면 구도보다는 기립 중인 인물을 넓게 잡은 구도로 읽힌다.",
        "entities": "금발과 밝은 피부, 어린 얼굴의 여자아이 한 명만 있으며 앰버 참고의 외형과 대체로 맞는다. 혼혈 배경 자체를 외형만으로 확정할 수는 없다. 기름때 묻은 카키 작업복, 공구가 달린 가죽 허리띠, 정교한 금속성 방진 마스크가 있다. 마스크는 참고처럼 목에 내려와 있고 얼굴을 덮지는 않는다. 찰리나 이전 숏의 아이는 없다.",
        "hard_violations": [],
        "physics": "하체는 화면 아래로 이어지고 발은 잘려 있어 지면 접촉을 직접 확인할 수 없다. 몸은 다리 위에 놓인 자연스러운 반기립 자세이며 공중에 떠 있다는 증거는 없다. 두 손은 내려와 있어 침구를 밀어 일어나는 지지 동작은 보이지 않는다. 마스크는 목의 끈으로, 공구와 주머니는 허리띠로 지지된다. 침구와 수납물은 바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "앰버의 코와 눈이 오른쪽을 향하고, 시선이 향하는 중간 오른쪽 배경에는 뒤쪽 벽과 오른쪽 벽이 만나는 빈 구석이 있다. 얼굴을 측면에서 직접 볼 수 있고 그 앞의 빈 공간도 함께 읽힌다. 몸통을 기울인 상태에서 머리를 오른쪽으로 돌려 찾는 동작이 드러난다.",
        "built_space": "왼쪽 열린 출입구 하나, 바닥 침구 하나와 체크무늬 베개 하나, 뒤쪽 수납장 하나와 검은 선반 하나, 낮은 상자 하나가 보인다. 오른쪽에는 내부 문구멍 하나와 칸막이, 걸린 천과 벽 부착물 하나가 있다. 참고의 낡은 금속 벽, 침구, 수납 배치와 낮빛이 유지되며 중복된 고정 설비는 보이지 않는다. 앰버는 침구 앞부분에 앉아 왼쪽 영역을 차지하고, 낮은 측면 시점에서 상체와 오른쪽 빈 구석을 함께 보여 준다.",
        "entities": "금발의 어린 여자아이 한 명이며 밝은 피부, 얼굴 윤곽과 체격이 앰버 참고에 가깝다. 명시된 혈통 자체는 영상만으로 검증할 수 없다. 얼룩진 카키 작업복과 가죽 공구 벨트, 목에 걸린 미래형 방진 마스크가 보이며 참고 의상과 일치한다. 마스크는 얼굴을 가리지 않는다. 빈 구석에는 찰리나 다른 인물이 추가되지 않았다.",
        "hard_violations": [],
        "physics": "골반과 접힌 다리가 침구에 놓이고, 화면 중앙 아래의 손바닥이 침구를 눌러 상체를 지지한다. 반대쪽 손도 왼쪽 가장자리에서 침구에 닿아 있다. 이 접촉점과 앞으로 기울어진 몸통은 앉은 상태에서 몸을 일으키려는 동작으로 자연스럽다. 침구는 바닥에, 마스크는 목의 끈에, 공구는 허리띠에 지지되어 있으며 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽 빈 구석을 보는 측면과 장소는 맞지만, 허벅지까지 넓어진 구도와 이미 일어선 듯한 자세가 앉은 눈높이의 미디엄 숏 및 일어나려는 순간에서 벗어난다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "침구에 앉아 손으로 몸을 밀어 올리며 오른쪽 빈 구석을 돌아보는 순간을 보여 주어, 왼쪽 측면 배치와 앉은 눈높이의 미디엄 숏 지시를 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 얼굴과 눈은 화면 오른쪽을 향하며, 시선 앞에는 칸막이와 오른쪽 벽이 만나는 비어 있는 구석이 보인다. 카메라를 보지 않으며 다른 사람이나 시선을 가로막는 인물도 없다. 얼굴은 측면이지만 몸통은 카메라 쪽으로 더 열려 있다.",
        "built_space": "왼쪽에 열린 출입구 하나와 커튼이 있는 창, 바닥에 침구 하나와 베개 하나가 보인다. 뒤쪽에는 수납장 하나, 검은 선반 하나, 낮은 상자 하나가 있고 오른쪽에는 칸막이와 내부 문구멍 하나, 걸린 천과 벽 부착물이 있다. 낡은 밝은 금속 벽과 마모된 바닥은 장소 참고와 대체로 이어진다. 앰버는 침구 앞쪽 가장자리 부근에서 상당히 높이 올라온 상태라 앉은 눈높이의 측면 구도보다는 기립 중인 인물을 넓게 잡은 구도로 읽힌다.",
        "entities": "금발과 밝은 피부, 어린 얼굴의 여자아이 한 명만 있으며 앰버 참고의 외형과 대체로 맞는다. 혼혈 배경 자체를 외형만으로 확정할 수는 없다. 기름때 묻은 카키 작업복, 공구가 달린 가죽 허리띠, 정교한 금속성 방진 마스크가 있다. 마스크는 참고처럼 목에 내려와 있고 얼굴을 덮지는 않는다. 찰리나 이전 숏의 아이는 없다.",
        "hard_violations": [],
        "physics": "하체는 화면 아래로 이어지고 발은 잘려 있어 지면 접촉을 직접 확인할 수 없다. 몸은 다리 위에 놓인 자연스러운 반기립 자세이며 공중에 떠 있다는 증거는 없다. 두 손은 내려와 있어 침구를 밀어 일어나는 지지 동작은 보이지 않는다. 마스크는 목의 끈으로, 공구와 주머니는 허리띠로 지지된다. 침구와 수납물은 바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "앰버의 코와 눈이 오른쪽을 향하고, 시선이 향하는 중간 오른쪽 배경에는 뒤쪽 벽과 오른쪽 벽이 만나는 빈 구석이 있다. 얼굴을 측면에서 직접 볼 수 있고 그 앞의 빈 공간도 함께 읽힌다. 몸통을 기울인 상태에서 머리를 오른쪽으로 돌려 찾는 동작이 드러난다.",
        "built_space": "왼쪽 열린 출입구 하나, 바닥 침구 하나와 체크무늬 베개 하나, 뒤쪽 수납장 하나와 검은 선반 하나, 낮은 상자 하나가 보인다. 오른쪽에는 내부 문구멍 하나와 칸막이, 걸린 천과 벽 부착물 하나가 있다. 참고의 낡은 금속 벽, 침구, 수납 배치와 낮빛이 유지되며 중복된 고정 설비는 보이지 않는다. 앰버는 침구 앞부분에 앉아 왼쪽 영역을 차지하고, 낮은 측면 시점에서 상체와 오른쪽 빈 구석을 함께 보여 준다.",
        "entities": "금발의 어린 여자아이 한 명이며 밝은 피부, 얼굴 윤곽과 체격이 앰버 참고에 가깝다. 명시된 혈통 자체는 영상만으로 검증할 수 없다. 얼룩진 카키 작업복과 가죽 공구 벨트, 목에 걸린 미래형 방진 마스크가 보이며 참고 의상과 일치한다. 마스크는 얼굴을 가리지 않는다. 빈 구석에는 찰리나 다른 인물이 추가되지 않았다.",
        "hard_violations": [],
        "physics": "골반과 접힌 다리가 침구에 놓이고, 화면 중앙 아래의 손바닥이 침구를 눌러 상체를 지지한다. 반대쪽 손도 왼쪽 가장자리에서 침구에 닿아 있다. 이 접촉점과 앞으로 기울어진 몸통은 앉은 상태에서 몸을 일으키려는 동작으로 자연스럽다. 침구는 바닥에, 마스크는 목의 끈에, 공구는 허리띠에 지지되어 있으며 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.206
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.956
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 스테이징(매트리스에서 일어나 앉은 상태)을 무시하고 문가에 엉거주춤하게 서 있는 자세로 배치됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 956
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 '매트리스 위에 일어나 앉은' 자세와 시선 방향, 측면 구도를 매우 충실히 구현함."
   },
   {
    "label": "B",
    "score": 956,
    "verdict_ko": "캐릭터의 위치와 자세가 '앉아있는 상태(seated)'라는 프롬프트의 명확한 연출 지시를 어기고 어정쩡하게 서 있음.  ★위반: [gemini-pro] 지정된 스테이징(매트리스에서 일어나 앉은 상태)을 무시하고 문가에 엉거주춤하게 서 있는 자세로 배치됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S21sh2_sel.png",
    "asset_id": "168ee888-d912-44c0-a24f-dbf0abcff7e9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab092b-6239-71c4-a2e3-7e22df263033",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh2"
  }
 },
 "S22sh2::signage": {
  "fp": "0bc536cf36114d04",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::banana_market": {
  "input_fingerprint": "082d7569eab5cbab",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "banana_market",
    "tags": [
     "S22sh2",
     "S22sh6",
     "S22sh9"
    ]
   },
   "context_sig": "7a09e42d77ce854d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately in front of a banana stall on the refugee settlement's open-air market street.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그러다 바나나를 파는 가게 앞에 멈추는 찰리.\n- 인파 틈으로 사라지는 찰리. 그 자리로 뛰어오는 앰버와 라울.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately in front of a banana stall on the refugee settlement's open-air market street.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그러다 바나나를 파는 가게 앞에 멈추는 찰리.\n- 인파 틈으로 사라지는 찰리. 그 자리로 뛰어오는 앰버와 라울.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_banana_market_5128db.png",
  "asset_id": "61f139c6-d230-4193-9a4e-b4aa2a250efb",
  "input_asset_ids": [
   "d92489e2-d827-4a66-be8b-a86dfc85890d"
  ],
  "origin_tag": "S22sh2",
  "place_text": "Immediately in front of a banana stall on the refugee settlement's open-air market street.",
  "origin_inputs": {
   "place_text": "Immediately in front of a banana stall on the refugee settlement's open-air market street.",
   "time_of_day_en": "day",
   "conti_asset_id": "d92489e2-d827-4a66-be8b-a86dfc85890d"
  }
 },
 "S22sh2::bgfirst_bg": {
  "input_fingerprint": "9998602288cdb8fc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노란 바나나가 가득 쌓인 가판대 앞으로 찰리의 두꺼운 기계 손이 반쯤 뻗은 채 허공에 머물러 있는 근접 구도.\n\nLOCATION (lock): Immediately in front of a banana stall on the refugee settlement's open-air market street.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the rear-three-quarter approach almost to a stop beside 찰리's reaching hand, with the lens just above hand height and angled gently downward toward the bananas. His thick mechanical hand hangs across the lower center, occupying less than a third of the frame, while yellow bananas sit to the right and a partial forearm and torso at the left preserve scale and bodily context. Show the suspended reach directly, with his head outside the crop and his downward attention continuing along the arm toward the fruit; keep a visible interval between fingertips and bananas.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banana stall (Piled with yellow bananas) — Viewed obliquely across the customer-facing display toward the fruit; used as Places the desired fruit along the hand's reaching direction while keeping the display below forty percent of the frame; Market aisle (Open beside the stall) — A narrow strip extends behind the partial robot torso; used as Preserves environmental context so the hand does not become an isolated, oversized object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled highlights retain precise mechanical articulation, while the bananas' stated yellow remains a restrained local color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노란 바나나가 가득 쌓인 가판대 앞으로 찰리의 두꺼운 기계 손이 반쯤 뻗은 채 허공에 머물러 있는 근접 구도.\n\nLOCATION (lock): Immediately in front of a banana stall on the refugee settlement's open-air market street.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the rear-three-quarter approach almost to a stop beside 찰리's reaching hand, with the lens just above hand height and angled gently downward toward the bananas. His thick mechanical hand hangs across the lower center, occupying less than a third of the frame, while yellow bananas sit to the right and a partial forearm and torso at the left preserve scale and bodily context. Show the suspended reach directly, with his head outside the crop and his downward attention continuing along the arm toward the fruit; keep a visible interval between fingertips and bananas.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banana stall (Piled with yellow bananas) — Viewed obliquely across the customer-facing display toward the fruit; used as Places the desired fruit along the hand's reaching direction while keeping the display below forty percent of the frame; Market aisle (Open beside the stall) — A narrow strip extends behind the partial robot torso; used as Preserves environmental context so the hand does not become an isolated, oversized object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled highlights retain precise mechanical articulation, while the bananas' stated yellow remains a restrained local color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S22sh2__bgfirst_bg.png",
  "asset_id": "57b0674d-990c-4bb1-b712-fd8c49c2c706",
  "input_asset_ids": [
   "d92489e2-d827-4a66-be8b-a86dfc85890d",
   "61f139c6-d230-4193-9a4e-b4aa2a250efb"
  ]
 },
 "S22sh2": {
  "input_fingerprint": "6010ea42bd4ff8b8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노란 바나나가 가득 쌓인 가판대 앞으로 찰리의 두꺼운 기계 손이 반쯤 뻗은 채 허공에 머물러 있는 근접 구도.\n\nLOCATION (lock): Immediately in front of a banana stall on the refugee settlement's open-air market street. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the rear-three-quarter approach almost to a stop beside 찰리's reaching hand, with the lens just above hand height and angled gently downward toward the bananas. His thick mechanical hand hangs across the lower center, occupying less than a third of the frame, while yellow bananas sit to the right and a partial forearm and torso at the left preserve scale and bodily context. Show the suspended reach directly, with his head outside the crop and his downward attention continuing along the arm toward the fruit; keep a visible interval between fingertips and bananas.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banana stall (Piled with yellow bananas) — Viewed obliquely across the customer-facing display toward the fruit; used as Places the desired fruit along the hand's reaching direction while keeping the display below forty percent of the frame; Market aisle (Open beside the stall) — A narrow strip extends behind the partial robot torso; used as Preserves environmental context so the hand does not become an isolated, oversized object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled highlights retain precise mechanical articulation, while the bananas' stated yellow remains a restrained local color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bananas are displayed at a stall on the busy market street. Charlie retains the old coat and hat, washed body, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노란 바나나가 가득 쌓인 가판대 앞으로 찰리의 두꺼운 기계 손이 반쯤 뻗은 채 허공에 머물러 있는 근접 구도.\n\nLOCATION (lock): Immediately in front of a banana stall on the refugee settlement's open-air market street. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the rear-three-quarter approach almost to a stop beside 찰리's reaching hand, with the lens just above hand height and angled gently downward toward the bananas. His thick mechanical hand hangs across the lower center, occupying less than a third of the frame, while yellow bananas sit to the right and a partial forearm and torso at the left preserve scale and bodily context. Show the suspended reach directly, with his head outside the crop and his downward attention continuing along the arm toward the fruit; keep a visible interval between fingertips and bananas.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banana stall (Piled with yellow bananas) — Viewed obliquely across the customer-facing display toward the fruit; used as Places the desired fruit along the hand's reaching direction while keeping the display below forty percent of the frame; Market aisle (Open beside the stall) — A narrow strip extends behind the partial robot torso; used as Preserves environmental context so the hand does not become an isolated, oversized object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled highlights retain precise mechanical articulation, while the bananas' stated yellow remains a restrained local color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bananas are displayed at a stall on the busy market street. Charlie retains the old coat and hat, washed body, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노란 바나나가 가득 쌓인 가판대 앞으로 찰리의 두꺼운 기계 손이 반쯤 뻗은 채 허공에 머물러 있는 근접 구도.\n\nLOCATION (lock): Immediately in front of a banana stall on the refugee settlement's open-air market street. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the rear-three-quarter approach almost to a stop beside 찰리's reaching hand, with the lens just above hand height and angled gently downward toward the bananas. His thick mechanical hand hangs across the lower center, occupying less than a third of the frame, while yellow bananas sit to the right and a partial forearm and torso at the left preserve scale and bodily context. Show the suspended reach directly, with his head outside the crop and his downward attention continuing along the arm toward the fruit; keep a visible interval between fingertips and bananas.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Banana stall (Piled with yellow bananas) — Viewed obliquely across the customer-facing display toward the fruit; used as Places the desired fruit along the hand's reaching direction while keeping the display below forty percent of the frame; Market aisle (Open beside the stall) — A narrow strip extends behind the partial robot torso; used as Preserves environmental context so the hand does not become an isolated, oversized object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled highlights retain precise mechanical articulation, while the bananas' stated yellow remains a restrained local color accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bananas are displayed at a stall on the busy market street. Charlie retains the old coat and hat, washed body, blue-lit eyes and worn UBIK chest logo.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S22sh2__bgfirst_bg.png",
     "asset_id": "57b0674d-990c-4bb1-b712-fd8c49c2c706",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S22sh2.png",
     "asset_id": "d92489e2-d827-4a66-be8b-a86dfc85890d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_banana_market_5128db.png",
     "asset_id": "61f139c6-d230-4193-9a4e-b4aa2a250efb",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "기계 손의 손가락들이 우측 하단의 바나나를 향해 뻗어 있음.",
    "built_space": "우측에 바나나 매대, 좌측에 시장 통로와 천막, 구조물들이 보이는 야외 시장 배경임.",
    "entities": "좌측에 코트를 걸친 기계 몸통이 있으나 장갑판이 흰색에 가깝고 UBIK 로고가 없음. 우측에는 노란 바나나가 있음.",
    "hard_violations": [
     "[gpt-high] 찰리만 허용한 인물 조건을 어기고 진열대 뒤 상인과 통로의 다수 행인을 추가했다."
    ],
    "physics": "허공에 멈춘 기계 손은 몸통과 연결된 팔에 의해 지지되며, 주변 사물들의 배치와 무게감이 자연스러움."
   },
   {
    "label": "B",
    "direction": "기계 손이 우측 하단의 바나나 무더기를 향해 뻗어 있음.",
    "built_space": "야외 시장 거리로, 우측에 바나나 가판대가 있고 좌측 배경으로 컨테이너와 천막이 늘어선 통로가 이어짐.",
    "entities": "좌측에 코트를 입은 찰리의 기계 팔과 몸통 일부가 보임(샌드 베이지 장갑판, 푸른 원자로, UBIK 로고 일치). 우측 가판대에는 노란 바나나들이 있음.",
    "hard_violations": [
     "[gpt-high] 찰리 외에는 인물을 보여 주지 말라는 조건과 달리 오른쪽 중경의 과일 상자 뒤에 별도 인물을 추가했다."
    ],
    "physics": "뻗은 기계 팔은 몸통에 자연스럽게 지지되어 있고, 바나나들은 가판대 나무 상자 안에 안정적으로 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 근접 구도와 배경을 훌륭히 재현했으며, 샌드 베이지 장갑판과 명시된 UBIK 로고 등 캐릭터의 세부 설정을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 구도와 동작은 잘 따랐으나, 찰리의 장갑판 색상이 기준과 다르고 가슴의 UBIK 로고가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "기계 손이 우측 하단의 바나나 무더기를 향해 뻗어 있음.",
        "built_space": "야외 시장 거리로, 우측에 바나나 가판대가 있고 좌측 배경으로 컨테이너와 천막이 늘어선 통로가 이어짐.",
        "entities": "좌측에 코트를 입은 찰리의 기계 팔과 몸통 일부가 보임(샌드 베이지 장갑판, 푸른 원자로, UBIK 로고 일치). 우측 가판대에는 노란 바나나들이 있음.",
        "hard_violations": [],
        "physics": "뻗은 기계 팔은 몸통에 자연스럽게 지지되어 있고, 바나나들은 가판대 나무 상자 안에 안정적으로 놓여 있음."
       },
       {
        "label": "A",
        "direction": "기계 손의 손가락들이 우측 하단의 바나나를 향해 뻗어 있음.",
        "built_space": "우측에 바나나 매대, 좌측에 시장 통로와 천막, 구조물들이 보이는 야외 시장 배경임.",
        "entities": "좌측에 코트를 걸친 기계 몸통이 있으나 장갑판이 흰색에 가깝고 UBIK 로고가 없음. 우측에는 노란 바나나가 있음.",
        "hard_violations": [],
        "physics": "허공에 멈춘 기계 손은 몸통과 연결된 팔에 의해 지지되며, 주변 사물들의 배치와 무게감이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 근접 구도와 배경을 훌륭히 재현했으며, 샌드 베이지 장갑판과 명시된 UBIK 로고 등 캐릭터의 세부 설정을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 구도와 동작은 잘 따랐으나, 찰리의 장갑판 색상이 기준과 다르고 가슴의 UBIK 로고가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "기계 손이 우측 하단의 바나나 무더기를 향해 뻗어 있음.",
        "built_space": "야외 시장 거리로, 우측에 바나나 가판대가 있고 좌측 배경으로 컨테이너와 천막이 늘어선 통로가 이어짐.",
        "entities": "좌측에 코트를 입은 찰리의 기계 팔과 몸통 일부가 보임(샌드 베이지 장갑판, 푸른 원자로, UBIK 로고 일치). 우측 가판대에는 노란 바나나들이 있음.",
        "hard_violations": [],
        "physics": "뻗은 기계 팔은 몸통에 자연스럽게 지지되어 있고, 바나나들은 가판대 나무 상자 안에 안정적으로 놓여 있음."
       },
       {
        "label": "A",
        "direction": "기계 손의 손가락들이 우측 하단의 바나나를 향해 뻗어 있음.",
        "built_space": "우측에 바나나 매대, 좌측에 시장 통로와 천막, 구조물들이 보이는 야외 시장 배경임.",
        "entities": "좌측에 코트를 걸친 기계 몸통이 있으나 장갑판이 흰색에 가깝고 UBIK 로고가 없음. 우측에는 노란 바나나가 있음.",
        "hard_violations": [],
        "physics": "허공에 멈춘 기계 손은 몸통과 연결된 팔에 의해 지지되며, 주변 사물들의 배치와 무게감이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두꺼운 기계 손의 멈춘 뻗기와 바나나 사이 간격은 잘 맞지만, 금지된 배경 인물이 있고 후방 사선 대신 가슴 정면이 보이는 구도다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "손끝이 바나나를 향하고 접촉 전 간격도 있으나, 상인과 다수 행인을 추가했으며 가슴 정면과 넓은 시장 통로를 보여 지정 구도에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 몸통에서 나온 팔이 오른쪽 아래 바나나 진열대로 이어지고, 굽힌 손가락 끝은 가장 가까운 바나나 위의 빈 공간을 향한다. 과일과 손끝은 떨어져 있다. 머리는 잘려 시선 자체는 확인할 수 없다. 가슴의 원자로와 로고가 카메라를 향해 있어 지정된 후방 사선 접근은 아니다.",
        "built_space": "오른쪽 아래에 목재 바나나 진열대 한 곳, 오른쪽 위에 매달린 바나나 묶음 한 곳과 갓등 한 개가 보인다. 천막, 금속 지주, 과일 상자, 뒤쪽 컨테이너와 전선은 장소 참고의 시장 구조에 부합한다. 왼쪽에는 코트 입은 몸통과 팔이 있지만 통로가 좁은 띠가 아니라 화면 중앙을 넓게 차지한다. 진열대는 대략 화면의 40% 이내이며 손도 3분의 1보다 작다. 오른쪽 중경 과일 상자 뒤에 흰 상의 인물이 보인다.",
        "entities": "주체는 사람 피부가 아닌 샌드 베이지 장갑판과 노출 관절로 된 기계이며, 두꺼운 손과 낡은 코트가 요구에 맞는다. 보이는 가슴에는 푸른 원자로와 UBIK 표기가 있다. 노란 바나나는 실제 과일의 반점과 곡면 질감으로 표현된다. 얼굴, 눈, 모자는 화면 밖이므로 평가 대상이 아니다. 배경 인물은 찰리만 허용한 조건에 어긋난다.",
        "hard_violations": [
         "찰리 외에는 인물을 보여 주지 말라는 조건과 달리 오른쪽 중경의 과일 상자 뒤에 별도 인물을 추가했다."
        ],
        "physics": "손은 손목·전완·팔꿈치를 통해 몸통에 연결되어 있으며 관절이 반쯤 뻗은 자세를 지탱한다. 허공에 있는 손이 독립적으로 떠 있는 것은 아니다. 바나나는 목재 진열대에 쌓여 있고, 위쪽 묶음은 천막 구조의 매달림 위치에 있다. 코트는 몸과 팔 위에 걸쳐 자연스럽게 처진다. 명백한 무지지 부유나 불가능한 관절 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽에서 나온 기계 팔과 펼친 손가락은 오른쪽 아래의 바나나를 향하고 손끝 앞에 명확한 간격이 남아 있다. 머리가 잘려 찰리의 눈 방향은 보이지 않는다. 오른쪽 상인은 진열대 쪽으로 고개를 숙이고, 중앙 행인들은 주로 통로 안쪽으로 향한다. 찰리의 가슴 앞면이 드러나 후방 사선 시점과 다르다.",
        "built_space": "오른쪽의 목재 바나나 진열대 한 곳, 상단의 매달린 바나나 묶음 한 곳, 갓등 한 개, 천막 지주와 뒤쪽 과일 상자가 보인다. 컨테이너, 전선과 시장 바닥도 장소 참고에 가깝다. 왼쪽 몸통과 하단 중앙 손이라는 배치는 맞지만 중앙 통로가 넓고 깊게 열려 있다. 진열대 뒤에는 상인이 서 있고 통로에는 다수 행인이 배치되어 있다. 손과 진열대의 화면 점유율 자체는 지정 범위에 대체로 들어간다.",
        "entities": "찰리의 보이는 부분은 밝은 베이지색 기계 장갑판과 낡은 코트다. 손가락은 참고의 육중한 손보다 가늘고 길며, 보이는 푸른 가슴 장치는 참고의 거대한 원자로보다 작다. 가슴의 보이는 부분에서 UBIK 표기는 확인되지 않는다. 노란 바나나는 요구한 과일에 맞는다. 얼굴과 모자는 화면 밖이다. 오른쪽 성인 남성으로 보이는 상인과 여러 행인은 허용되지 않은 추가 인물이다.",
        "hard_violations": [
         "찰리만 허용한 인물 조건을 어기고 진열대 뒤 상인과 통로의 다수 행인을 추가했다."
        ],
        "physics": "기계 손은 손목과 전완을 통해 몸통에 연결되어 있어 멈춘 뻗기 자세가 물리적으로 가능하다. 바나나는 목재 상자와 진열면이 받치며, 상단 묶음은 위쪽 구조에 매달려 있다. 행인들은 발을 시장 바닥에 딛고 있고 중앙 인물의 봉투는 양손에 들려 있다. 명백하게 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두꺼운 기계 손의 멈춘 뻗기와 바나나 사이 간격은 잘 맞지만, 금지된 배경 인물이 있고 후방 사선 대신 가슴 정면이 보이는 구도다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "손끝이 바나나를 향하고 접촉 전 간격도 있으나, 상인과 다수 행인을 추가했으며 가슴 정면과 넓은 시장 통로를 보여 지정 구도에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 몸통에서 나온 팔이 오른쪽 아래 바나나 진열대로 이어지고, 굽힌 손가락 끝은 가장 가까운 바나나 위의 빈 공간을 향한다. 과일과 손끝은 떨어져 있다. 머리는 잘려 시선 자체는 확인할 수 없다. 가슴의 원자로와 로고가 카메라를 향해 있어 지정된 후방 사선 접근은 아니다.",
        "built_space": "오른쪽 아래에 목재 바나나 진열대 한 곳, 오른쪽 위에 매달린 바나나 묶음 한 곳과 갓등 한 개가 보인다. 천막, 금속 지주, 과일 상자, 뒤쪽 컨테이너와 전선은 장소 참고의 시장 구조에 부합한다. 왼쪽에는 코트 입은 몸통과 팔이 있지만 통로가 좁은 띠가 아니라 화면 중앙을 넓게 차지한다. 진열대는 대략 화면의 40% 이내이며 손도 3분의 1보다 작다. 오른쪽 중경 과일 상자 뒤에 흰 상의 인물이 보인다.",
        "entities": "주체는 사람 피부가 아닌 샌드 베이지 장갑판과 노출 관절로 된 기계이며, 두꺼운 손과 낡은 코트가 요구에 맞는다. 보이는 가슴에는 푸른 원자로와 UBIK 표기가 있다. 노란 바나나는 실제 과일의 반점과 곡면 질감으로 표현된다. 얼굴, 눈, 모자는 화면 밖이므로 평가 대상이 아니다. 배경 인물은 찰리만 허용한 조건에 어긋난다.",
        "hard_violations": [
         "찰리 외에는 인물을 보여 주지 말라는 조건과 달리 오른쪽 중경의 과일 상자 뒤에 별도 인물을 추가했다."
        ],
        "physics": "손은 손목·전완·팔꿈치를 통해 몸통에 연결되어 있으며 관절이 반쯤 뻗은 자세를 지탱한다. 허공에 있는 손이 독립적으로 떠 있는 것은 아니다. 바나나는 목재 진열대에 쌓여 있고, 위쪽 묶음은 천막 구조의 매달림 위치에 있다. 코트는 몸과 팔 위에 걸쳐 자연스럽게 처진다. 명백한 무지지 부유나 불가능한 관절 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽에서 나온 기계 팔과 펼친 손가락은 오른쪽 아래의 바나나를 향하고 손끝 앞에 명확한 간격이 남아 있다. 머리가 잘려 찰리의 눈 방향은 보이지 않는다. 오른쪽 상인은 진열대 쪽으로 고개를 숙이고, 중앙 행인들은 주로 통로 안쪽으로 향한다. 찰리의 가슴 앞면이 드러나 후방 사선 시점과 다르다.",
        "built_space": "오른쪽의 목재 바나나 진열대 한 곳, 상단의 매달린 바나나 묶음 한 곳, 갓등 한 개, 천막 지주와 뒤쪽 과일 상자가 보인다. 컨테이너, 전선과 시장 바닥도 장소 참고에 가깝다. 왼쪽 몸통과 하단 중앙 손이라는 배치는 맞지만 중앙 통로가 넓고 깊게 열려 있다. 진열대 뒤에는 상인이 서 있고 통로에는 다수 행인이 배치되어 있다. 손과 진열대의 화면 점유율 자체는 지정 범위에 대체로 들어간다.",
        "entities": "찰리의 보이는 부분은 밝은 베이지색 기계 장갑판과 낡은 코트다. 손가락은 참고의 육중한 손보다 가늘고 길며, 보이는 푸른 가슴 장치는 참고의 거대한 원자로보다 작다. 가슴의 보이는 부분에서 UBIK 표기는 확인되지 않는다. 노란 바나나는 요구한 과일에 맞는다. 얼굴과 모자는 화면 밖이다. 오른쪽 성인 남성으로 보이는 상인과 여러 행인은 허용되지 않은 추가 인물이다.",
        "hard_violations": [
         "찰리만 허용한 인물 조건을 어기고 진열대 뒤 상인과 통로의 다수 행인을 추가했다."
        ],
        "physics": "기계 손은 손목과 전완을 통해 몸통에 연결되어 있어 멈춘 뻗기 자세가 물리적으로 가능하다. 바나나는 목재 상자와 진열면이 받치며, 상단 묶음은 위쪽 구조에 매달려 있다. 행인들은 발을 시장 바닥에 딛고 있고 중앙 인물의 봉투는 양손에 들려 있다. 명백하게 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.381,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.131,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gpt-high] 찰리 외에는 인물을 보여 주지 말라는 조건과 달리 오른쪽 중경의 과일 상자 뒤에 별도 인물을 추가했다."
    ],
    "A": [
     "[gpt-high] 찰리만 허용한 인물 조건을 어기고 진열대 뒤 상인과 통로의 다수 행인을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 1131
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "요구된 근접 구도와 배경을 훌륭히 재현했으며, 샌드 베이지 장갑판과 명시된 UBIK 로고 등 캐릭터의 세부 설정을 정확히 구현했습니다.  ★위반: [gpt-high] 찰리 외에는 인물을 보여 주지 말라는 조건과 달리 오른쪽 중경의 과일 상자 뒤에 별도 인물을 추가했다."
   },
   {
    "label": "A",
    "score": 1131,
    "verdict_ko": "지정된 구도와 동작은 잘 따랐으나, 찰리의 장갑판 색상이 기준과 다르고 가슴의 UBIK 로고가 누락되었습니다.  ★위반: [gpt-high] 찰리만 허용한 인물 조건을 어기고 진열대 뒤 상인과 통로의 다수 행인을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_banana_market_5128db.png",
    "asset_id": "61f139c6-d230-4193-9a4e-b4aa2a250efb",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0931-65ce-7486-8215-f9eb313dfcea",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S22sh2__bgfirst_bg.png",
   "bg_asset_id": "57b0674d-990c-4bb1-b712-fd8c49c2c706",
   "bg_record_key": "S22sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "banana_market",
   "groupbg_asset_id": "61f139c6-d230-4193-9a4e-b4aa2a250efb"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S22sh6::signage": {
  "fp": "86d04cb1e4878a3e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh6": {
  "input_fingerprint": "358e6df2ba316c26",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 난민들 인파 속으로 반쯤 가려진 채 걸음을 내딛는 찰리의 거대한 뒷모습.\n\nLOCATION (lock): Within the dense pedestrian crowd on the refugee settlement's outdoor market street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue behind 찰리 from the established offset line with the lens below his shoulders and tilted gently upward, then let his increasing distance pull him deeper into a wide market view as the camera remains near the stall. His large back advances through the middle of the frame toward the market passage ahead, while intervening refugees obscure roughly half of it; do not sidestep to regain a clean silhouette. Observe the crowd directly as individuals at different walking phases, with uneven shoulder angles and small variations in spacing and distant gazes along the aisle rather than synchronized movement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Crowded market passage (Busy with refugees moving between the camera and 찰리) — The passage recedes ahead of 찰리 through the center of the image; used as Creates layered, irregular occlusion through differing stride phases, body angles, and natural gaps rather than a uniform wall of people; Banana stall (Remains beside the camera as 찰리 moves away) — Only a small oblique edge of the display remains at the lower side of the image; used as Provides a stationary location reference during the release from following movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep subdued daylight and consistent exposure across people and robot, allowing ordinary crowd occlusion rather than a lighting effect to conceal 찰리.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains on the crowded market street. Charlie is moving into the crowd with the old coat and hat still on, his washed but worn body unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 난민들 인파 속으로 반쯤 가려진 채 걸음을 내딛는 찰리의 거대한 뒷모습.\n\nLOCATION (lock): Within the dense pedestrian crowd on the refugee settlement's outdoor market street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue behind 찰리 from the established offset line with the lens below his shoulders and tilted gently upward, then let his increasing distance pull him deeper into a wide market view as the camera remains near the stall. His large back advances through the middle of the frame toward the market passage ahead, while intervening refugees obscure roughly half of it; do not sidestep to regain a clean silhouette. Observe the crowd directly as individuals at different walking phases, with uneven shoulder angles and small variations in spacing and distant gazes along the aisle rather than synchronized movement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Crowded market passage (Busy with refugees moving between the camera and 찰리) — The passage recedes ahead of 찰리 through the center of the image; used as Creates layered, irregular occlusion through differing stride phases, body angles, and natural gaps rather than a uniform wall of people; Banana stall (Remains beside the camera as 찰리 moves away) — Only a small oblique edge of the display remains at the lower side of the image; used as Provides a stationary location reference during the release from following movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep subdued daylight and consistent exposure across people and robot, allowing ordinary crowd occlusion rather than a lighting effect to conceal 찰리.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains on the crowded market street. Charlie is moving into the crowd with the old coat and hat still on, his washed but worn body unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 난민들 인파 속으로 반쯤 가려진 채 걸음을 내딛는 찰리의 거대한 뒷모습.\n\nLOCATION (lock): Within the dense pedestrian crowd on the refugee settlement's outdoor market street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue behind 찰리 from the established offset line with the lens below his shoulders and tilted gently upward, then let his increasing distance pull him deeper into a wide market view as the camera remains near the stall. His large back advances through the middle of the frame toward the market passage ahead, while intervening refugees obscure roughly half of it; do not sidestep to regain a clean silhouette. Observe the crowd directly as individuals at different walking phases, with uneven shoulder angles and small variations in spacing and distant gazes along the aisle rather than synchronized movement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Crowded market passage (Busy with refugees moving between the camera and 찰리) — The passage recedes ahead of 찰리 through the center of the image; used as Creates layered, irregular occlusion through differing stride phases, body angles, and natural gaps rather than a uniform wall of people; Banana stall (Remains beside the camera as 찰리 moves away) — Only a small oblique edge of the display remains at the lower side of the image; used as Provides a stationary location reference during the release from following movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep subdued daylight and consistent exposure across people and robot, allowing ordinary crowd occlusion rather than a lighting effect to conceal 찰리.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains on the crowded market street. Charlie is moving into the crowd with the old coat and hat still on, his washed but worn body unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리와 군중 대부분이 화면 안쪽의 시장 통로를 향해 걸어가고 있으며, 시선과 움직임의 방향이 올바르게 일치함.",
    "built_space": "야외 시장 통로로, 우측 하단에 이전 샷과 동일한 바나나 가판대가 정확한 디테일로 배치되어 있으며 배경의 천막과 컨테이너 구조도 자연스러움.",
    "entities": "찰리는 이전 샷과 완벽히 일치하는 질감의 코트를 입고 샌드 베이지색 기계 부품을 노출한 채 걷고 있음. 다민족 난민 군중이 화면을 적절히 가리고 있으며, 바나나의 상표 스티커까지 동일하게 묘사됨.",
    "hard_violations": [],
    "physics": "모든 인물이 땅에 정상적으로 발을 딛고 걷고 있으며, 의상과 가판대의 사물들이 중력에 맞게 자연스럽게 놓여 있음."
   },
   {
    "label": "B",
    "direction": "찰리와 군중이 시장 통로 안쪽을 향해 이동하고 있으며 방향성은 프롬프트와 부합함.",
    "built_space": "야외 시장 구도와 우측 하단의 바나나 가판대가 존재하지만, 이전 샷과 비교해 가판대와 주변 구조물의 세부적인 형태가 다름.",
    "entities": "찰리의 코트에 이전 샷에 없던 후드가 추가되었고, 뒤통수에 레퍼런스 시트에 없는 임의의 원형 문양이 생성됨. 바나나 가판대의 상표 스티커 디테일이 사라짐.",
    "hard_violations": [],
    "physics": "인물들의 보행 자세와 체중 이동이 자연스럽게 땅에 지지되고 있으며 물리적 오류는 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 낡은 코트 질감과 바나나 가판대의 디테일(스티커 등)을 완벽하게 유지했으며, 프롬프트에서 요구한 로봇의 뒷모습과 군중 사이의 구도를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스에 없는 후드가 코트에 추가되었고 로봇의 뒤통수에 불필요한 문양이 들어갔으며, 이전 샷에 있던 바나나 스티커 디테일이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리와 군중 대부분이 화면 안쪽의 시장 통로를 향해 걸어가고 있으며, 시선과 움직임의 방향이 올바르게 일치함.",
        "built_space": "야외 시장 통로로, 우측 하단에 이전 샷과 동일한 바나나 가판대가 정확한 디테일로 배치되어 있으며 배경의 천막과 컨테이너 구조도 자연스러움.",
        "entities": "찰리는 이전 샷과 완벽히 일치하는 질감의 코트를 입고 샌드 베이지색 기계 부품을 노출한 채 걷고 있음. 다민족 난민 군중이 화면을 적절히 가리고 있으며, 바나나의 상표 스티커까지 동일하게 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 땅에 정상적으로 발을 딛고 걷고 있으며, 의상과 가판대의 사물들이 중력에 맞게 자연스럽게 놓여 있음."
       },
       {
        "label": "B",
        "direction": "찰리와 군중이 시장 통로 안쪽을 향해 이동하고 있으며 방향성은 프롬프트와 부합함.",
        "built_space": "야외 시장 구도와 우측 하단의 바나나 가판대가 존재하지만, 이전 샷과 비교해 가판대와 주변 구조물의 세부적인 형태가 다름.",
        "entities": "찰리의 코트에 이전 샷에 없던 후드가 추가되었고, 뒤통수에 레퍼런스 시트에 없는 임의의 원형 문양이 생성됨. 바나나 가판대의 상표 스티커 디테일이 사라짐.",
        "hard_violations": [],
        "physics": "인물들의 보행 자세와 체중 이동이 자연스럽게 땅에 지지되고 있으며 물리적 오류는 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 낡은 코트 질감과 바나나 가판대의 디테일(스티커 등)을 완벽하게 유지했으며, 프롬프트에서 요구한 로봇의 뒷모습과 군중 사이의 구도를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스에 없는 후드가 코트에 추가되었고 로봇의 뒤통수에 불필요한 문양이 들어갔으며, 이전 샷에 있던 바나나 스티커 디테일이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리와 군중 대부분이 화면 안쪽의 시장 통로를 향해 걸어가고 있으며, 시선과 움직임의 방향이 올바르게 일치함.",
        "built_space": "야외 시장 통로로, 우측 하단에 이전 샷과 동일한 바나나 가판대가 정확한 디테일로 배치되어 있으며 배경의 천막과 컨테이너 구조도 자연스러움.",
        "entities": "찰리는 이전 샷과 완벽히 일치하는 질감의 코트를 입고 샌드 베이지색 기계 부품을 노출한 채 걷고 있음. 다민족 난민 군중이 화면을 적절히 가리고 있으며, 바나나의 상표 스티커까지 동일하게 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 땅에 정상적으로 발을 딛고 걷고 있으며, 의상과 가판대의 사물들이 중력에 맞게 자연스럽게 놓여 있음."
       },
       {
        "label": "B",
        "direction": "찰리와 군중이 시장 통로 안쪽을 향해 이동하고 있으며 방향성은 프롬프트와 부합함.",
        "built_space": "야외 시장 구도와 우측 하단의 바나나 가판대가 존재하지만, 이전 샷과 비교해 가판대와 주변 구조물의 세부적인 형태가 다름.",
        "entities": "찰리의 코트에 이전 샷에 없던 후드가 추가되었고, 뒤통수에 레퍼런스 시트에 없는 임의의 원형 문양이 생성됨. 바나나 가판대의 상표 스티커 디테일이 사라짐.",
        "hard_violations": [],
        "physics": "인물들의 보행 자세와 체중 이동이 자연스럽게 땅에 지지되고 있으며 물리적 오류는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인파 사이로 내딛는 동작은 분명하지만 찰리가 더 가깝고 크게 남으며, 등 절반의 가림이 부족하고 착용을 유지해야 할 모자가 없다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리가 조금 더 깊이 들어간 중앙 후방 와이드 구도와 낡은 코트·모자 유지가 더 충실하지만, 등 절반 가림과 바나나 진열대의 작은 가장자리만 남기는 조건은 미달한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 카메라에 등을 보이고 화면 중앙의 시장 통로 안쪽으로 전진한다. 왼쪽 전경 사람들은 카메라 쪽 또는 화면 왼쪽으로 지나가고, 오른쪽 검은 비니 남자는 화면 오른쪽을 본다. 배낭을 멘 중경 인물들은 통로 안쪽을 향해 있어 이동 방향이 획일적이지 않다. 무기나 조준 대상은 없다.",
        "built_space": "중앙 통로 양옆에 천막 노점과 금속 지지대가 이어지고, 멀리 적층 컨테이너와 전선이 보인다. 오른쪽 노점의 갓등 한 개와 왼쪽 줄의 여러 전구가 보이며, 오른쪽 아래에는 바나나 진열대 한 개가 있다. 젖은 노면과 천막 재질은 이전 장소와 이어지지만 진열대는 작은 가장자리가 아니라 상당한 전경 면적을 차지한다. 찰리의 왼쪽 몸통은 가까운 사람에게 가려져 있으나 등 중앙 대부분은 드러나며, 등 절반 정도를 가리는 배치는 아니다.",
        "entities": "찰리는 한 명이며 낡은 갈색 코트, 샌드 베이지 기계 머리와 팔·다리가 보인다. 긴 팔과 육중한 체형은 맞지만 주변 성인보다 상당히 크게 읽힌다. 머리 장갑이 노출되어 있어 유지해야 할 모자는 보이지 않는다. 얼굴과 가슴 원자로는 후면 구도에 맞게 숨겨져 있다. 군중은 서로 다른 피부색과 복장의 성인 남녀로 보이며, 구체적인 국적은 확인할 수 없다. 바나나와 나무 진열대가 있고 새로 삽입된 문구나 화면 표식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 한쪽 기계발을 젖은 지면에 대고 다른 다리를 인파 뒤로 옮기는 보행 순간으로 읽힌다. 노출된 팔과 손은 어깨·팔꿈치 관절에 연결되어 있다. 사람들의 배낭은 어깨끈으로 지지되고 바나나는 나무 진열대에 놓여 있다. 천막과 등은 기둥 및 상부 구조에 달려 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 머리와 등은 중앙 시장 통로의 먼 쪽을 향한다. 중앙의 모자 쓴 배낭 인물과 오른쪽 보라색 옷 인물도 통로 안쪽으로 이동한다. 왼쪽 전경 여성은 아래쪽을 보고 반대 방향으로 지나가며, 오른쪽 비니 남자와 두건 여성은 화면 오른쪽을 본다. 시선과 어깨 방향에 차이가 있고 무기나 조준 대상은 없다.",
        "built_space": "중앙의 젖은 통로 양쪽으로 천막 노점, 금속 기둥, 상자들이 이어지고 뒤에는 컨테이너 건물과 전선이 있다. 오른쪽에 갓등 한 개, 상단 가장자리에 매달린 바나나 묶음 한 곳, 아래쪽에 나무 바나나 진열대 한 개가 보여 이전 장소의 주요 요소를 유지한다. 찰리는 A보다 조금 작고 통로 안쪽에 위치하지만, 왼쪽 인물과 오른쪽 배낭 인물 사이로 등 대부분이 드러난다. 바나나 진열대도 요구된 작은 사선 가장자리보다 넓게 남는다.",
        "entities": "찰리는 한 명이며 낡은 갈색 코트와 어두운 낡은 모자를 착용하고, 샌드 베이지의 관절식 팔과 짧은 기계 다리가 노출된다. 얼굴과 가슴은 뒤를 향한 구도 때문에 보이지 않아 적절하다. 체격은 넓고 육중하지만 주변 성인보다 커 보여 키 작은 성인 남성 정도라는 크기 조건에는 덜 맞는다. 군중에는 다양한 피부색과 복장의 성인 남녀가 보이며 국적은 판별할 수 없다. 오른쪽에는 바나나와 상품 상자가 있고 추가 문구나 그래픽 표식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 기계발은 앞뒤로 어긋나 지면에 닿아 있어 체중 지지가 가능하다. 다만 A보다 보폭과 추진 동작은 약하게 읽힌다. 군중의 배낭은 어깨끈에 걸려 있고, 오른쪽 남자의 상자는 손과 팔 앞에서 받쳐진다. 진열 바나나는 나무 상자 위에 놓이고 상단 바나나는 노점 상부에 매달려 있다. 지지 없는 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인파 사이로 내딛는 동작은 분명하지만 찰리가 더 가깝고 크게 남으며, 등 절반의 가림이 부족하고 착용을 유지해야 할 모자가 없다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리가 조금 더 깊이 들어간 중앙 후방 와이드 구도와 낡은 코트·모자 유지가 더 충실하지만, 등 절반 가림과 바나나 진열대의 작은 가장자리만 남기는 조건은 미달한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 카메라에 등을 보이고 화면 중앙의 시장 통로 안쪽으로 전진한다. 왼쪽 전경 사람들은 카메라 쪽 또는 화면 왼쪽으로 지나가고, 오른쪽 검은 비니 남자는 화면 오른쪽을 본다. 배낭을 멘 중경 인물들은 통로 안쪽을 향해 있어 이동 방향이 획일적이지 않다. 무기나 조준 대상은 없다.",
        "built_space": "중앙 통로 양옆에 천막 노점과 금속 지지대가 이어지고, 멀리 적층 컨테이너와 전선이 보인다. 오른쪽 노점의 갓등 한 개와 왼쪽 줄의 여러 전구가 보이며, 오른쪽 아래에는 바나나 진열대 한 개가 있다. 젖은 노면과 천막 재질은 이전 장소와 이어지지만 진열대는 작은 가장자리가 아니라 상당한 전경 면적을 차지한다. 찰리의 왼쪽 몸통은 가까운 사람에게 가려져 있으나 등 중앙 대부분은 드러나며, 등 절반 정도를 가리는 배치는 아니다.",
        "entities": "찰리는 한 명이며 낡은 갈색 코트, 샌드 베이지 기계 머리와 팔·다리가 보인다. 긴 팔과 육중한 체형은 맞지만 주변 성인보다 상당히 크게 읽힌다. 머리 장갑이 노출되어 있어 유지해야 할 모자는 보이지 않는다. 얼굴과 가슴 원자로는 후면 구도에 맞게 숨겨져 있다. 군중은 서로 다른 피부색과 복장의 성인 남녀로 보이며, 구체적인 국적은 확인할 수 없다. 바나나와 나무 진열대가 있고 새로 삽입된 문구나 화면 표식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 한쪽 기계발을 젖은 지면에 대고 다른 다리를 인파 뒤로 옮기는 보행 순간으로 읽힌다. 노출된 팔과 손은 어깨·팔꿈치 관절에 연결되어 있다. 사람들의 배낭은 어깨끈으로 지지되고 바나나는 나무 진열대에 놓여 있다. 천막과 등은 기둥 및 상부 구조에 달려 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 머리와 등은 중앙 시장 통로의 먼 쪽을 향한다. 중앙의 모자 쓴 배낭 인물과 오른쪽 보라색 옷 인물도 통로 안쪽으로 이동한다. 왼쪽 전경 여성은 아래쪽을 보고 반대 방향으로 지나가며, 오른쪽 비니 남자와 두건 여성은 화면 오른쪽을 본다. 시선과 어깨 방향에 차이가 있고 무기나 조준 대상은 없다.",
        "built_space": "중앙의 젖은 통로 양쪽으로 천막 노점, 금속 기둥, 상자들이 이어지고 뒤에는 컨테이너 건물과 전선이 있다. 오른쪽에 갓등 한 개, 상단 가장자리에 매달린 바나나 묶음 한 곳, 아래쪽에 나무 바나나 진열대 한 개가 보여 이전 장소의 주요 요소를 유지한다. 찰리는 A보다 조금 작고 통로 안쪽에 위치하지만, 왼쪽 인물과 오른쪽 배낭 인물 사이로 등 대부분이 드러난다. 바나나 진열대도 요구된 작은 사선 가장자리보다 넓게 남는다.",
        "entities": "찰리는 한 명이며 낡은 갈색 코트와 어두운 낡은 모자를 착용하고, 샌드 베이지의 관절식 팔과 짧은 기계 다리가 노출된다. 얼굴과 가슴은 뒤를 향한 구도 때문에 보이지 않아 적절하다. 체격은 넓고 육중하지만 주변 성인보다 커 보여 키 작은 성인 남성 정도라는 크기 조건에는 덜 맞는다. 군중에는 다양한 피부색과 복장의 성인 남녀가 보이며 국적은 판별할 수 없다. 오른쪽에는 바나나와 상품 상자가 있고 추가 문구나 그래픽 표식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 기계발은 앞뒤로 어긋나 지면에 닿아 있어 체중 지지가 가능하다. 다만 A보다 보폭과 추진 동작은 약하게 읽힌다. 군중의 배낭은 어깨끈에 걸려 있고, 오른쪽 남자의 상자는 손과 팔 앞에서 받쳐진다. 진열 바나나는 나무 상자 위에 놓이고 상단 바나나는 노점 상부에 매달려 있다. 지지 없는 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.429
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.429
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이전 샷의 낡은 코트 질감과 바나나 가판대의 디테일(스티커 등)을 완벽하게 유지했으며, 프롬프트에서 요구한 로봇의 뒷모습과 군중 사이의 구도를 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "레퍼런스에 없는 후드가 코트에 추가되었고 로봇의 뒤통수에 불필요한 문양이 들어갔으며, 이전 샷에 있던 바나나 스티커 디테일이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S22sh2_sel.png",
    "asset_id": "c85296a8-5ae8-44c2-81cb-06d4751c4846",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab093a-2d8e-7fbb-a92c-4b5ade8bb52c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh2"
  }
 },
 "S22sh9::signage": {
  "fp": "3bcae16e87865fd6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh9": {
  "input_fingerprint": "1dc842cded8c1790",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 음악 소리가 울려 퍼지는 시장 인파 쪽을 향해 팔을 힘껏 뻗어 가리키는 앰버의 상체.\n\nLOCATION (lock): On the outdoor market lane beside the banana stall, looking toward the crowd and the music farther along the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the short lateral track beside the stall at 앰버's shoulder height, keeping a three-quarter side view outside the line of her pointing arm. Place her upper body on the left and carry the fully extended arm diagonally toward the upper-right background, leaving visible space beyond her fingertips toward the crowd where 찰리 disappeared. Her weight is still forward from arriving at a run and her distant gaze follows that same market direction; the directly observed background figures continue at varied stride phases rather than turning toward the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Market crowd beyond 앰버's pointing fingertips in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market crowd beyond the pointing hand (찰리 is no longer distinctly visible among the people) — The occupied passage recedes into the upper-right background; used as Supplies the gesture's destination; varied shoulders, walking phases, and spacing preserve the crowd as individuals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's subdued daylight and restrained contrast, keeping the pointing hand distinct without adding a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains beside the crowded market route. Charlie has moved into the crowd, still in the old coat and hat, with the washed metal body and worn chest logo unchanged. 앰버: She has reached the banana-stall area and is directing the continuing search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 음악 소리가 울려 퍼지는 시장 인파 쪽을 향해 팔을 힘껏 뻗어 가리키는 앰버의 상체.\n\nLOCATION (lock): On the outdoor market lane beside the banana stall, looking toward the crowd and the music farther along the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the short lateral track beside the stall at 앰버's shoulder height, keeping a three-quarter side view outside the line of her pointing arm. Place her upper body on the left and carry the fully extended arm diagonally toward the upper-right background, leaving visible space beyond her fingertips toward the crowd where 찰리 disappeared. Her weight is still forward from arriving at a run and her distant gaze follows that same market direction; the directly observed background figures continue at varied stride phases rather than turning toward the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Market crowd beyond 앰버's pointing fingertips in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market crowd beyond the pointing hand (찰리 is no longer distinctly visible among the people) — The occupied passage recedes into the upper-right background; used as Supplies the gesture's destination; varied shoulders, walking phases, and spacing preserve the crowd as individuals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's subdued daylight and restrained contrast, keeping the pointing hand distinct without adding a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains beside the crowded market route. Charlie has moved into the crowd, still in the old coat and hat, with the washed metal body and worn chest logo unchanged. 앰버: She has reached the banana-stall area and is directing the continuing search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 음악 소리가 울려 퍼지는 시장 인파 쪽을 향해 팔을 힘껏 뻗어 가리키는 앰버의 상체.\n\nLOCATION (lock): On the outdoor market lane beside the banana stall, looking toward the crowd and the music farther along the street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the short lateral track beside the stall at 앰버's shoulder height, keeping a three-quarter side view outside the line of her pointing arm. Place her upper body on the left and carry the fully extended arm diagonally toward the upper-right background, leaving visible space beyond her fingertips toward the crowd where 찰리 disappeared. Her weight is still forward from arriving at a run and her distant gaze follows that same market direction; the directly observed background figures continue at varied stride phases rather than turning toward the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Market crowd beyond 앰버's pointing fingertips in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Market crowd beyond the pointing hand (찰리 is no longer distinctly visible among the people) — The occupied passage recedes into the upper-right background; used as Supplies the gesture's destination; varied shoulders, walking phases, and spacing preserve the crowd as individuals.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's subdued daylight and restrained contrast, keeping the pointing hand distinct without adding a separate accent light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The banana stall remains beside the crowded market route. Charlie has moved into the crowd, still in the old coat and hat, with the washed metal body and worn chest logo unchanged. 앰버: She has reached the banana-stall area and is directing the continuing search.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "앰버가 우측 상단의 인파를 향해 왼팔을 뻗어 가리키고 있으며, 시선도 같은 곳을 향함.",
    "built_space": "좌측에 바나나 가판대가 위치하며 야외 시장 골목이 배경으로 넓게 이어짐.",
    "entities": "앰버는 금발, 마스크, 작업복 등 기준에 부합하나, 레퍼런스의 배경 인물들이 복사되어 나타남.",
    "hard_violations": [
     "[gemini-pro] 이전 샷의 인물들(꽃무늬 배낭을 멘 여성, 회색 후드 남성 등)을 이 샷에 포함하지 말라는 명시적 지시를 어기고 그대로 복사하여 배치함.",
     "[gpt-high] 군중 속에서 더 이상 뚜렷이 보이지 않아야 하는 찰리를 중앙에 코트·모자·기계 손발로 식별 가능하게 배치했다.",
     "[gpt-high] 이전 장면의 인물과 의상을 재사용하지 말라는 금지에도 꽃무늬 머릿수건 인물과 야구모자·회색 후드·대형 배낭 인물을 재현했다."
    ],
    "physics": "앰버는 상체를 앞으로 기울인 자세로 두 발이 바닥에 안정적으로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "앰버가 우측 상단의 인파 쪽을 향해 손가락으로 가리키며 시선이 방향과 일치함.",
    "built_space": "좌측의 바나나 가판대와 시장 통로 공간이 구조에 맞게 구현됨.",
    "entities": "앰버의 외형과 복장은 일치하지만 레퍼런스 이미지의 특정 군중이 동일하게 등장함.",
    "hard_violations": [
     "[gemini-pro] 금지된 이전 샷의 군중(꽃무늬 배낭을 멘 여성, 비니를 쓴 남성 등)의 외형과 복장을 그대로 복사하여 재사용함.",
     "[gpt-high] 이전 장면 인물의 의상을 누구에게도 옮기지 말라는 명시적 금지에도 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 자주색 상의·올림머리·배낭 인물을 배경에 재현했다."
    ],
    "physics": "몸을 앞으로 숙인 자세를 취하며 발목과 다리가 땅에 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트가 엄격히 금지한 이전 샷의 특정 인물들(꽃무늬 배낭을 멘 여성 등)을 그대로 복사해 배치하는 치명적 위반을 범함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 마찬가지로 금지된 이전 샷의 군중을 그대로 재사용하는 치명적인 오류를 범하여 지시를 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버가 우측 상단의 인파를 향해 왼팔을 뻗어 가리키고 있으며, 시선도 같은 곳을 향함.",
        "built_space": "좌측에 바나나 가판대가 위치하며 야외 시장 골목이 배경으로 넓게 이어짐.",
        "entities": "앰버는 금발, 마스크, 작업복 등 기준에 부합하나, 레퍼런스의 배경 인물들이 복사되어 나타남.",
        "hard_violations": [
         "이전 샷의 인물들(꽃무늬 배낭을 멘 여성, 회색 후드 남성 등)을 이 샷에 포함하지 말라는 명시적 지시를 어기고 그대로 복사하여 배치함."
        ],
        "physics": "앰버는 상체를 앞으로 기울인 자세로 두 발이 바닥에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "앰버가 우측 상단의 인파 쪽을 향해 손가락으로 가리키며 시선이 방향과 일치함.",
        "built_space": "좌측의 바나나 가판대와 시장 통로 공간이 구조에 맞게 구현됨.",
        "entities": "앰버의 외형과 복장은 일치하지만 레퍼런스 이미지의 특정 군중이 동일하게 등장함.",
        "hard_violations": [
         "금지된 이전 샷의 군중(꽃무늬 배낭을 멘 여성, 비니를 쓴 남성 등)의 외형과 복장을 그대로 복사하여 재사용함."
        ],
        "physics": "몸을 앞으로 숙인 자세를 취하며 발목과 다리가 땅에 지지됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트가 엄격히 금지한 이전 샷의 특정 인물들(꽃무늬 배낭을 멘 여성 등)을 그대로 복사해 배치하는 치명적 위반을 범함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 마찬가지로 금지된 이전 샷의 군중을 그대로 재사용하는 치명적인 오류를 범하여 지시를 위반함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "앰버가 우측 상단의 인파를 향해 왼팔을 뻗어 가리키고 있으며, 시선도 같은 곳을 향함.",
        "built_space": "좌측에 바나나 가판대가 위치하며 야외 시장 골목이 배경으로 넓게 이어짐.",
        "entities": "앰버는 금발, 마스크, 작업복 등 기준에 부합하나, 레퍼런스의 배경 인물들이 복사되어 나타남.",
        "hard_violations": [
         "이전 샷의 인물들(꽃무늬 배낭을 멘 여성, 회색 후드 남성 등)을 이 샷에 포함하지 말라는 명시적 지시를 어기고 그대로 복사하여 배치함."
        ],
        "physics": "앰버는 상체를 앞으로 기울인 자세로 두 발이 바닥에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "앰버가 우측 상단의 인파 쪽을 향해 손가락으로 가리키며 시선이 방향과 일치함.",
        "built_space": "좌측의 바나나 가판대와 시장 통로 공간이 구조에 맞게 구현됨.",
        "entities": "앰버의 외형과 복장은 일치하지만 레퍼런스 이미지의 특정 군중이 동일하게 등장함.",
        "hard_violations": [
         "금지된 이전 샷의 군중(꽃무늬 배낭을 멘 여성, 비니를 쓴 남성 등)의 외형과 복장을 그대로 복사하여 재사용함."
        ],
        "physics": "몸을 앞으로 숙인 자세를 취하며 발목과 다리가 땅에 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 상체와 우상향으로 뻗은 팔, 손끝 너머 군중이라는 구도는 잘 맞지만, 이전 장면의 군중 의상을 재사용해 인물 연속성 금지 조건을 위반했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "앞으로 쏠린 상체와 지시 동작은 맞지만, 이전 군중을 재사용하고 코트·모자·기계 손발의 찰리까지 뚜렷하게 남겨 찰리가 식별되지 않아야 한다는 조건을 위반했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 검지는 오른쪽 위로 뻗으며, 그 너머에는 시장 통로와 멀어지는 군중이 있다. 눈도 같은 오른쪽 시장 방향을 본다. 손끝은 화면상 군중 머리보다 높지만, 통로 깊숙한 곳을 가리키는 동작으로 읽힌다. 배경 인물들은 대부분 카메라 반대쪽으로 걷거나 옆을 보며, 일제히 카메라를 향하지 않는다.",
        "built_space": "왼쪽에 바나나 진열대 한 구역과 매달린 바나나, 큰 갓등 한 개가 보이고 양쪽에는 금속 기둥과 낡은 천막이 이어진다. 뒤로 컨테이너, 전봇대와 전선이 있으며 젖은 통로가 오른쪽 배경으로 이어진다. 앰버는 진열대 옆 전경에 있고 군중은 통로를 차지한다. 참고의 시장 재료와 설비 유형은 유지되지만, 바나나 가판대가 화면 반대편에 있어 동일 가판대의 정확한 공간 연결은 확인하기 어렵다. 명백히 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "앰버는 금발과 밝은 피부, 어린 얼굴을 가진 여자아이로 보이며 참고 인물과 대체로 닮았다. 혼혈 배경 자체는 외모만으로 확정할 수 없다. 기름때 묻은 카키 작업복과 가죽 공구 벨트가 보인다. 방진 마스크는 참고처럼 목에 걸려 있으나 필터와 배관 형태가 다르다. 바나나와 개별 인물로 구성된 군중이 있고 찰리의 기계 몸은 명확히 드러나지 않는다. 다만 중앙의 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 오른쪽의 자주색 상의·올림머리·배낭 인물은 이전 장면의 의상 조합을 재사용했다.",
        "hard_violations": [
         "이전 장면 인물의 의상을 누구에게도 옮기지 말라는 명시적 금지에도 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 자주색 상의·올림머리·배낭 인물을 배경에 재현했다."
        ],
        "physics": "앰버의 팔은 어깨에서 손끝까지 자연스럽게 연결되어 있고 팔꿈치를 펴서 가리키는 자세가 가능하다. 상체는 약간 앞으로 기울어 있으나 발은 프레임 밖이므로 접지 상태는 판단할 수 없다. 이것만으로 공중에 떠 있다고 볼 근거는 없다. 마스크는 목끈, 공구는 허리 벨트, 바나나는 진열대와 매달림 장치로 지지된다. 발이 보이는 군중은 노면에 접지하며 걷고 있다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선과 곧게 뻗은 검지가 모두 오른쪽 먼 시장 방향을 향한다. 손끝 너머에 군중과 통로가 남아 있어 지시 대상은 읽힌다. 다만 바로 그 방향의 중앙에 코트와 모자를 쓴 찰리로 식별되는 인물이 서 있어, 이미 사람들 속에서 식별되지 않는 대상을 찾는 장면보다는 보이는 찰리를 가리키는 장면에 가까워진다. 군중은 주로 화면 안쪽으로 걸어간다.",
        "built_space": "왼쪽에 매달린 바나나와 목재 상자형 진열대 한 구역, 큰 갓등 한 개가 있고 오른쪽 가판대에도 갓등 한 개가 보인다. 양쪽 금속 지주와 천막 사이로 젖은 통로가 중앙 오른쪽으로 후퇴하며, 컨테이너와 전선이 배경을 채운다. 앰버는 왼쪽 가판대 바로 옆에 있다. 참고 시장의 재료와 마모는 유사하고 구조적으로 불가능한 배치는 보이지 않지만, 참고에서 오른쪽이던 바나나 가판대와의 정확한 위치 관계는 확정하기 어렵다.",
        "entities": "금발의 어린 여자아이, 밝은 피부, 카키 작업복과 공구 벨트는 앰버 설정에 대체로 맞는다. 목에 걸린 배관형 마스크도 참고 디자인과 유사하다. 중앙에는 낡은 모자와 긴 코트 아래로 기계 손과 다리가 드러난 찰리가 별개의 인물로 보인다. 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 오른쪽 야구모자·회색 후드·대형 배낭 인물 역시 이전 장면의 의상을 직접적으로 재현한다. 바나나와 시장 군중은 존재하며 별도 음악 장치는 보이지 않지만, 화면에 음악 장치가 나와야 한다는 조건은 없다.",
        "hard_violations": [
         "군중 속에서 더 이상 뚜렷이 보이지 않아야 하는 찰리를 중앙에 코트·모자·기계 손발로 식별 가능하게 배치했다.",
         "이전 장면의 인물과 의상을 재사용하지 말라는 금지에도 꽃무늬 머릿수건 인물과 야구모자·회색 후드·대형 배낭 인물을 재현했다."
        ],
        "physics": "앰버의 상체가 앞으로 쏠리고 팔이 어깨에서 대각선으로 뻗어 있어 달려온 직후 지시하는 동작으로 자연스럽게 읽힌다. 하체 접지는 프레임 밖이라 확인할 수 없으며, 부유를 시사하는 장면은 아니다. 마스크는 목끈에, 공구 주머니는 벨트에 매달려 있고 바나나는 가판대나 상부 걸이에 지지된다. 군중과 중앙 기계 인물의 보이는 발은 노면과 접촉하며, 일부 발을 들어 올린 자세도 보행 단계로 설명된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 상체와 우상향으로 뻗은 팔, 손끝 너머 군중이라는 구도는 잘 맞지만, 이전 장면의 군중 의상을 재사용해 인물 연속성 금지 조건을 위반했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "앞으로 쏠린 상체와 지시 동작은 맞지만, 이전 군중을 재사용하고 코트·모자·기계 손발의 찰리까지 뚜렷하게 남겨 찰리가 식별되지 않아야 한다는 조건을 위반했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 검지는 오른쪽 위로 뻗으며, 그 너머에는 시장 통로와 멀어지는 군중이 있다. 눈도 같은 오른쪽 시장 방향을 본다. 손끝은 화면상 군중 머리보다 높지만, 통로 깊숙한 곳을 가리키는 동작으로 읽힌다. 배경 인물들은 대부분 카메라 반대쪽으로 걷거나 옆을 보며, 일제히 카메라를 향하지 않는다.",
        "built_space": "왼쪽에 바나나 진열대 한 구역과 매달린 바나나, 큰 갓등 한 개가 보이고 양쪽에는 금속 기둥과 낡은 천막이 이어진다. 뒤로 컨테이너, 전봇대와 전선이 있으며 젖은 통로가 오른쪽 배경으로 이어진다. 앰버는 진열대 옆 전경에 있고 군중은 통로를 차지한다. 참고의 시장 재료와 설비 유형은 유지되지만, 바나나 가판대가 화면 반대편에 있어 동일 가판대의 정확한 공간 연결은 확인하기 어렵다. 명백히 중복된 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "앰버는 금발과 밝은 피부, 어린 얼굴을 가진 여자아이로 보이며 참고 인물과 대체로 닮았다. 혼혈 배경 자체는 외모만으로 확정할 수 없다. 기름때 묻은 카키 작업복과 가죽 공구 벨트가 보인다. 방진 마스크는 참고처럼 목에 걸려 있으나 필터와 배관 형태가 다르다. 바나나와 개별 인물로 구성된 군중이 있고 찰리의 기계 몸은 명확히 드러나지 않는다. 다만 중앙의 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 오른쪽의 자주색 상의·올림머리·배낭 인물은 이전 장면의 의상 조합을 재사용했다.",
        "hard_violations": [
         "이전 장면 인물의 의상을 누구에게도 옮기지 말라는 명시적 금지에도 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 자주색 상의·올림머리·배낭 인물을 배경에 재현했다."
        ],
        "physics": "앰버의 팔은 어깨에서 손끝까지 자연스럽게 연결되어 있고 팔꿈치를 펴서 가리키는 자세가 가능하다. 상체는 약간 앞으로 기울어 있으나 발은 프레임 밖이므로 접지 상태는 판단할 수 없다. 이것만으로 공중에 떠 있다고 볼 근거는 없다. 마스크는 목끈, 공구는 허리 벨트, 바나나는 진열대와 매달림 장치로 지지된다. 발이 보이는 군중은 노면에 접지하며 걷고 있다."
       },
       {
        "label": "A",
        "direction": "앰버의 시선과 곧게 뻗은 검지가 모두 오른쪽 먼 시장 방향을 향한다. 손끝 너머에 군중과 통로가 남아 있어 지시 대상은 읽힌다. 다만 바로 그 방향의 중앙에 코트와 모자를 쓴 찰리로 식별되는 인물이 서 있어, 이미 사람들 속에서 식별되지 않는 대상을 찾는 장면보다는 보이는 찰리를 가리키는 장면에 가까워진다. 군중은 주로 화면 안쪽으로 걸어간다.",
        "built_space": "왼쪽에 매달린 바나나와 목재 상자형 진열대 한 구역, 큰 갓등 한 개가 있고 오른쪽 가판대에도 갓등 한 개가 보인다. 양쪽 금속 지주와 천막 사이로 젖은 통로가 중앙 오른쪽으로 후퇴하며, 컨테이너와 전선이 배경을 채운다. 앰버는 왼쪽 가판대 바로 옆에 있다. 참고 시장의 재료와 마모는 유사하고 구조적으로 불가능한 배치는 보이지 않지만, 참고에서 오른쪽이던 바나나 가판대와의 정확한 위치 관계는 확정하기 어렵다.",
        "entities": "금발의 어린 여자아이, 밝은 피부, 카키 작업복과 공구 벨트는 앰버 설정에 대체로 맞는다. 목에 걸린 배관형 마스크도 참고 디자인과 유사하다. 중앙에는 낡은 모자와 긴 코트 아래로 기계 손과 다리가 드러난 찰리가 별개의 인물로 보인다. 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 오른쪽 야구모자·회색 후드·대형 배낭 인물 역시 이전 장면의 의상을 직접적으로 재현한다. 바나나와 시장 군중은 존재하며 별도 음악 장치는 보이지 않지만, 화면에 음악 장치가 나와야 한다는 조건은 없다.",
        "hard_violations": [
         "군중 속에서 더 이상 뚜렷이 보이지 않아야 하는 찰리를 중앙에 코트·모자·기계 손발로 식별 가능하게 배치했다.",
         "이전 장면의 인물과 의상을 재사용하지 말라는 금지에도 꽃무늬 머릿수건 인물과 야구모자·회색 후드·대형 배낭 인물을 재현했다."
        ],
        "physics": "앰버의 상체가 앞으로 쏠리고 팔이 어깨에서 대각선으로 뻗어 있어 달려온 직후 지시하는 동작으로 자연스럽게 읽힌다. 하체 접지는 프레임 밖이라 확인할 수 없으며, 부유를 시사하는 장면은 아니다. 마스크는 목끈에, 공구 주머니는 벨트에 매달려 있고 바나나는 가판대나 상부 걸이에 지지된다. 군중과 중앙 기계 인물의 보이는 발은 노면과 접촉하며, 일부 발을 들어 올린 자세도 보행 단계로 설명된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷의 인물들(꽃무늬 배낭을 멘 여성, 회색 후드 남성 등)을 이 샷에 포함하지 말라는 명시적 지시를 어기고 그대로 복사하여 배치함.",
     "[gpt-high] 군중 속에서 더 이상 뚜렷이 보이지 않아야 하는 찰리를 중앙에 코트·모자·기계 손발로 식별 가능하게 배치했다.",
     "[gpt-high] 이전 장면의 인물과 의상을 재사용하지 말라는 금지에도 꽃무늬 머릿수건 인물과 야구모자·회색 후드·대형 배낭 인물을 재현했다."
    ],
    "B": [
     "[gemini-pro] 금지된 이전 샷의 군중(꽃무늬 배낭을 멘 여성, 비니를 쓴 남성 등)의 외형과 복장을 그대로 복사하여 재사용함.",
     "[gpt-high] 이전 장면 인물의 의상을 누구에게도 옮기지 말라는 명시적 금지에도 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 자주색 상의·올림머리·배낭 인물을 배경에 재현했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1250,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "프롬프트가 엄격히 금지한 이전 샷의 특정 인물들(꽃무늬 배낭을 멘 여성 등)을 그대로 복사해 배치하는 치명적 위반을 범함.  ★위반: [gemini-pro] 이전 샷의 인물들(꽃무늬 배낭을 멘 여성, 회색 후드 남성 등)을 이 샷에 포함하지 말라는 명시적 지시를 어기고 그대로 복사하여 배치함. / [gpt-high] 군중 속에서 더 이상 뚜렷이 보이지 않아야 하는 찰리를 중앙에 코트·모자·기계 손발로 식별 가능하게 배치했다. / [gpt-high] 이전 장면의 인물과 의상을 재사용하지 말라는 금지에도 꽃무늬 머릿수건 인물과 야구모자·회색 후드·대형 배낭 인물을 재현했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "A와 마찬가지로 금지된 이전 샷의 군중을 그대로 재사용하는 치명적인 오류를 범하여 지시를 위반함.  ★위반: [gemini-pro] 금지된 이전 샷의 군중(꽃무늬 배낭을 멘 여성, 비니를 쓴 남성 등)의 외형과 복장을 그대로 복사하여 재사용함. / [gpt-high] 이전 장면 인물의 의상을 누구에게도 옮기지 말라는 명시적 금지에도 꽃무늬 머릿수건·녹색 외투·꽃무늬 배낭 인물과 자주색 상의·올림머리·배낭 인물을 배경에 재현했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S22sh6_sel.png",
    "asset_id": "ed349e4a-2785-4640-9ad1-75bb2355ab08",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab093f-3c1f-746e-901c-80eab1f1051f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh6"
  }
 },
 "S23sh1::signage": {
  "fp": "02f4e8a1f8fc483d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::a6edc64f6a03cae2": {
  "subjects": [],
  "subject_text": "구치소 면회실\n넓고 단순한 면회 공간. 실내 한가운데 테이블과 마주 보는 의자가 놓여 있고, 한쪽에 출입문이 있다.",
  "identity": "canonical",
  "scope_id": "L33",
  "scope_role": "location_interior",
  "scope_sha": "348ebaa1090709ce"
 },
 "S23sh1::bgfirst_bg": {
  "input_fingerprint": "7302bfcf0bee373c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차가운 면회실 철제 테이블 앞에 두 손목에 수갑이 채워진 채 고개를 숙인 이현우의 피투성이 상체.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, illuminated for daytime visits.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at the establishing stage's held entry position beside a long edge of the table, above 이현우's seated eye level and looking diagonally downward at his front three-quarter view before the crane descends. Place his bowed head and bloodied, bruised upper body just right of center, with both cuffed wrists visible low in the composition and his gaze lowered toward his hands. Observe him directly, using the table's diagonal edge and a restrained area of the spacious room behind him to express isolation without enlarging the table beyond the human scale.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Metal visitation table (Positioned in the middle of the room with 이현우 seated at it) — Its top and one long edge are seen obliquely from above; used as The diagonal edge leads toward the wrists while the tabletop occupies less than forty percent of the image; Handcuffs (Fastened around both wrists) — Seen from above at the lower center beside the table edge; used as Makes his confinement readable within the upper-body composition; Spacious visitation-room interior (Visible beyond the centrally placed table) — A limited expanse of the room recedes behind 이현우; used as Provides negative space around the isolated seated figure without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the visitation room stays subdued and controlled, retaining injury detail without sensational highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차가운 면회실 철제 테이블 앞에 두 손목에 수갑이 채워진 채 고개를 숙인 이현우의 피투성이 상체.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, illuminated for daytime visits.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at the establishing stage's held entry position beside a long edge of the table, above 이현우's seated eye level and looking diagonally downward at his front three-quarter view before the crane descends. Place his bowed head and bloodied, bruised upper body just right of center, with both cuffed wrists visible low in the composition and his gaze lowered toward his hands. Observe him directly, using the table's diagonal edge and a restrained area of the spacious room behind him to express isolation without enlarging the table beyond the human scale.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Metal visitation table (Positioned in the middle of the room with 이현우 seated at it) — Its top and one long edge are seen obliquely from above; used as The diagonal edge leads toward the wrists while the tabletop occupies less than forty percent of the image; Handcuffs (Fastened around both wrists) — Seen from above at the lower center beside the table edge; used as Makes his confinement readable within the upper-body composition; Spacious visitation-room interior (Visible beyond the centrally placed table) — A limited expanse of the room recedes behind 이현우; used as Provides negative space around the isolated seated figure without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the visitation room stays subdued and controlled, retaining injury detail without sensational highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S23sh1__bgfirst_bg.png",
  "asset_id": "6b2bfcad-7ccf-476c-a01d-50ee349e5a7e",
  "input_asset_ids": [
   "14df2182-3f90-4ddb-ad5f-5cf99e0be15d",
   "7a77b530-d055-46cc-8a2e-a9f1350658ff"
  ]
 },
 "S23sh1": {
  "input_fingerprint": "77e78c9708799b5e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차가운 면회실 철제 테이블 앞에 두 손목에 수갑이 채워진 채 고개를 숙인 이현우의 피투성이 상체.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, illuminated for daytime visits. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at the establishing stage's held entry position beside a long edge of the table, above 이현우's seated eye level and looking diagonally downward at his front three-quarter view before the crane descends. Place his bowed head and bloodied, bruised upper body just right of center, with both cuffed wrists visible low in the composition and his gaze lowered toward his hands. Observe him directly, using the table's diagonal edge and a restrained area of the spacious room behind him to express isolation without enlarging the table beyond the human scale.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Metal visitation table (Positioned in the middle of the room with 이현우 seated at it) — Its top and one long edge are seen obliquely from above; used as The diagonal edge leads toward the wrists while the tabletop occupies less than forty percent of the image; Handcuffs (Fastened around both wrists) — Seen from above at the lower center beside the table edge; used as Makes his confinement readable within the upper-body composition; Spacious visitation-room interior (Visible beyond the centrally placed table) — A limited expanse of the room recedes behind 이현우; used as Provides negative space around the isolated seated figure without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the visitation room stays subdued and controlled, retaining injury detail without sensational highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the large visitation room. 이현우: He sits at the central table with handcuffs still on and bruising across his face. His earlier leg bite remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차가운 면회실 철제 테이블 앞에 두 손목에 수갑이 채워진 채 고개를 숙인 이현우의 피투성이 상체.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, illuminated for daytime visits. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at the establishing stage's held entry position beside a long edge of the table, above 이현우's seated eye level and looking diagonally downward at his front three-quarter view before the crane descends. Place his bowed head and bloodied, bruised upper body just right of center, with both cuffed wrists visible low in the composition and his gaze lowered toward his hands. Observe him directly, using the table's diagonal edge and a restrained area of the spacious room behind him to express isolation without enlarging the table beyond the human scale.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Metal visitation table (Positioned in the middle of the room with 이현우 seated at it) — Its top and one long edge are seen obliquely from above; used as The diagonal edge leads toward the wrists while the tabletop occupies less than forty percent of the image; Handcuffs (Fastened around both wrists) — Seen from above at the lower center beside the table edge; used as Makes his confinement readable within the upper-body composition; Spacious visitation-room interior (Visible beyond the centrally placed table) — A limited expanse of the room recedes behind 이현우; used as Provides negative space around the isolated seated figure without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the visitation room stays subdued and controlled, retaining injury detail without sensational highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the large visitation room. 이현우: He sits at the central table with handcuffs still on and bruising across his face. His earlier leg bite remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차가운 면회실 철제 테이블 앞에 두 손목에 수갑이 채워진 채 고개를 숙인 이현우의 피투성이 상체.\n\nLOCATION (lock): At the central table inside a spacious detention-center visiting room, illuminated for daytime visits. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at the establishing stage's held entry position beside a long edge of the table, above 이현우's seated eye level and looking diagonally downward at his front three-quarter view before the crane descends. Place his bowed head and bloodied, bruised upper body just right of center, with both cuffed wrists visible low in the composition and his gaze lowered toward his hands. Observe him directly, using the table's diagonal edge and a restrained area of the spacious room behind him to express isolation without enlarging the table beyond the human scale.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Metal visitation table (Positioned in the middle of the room with 이현우 seated at it) — Its top and one long edge are seen obliquely from above; used as The diagonal edge leads toward the wrists while the tabletop occupies less than forty percent of the image; Handcuffs (Fastened around both wrists) — Seen from above at the lower center beside the table edge; used as Makes his confinement readable within the upper-body composition; Spacious visitation-room interior (Visible beyond the centrally placed table) — A limited expanse of the room recedes behind 이현우; used as Provides negative space around the isolated seated figure without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the visitation room stays subdued and controlled, retaining injury detail without sensational highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table stands in the middle of the large visitation room. 이현우: He sits at the central table with handcuffs still on and bruising across his face. His earlier leg bite remains untreated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S23sh1__bgfirst_bg.png",
     "asset_id": "6b2bfcad-7ccf-476c-a01d-50ee349e5a7e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S23sh1.png",
     "asset_id": "14df2182-3f90-4ddb-ad5f-5cf99e0be15d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L33B01.png",
     "asset_id": "7a77b530-d055-46cc-8a2e-a9f1350658ff",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 아래로 향해 수갑이 채워진 두 손을 바라보고 있음.",
    "built_space": "면회실 창문, 텍스트가 정확한 포스터, 금속 테이블, 배경의 문이 레퍼런스와 일치하며 지정된 하이앵글 구도를 반영함.",
    "entities": "이현우(피투성이 셔츠, 헝클어진 머리, 인이어 무전기), 수갑 모두 프롬프트와 정확히 일치함.",
    "hard_violations": [],
    "physics": "의자에 앉아 체중을 지탱하고 있으며 두 손은 테이블 위에 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "시선은 아래로 향해 테이블 위의 손을 바라봄.",
    "built_space": "면회실 구조는 유사하나 정면 벽 포스터의 텍스트가 완전히 누락되어 빈 사각형으로 보임.",
    "entities": "이현우(피투성이 셔츠, 인이어), 수갑은 존재하나 손가락 형태가 뭉개져 있음.",
    "hard_violations": [],
    "physics": "앉은 자세로 테이블 위에 두 손을 올리고 지탱함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스의 배경 디테일(정확한 텍스트 포함), 카메라 앵글, 피투성이 상체와 수갑 등 인물 외양을 지시문대로 매우 충실히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 인물 묘사는 지시문을 따랐으나, 배경 안내문의 텍스트가 누락되고 손의 묘사가 다소 부자연스러움."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래로 향해 수갑이 채워진 두 손을 바라보고 있음.",
        "built_space": "면회실 창문, 텍스트가 정확한 포스터, 금속 테이블, 배경의 문이 레퍼런스와 일치하며 지정된 하이앵글 구도를 반영함.",
        "entities": "이현우(피투성이 셔츠, 헝클어진 머리, 인이어 무전기), 수갑 모두 프롬프트와 정확히 일치함.",
        "hard_violations": [],
        "physics": "의자에 앉아 체중을 지탱하고 있으며 두 손은 테이블 위에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "시선은 아래로 향해 테이블 위의 손을 바라봄.",
        "built_space": "면회실 구조는 유사하나 정면 벽 포스터의 텍스트가 완전히 누락되어 빈 사각형으로 보임.",
        "entities": "이현우(피투성이 셔츠, 인이어), 수갑은 존재하나 손가락 형태가 뭉개져 있음.",
        "hard_violations": [],
        "physics": "앉은 자세로 테이블 위에 두 손을 올리고 지탱함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스의 배경 디테일(정확한 텍스트 포함), 카메라 앵글, 피투성이 상체와 수갑 등 인물 외양을 지시문대로 매우 충실히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 인물 묘사는 지시문을 따랐으나, 배경 안내문의 텍스트가 누락되고 손의 묘사가 다소 부자연스러움."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래로 향해 수갑이 채워진 두 손을 바라보고 있음.",
        "built_space": "면회실 창문, 텍스트가 정확한 포스터, 금속 테이블, 배경의 문이 레퍼런스와 일치하며 지정된 하이앵글 구도를 반영함.",
        "entities": "이현우(피투성이 셔츠, 헝클어진 머리, 인이어 무전기), 수갑 모두 프롬프트와 정확히 일치함.",
        "hard_violations": [],
        "physics": "의자에 앉아 체중을 지탱하고 있으며 두 손은 테이블 위에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "시선은 아래로 향해 테이블 위의 손을 바라봄.",
        "built_space": "면회실 구조는 유사하나 정면 벽 포스터의 텍스트가 완전히 누락되어 빈 사각형으로 보임.",
        "entities": "이현우(피투성이 셔츠, 인이어), 수갑은 존재하나 손가락 형태가 뭉개져 있음.",
        "hard_violations": [],
        "physics": "앉은 자세로 테이블 위에 두 손을 올리고 지탱함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상체 중심의 미디엄 숏과 높은 사선 시점, 중앙 오른쪽의 숙인 머리, 하단의 양손 수갑을 더 정확히 구현하며 피투성이 상체도 선명하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "장소와 손을 향한 시선, 양손 수갑은 충실하지만 A보다 인물이 작고 배경이 넓어 지정된 상체 중심 미디엄 숏의 밀도가 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리를 깊이 숙이고 얼굴과 시선을 테이블 위 두 손 쪽으로 향한다. 카메라를 보지 않는다. 테이블의 사선 가장자리가 화면 왼쪽에서 하단의 수갑 찬 손목 부근으로 이어진다.",
        "built_space": "중앙 오른쪽 인물 앞에 철제 테이블 하나, 등 뒤에 검은 의자 하나가 보인다. 배경에는 왼쪽 창 한 구획과 그 아래 환기구 하나, 뒤쪽 수납장 하나, 오른쪽 열린 문 하나가 있으며 벽 게시판 일부가 상단에 잘려 보인다. 수납장 위 서류철과 보관함도 장소 참조와 대응한다. 회백색 투톤 벽과 콘크리트 바닥이 유지되고, 높은 사선 시점에서 보이는 상판은 화면의 약 4분의 1로 사람보다 과장되지 않는다. 금속 표면의 확산 반사에 광학적 모순은 보이지 않는다.",
        "entities": "등장인물은 한 명이며, 짧고 헝클어진 검은 머리와 마른 체격의 동아시아계 젊은 남성으로 참조의 이현우와 대체로 부합한다. 숙인 얼굴 탓에 세부 동일성 확인은 제한되며 국적은 외모로 확인할 수 없다. 어두운 낡은 셔츠에 흙먼지와 다량의 핏자국이 있고 얼굴에도 상처와 피가 보인다. 귀의 작은 검은 인이어 장치, 양쪽 손목을 각각 감싼 금속 수갑과 연결 사슬이 확인된다. 다리의 물린 상처와 바지는 이 상체 구도에서 판정할 수 없다.",
        "hard_violations": [],
        "physics": "몸 뒤와 옆으로 의자의 등받이와 좌판 일부가 보여 앉은 몸의 지지가 설명된다. 상체는 앞으로 숙여져 있고 두 손은 상판에 닿아 있다. 수갑은 각 손목에 걸려 있으며 짧은 사슬이 두 수갑을 연결한다. 머리와 팔의 자세는 자연스러운 관절 범위이며 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "고개와 얼굴을 아래로 떨어뜨리고 시선을 하단 중앙의 두 손 쪽으로 향한다. 양팔도 그 손 위치로 모인다. 테이블 가장자리는 왼쪽에서 손목 부근을 거쳐 오른쪽 아래로 뻗어 구속된 손으로 시선을 유도한다.",
        "built_space": "철제 테이블 하나와 인물이 앉은 검은 의자 하나가 보인다. 왼쪽 창 한 구획, 환기구 하나, 뒤쪽 양문 수납장 하나, 오른쪽 열린 문 하나, 벽 게시물 두 개가 장소 참조의 배치와 대응한다. 수납장 위 서류철과 상자류도 유지된다. 상판은 화면의 약 30퍼센트로 제한을 지키며 사람과 가구의 크기 관계도 자연스럽다. 다만 A보다 벽과 바닥을 넓게 보여 주며, 카메라의 하향 각도도 조금 완만하게 읽힌다. 상판 반사는 창 쪽의 밝기와 손의 위치에 부합한다.",
        "entities": "한 명의 젊은 동아시아계 남성이 보이며 검은 헝클어진 머리, 마른 체격, 어두운 셔츠가 참조와 대체로 맞는다. 얼굴은 숙여져 세부 동일성 판정에 한계가 있다. 작은 검은 인이어 장치가 귀에 있고 양 손목에 사슬로 연결된 수갑이 채워져 있다. 얼굴과 셔츠, 노출된 팔에 피와 상처가 보인다. 셔츠 소매는 참조보다 많이 걷혀 있다. 하체 대부분과 다리 상처는 프레임 밖이다. 배경의 한글 게시물은 장소 참조에 있는 것으로 별도의 자막이나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "인물 뒤의 의자 등받이와 좌판 일부가 앉은 몸을 지지한다. 앞으로 기운 상체 아래에서 팔과 손이 테이블에 닿아 무게를 받는다. 두 수갑은 실제로 손목을 감싸고 사슬은 그 사이에 연결되어 있다. 지지 없는 부유나 불가능한 관절 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상체 중심의 미디엄 숏과 높은 사선 시점, 중앙 오른쪽의 숙인 머리, 하단의 양손 수갑을 더 정확히 구현하며 피투성이 상체도 선명하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "장소와 손을 향한 시선, 양손 수갑은 충실하지만 A보다 인물이 작고 배경이 넓어 지정된 상체 중심 미디엄 숏의 밀도가 떨어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리를 깊이 숙이고 얼굴과 시선을 테이블 위 두 손 쪽으로 향한다. 카메라를 보지 않는다. 테이블의 사선 가장자리가 화면 왼쪽에서 하단의 수갑 찬 손목 부근으로 이어진다.",
        "built_space": "중앙 오른쪽 인물 앞에 철제 테이블 하나, 등 뒤에 검은 의자 하나가 보인다. 배경에는 왼쪽 창 한 구획과 그 아래 환기구 하나, 뒤쪽 수납장 하나, 오른쪽 열린 문 하나가 있으며 벽 게시판 일부가 상단에 잘려 보인다. 수납장 위 서류철과 보관함도 장소 참조와 대응한다. 회백색 투톤 벽과 콘크리트 바닥이 유지되고, 높은 사선 시점에서 보이는 상판은 화면의 약 4분의 1로 사람보다 과장되지 않는다. 금속 표면의 확산 반사에 광학적 모순은 보이지 않는다.",
        "entities": "등장인물은 한 명이며, 짧고 헝클어진 검은 머리와 마른 체격의 동아시아계 젊은 남성으로 참조의 이현우와 대체로 부합한다. 숙인 얼굴 탓에 세부 동일성 확인은 제한되며 국적은 외모로 확인할 수 없다. 어두운 낡은 셔츠에 흙먼지와 다량의 핏자국이 있고 얼굴에도 상처와 피가 보인다. 귀의 작은 검은 인이어 장치, 양쪽 손목을 각각 감싼 금속 수갑과 연결 사슬이 확인된다. 다리의 물린 상처와 바지는 이 상체 구도에서 판정할 수 없다.",
        "hard_violations": [],
        "physics": "몸 뒤와 옆으로 의자의 등받이와 좌판 일부가 보여 앉은 몸의 지지가 설명된다. 상체는 앞으로 숙여져 있고 두 손은 상판에 닿아 있다. 수갑은 각 손목에 걸려 있으며 짧은 사슬이 두 수갑을 연결한다. 머리와 팔의 자세는 자연스러운 관절 범위이며 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "고개와 얼굴을 아래로 떨어뜨리고 시선을 하단 중앙의 두 손 쪽으로 향한다. 양팔도 그 손 위치로 모인다. 테이블 가장자리는 왼쪽에서 손목 부근을 거쳐 오른쪽 아래로 뻗어 구속된 손으로 시선을 유도한다.",
        "built_space": "철제 테이블 하나와 인물이 앉은 검은 의자 하나가 보인다. 왼쪽 창 한 구획, 환기구 하나, 뒤쪽 양문 수납장 하나, 오른쪽 열린 문 하나, 벽 게시물 두 개가 장소 참조의 배치와 대응한다. 수납장 위 서류철과 상자류도 유지된다. 상판은 화면의 약 30퍼센트로 제한을 지키며 사람과 가구의 크기 관계도 자연스럽다. 다만 A보다 벽과 바닥을 넓게 보여 주며, 카메라의 하향 각도도 조금 완만하게 읽힌다. 상판 반사는 창 쪽의 밝기와 손의 위치에 부합한다.",
        "entities": "한 명의 젊은 동아시아계 남성이 보이며 검은 헝클어진 머리, 마른 체격, 어두운 셔츠가 참조와 대체로 맞는다. 얼굴은 숙여져 세부 동일성 판정에 한계가 있다. 작은 검은 인이어 장치가 귀에 있고 양 손목에 사슬로 연결된 수갑이 채워져 있다. 얼굴과 셔츠, 노출된 팔에 피와 상처가 보인다. 셔츠 소매는 참조보다 많이 걷혀 있다. 하체 대부분과 다리 상처는 프레임 밖이다. 배경의 한글 게시물은 장소 참조에 있는 것으로 별도의 자막이나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "인물 뒤의 의자 등받이와 좌판 일부가 앉은 몸을 지지한다. 앞으로 기운 상체 아래에서 팔과 손이 테이블에 닿아 무게를 받는다. 두 수갑은 실제로 손목을 감싸고 사슬은 그 사이에 연결되어 있다. 지지 없는 부유나 불가능한 관절 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "레퍼런스의 배경 디테일(정확한 텍스트 포함), 카메라 앵글, 피투성이 상체와 수갑 등 인물 외양을 지시문대로 매우 충실히 구현함."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "전반적인 구도와 인물 묘사는 지시문을 따랐으나, 배경 안내문의 텍스트가 누락되고 손의 묘사가 다소 부자연스러움."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L33B01.png",
    "asset_id": "7a77b530-d055-46cc-8a2e-a9f1350658ff",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0944-d112-75ed-a842-678f11422263",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S23sh1__bgfirst_bg.png",
   "bg_asset_id": "6b2bfcad-7ccf-476c-a01d-50ee349e5a7e",
   "bg_record_key": "S23sh1::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S23sh7::signage": {
  "fp": "0d807791ef66d98e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh7": {
  "input_fingerprint": "d2fa3f359306ad5d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 찰리의 사진을 내려다보며 충격으로 두 눈을 커다랗게 뜬 이현우의 얼굴.\n\nLOCATION (lock): At the central visiting table inside the spacious detention-center room, under daytime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the inherited dolly inward from the same side of the table, slightly above 이현우's seated eyes and obliquely downward onto his bruised face. His face occupies the upper central frame as he inclines his head toward 찰리's photograph, eyes widening downward; retain the photograph's image-bearing face as a small, softly resolved element along the lower edge. Let camera distance carry the emphasis, without anticipating his later upward glance.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 (현우 앞에 놓여 있으며 사진을 받치고 있음) — The tabletop is viewed obliquely from above along the lower frame; used as Connects the downward eyeline to the photograph without competing with the face; 찰리의 수배 사진 (테이블 위에 놓여 있음; 찰리는 실물이 아니라 사진 속에만 나타남) — The image-bearing face showing Charlie is turned upward and partly visible to the camera; used as Small foreground evidence of what has shocked Hyunwoo.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room's ambient illumination restrained and evenly controlled, preserving the bruising and widened eyes without a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table, surrounding interior surfaces, and cool daytime appearance from the reference. Exclude the fastened handcuffs, which have been unlocked before this exchange.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted photograph depicts Charlie; a table occupies the middle of the visitation room. 이현우: He is seated at the table with his handcuffs removed. His face remains bruised, and his dog-bitten leg is still injured.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 찰리의 사진을 내려다보며 충격으로 두 눈을 커다랗게 뜬 이현우의 얼굴.\n\nLOCATION (lock): At the central visiting table inside the spacious detention-center room, under daytime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the inherited dolly inward from the same side of the table, slightly above 이현우's seated eyes and obliquely downward onto his bruised face. His face occupies the upper central frame as he inclines his head toward 찰리's photograph, eyes widening downward; retain the photograph's image-bearing face as a small, softly resolved element along the lower edge. Let camera distance carry the emphasis, without anticipating his later upward glance.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 (현우 앞에 놓여 있으며 사진을 받치고 있음) — The tabletop is viewed obliquely from above along the lower frame; used as Connects the downward eyeline to the photograph without competing with the face; 찰리의 수배 사진 (테이블 위에 놓여 있음; 찰리는 실물이 아니라 사진 속에만 나타남) — The image-bearing face showing Charlie is turned upward and partly visible to the camera; used as Small foreground evidence of what has shocked Hyunwoo.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room's ambient illumination restrained and evenly controlled, preserving the bruising and widened eyes without a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table, surrounding interior surfaces, and cool daytime appearance from the reference. Exclude the fastened handcuffs, which have been unlocked before this exchange.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted photograph depicts Charlie; a table occupies the middle of the visitation room. 이현우: He is seated at the table with his handcuffs removed. His face remains bruised, and his dog-bitten leg is still injured.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 찰리의 사진을 내려다보며 충격으로 두 눈을 커다랗게 뜬 이현우의 얼굴.\n\nLOCATION (lock): At the central visiting table inside the spacious detention-center room, under daytime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the inherited dolly inward from the same side of the table, slightly above 이현우's seated eyes and obliquely downward onto his bruised face. His face occupies the upper central frame as he inclines his head toward 찰리's photograph, eyes widening downward; retain the photograph's image-bearing face as a small, softly resolved element along the lower edge. Let camera distance carry the emphasis, without anticipating his later upward glance.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 (현우 앞에 놓여 있으며 사진을 받치고 있음) — The tabletop is viewed obliquely from above along the lower frame; used as Connects the downward eyeline to the photograph without competing with the face; 찰리의 수배 사진 (테이블 위에 놓여 있음; 찰리는 실물이 아니라 사진 속에만 나타남) — The image-bearing face showing Charlie is turned upward and partly visible to the camera; used as Small foreground evidence of what has shocked Hyunwoo.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room's ambient illumination restrained and evenly controlled, preserving the bruising and widened eyes without a stylized lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table, surrounding interior surfaces, and cool daytime appearance from the reference. Exclude the fastened handcuffs, which have been unlocked before this exchange.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted photograph depicts Charlie; a table occupies the middle of the visitation room. 이현우: He is seated at the table with his handcuffs removed. His face remains bruised, and his dog-bitten leg is still injured.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선이 테이블 위에 놓인 사진을 정확히 향하고 있습니다.",
    "built_space": "레퍼런스와 동일한 구치소 면회실 내부이며, 창문과 문, 철제 테이블이 올바른 위치에 있습니다.",
    "entities": "이현우의 외모, 상처, 의상 및 인이어 무전기가 레퍼런스와 일치하며, 테이블 위에 찰리의 얼굴이 담긴 사진이 있습니다.",
    "hard_violations": [],
    "physics": "인물은 테이블 앞에 자연스럽게 앉아 있으며, 특별히 어색한 물리적 지지나 부자연스러운 포즈는 없습니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선이 테이블 아래의 사진을 향하고 있습니다.",
    "built_space": "레퍼런스와 일치하는 면회실 배경 및 테이블이 존재합니다.",
    "entities": "이현우의 인상착의와 의상, 상처가 일치하며, 테이블 위에 사진이 놓여 있습니다.",
    "hard_violations": [],
    "physics": "의자에 앉아 몸을 약간 숙인 자세가 자연스럽게 유지되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도(살짝 위에서 내려다보는 앵글)와 충격으로 크게 뜬 눈의 표정이 프롬프트에 매우 충실하게 표현되었습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트의 요구사항을 잘 따랐으나, A에 비해 충격받은 표정의 강도가 다소 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 테이블 위에 놓인 사진을 정확히 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 구치소 면회실 내부이며, 창문과 문, 철제 테이블이 올바른 위치에 있습니다.",
        "entities": "이현우의 외모, 상처, 의상 및 인이어 무전기가 레퍼런스와 일치하며, 테이블 위에 찰리의 얼굴이 담긴 사진이 있습니다.",
        "hard_violations": [],
        "physics": "인물은 테이블 앞에 자연스럽게 앉아 있으며, 특별히 어색한 물리적 지지나 부자연스러운 포즈는 없습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선이 테이블 아래의 사진을 향하고 있습니다.",
        "built_space": "레퍼런스와 일치하는 면회실 배경 및 테이블이 존재합니다.",
        "entities": "이현우의 인상착의와 의상, 상처가 일치하며, 테이블 위에 사진이 놓여 있습니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 몸을 약간 숙인 자세가 자연스럽게 유지되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도(살짝 위에서 내려다보는 앵글)와 충격으로 크게 뜬 눈의 표정이 프롬프트에 매우 충실하게 표현되었습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트의 요구사항을 잘 따랐으나, A에 비해 충격받은 표정의 강도가 다소 약합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선이 테이블 위에 놓인 사진을 정확히 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 구치소 면회실 내부이며, 창문과 문, 철제 테이블이 올바른 위치에 있습니다.",
        "entities": "이현우의 외모, 상처, 의상 및 인이어 무전기가 레퍼런스와 일치하며, 테이블 위에 찰리의 얼굴이 담긴 사진이 있습니다.",
        "hard_violations": [],
        "physics": "인물은 테이블 앞에 자연스럽게 앉아 있으며, 특별히 어색한 물리적 지지나 부자연스러운 포즈는 없습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선이 테이블 아래의 사진을 향하고 있습니다.",
        "built_space": "레퍼런스와 일치하는 면회실 배경 및 테이블이 존재합니다.",
        "entities": "이현우의 인상착의와 의상, 상처가 일치하며, 테이블 위에 사진이 놓여 있습니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 몸을 약간 숙인 자세가 자연스럽게 유지되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "상단 중앙의 얼굴과 사진을 향해 커진 눈이 지정된 클로즈업에 더 충실하지만, 사진에 닿아 있어야 할 현우의 손은 두 후보 모두 누락했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "사진을 내려다보는 충격과 장소의 연속성은 맞지만, 얼굴이 오른쪽으로 치우치고 사진에 닿는 손도 없어 B보다 구도 충실도가 낮다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 고개를 숙이고 두 눈을 크게 뜬 채 하단 중앙의 사진 쪽을 내려다본다. 카메라나 위쪽을 보지 않는다. 사진의 인화면은 위를 향해 카메라에도 일부 보인다.",
        "built_space": "하단에 긁힌 금속 테이블 하나, 왼쪽에 창 하나와 벽 환기구 하나, 뒤쪽에 수납장 하나, 오른쪽에 열린 문 하나가 보인다. 회백색 투톤 벽과 차가운 주간광은 이전 장면과 이어진다. 현우는 테이블 뒤에 있으며, 얼굴은 상단 중앙보다 오른쪽에 치우쳐 있다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "실물 인물은 젊은 동아시아계 남성 현우 한 명이며, 헝클어진 검은 머리, 마른 체형, 피와 먼지가 묻은 어두운 셔츠, 얼굴 상처와 검은 인이어가 참조와 대체로 맞는다. 국적은 외형만으로 확인할 수 없다. 찰리에 해당하는 성인 남성은 하단 사진 한 장 안에만 등장하며, 찰리의 별도 외모 기준은 주어지지 않았다. 수갑은 보이지 않는다. 다리 부상은 구도 밖이라 평가하지 않는다. 명시적으로 요구된 현우의 손과 사진의 접촉은 보이지 않는다.",
        "hard_violations": [],
        "physics": "사진은 금속 상판에 평평하게 놓여 상판이 받친다. 현우는 테이블 뒤에서 상체를 앞으로 기울인 자세이며, 좌판과 하체는 화면 밖이다. 떠 있는 신체나 물체는 보이지 않는다. 사진을 받치는 손이 없는 것은 부유 문제가 아니라 손을 화면에 넣으라는 동작 조건의 누락이다."
       },
       {
        "label": "B",
        "direction": "숙인 얼굴의 두 눈이 크게 열려 있고 시선은 하단 사진 쪽으로 내려간다. 이후의 위쪽 시선을 미리 보여주지 않는다. 사진은 인화면이 위를 향해 현우와 카메라 양쪽에서 볼 수 있는 상태다.",
        "built_space": "하단의 금속 테이블 하나, 왼쪽 창 하나와 그 아래 환기구 하나, 후방 수납장 하나, 오른쪽 열린 문 하나가 보인다. 현우 오른쪽에는 의자 등받이의 금속 테두리가 하나 드러난다. 기존의 투톤 벽과 주간광을 유지한다. 얼굴이 상단 중앙에 놓이고 상판과 작은 사진이 하단을 차지해 지정된 배치에 더 가깝다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "실물 인물은 현우 한 명뿐이다. 젊은 동아시아계 남성의 얼굴, 검은 헝클어진 머리, 마른 상체, 오염된 어두운 셔츠, 얼굴의 피와 상처, 소형 인이어가 참조의 외형과 대체로 일치한다. 찰리는 하단의 인화 사진 속 성인 남성으로만 나타난다. 수갑은 보이지 않으며, 바지와 부상당한 다리는 클로즈업 밖이다. 사진에 놓이거나 사진을 잡아야 하는 현우의 손은 나오지 않는다.",
        "hard_violations": [],
        "physics": "사진은 테이블 표면에 놓여 지지된다. 현우는 뒤에 보이는 의자 등받이 앞에서 상체를 기울이고 있어 앉은 자세로 자연스럽게 읽힌다. 좌판 접촉은 화면 밖이지만 부유하거나 불가능한 자세의 증거는 없다. 사진에 손을 대는 명시적 동작은 누락되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "상단 중앙의 얼굴과 사진을 향해 커진 눈이 지정된 클로즈업에 더 충실하지만, 사진에 닿아 있어야 할 현우의 손은 두 후보 모두 누락했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "사진을 내려다보는 충격과 장소의 연속성은 맞지만, 얼굴이 오른쪽으로 치우치고 사진에 닿는 손도 없어 B보다 구도 충실도가 낮다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 고개를 숙이고 두 눈을 크게 뜬 채 하단 중앙의 사진 쪽을 내려다본다. 카메라나 위쪽을 보지 않는다. 사진의 인화면은 위를 향해 카메라에도 일부 보인다.",
        "built_space": "하단에 긁힌 금속 테이블 하나, 왼쪽에 창 하나와 벽 환기구 하나, 뒤쪽에 수납장 하나, 오른쪽에 열린 문 하나가 보인다. 회백색 투톤 벽과 차가운 주간광은 이전 장면과 이어진다. 현우는 테이블 뒤에 있으며, 얼굴은 상단 중앙보다 오른쪽에 치우쳐 있다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "실물 인물은 젊은 동아시아계 남성 현우 한 명이며, 헝클어진 검은 머리, 마른 체형, 피와 먼지가 묻은 어두운 셔츠, 얼굴 상처와 검은 인이어가 참조와 대체로 맞는다. 국적은 외형만으로 확인할 수 없다. 찰리에 해당하는 성인 남성은 하단 사진 한 장 안에만 등장하며, 찰리의 별도 외모 기준은 주어지지 않았다. 수갑은 보이지 않는다. 다리 부상은 구도 밖이라 평가하지 않는다. 명시적으로 요구된 현우의 손과 사진의 접촉은 보이지 않는다.",
        "hard_violations": [],
        "physics": "사진은 금속 상판에 평평하게 놓여 상판이 받친다. 현우는 테이블 뒤에서 상체를 앞으로 기울인 자세이며, 좌판과 하체는 화면 밖이다. 떠 있는 신체나 물체는 보이지 않는다. 사진을 받치는 손이 없는 것은 부유 문제가 아니라 손을 화면에 넣으라는 동작 조건의 누락이다."
       },
       {
        "label": "A",
        "direction": "숙인 얼굴의 두 눈이 크게 열려 있고 시선은 하단 사진 쪽으로 내려간다. 이후의 위쪽 시선을 미리 보여주지 않는다. 사진은 인화면이 위를 향해 현우와 카메라 양쪽에서 볼 수 있는 상태다.",
        "built_space": "하단의 금속 테이블 하나, 왼쪽 창 하나와 그 아래 환기구 하나, 후방 수납장 하나, 오른쪽 열린 문 하나가 보인다. 현우 오른쪽에는 의자 등받이의 금속 테두리가 하나 드러난다. 기존의 투톤 벽과 주간광을 유지한다. 얼굴이 상단 중앙에 놓이고 상판과 작은 사진이 하단을 차지해 지정된 배치에 더 가깝다. 설비 중복이나 불가능한 반사는 없다.",
        "entities": "실물 인물은 현우 한 명뿐이다. 젊은 동아시아계 남성의 얼굴, 검은 헝클어진 머리, 마른 상체, 오염된 어두운 셔츠, 얼굴의 피와 상처, 소형 인이어가 참조의 외형과 대체로 일치한다. 찰리는 하단의 인화 사진 속 성인 남성으로만 나타난다. 수갑은 보이지 않으며, 바지와 부상당한 다리는 클로즈업 밖이다. 사진에 놓이거나 사진을 잡아야 하는 현우의 손은 나오지 않는다.",
        "hard_violations": [],
        "physics": "사진은 테이블 표면에 놓여 지지된다. 현우는 뒤에 보이는 의자 등받이 앞에서 상체를 기울이고 있어 앉은 자세로 자연스럽게 읽힌다. 좌판 접촉은 화면 밖이지만 부유하거나 불가능한 자세의 증거는 없다. 사진에 손을 대는 명시적 동작은 누락되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.75
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 카메라 구도(살짝 위에서 내려다보는 앵글)와 충격으로 크게 뜬 눈의 표정이 프롬프트에 매우 충실하게 표현되었습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "프롬프트의 요구사항을 잘 따랐으나, A에 비해 충격받은 표정의 강도가 다소 약합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S23sh1_sel.png",
    "asset_id": "459eaef7-3a2b-4d9e-92cb-388c5a958d6e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab094b-ad46-7ae3-8ead-9e84ada59f30",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh1"
  }
 },
 "S23sh11::signage": {
  "fp": "b7b10b6fc85918e6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh11": {
  "input_fingerprint": "c19e82934bf3b195",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 당돌한 제안에 흥미롭다는 듯 고개를 살짝 기울인 채 비릿하게 웃고 있는 윤성찬의 얼굴 클로즈업.\n\nLOCATION (lock): On the visitor's side of the central table in the spacious detention-center visiting room, under daytime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inherited track just behind and outside 이현우's shoulder, looking slightly upward at 윤성찬 from an oblique angle without crossing the table's dialogue axis. 윤성찬's face occupies the right-center, his head subtly tilted and smile held as he watches 이현우's face just beyond the left crop; 이현우 contributes only a receding shoulder and a sliver of the back of his head. Make the changed eyeline ownership the emphasis, retaining the restrained exposure and close facial scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 가장자리 (두 사람 사이에 놓여 있음) — A short oblique segment remains below the faces; used as Maintains the established spatial relationship across the conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's subdued ambient illumination, with enough facial detail to distinguish the smile from the calculating eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table and established interior lighting from the reference. Exclude the earlier restraint setup with locked handcuffs and any furnishings from a private office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted photograph of Charlie remains available in the visitation room, whose central table has not changed. 이현우: He remains at the table without handcuffs, with facial bruising and an untreated dog-bite injury to his leg. 윤성찬: He remains in his suit, smiling during the negotiation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 당돌한 제안에 흥미롭다는 듯 고개를 살짝 기울인 채 비릿하게 웃고 있는 윤성찬의 얼굴 클로즈업.\n\nLOCATION (lock): On the visitor's side of the central table in the spacious detention-center visiting room, under daytime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inherited track just behind and outside 이현우's shoulder, looking slightly upward at 윤성찬 from an oblique angle without crossing the table's dialogue axis. 윤성찬's face occupies the right-center, his head subtly tilted and smile held as he watches 이현우's face just beyond the left crop; 이현우 contributes only a receding shoulder and a sliver of the back of his head. Make the changed eyeline ownership the emphasis, retaining the restrained exposure and close facial scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 가장자리 (두 사람 사이에 놓여 있음) — A short oblique segment remains below the faces; used as Maintains the established spatial relationship across the conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's subdued ambient illumination, with enough facial detail to distinguish the smile from the calculating eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table and established interior lighting from the reference. Exclude the earlier restraint setup with locked handcuffs and any furnishings from a private office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted photograph of Charlie remains available in the visitation room, whose central table has not changed. 이현우: He remains at the table without handcuffs, with facial bruising and an untreated dog-bite injury to his leg. 윤성찬: He remains in his suit, smiling during the negotiation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 당돌한 제안에 흥미롭다는 듯 고개를 살짝 기울인 채 비릿하게 웃고 있는 윤성찬의 얼굴 클로즈업.\n\nLOCATION (lock): On the visitor's side of the central table in the spacious detention-center visiting room, under daytime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inherited track just behind and outside 이현우's shoulder, looking slightly upward at 윤성찬 from an oblique angle without crossing the table's dialogue axis. 윤성찬's face occupies the right-center, his head subtly tilted and smile held as he watches 이현우's face just beyond the left crop; 이현우 contributes only a receding shoulder and a sliver of the back of his head. Make the changed eyeline ownership the emphasis, retaining the restrained exposure and close facial scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 테이블 가장자리 (두 사람 사이에 놓여 있음) — A short oblique segment remains below the faces; used as Maintains the established spatial relationship across the conversation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the visitation room's subdued ambient illumination, with enough facial detail to distinguish the smile from the calculating eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the visitation room's table and established interior lighting from the reference. Exclude the earlier restraint setup with locked handcuffs and any furnishings from a private office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted photograph of Charlie remains available in the visitation room, whose central table has not changed. 이현우: He remains at the table without handcuffs, with facial bruising and an untreated dog-bite injury to his leg. 윤성찬: He remains in his suit, smiling during the negotiation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬의 시선이 화면 좌측에 걸친 이현우의 얼굴 방향을 정확히 향하고 있습니다.",
    "built_space": "금속 테이블을 사이에 두고 마주 앉은 구조이나, 카메라가 반대편을 향했음에도 이전 샷의 이현우 뒤 배경(창문, 캐비닛, 문)이 그대로 렌더링된 공간적 모순이 있습니다.",
    "entities": "윤성찬은 70대 후반의 외모와 챠콜 그레이 맞춤 정장을 정확히 입고 있으며, 이현우는 헝클어진 머리와 텍스처가 있는 셔츠 차림의 뒷모습으로 올바르게 묘사되었습니다.",
    "hard_violations": [],
    "physics": "인물들의 자세와 옷주름이 자연스러우며, 테이블 앞에서의 신체적 접촉과 중력에 따른 어색함이 없습니다."
   },
   {
    "label": "B",
    "direction": "윤성찬이 화면 좌측의 이현우를 향해 시선을 고정하고 있습니다.",
    "built_space": "후보 A와 마찬가지로 반대편 앵글임에도 이전 샷의 벽면과 구조물(창문, 문, 수납장)이 똑같이 나타나는 공간 오류가 발생했습니다.",
    "entities": "윤성찬의 정장과 헤어스타일, 이현우의 뒷모습 및 피 묻은 셔츠 등 레퍼런스의 인물 설정이 정확하게 반영되었습니다.",
    "hard_violations": [],
    "physics": "의자에 앉은 자세와 중력, 의복의 형태 등 물리적으로 안정적인 상태를 유지하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 '비릿하게 웃고 있는' 표정과 미세하게 기울어진 고개의 디테일을 훌륭하게 포착하여 상황의 분위기를 더 잘 살렸습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 구도는 지시사항을 잘 따랐으나, 미소가 다소 온화하여 텍스트가 요구한 비릿한 느낌이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 시선이 화면 좌측에 걸친 이현우의 얼굴 방향을 정확히 향하고 있습니다.",
        "built_space": "금속 테이블을 사이에 두고 마주 앉은 구조이나, 카메라가 반대편을 향했음에도 이전 샷의 이현우 뒤 배경(창문, 캐비닛, 문)이 그대로 렌더링된 공간적 모순이 있습니다.",
        "entities": "윤성찬은 70대 후반의 외모와 챠콜 그레이 맞춤 정장을 정확히 입고 있으며, 이현우는 헝클어진 머리와 텍스처가 있는 셔츠 차림의 뒷모습으로 올바르게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "인물들의 자세와 옷주름이 자연스러우며, 테이블 앞에서의 신체적 접촉과 중력에 따른 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "윤성찬이 화면 좌측의 이현우를 향해 시선을 고정하고 있습니다.",
        "built_space": "후보 A와 마찬가지로 반대편 앵글임에도 이전 샷의 벽면과 구조물(창문, 문, 수납장)이 똑같이 나타나는 공간 오류가 발생했습니다.",
        "entities": "윤성찬의 정장과 헤어스타일, 이현우의 뒷모습 및 피 묻은 셔츠 등 레퍼런스의 인물 설정이 정확하게 반영되었습니다.",
        "hard_violations": [],
        "physics": "의자에 앉은 자세와 중력, 의복의 형태 등 물리적으로 안정적인 상태를 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 '비릿하게 웃고 있는' 표정과 미세하게 기울어진 고개의 디테일을 훌륭하게 포착하여 상황의 분위기를 더 잘 살렸습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 구도는 지시사항을 잘 따랐으나, 미소가 다소 온화하여 텍스트가 요구한 비릿한 느낌이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 시선이 화면 좌측에 걸친 이현우의 얼굴 방향을 정확히 향하고 있습니다.",
        "built_space": "금속 테이블을 사이에 두고 마주 앉은 구조이나, 카메라가 반대편을 향했음에도 이전 샷의 이현우 뒤 배경(창문, 캐비닛, 문)이 그대로 렌더링된 공간적 모순이 있습니다.",
        "entities": "윤성찬은 70대 후반의 외모와 챠콜 그레이 맞춤 정장을 정확히 입고 있으며, 이현우는 헝클어진 머리와 텍스처가 있는 셔츠 차림의 뒷모습으로 올바르게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "인물들의 자세와 옷주름이 자연스러우며, 테이블 앞에서의 신체적 접촉과 중력에 따른 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "윤성찬이 화면 좌측의 이현우를 향해 시선을 고정하고 있습니다.",
        "built_space": "후보 A와 마찬가지로 반대편 앵글임에도 이전 샷의 벽면과 구조물(창문, 문, 수납장)이 똑같이 나타나는 공간 오류가 발생했습니다.",
        "entities": "윤성찬의 정장과 헤어스타일, 이현우의 뒷모습 및 피 묻은 셔츠 등 레퍼런스의 인물 설정이 정확하게 반영되었습니다.",
        "hard_violations": [],
        "physics": "의자에 앉은 자세와 중력, 의복의 형태 등 물리적으로 안정적인 상태를 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "고개 기울임과 상대를 향한 미소는 맞지만, 윤성찬의 상반신과 이현우의 뒷머리·등을 너무 넓게 보여 얼굴 클로즈업과 최소한의 전경이라는 핵심 구도에서 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "윤성찬의 얼굴을 더 크게 오른쪽 중앙에 배치하고 계산적인 눈빛과 억제된 미소를 살려 우세하지만, 이현우의 전경 비중은 여전히 크고 약한 올려다보기는 뚜렷하지 않다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 고개를 화면 오른쪽으로 기울인 채 왼쪽 전경의 이현우 얼굴을 바라본다. 입꼬리가 올라가 있어 제안을 흥미롭게 듣는 반응으로 읽힌다. 이현우는 윤성찬 쪽으로 몸과 머리를 돌렸으나 눈은 보이지 않는다. 카메라는 이현우의 어깨 뒤에서 비스듬히 보지만, 요구한 약한 올려다보기보다는 눈높이에 가깝다.",
        "built_space": "아래에 긁힌 금속 테이블 하나, 왼쪽에 격자창 하나와 하단 환기구 하나, 뒤쪽에 수납장 하나, 오른쪽에 열린 문 하나와 벽 스위치 하나가 보인다. 윤성찬 뒤로 의자 등받이 일부가 있으며 두 사람 사이에 테이블이 놓인다. 회백색 이중 도장 벽과 낮의 창빛은 이전 장소와 유사하다. 다만 배경과 윤성찬의 몸통을 넓게 담아 요구한 얼굴 중심의 좁은 화면보다 공간 설명이 많다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "등장인물은 두 명뿐이다. 윤성찬은 짧게 정돈한 은발과 콧수염, 깊은 주름이 있는 고령의 동아시아계 남성으로 참고 인물에 가깝고, 차콜 정장·흰 셔츠·어두운 넥타이·흰 포켓스퀘어를 착용한다. 이현우의 보이는 부분은 헝클어진 검은 머리와 상처 있는 뺨 가장자리, 피와 때가 묻은 회갈색 셔츠로 이전 장면과 부합한다. 얼굴 대부분이 가려져 정확한 신원과 나이 대조에는 한계가 있다. 수갑은 보이지 않으며 다리 부상과 찰리의 사진은 이 구도에서 확인할 수 없다. 추가 인물이나 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 뒤에 보이는 의자에 앉아 상체를 조금 앞으로 기울인 자세로 읽히며 머리는 목과 몸통에 자연스럽게 연결된다. 좌판과 발은 화면 밖이지만 공중에 뜬 징후는 없다. 이현우 역시 테이블 앞에 위치한 연속적인 머리·목·어깨를 보인다. 테이블은 하단을 가로지르는 정상적인 수평 구조물이며, 떠 있는 물체나 비현실적인 관절은 없다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 눈은 왼쪽 전경에 있는 이현우의 얼굴을 향하고, 기울어진 머리와 살짝 다문 미소가 상대의 제안을 가늠하는 반응으로 읽힌다. 렌즈를 직접 응시하지 않는다. 이현우의 머리도 윤성찬 쪽을 향하지만 눈은 가려져 있다. 어깨 너머의 사선 시점은 맞으며, 낮은 카메라 각도는 강하게 드러나지 않는다.",
        "built_space": "왼쪽의 격자창 하나와 환기구 하나, 뒤쪽 수납장 하나, 오른쪽 열린 문 하나와 스위치 하나, 하단의 금속 테이블 하나가 보인다. 윤성찬 뒤 오른쪽에는 의자 등받이 일부가 보인다. 두 사람은 테이블을 사이에 두고 마주하며, 테이블 가장자리는 얼굴 아래에서 비스듬한 구간으로 남는다. 벽의 색 분할과 금속 재질, 주간 실내광은 참고 장소와 이어진다. A보다 얼굴이 커지고 배경 비중이 줄었지만 이현우의 뒷머리와 어깨는 요구한 가느다란 일부보다 많이 보인다.",
        "entities": "두 인물만 보인다. 윤성찬의 고령 남성 얼굴, 은발, 콧수염과 차콜 맞춤 정장·넥타이는 참고 이미지에 가깝다. 정상적인 눈과 미세한 입가 움직임으로 표정을 표현한다. 이현우는 흐릿한 검은 머리, 귀와 상처 있는 뺨 일부, 낡은 회갈색 셔츠로 나타나 이전 장면의 외형을 유지한다. 가려진 얼굴만으로 정확한 신원 전체를 확인할 수는 없다. 손목과 다리가 잘려 있어 수갑 부재와 다리 상처는 직접 검증할 수 없고, 찰리의 사진도 보이지 않는다. 불필요한 인물·문구·사무실 소품은 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 의자 등받이 앞에서 자연스럽게 상체를 기울이고 있으며, 목이 기울어진 머리를 정상적으로 지탱한다. 앉은 자세의 하체 접점은 화면 밖이지만 부유나 신체 단절은 보이지 않는다. 이현우의 목과 어깨도 자연스럽게 연결된다. 테이블은 두 사람 사이를 가로지르는 안정적인 구조로 보이며, 손에 들린 물체나 지지 없이 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "고개 기울임과 상대를 향한 미소는 맞지만, 윤성찬의 상반신과 이현우의 뒷머리·등을 너무 넓게 보여 얼굴 클로즈업과 최소한의 전경이라는 핵심 구도에서 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "윤성찬의 얼굴을 더 크게 오른쪽 중앙에 배치하고 계산적인 눈빛과 억제된 미소를 살려 우세하지만, 이현우의 전경 비중은 여전히 크고 약한 올려다보기는 뚜렷하지 않다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬은 고개를 화면 오른쪽으로 기울인 채 왼쪽 전경의 이현우 얼굴을 바라본다. 입꼬리가 올라가 있어 제안을 흥미롭게 듣는 반응으로 읽힌다. 이현우는 윤성찬 쪽으로 몸과 머리를 돌렸으나 눈은 보이지 않는다. 카메라는 이현우의 어깨 뒤에서 비스듬히 보지만, 요구한 약한 올려다보기보다는 눈높이에 가깝다.",
        "built_space": "아래에 긁힌 금속 테이블 하나, 왼쪽에 격자창 하나와 하단 환기구 하나, 뒤쪽에 수납장 하나, 오른쪽에 열린 문 하나와 벽 스위치 하나가 보인다. 윤성찬 뒤로 의자 등받이 일부가 있으며 두 사람 사이에 테이블이 놓인다. 회백색 이중 도장 벽과 낮의 창빛은 이전 장소와 유사하다. 다만 배경과 윤성찬의 몸통을 넓게 담아 요구한 얼굴 중심의 좁은 화면보다 공간 설명이 많다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "등장인물은 두 명뿐이다. 윤성찬은 짧게 정돈한 은발과 콧수염, 깊은 주름이 있는 고령의 동아시아계 남성으로 참고 인물에 가깝고, 차콜 정장·흰 셔츠·어두운 넥타이·흰 포켓스퀘어를 착용한다. 이현우의 보이는 부분은 헝클어진 검은 머리와 상처 있는 뺨 가장자리, 피와 때가 묻은 회갈색 셔츠로 이전 장면과 부합한다. 얼굴 대부분이 가려져 정확한 신원과 나이 대조에는 한계가 있다. 수갑은 보이지 않으며 다리 부상과 찰리의 사진은 이 구도에서 확인할 수 없다. 추가 인물이나 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 뒤에 보이는 의자에 앉아 상체를 조금 앞으로 기울인 자세로 읽히며 머리는 목과 몸통에 자연스럽게 연결된다. 좌판과 발은 화면 밖이지만 공중에 뜬 징후는 없다. 이현우 역시 테이블 앞에 위치한 연속적인 머리·목·어깨를 보인다. 테이블은 하단을 가로지르는 정상적인 수평 구조물이며, 떠 있는 물체나 비현실적인 관절은 없다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 눈은 왼쪽 전경에 있는 이현우의 얼굴을 향하고, 기울어진 머리와 살짝 다문 미소가 상대의 제안을 가늠하는 반응으로 읽힌다. 렌즈를 직접 응시하지 않는다. 이현우의 머리도 윤성찬 쪽을 향하지만 눈은 가려져 있다. 어깨 너머의 사선 시점은 맞으며, 낮은 카메라 각도는 강하게 드러나지 않는다.",
        "built_space": "왼쪽의 격자창 하나와 환기구 하나, 뒤쪽 수납장 하나, 오른쪽 열린 문 하나와 스위치 하나, 하단의 금속 테이블 하나가 보인다. 윤성찬 뒤 오른쪽에는 의자 등받이 일부가 보인다. 두 사람은 테이블을 사이에 두고 마주하며, 테이블 가장자리는 얼굴 아래에서 비스듬한 구간으로 남는다. 벽의 색 분할과 금속 재질, 주간 실내광은 참고 장소와 이어진다. A보다 얼굴이 커지고 배경 비중이 줄었지만 이현우의 뒷머리와 어깨는 요구한 가느다란 일부보다 많이 보인다.",
        "entities": "두 인물만 보인다. 윤성찬의 고령 남성 얼굴, 은발, 콧수염과 차콜 맞춤 정장·넥타이는 참고 이미지에 가깝다. 정상적인 눈과 미세한 입가 움직임으로 표정을 표현한다. 이현우는 흐릿한 검은 머리, 귀와 상처 있는 뺨 일부, 낡은 회갈색 셔츠로 나타나 이전 장면의 외형을 유지한다. 가려진 얼굴만으로 정확한 신원 전체를 확인할 수는 없다. 손목과 다리가 잘려 있어 수갑 부재와 다리 상처는 직접 검증할 수 없고, 찰리의 사진도 보이지 않는다. 불필요한 인물·문구·사무실 소품은 없다.",
        "hard_violations": [],
        "physics": "윤성찬은 의자 등받이 앞에서 자연스럽게 상체를 기울이고 있으며, 목이 기울어진 머리를 정상적으로 지탱한다. 앉은 자세의 하체 접점은 화면 밖이지만 부유나 신체 단절은 보이지 않는다. 이현우의 목과 어깨도 자연스럽게 연결된다. 테이블은 두 사람 사이를 가로지르는 안정적인 구조로 보이며, 손에 들린 물체나 지지 없이 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.607
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.607
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 '비릿하게 웃고 있는' 표정과 미세하게 기울어진 고개의 디테일을 훌륭하게 포착하여 상황의 분위기를 더 잘 살렸습니다."
   },
   {
    "label": "B",
    "score": 1607,
    "verdict_ko": "인물과 구도는 지시사항을 잘 따랐으나, 미소가 다소 온화하여 텍스트가 요구한 비릿한 느낌이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S23sh1_sel.png",
    "asset_id": "459eaef7-3a2b-4d9e-92cb-388c5a958d6e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab094f-fd0c-7f5b-a40a-9194c16131e4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh1"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S24sh4::signage": {
  "fp": "4ebb81ee3a6d9edf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::371e4115b3716604": {
  "subjects": [],
  "subject_text": "이현우의 컨테이너 외부·출입문 앞\n조밀한 철제 주거동 사이 도로변 끝에 있는 낡은 컨테이너 외벽. 녹슨 표면과 번호가 표시된 출입문, 문고리가 보인다.",
  "identity": "canonical",
  "scope_id": "L24",
  "scope_role": "location_exterior",
  "scope_sha": "b717c300c07e6861"
 },
 "S24sh4::bgfirst_bg": {
  "input_fingerprint": "dc62773883aa77d2",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉게 충혈된 눈으로 페드로의 멱살을 향해 한 손을 뻗은 채 거친 얼굴을 바짝 들이민 이현우.\n\nLOCATION (lock): Inside the shared living area of the searched container home, just beyond its entrance. Daylight comes through the doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just outside 페드로's shoulder at his upper-chest height, looking slightly upward toward 이현우 while retaining the established confrontation axis. Place 페드로's shoulder and collar narrowly along the right foreground, with 이현우's upper body at left-center and his reaching hand fully visible in the lower frame. 이현우 leans into the gap, bloodshot eyes fixed on 페드로's face immediately beyond the right crop, while 페드로's cropped head remains turned toward him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (두 사람이 들어와 대치하고 있는 실내) — Only a narrow, unresolved portion of the interior remains beyond Hyunwoo; used as Provides spatial enclosure without introducing unsupported furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the interior, keeping the bloodshot eyes and reaching hand readable without an added dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉게 충혈된 눈으로 페드로의 멱살을 향해 한 손을 뻗은 채 거친 얼굴을 바짝 들이민 이현우.\n\nLOCATION (lock): Inside the shared living area of the searched container home, just beyond its entrance. Daylight comes through the doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just outside 페드로's shoulder at his upper-chest height, looking slightly upward toward 이현우 while retaining the established confrontation axis. Place 페드로's shoulder and collar narrowly along the right foreground, with 이현우's upper body at left-center and his reaching hand fully visible in the lower frame. 이현우 leans into the gap, bloodshot eyes fixed on 페드로's face immediately beyond the right crop, while 페드로's cropped head remains turned toward him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (두 사람이 들어와 대치하고 있는 실내) — Only a narrow, unresolved portion of the interior remains beyond Hyunwoo; used as Provides spatial enclosure without introducing unsupported furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the interior, keeping the bloodshot eyes and reaching hand readable without an added dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S24sh4__bgfirst_bg.png",
  "asset_id": "1e75d314-aff0-417c-9096-0940ab4b2c40",
  "input_asset_ids": [
   "835ca9be-da8c-48a5-bf21-995533ce4a6b",
   "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  ]
 },
 "S24sh4": {
  "input_fingerprint": "e524b82f84cd8e83",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 붉게 충혈된 눈으로 페드로의 멱살을 향해 한 손을 뻗은 채 거친 얼굴을 바짝 들이민 이현우.\n\nLOCATION (lock): Inside the shared living area of the searched container home, just beyond its entrance. Daylight comes through the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just outside 페드로's shoulder at his upper-chest height, looking slightly upward toward 이현우 while retaining the established confrontation axis. Place 페드로's shoulder and collar narrowly along the right foreground, with 이현우's upper body at left-center and his reaching hand fully visible in the lower frame. 이현우 leans into the gap, bloodshot eyes fixed on 페드로's face immediately beyond the right crop, while 페드로's cropped head remains turned toward him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (두 사람이 들어와 대치하고 있는 실내) — Only a narrow, unresolved portion of the interior remains beyond Hyunwoo; used as Provides spatial enclosure without introducing unsupported furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the interior, keeping the bloodshot eyes and reaching hand readable without an added dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is container 7-31, with the disturbance from the militia's search still unresolved. 이현우: He is inside the container, still wearing his outer top. His facial bruises and injured leg remain, and his wrists are no longer cuffed. 페드로: He has hurried into the container and is out of breath.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 붉게 충혈된 눈으로 페드로의 멱살을 향해 한 손을 뻗은 채 거친 얼굴을 바짝 들이민 이현우.\n\nLOCATION (lock): Inside the shared living area of the searched container home, just beyond its entrance. Daylight comes through the doorway. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just outside 페드로's shoulder at his upper-chest height, looking slightly upward toward 이현우 while retaining the established confrontation axis. Place 페드로's shoulder and collar narrowly along the right foreground, with 이현우's upper body at left-center and his reaching hand fully visible in the lower frame. 이현우 leans into the gap, bloodshot eyes fixed on 페드로's face immediately beyond the right crop, while 페드로's cropped head remains turned toward him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (두 사람이 들어와 대치하고 있는 실내) — Only a narrow, unresolved portion of the interior remains beyond Hyunwoo; used as Provides spatial enclosure without introducing unsupported furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the interior, keeping the bloodshot eyes and reaching hand readable without an added dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is container 7-31, with the disturbance from the militia's search still unresolved. 이현우: He is inside the container, still wearing his outer top. His facial bruises and injured leg remain, and his wrists are no longer cuffed. 페드로: He has hurried into the container and is out of breath.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 붉게 충혈된 눈으로 페드로의 멱살을 향해 한 손을 뻗은 채 거친 얼굴을 바짝 들이민 이현우.\n\nLOCATION (lock): Inside the shared living area of the searched container home, just beyond its entrance. Daylight comes through the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just outside 페드로's shoulder at his upper-chest height, looking slightly upward toward 이현우 while retaining the established confrontation axis. Place 페드로's shoulder and collar narrowly along the right foreground, with 이현우's upper body at left-center and his reaching hand fully visible in the lower frame. 이현우 leans into the gap, bloodshot eyes fixed on 페드로's face immediately beyond the right crop, while 페드로's cropped head remains turned toward him.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (두 사람이 들어와 대치하고 있는 실내) — Only a narrow, unresolved portion of the interior remains beyond Hyunwoo; used as Provides spatial enclosure without introducing unsupported furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the interior, keeping the bloodshot eyes and reaching hand readable without an added dramatic light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The home is container 7-31, with the disturbance from the militia's search still unresolved. 이현우: He is inside the container, still wearing his outer top. His facial bruises and injured leg remain, and his wrists are no longer cuffed. 페드로: He has hurried into the container and is out of breath.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S24sh4__bgfirst_bg.png",
     "asset_id": "1e75d314-aff0-417c-9096-0940ab4b2c40",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S24sh4.png",
     "asset_id": "835ca9be-da8c-48a5-bf21-995533ce4a6b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
     "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 우측에 위치한 페드로의 얼굴을 똑바로 향하고 있으며, 뻗은 손은 페드로의 멱살을 쥐고 있음.",
    "built_space": "컨테이너 내부. 좌측 벽면에 걸린 수건, 화면 좌측 하단의 나무 의자 등 레퍼런스의 구조가 일치하며, 뒤쪽 문을 통해 채광이 들어옴.",
    "entities": "이현우는 검은 머리, 충혈된 눈, 멍든 얼굴, 어두운 셔츠, 인이어 무전기를 착용함. 페드로는 우측에 크롭되어 있으며 비니와 자켓을 착용함.",
    "hard_violations": [],
    "physics": "이현우가 앞으로 몸을 기울인 자세이며, 페드로의 옷을 쥔 손이 동작과 무게 중심을 자연스럽게 지지함."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 페드로를 향하고, 오른손은 페드로의 가슴 쪽을 향해 허공으로 뻗어 있음.",
    "built_space": "컨테이너 내부. 좌측 벽에 수건과 사진이 있으며 배경 깊은 곳에 문이 배치됨.",
    "entities": "이현우는 멍든 얼굴, 충혈된 눈, 무전기 등 지시문을 충족함. 페드로는 자켓과 비니를 썼으나, 지시문에 없는 무전기를 귀에 착용하고 있음.",
    "hard_violations": [],
    "physics": "앞으로 몸을 기울인 채 손을 뻗는 동작 중이며, 허공에 뜬 손은 뻗는 행위의 일부로 자연스럽게 유지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "멱살을 쥔 손과 충혈된 눈의 묘사가 훌륭하며, 레퍼런스의 실내 구조와 전경의 의자까지 화면에 정확하게 통합되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손을 뻗은 동작은 지시문에 맞으나, 페드로에게 지시되지 않은 무전기가 나타났고 인물 배치가 덜 정확합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 우측에 위치한 페드로의 얼굴을 똑바로 향하고 있으며, 뻗은 손은 페드로의 멱살을 쥐고 있음.",
        "built_space": "컨테이너 내부. 좌측 벽면에 걸린 수건, 화면 좌측 하단의 나무 의자 등 레퍼런스의 구조가 일치하며, 뒤쪽 문을 통해 채광이 들어옴.",
        "entities": "이현우는 검은 머리, 충혈된 눈, 멍든 얼굴, 어두운 셔츠, 인이어 무전기를 착용함. 페드로는 우측에 크롭되어 있으며 비니와 자켓을 착용함.",
        "hard_violations": [],
        "physics": "이현우가 앞으로 몸을 기울인 자세이며, 페드로의 옷을 쥔 손이 동작과 무게 중심을 자연스럽게 지지함."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 페드로를 향하고, 오른손은 페드로의 가슴 쪽을 향해 허공으로 뻗어 있음.",
        "built_space": "컨테이너 내부. 좌측 벽에 수건과 사진이 있으며 배경 깊은 곳에 문이 배치됨.",
        "entities": "이현우는 멍든 얼굴, 충혈된 눈, 무전기 등 지시문을 충족함. 페드로는 자켓과 비니를 썼으나, 지시문에 없는 무전기를 귀에 착용하고 있음.",
        "hard_violations": [],
        "physics": "앞으로 몸을 기울인 채 손을 뻗는 동작 중이며, 허공에 뜬 손은 뻗는 행위의 일부로 자연스럽게 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "멱살을 쥔 손과 충혈된 눈의 묘사가 훌륭하며, 레퍼런스의 실내 구조와 전경의 의자까지 화면에 정확하게 통합되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "손을 뻗은 동작은 지시문에 맞으나, 페드로에게 지시되지 않은 무전기가 나타났고 인물 배치가 덜 정확합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 우측에 위치한 페드로의 얼굴을 똑바로 향하고 있으며, 뻗은 손은 페드로의 멱살을 쥐고 있음.",
        "built_space": "컨테이너 내부. 좌측 벽면에 걸린 수건, 화면 좌측 하단의 나무 의자 등 레퍼런스의 구조가 일치하며, 뒤쪽 문을 통해 채광이 들어옴.",
        "entities": "이현우는 검은 머리, 충혈된 눈, 멍든 얼굴, 어두운 셔츠, 인이어 무전기를 착용함. 페드로는 우측에 크롭되어 있으며 비니와 자켓을 착용함.",
        "hard_violations": [],
        "physics": "이현우가 앞으로 몸을 기울인 자세이며, 페드로의 옷을 쥔 손이 동작과 무게 중심을 자연스럽게 지지함."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 페드로를 향하고, 오른손은 페드로의 가슴 쪽을 향해 허공으로 뻗어 있음.",
        "built_space": "컨테이너 내부. 좌측 벽에 수건과 사진이 있으며 배경 깊은 곳에 문이 배치됨.",
        "entities": "이현우는 멍든 얼굴, 충혈된 눈, 무전기 등 지시문을 충족함. 페드로는 자켓과 비니를 썼으나, 지시문에 없는 무전기를 귀에 착용하고 있음.",
        "hard_violations": [],
        "physics": "앞으로 몸을 기울인 채 손을 뻗는 동작 중이며, 허공에 뜬 손은 뻗는 행위의 일부로 자연스럽게 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "페드로를 응시하며 열린 손을 멱살 쪽으로 뻗는 순간과 하단의 손 노출이 더 정확하지만, 오른쪽 전경의 페드로와 배경 내부가 지정보다 넓게 보인다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "충혈된 시선과 근접 대치는 맞지만 이미 멱살을 움켜쥔 다음 순간이며, 넓게 드러난 출입구와 실내가 좁고 불분명해야 하는 배경 조건에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 두 눈은 오른쪽 페드로의 얼굴을 향하고, 페드로의 잘린 옆얼굴도 이현우 쪽으로 돌아 있다. 이현우의 열린 손은 페드로의 깃 아래 윗가슴·어깨 부근을 향해 뻗어 있으며 아직 옷을 움켜쥐지는 않았다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 금속 벽에 수건 한 장과 그 뒤의 검은 걸이 물품, 작은 벽 사진 세 장, 전기함 하나가 보인다. 왼쪽 안쪽 문틀 하나, 후면 닫힌 문 하나, 오른쪽 창 하나와 천장 등기구 하나가 식별된다. 참고 장소의 낡은 금속 벽과 좁은 실내 통로에 대체로 부합한다. 이현우는 왼쪽 중앙, 페드로는 오른쪽 전경에 있지만 페드로의 어깨가 화면을 상당히 넓게 차지하고 배경도 좁고 불분명한 조각 이상으로 드러난다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "보이는 인물은 두 명뿐이다. 이현우는 젊은 동아시아계 남성 외관으로, 헝클어진 짧은 검은 머리와 마른 체격, 어두운 오염된 셔츠가 참고와 부합한다. 귀의 검은 인이어 장치, 얼굴의 타박상과 입가 출혈, 자연스러운 홍채·동공을 유지한 충혈된 눈이 보인다. 손목에는 수갑이 없다. 페드로는 참고의 검은 비니와 낡은 녹색 집업 재킷을 착용했으나 얼굴이 일부만 보여 정확한 얼굴 일치와 연령·혼혈 외관은 제한적으로만 확인된다. 다리 부상과 바지는 구도 밖이고, 페드로의 가쁜 호흡은 이 정지 화면에서 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우의 상체는 앞으로 기울고, 뻗은 손은 손목과 팔을 통해 어깨에 자연스럽게 연결된다. 손가락을 구부려 옷깃을 잡기 직전인 동작으로 가능한 자세다. 두 사람의 하체와 발은 화면 밖이므로 바닥 접촉 자체는 보이지 않지만 공중에 떠 있다는 징후는 없다. 수건은 벽걸이에 걸려 있으며 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 페드로의 얼굴을 올려다보고, 페드로의 잘린 머리도 그를 향한다. 이현우의 손은 페드로의 목 아래 재킷 깃에 정확히 닿아 천을 움켜쥐고 있다. 목표 방향은 맞지만 멱살을 향해 손을 뻗는 단계보다 행동이 진행되어 있다.",
        "built_space": "왼쪽에 수건 한 장과 검은 걸이 물품, 크게 열린 외부 출입구 하나, 아래쪽 의자 등받이 하나가 보인다. 뒤쪽에는 냉장고 하나와 그 위 용기 여러 개, 작은 수납 선반 하나, 천장 등기구 일부가 드러난다. 금속 벽과 냉장고 등 재료·설비는 참고 장소와 유사하지만, 왼쪽의 밝은 외부 통로까지 선명하게 보여 요구된 좁고 미해결된 배경보다 공간 설명이 훨씬 많다. 두 사람의 좌우 배치는 맞지만 페드로의 어깨와 머리가 오른쪽 전경을 넓게 차지한다. 명백한 설비 중복이나 반사 오류는 보이지 않는다.",
        "entities": "두 인물만 등장한다. 이현우의 젊은 동아시아계 남성 외관, 헝클어진 검은 머리, 마른 체격, 더러워진 어두운 셔츠와 인이어 장치는 참고에 대체로 맞는다. 얼굴의 멍과 입가 상처, 충혈된 정상 형태의 눈이 명확하며 드러난 손목에 수갑은 없다. 페드로의 검은 비니와 녹색 집업 재킷은 참고에 맞지만 잘린 얼굴만으로 정확한 정체성이나 숨찬 상태를 확정하기 어렵다. 하반신과 부상당한 다리는 화면 밖이므로 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "이현우의 손과 팔은 자연스럽게 연결되고, 주먹 안에 잡힌 페드로의 재킷 천이 당겨져 있어 실제 멱살잡이로 가능한 접촉이다. 다만 손의 일부는 옷깃과 전경 어깨에 겹쳐 열린 손을 온전히 보여주는 구성과 다르다. 두 상체는 서서 몸을 기울인 자세로 읽히며 발은 잘려 있다. 냉장고 위 용기는 상판에, 수건은 걸이에 지지되어 있고 부유하는 인물이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "페드로를 응시하며 열린 손을 멱살 쪽으로 뻗는 순간과 하단의 손 노출이 더 정확하지만, 오른쪽 전경의 페드로와 배경 내부가 지정보다 넓게 보인다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "충혈된 시선과 근접 대치는 맞지만 이미 멱살을 움켜쥔 다음 순간이며, 넓게 드러난 출입구와 실내가 좁고 불분명해야 하는 배경 조건에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 두 눈은 오른쪽 페드로의 얼굴을 향하고, 페드로의 잘린 옆얼굴도 이현우 쪽으로 돌아 있다. 이현우의 열린 손은 페드로의 깃 아래 윗가슴·어깨 부근을 향해 뻗어 있으며 아직 옷을 움켜쥐지는 않았다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 금속 벽에 수건 한 장과 그 뒤의 검은 걸이 물품, 작은 벽 사진 세 장, 전기함 하나가 보인다. 왼쪽 안쪽 문틀 하나, 후면 닫힌 문 하나, 오른쪽 창 하나와 천장 등기구 하나가 식별된다. 참고 장소의 낡은 금속 벽과 좁은 실내 통로에 대체로 부합한다. 이현우는 왼쪽 중앙, 페드로는 오른쪽 전경에 있지만 페드로의 어깨가 화면을 상당히 넓게 차지하고 배경도 좁고 불분명한 조각 이상으로 드러난다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "보이는 인물은 두 명뿐이다. 이현우는 젊은 동아시아계 남성 외관으로, 헝클어진 짧은 검은 머리와 마른 체격, 어두운 오염된 셔츠가 참고와 부합한다. 귀의 검은 인이어 장치, 얼굴의 타박상과 입가 출혈, 자연스러운 홍채·동공을 유지한 충혈된 눈이 보인다. 손목에는 수갑이 없다. 페드로는 참고의 검은 비니와 낡은 녹색 집업 재킷을 착용했으나 얼굴이 일부만 보여 정확한 얼굴 일치와 연령·혼혈 외관은 제한적으로만 확인된다. 다리 부상과 바지는 구도 밖이고, 페드로의 가쁜 호흡은 이 정지 화면에서 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우의 상체는 앞으로 기울고, 뻗은 손은 손목과 팔을 통해 어깨에 자연스럽게 연결된다. 손가락을 구부려 옷깃을 잡기 직전인 동작으로 가능한 자세다. 두 사람의 하체와 발은 화면 밖이므로 바닥 접촉 자체는 보이지 않지만 공중에 떠 있다는 징후는 없다. 수건은 벽걸이에 걸려 있으며 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 페드로의 얼굴을 올려다보고, 페드로의 잘린 머리도 그를 향한다. 이현우의 손은 페드로의 목 아래 재킷 깃에 정확히 닿아 천을 움켜쥐고 있다. 목표 방향은 맞지만 멱살을 향해 손을 뻗는 단계보다 행동이 진행되어 있다.",
        "built_space": "왼쪽에 수건 한 장과 검은 걸이 물품, 크게 열린 외부 출입구 하나, 아래쪽 의자 등받이 하나가 보인다. 뒤쪽에는 냉장고 하나와 그 위 용기 여러 개, 작은 수납 선반 하나, 천장 등기구 일부가 드러난다. 금속 벽과 냉장고 등 재료·설비는 참고 장소와 유사하지만, 왼쪽의 밝은 외부 통로까지 선명하게 보여 요구된 좁고 미해결된 배경보다 공간 설명이 훨씬 많다. 두 사람의 좌우 배치는 맞지만 페드로의 어깨와 머리가 오른쪽 전경을 넓게 차지한다. 명백한 설비 중복이나 반사 오류는 보이지 않는다.",
        "entities": "두 인물만 등장한다. 이현우의 젊은 동아시아계 남성 외관, 헝클어진 검은 머리, 마른 체격, 더러워진 어두운 셔츠와 인이어 장치는 참고에 대체로 맞는다. 얼굴의 멍과 입가 상처, 충혈된 정상 형태의 눈이 명확하며 드러난 손목에 수갑은 없다. 페드로의 검은 비니와 녹색 집업 재킷은 참고에 맞지만 잘린 얼굴만으로 정확한 정체성이나 숨찬 상태를 확정하기 어렵다. 하반신과 부상당한 다리는 화면 밖이므로 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "이현우의 손과 팔은 자연스럽게 연결되고, 주먹 안에 잡힌 페드로의 재킷 천이 당겨져 있어 실제 멱살잡이로 가능한 접촉이다. 다만 손의 일부는 옷깃과 전경 어깨에 겹쳐 열린 손을 온전히 보여주는 구성과 다르다. 두 상체는 서서 몸을 기울인 자세로 읽히며 발은 잘려 있다. 냉장고 위 용기는 상판에, 수건은 걸이에 지지되어 있고 부유하는 인물이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "멱살을 쥔 손과 충혈된 눈의 묘사가 훌륭하며, 레퍼런스의 실내 구조와 전경의 의자까지 화면에 정확하게 통합되었습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "손을 뻗은 동작은 지시문에 맞으나, 페드로에게 지시되지 않은 무전기가 나타났고 인물 배치가 덜 정확합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_container_room_12e5e2.png",
    "asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0954-85f4-7613-93f6-5872ca39816b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S24sh4__bgfirst_bg.png",
   "bg_asset_id": "1e75d314-aff0-417c-9096-0940ab4b2c40",
   "bg_record_key": "S24sh4::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "family_container_room",
   "groupbg_asset_id": "8cc72072-c8cf-4a83-b8ea-2836d963893e"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S24sh7::signage": {
  "fp": "545480404ee38ff7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S24sh7": {
  "input_fingerprint": "92fe1ece4ac099b6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리와 아이들이 모두 사라졌다는 페드로의 말에 충격을 받아 입을 반쯤 벌린 이현우의 굳은 얼굴.\n\nLOCATION (lock): Inside the daylit living area of the container home, where the two young men have stopped to speak. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the inherited eye-height position just outside 페드로's shoulder line, preserving the same side of the axis with his shoulder now beyond the near crop. 이현우's face sits left of center with open space to the right toward the off-screen 페드로; his forward tension arrests, lips half parted and eyes fixed on the person delivering the news. Emphasize only the closer camera distance, allowing the inherited rise and outward adjustment to remain unobtrusive.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (대화가 계속되는 실내) — A soft strip of the interior borders the face and its look-space; used as Retains location continuity while excluding Pedro from the crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the previous shot's restrained ambient treatment so the shock reads through facial stillness rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the militia's search. 이현우: He remains inside the container in his outer top, with facial bruises and the persistent leg injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리와 아이들이 모두 사라졌다는 페드로의 말에 충격을 받아 입을 반쯤 벌린 이현우의 굳은 얼굴.\n\nLOCATION (lock): Inside the daylit living area of the container home, where the two young men have stopped to speak. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the inherited eye-height position just outside 페드로's shoulder line, preserving the same side of the axis with his shoulder now beyond the near crop. 이현우's face sits left of center with open space to the right toward the off-screen 페드로; his forward tension arrests, lips half parted and eyes fixed on the person delivering the news. Emphasize only the closer camera distance, allowing the inherited rise and outward adjustment to remain unobtrusive.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (대화가 계속되는 실내) — A soft strip of the interior borders the face and its look-space; used as Retains location continuity while excluding Pedro from the crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the previous shot's restrained ambient treatment so the shock reads through facial stillness rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the militia's search. 이현우: He remains inside the container in his outer top, with facial bruises and the persistent leg injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리와 아이들이 모두 사라졌다는 페드로의 말에 충격을 받아 입을 반쯤 벌린 이현우의 굳은 얼굴.\n\nLOCATION (lock): Inside the daylit living area of the container home, where the two young men have stopped to speak. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the inherited eye-height position just outside 페드로's shoulder line, preserving the same side of the axis with his shoulder now beyond the near crop. 이현우's face sits left of center with open space to the right toward the off-screen 페드로; his forward tension arrests, lips half parted and eyes fixed on the person delivering the news. Emphasize only the closer camera distance, allowing the inherited rise and outward adjustment to remain unobtrusive.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 컨테이너 집 내부 (대화가 계속되는 실내) — A soft strip of the interior borders the face and its look-space; used as Retains location continuity while excluding Pedro from the crop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the previous shot's restrained ambient treatment so the shock reads through facial stillness rather than a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Container 7-31 remains disturbed by the militia's search. 이현우: He remains inside the container in his outer top, with facial bruises and the persistent leg injury.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선은 화면 우측 프레임 밖의 페드로를 향해 고정되어 있습니다.",
    "built_space": "컨테이너 내부. 화면 좌측에 열린 외부 문이 있고, 우측 배경에 냉장고와 선반이 이전 샷과 동일한 위치에 있습니다.",
    "entities": "이현우: 이전 샷과 일치하는 멍든 얼굴, 헝클어진 머리, 인이어 무전기, 오염된 셔츠를 착용하고 있습니다.",
    "hard_violations": [],
    "physics": "인물의 목과 어깨 자세가 지시된 행동의 멈춤 상태를 자연스럽게 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 화면 우측 밖 보이지 않는 대상을 향하고 있습니다.",
    "built_space": "컨테이너 내부. 좌측의 열린 문과 우측의 냉장고, 선반 등 이전 샷의 구조물들이 올바른 위치에 있습니다.",
    "entities": "이현우: 캐릭터 레퍼런스 및 이전 샷과 일치하는 외모, 상처, 인이어 무전기 및 의상을 보여줍니다.",
    "hard_violations": [],
    "physics": "어색함 없이 머리와 상체가 올바른 지지 상태를 보여줍니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 배경 연속성을 완벽하게 유지하며, 페드로를 프레임에서 제외하고 이현우의 충격받은 표정과 시선을 지시된 클로즈업 앵글로 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 프레이밍, 배경의 연속성, 인물의 외양 및 감정 표현(반쯤 벌린 입, 굳은 얼굴)을 매우 훌륭하게 따랐습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 화면 우측 프레임 밖의 페드로를 향해 고정되어 있습니다.",
        "built_space": "컨테이너 내부. 화면 좌측에 열린 외부 문이 있고, 우측 배경에 냉장고와 선반이 이전 샷과 동일한 위치에 있습니다.",
        "entities": "이현우: 이전 샷과 일치하는 멍든 얼굴, 헝클어진 머리, 인이어 무전기, 오염된 셔츠를 착용하고 있습니다.",
        "hard_violations": [],
        "physics": "인물의 목과 어깨 자세가 지시된 행동의 멈춤 상태를 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 우측 밖 보이지 않는 대상을 향하고 있습니다.",
        "built_space": "컨테이너 내부. 좌측의 열린 문과 우측의 냉장고, 선반 등 이전 샷의 구조물들이 올바른 위치에 있습니다.",
        "entities": "이현우: 캐릭터 레퍼런스 및 이전 샷과 일치하는 외모, 상처, 인이어 무전기 및 의상을 보여줍니다.",
        "hard_violations": [],
        "physics": "어색함 없이 머리와 상체가 올바른 지지 상태를 보여줍니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 샷의 배경 연속성을 완벽하게 유지하며, 페드로를 프레임에서 제외하고 이현우의 충격받은 표정과 시선을 지시된 클로즈업 앵글로 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 프레이밍, 배경의 연속성, 인물의 외양 및 감정 표현(반쯤 벌린 입, 굳은 얼굴)을 매우 훌륭하게 따랐습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선은 화면 우측 프레임 밖의 페드로를 향해 고정되어 있습니다.",
        "built_space": "컨테이너 내부. 화면 좌측에 열린 외부 문이 있고, 우측 배경에 냉장고와 선반이 이전 샷과 동일한 위치에 있습니다.",
        "entities": "이현우: 이전 샷과 일치하는 멍든 얼굴, 헝클어진 머리, 인이어 무전기, 오염된 셔츠를 착용하고 있습니다.",
        "hard_violations": [],
        "physics": "인물의 목과 어깨 자세가 지시된 행동의 멈춤 상태를 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 우측 밖 보이지 않는 대상을 향하고 있습니다.",
        "built_space": "컨테이너 내부. 좌측의 열린 문과 우측의 냉장고, 선반 등 이전 샷의 구조물들이 올바른 위치에 있습니다.",
        "entities": "이현우: 캐릭터 레퍼런스 및 이전 샷과 일치하는 외모, 상처, 인이어 무전기 및 의상을 보여줍니다.",
        "hard_violations": [],
        "physics": "어색함 없이 머리와 상체가 올바른 지지 상태를 보여줍니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 화면 밖 페드로를 보는 시선과 반쯤 열린 입, 상처와 실내 연속성은 맞지만, B보다 어깨와 배경을 넓게 담아 얼굴 중심의 밀착된 클로즈업이 덜하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙 왼쪽의 얼굴을 더 밀착해 담으면서 오른쪽 시선 공간과 페드로의 완전한 화면 제외를 유지해, 소식을 듣고 굳은 표정의 클로즈업을 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 두 눈이 화면 오른쪽의 약간 높은 지점을 향하며, 화면 밖 페드로를 바라보는 대화 방향과 맞는다. 렌즈를 직접 보지 않는다. 입은 반쯤 열려 있고 앞으로 기울어진 자세에서 움직임이 멎은 듯하다. 무기나 방향을 확인할 휴대 물건은 보이지 않는다.",
        "built_space": "왼쪽에 열린 출입구 하나와 걸린 수건 하나, 오른쪽 뒤에 냉장고 하나와 그 앞의 검은 수납 선반 하나가 보인다. 골이 있는 금속 벽과 천장, 출입구로 들어오는 낮빛이 이전 장면의 컨테이너 실내와 이어진다. 이현우는 이 설비들보다 앞에 있으며 페드로는 완전히 제외되었다. 설비 중복이나 불가능한 반사는 없다. 다만 오른쪽 실내와 상체가 비교적 넓게 드러나 배경을 부드러운 띠 정도로 제한하라는 지시에는 다소 느슨하다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며, 한국계 미국인이라는 국적은 외형만으로 확인할 수 없다. 앳된 얼굴, 헝클어진 검은 머리, 마른 목과 어깨, 먼지와 얼룩이 묻은 어두운 셔츠가 참조와 부합한다. 검은 인이어 무전기 하나, 볼과 눈 주변의 타박상, 찢어진 입술도 유지된다. 찰리와 아이들은 등장하지 않으며 페드로의 신체나 옷도 없다. 바지와 다리 부상은 적절한 상반신 크롭 밖이라 확인 대상이 아니다. 추가 문구나 도식은 없다.",
        "hard_violations": [],
        "physics": "머리는 목에, 목과 어깨는 화면 아래로 이어지는 몸통에 자연스럽게 연결된다. 상체가 앞으로 기울어 있지만 공중에 떠 있는 형상은 아니며, 발과 하체의 지지는 크롭 밖이다. 인이어 장치는 귀에 끼워져 있고 수건은 벽 쪽에 걸려 있으며 용기들은 선반이나 냉장고 위에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 눈과 얼굴이 화면 오른쪽 위의 화면 밖 상대를 향한다. 이전 장면에서 마주하던 페드로에게 시선이 고정된 것으로 읽히며 렌즈 응시는 아니다. 입술을 반쯤 벌린 채 얼굴의 움직임이 멎어 있어 충격을 받은 순간에 부합한다. 조준하거나 이동하는 별도 물체는 없다.",
        "built_space": "왼쪽 뒤에 출입구 하나와 벽에 걸린 수건 하나, 오른쪽 뒤에 냉장고 하나와 검은 선반 하나가 보인다. 냉장고 위의 용기들과 금속 벽·천장은 이전 장면의 배치와 재질을 유지하며 낮의 주변광도 이어진다. 얼굴은 중앙 왼쪽, 시선 공간은 오른쪽에 있고 페드로의 어깨까지 크롭 밖이다. A보다 얼굴이 크게 잡혀 실내가 주변 배경으로 물러난다. 고정 설비의 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "참조와 닮은 앳된 동아시아계 남성 한 명만 등장한다. 헝클어진 짧은 검은 머리, 마른 체격, 낡고 더러워진 어두운 셔츠, 귀의 검은 인이어 장치가 유지된다. 볼의 멍과 긁힌 자국, 눈가의 붉은 기운, 입술의 혈흔도 이전 장면과 연결된다. 눈은 정상적인 홍채와 동공을 가진 사람의 눈이다. 페드로와 찰리, 아이들은 보이지 않는다. 하의와 다리 부상은 클로즈업 밖이며 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 기울어진 어깨와 몸통의 연결이 자연스럽다. 앞으로 쏠린 상태에서 멈춘 상체로 읽히며, 하체가 보이지 않는다고 해서 부유한 자세로 볼 근거는 없다. 인이어 장치는 귀가 지지하고, 수건은 걸이에 걸려 있으며 배경 용기들은 가구 표면에 놓여 있다. 지지 없는 물체나 해부학적으로 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 화면 밖 페드로를 보는 시선과 반쯤 열린 입, 상처와 실내 연속성은 맞지만, B보다 어깨와 배경을 넓게 담아 얼굴 중심의 밀착된 클로즈업이 덜하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙 왼쪽의 얼굴을 더 밀착해 담으면서 오른쪽 시선 공간과 페드로의 완전한 화면 제외를 유지해, 소식을 듣고 굳은 표정의 클로즈업을 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 두 눈이 화면 오른쪽의 약간 높은 지점을 향하며, 화면 밖 페드로를 바라보는 대화 방향과 맞는다. 렌즈를 직접 보지 않는다. 입은 반쯤 열려 있고 앞으로 기울어진 자세에서 움직임이 멎은 듯하다. 무기나 방향을 확인할 휴대 물건은 보이지 않는다.",
        "built_space": "왼쪽에 열린 출입구 하나와 걸린 수건 하나, 오른쪽 뒤에 냉장고 하나와 그 앞의 검은 수납 선반 하나가 보인다. 골이 있는 금속 벽과 천장, 출입구로 들어오는 낮빛이 이전 장면의 컨테이너 실내와 이어진다. 이현우는 이 설비들보다 앞에 있으며 페드로는 완전히 제외되었다. 설비 중복이나 불가능한 반사는 없다. 다만 오른쪽 실내와 상체가 비교적 넓게 드러나 배경을 부드러운 띠 정도로 제한하라는 지시에는 다소 느슨하다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며, 한국계 미국인이라는 국적은 외형만으로 확인할 수 없다. 앳된 얼굴, 헝클어진 검은 머리, 마른 목과 어깨, 먼지와 얼룩이 묻은 어두운 셔츠가 참조와 부합한다. 검은 인이어 무전기 하나, 볼과 눈 주변의 타박상, 찢어진 입술도 유지된다. 찰리와 아이들은 등장하지 않으며 페드로의 신체나 옷도 없다. 바지와 다리 부상은 적절한 상반신 크롭 밖이라 확인 대상이 아니다. 추가 문구나 도식은 없다.",
        "hard_violations": [],
        "physics": "머리는 목에, 목과 어깨는 화면 아래로 이어지는 몸통에 자연스럽게 연결된다. 상체가 앞으로 기울어 있지만 공중에 떠 있는 형상은 아니며, 발과 하체의 지지는 크롭 밖이다. 인이어 장치는 귀에 끼워져 있고 수건은 벽 쪽에 걸려 있으며 용기들은 선반이나 냉장고 위에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 눈과 얼굴이 화면 오른쪽 위의 화면 밖 상대를 향한다. 이전 장면에서 마주하던 페드로에게 시선이 고정된 것으로 읽히며 렌즈 응시는 아니다. 입술을 반쯤 벌린 채 얼굴의 움직임이 멎어 있어 충격을 받은 순간에 부합한다. 조준하거나 이동하는 별도 물체는 없다.",
        "built_space": "왼쪽 뒤에 출입구 하나와 벽에 걸린 수건 하나, 오른쪽 뒤에 냉장고 하나와 검은 선반 하나가 보인다. 냉장고 위의 용기들과 금속 벽·천장은 이전 장면의 배치와 재질을 유지하며 낮의 주변광도 이어진다. 얼굴은 중앙 왼쪽, 시선 공간은 오른쪽에 있고 페드로의 어깨까지 크롭 밖이다. A보다 얼굴이 크게 잡혀 실내가 주변 배경으로 물러난다. 고정 설비의 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "참조와 닮은 앳된 동아시아계 남성 한 명만 등장한다. 헝클어진 짧은 검은 머리, 마른 체격, 낡고 더러워진 어두운 셔츠, 귀의 검은 인이어 장치가 유지된다. 볼의 멍과 긁힌 자국, 눈가의 붉은 기운, 입술의 혈흔도 이전 장면과 연결된다. 눈은 정상적인 홍채와 동공을 가진 사람의 눈이다. 페드로와 찰리, 아이들은 보이지 않는다. 하의와 다리 부상은 클로즈업 밖이며 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 기울어진 어깨와 몸통의 연결이 자연스럽다. 앞으로 쏠린 상태에서 멈춘 상체로 읽히며, 하체가 보이지 않는다고 해서 부유한 자세로 볼 근거는 없다. 인이어 장치는 귀가 지지하고, 수건은 걸이에 걸려 있으며 배경 용기들은 가구 표면에 놓여 있다. 지지 없는 물체나 해부학적으로 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.889
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.889
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1889
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이전 샷의 배경 연속성을 완벽하게 유지하며, 페드로를 프레임에서 제외하고 이현우의 충격받은 표정과 시선을 지시된 클로즈업 앵글로 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "프롬프트가 요구한 프레이밍, 배경의 연속성, 인물의 외양 및 감정 표현(반쯤 벌린 입, 굳은 얼굴)을 매우 훌륭하게 따랐습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S24sh4_sel.png",
    "asset_id": "4a6ca536-1857-46fc-b004-fac332356499",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab095b-0410-7928-92c0-fe3f07d53ec2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh4"
  }
 },
 "S24sh10::signage": {
  "fp": "c1f249c028532f1d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S24sh10": {
  "input_fingerprint": "00e30f0cea79db3e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖을 향해 뒷발로 바닥을 차고 앞발을 크게 뻗은 채 허공에 떠 있는 전력 질주 도중의 이현우의 뒷모습.\n\nLOCATION (lock): At the open front threshold of the container home, leading out into the refugee-settlement lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Follow the inherited pursuit path from inside the home, behind and slightly beside 이현우's running line at roughly waist height, with a gentle upward view retaining his whole body. Catch him at center in the airborne interval after his rear-foot push, his leading leg extended and both feet clear of the lower crop, with the open doorway ahead in the upper center. His head and attention remain directed through that opening, and the main emphasis is his movement away from the camera rather than a new lighting or eyeline effect.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open container doorway ahead of the runner in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 컨테이너 출입문 (열려 있으며 현우가 아직 문턱을 넘기 전) — The opening is seen from inside, ahead of Hyunwoo's running direction; used as Destination anchor above the full-body stride, occupying a limited portion of the composition; 집 안 바닥 (현우가 박차고 달리는 바닥); used as Visible beneath both feet to establish the airborne moment.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain controlled ambient exposure across the interior and visible doorway without inventing a pronounced exterior backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance to container 7-31 is open for the hurried exit; the earlier search disturbance has not been cleared. 이현우: He is rushing out in his outer top, still bearing facial bruises and the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖을 향해 뒷발로 바닥을 차고 앞발을 크게 뻗은 채 허공에 떠 있는 전력 질주 도중의 이현우의 뒷모습.\n\nLOCATION (lock): At the open front threshold of the container home, leading out into the refugee-settlement lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Follow the inherited pursuit path from inside the home, behind and slightly beside 이현우's running line at roughly waist height, with a gentle upward view retaining his whole body. Catch him at center in the airborne interval after his rear-foot push, his leading leg extended and both feet clear of the lower crop, with the open doorway ahead in the upper center. His head and attention remain directed through that opening, and the main emphasis is his movement away from the camera rather than a new lighting or eyeline effect.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open container doorway ahead of the runner in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 컨테이너 출입문 (열려 있으며 현우가 아직 문턱을 넘기 전) — The opening is seen from inside, ahead of Hyunwoo's running direction; used as Destination anchor above the full-body stride, occupying a limited portion of the composition; 집 안 바닥 (현우가 박차고 달리는 바닥); used as Visible beneath both feet to establish the airborne moment.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain controlled ambient exposure across the interior and visible doorway without inventing a pronounced exterior backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance to container 7-31 is open for the hurried exit; the earlier search disturbance has not been cleared. 이현우: He is rushing out in his outer top, still bearing facial bruises and the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖을 향해 뒷발로 바닥을 차고 앞발을 크게 뻗은 채 허공에 떠 있는 전력 질주 도중의 이현우의 뒷모습.\n\nLOCATION (lock): At the open front threshold of the container home, leading out into the refugee-settlement lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Follow the inherited pursuit path from inside the home, behind and slightly beside 이현우's running line at roughly waist height, with a gentle upward view retaining his whole body. Catch him at center in the airborne interval after his rear-foot push, his leading leg extended and both feet clear of the lower crop, with the open doorway ahead in the upper center. His head and attention remain directed through that opening, and the main emphasis is his movement away from the camera rather than a new lighting or eyeline effect.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open container doorway ahead of the runner in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 컨테이너 출입문 (열려 있으며 현우가 아직 문턱을 넘기 전) — The opening is seen from inside, ahead of Hyunwoo's running direction; used as Destination anchor above the full-body stride, occupying a limited portion of the composition; 집 안 바닥 (현우가 박차고 달리는 바닥); used as Visible beneath both feet to establish the airborne moment.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain controlled ambient exposure across the interior and visible doorway without inventing a pronounced exterior backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The entrance to container 7-31 is open for the hurried exit; the earlier search disturbance has not been cleared. 이현우: He is rushing out in his outer top, still bearing facial bruises and the injured leg.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 몸체와 진행 방향이 정면 상단 중앙의 열린 컨테이너 문을 향하고 있음.",
    "built_space": "컨테이너 내부. 왼쪽의 의자 1개와 수건 1개, 오른쪽의 선반 1개와 냉장고 1개가 레퍼런스와 정확한 위치로 일치함. 인물은 중앙에 위치함.",
    "entities": "이현우. 레퍼런스와 일치하는 검은 머리 및 오염된 어두운 셔츠와 바지가 확인됨. 무전기는 후면 각도상 보이지 않음.",
    "hard_violations": [],
    "physics": "전력 질주하는 굽힘과 뻗음(오른발 전진, 왼발 킥)이 도약을 만들어내어 두 발이 모두 허공에 떠 있는 상태를 뒷받침함."
   },
   {
    "label": "B",
    "direction": "인물의 몸체와 진행 방향이 정면의 열린 문을 향하고 있음.",
    "built_space": "컨테이너 내부. 왼쪽에 가방 1개가 추가되었고 오른쪽에 레퍼런스에 없던 나무 테이블 1개와 스툴 1개가 임의로 배치됨.",
    "entities": "이현우. 검은 머리와 어두운 셔츠 및 바지가 일치함.",
    "hard_violations": [],
    "physics": "달리는 팔다리의 굽힘과 뻗음(왼발 전진, 오른발 킥)이 두 발이 공중에 떠 있는 체공 상태를 물리적으로 설명함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 배경 요소(수건, 선반 위 기기)를 완벽히 유지하면서 지시된 공중 질주 동작과 프레이밍을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지시된 도약 동작은 잘 표현했으나 레퍼런스에 없는 테이블과 가방이 추가되어 공간 일관성이 어긋남."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 몸체와 진행 방향이 정면 상단 중앙의 열린 컨테이너 문을 향하고 있음.",
        "built_space": "컨테이너 내부. 왼쪽의 의자 1개와 수건 1개, 오른쪽의 선반 1개와 냉장고 1개가 레퍼런스와 정확한 위치로 일치함. 인물은 중앙에 위치함.",
        "entities": "이현우. 레퍼런스와 일치하는 검은 머리 및 오염된 어두운 셔츠와 바지가 확인됨. 무전기는 후면 각도상 보이지 않음.",
        "hard_violations": [],
        "physics": "전력 질주하는 굽힘과 뻗음(오른발 전진, 왼발 킥)이 도약을 만들어내어 두 발이 모두 허공에 떠 있는 상태를 뒷받침함."
       },
       {
        "label": "B",
        "direction": "인물의 몸체와 진행 방향이 정면의 열린 문을 향하고 있음.",
        "built_space": "컨테이너 내부. 왼쪽에 가방 1개가 추가되었고 오른쪽에 레퍼런스에 없던 나무 테이블 1개와 스툴 1개가 임의로 배치됨.",
        "entities": "이현우. 검은 머리와 어두운 셔츠 및 바지가 일치함.",
        "hard_violations": [],
        "physics": "달리는 팔다리의 굽힘과 뻗음(왼발 전진, 오른발 킥)이 두 발이 공중에 떠 있는 체공 상태를 물리적으로 설명함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 배경 요소(수건, 선반 위 기기)를 완벽히 유지하면서 지시된 공중 질주 동작과 프레이밍을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지시된 도약 동작은 잘 표현했으나 레퍼런스에 없는 테이블과 가방이 추가되어 공간 일관성이 어긋남."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 몸체와 진행 방향이 정면 상단 중앙의 열린 컨테이너 문을 향하고 있음.",
        "built_space": "컨테이너 내부. 왼쪽의 의자 1개와 수건 1개, 오른쪽의 선반 1개와 냉장고 1개가 레퍼런스와 정확한 위치로 일치함. 인물은 중앙에 위치함.",
        "entities": "이현우. 레퍼런스와 일치하는 검은 머리 및 오염된 어두운 셔츠와 바지가 확인됨. 무전기는 후면 각도상 보이지 않음.",
        "hard_violations": [],
        "physics": "전력 질주하는 굽힘과 뻗음(오른발 전진, 왼발 킥)이 도약을 만들어내어 두 발이 모두 허공에 떠 있는 상태를 뒷받침함."
       },
       {
        "label": "B",
        "direction": "인물의 몸체와 진행 방향이 정면의 열린 문을 향하고 있음.",
        "built_space": "컨테이너 내부. 왼쪽에 가방 1개가 추가되었고 오른쪽에 레퍼런스에 없던 나무 테이블 1개와 스툴 1개가 임의로 배치됨.",
        "entities": "이현우. 검은 머리와 어두운 셔츠 및 바지가 일치함.",
        "hard_violations": [],
        "physics": "달리는 팔다리의 굽힘과 뻗음(왼발 전진, 오른발 킥)이 두 발이 공중에 떠 있는 체공 상태를 물리적으로 설명함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "문턱 이전의 공중 질주와 전신 후면은 구현했지만, 인물이 더 크게 오른쪽에 치우치고 출입구가 왼쪽으로 밀려 지정된 중앙 추격 구도에서 더 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "중앙의 전신과 양발 아래 바닥, 문턱 이전의 질주를 더 충실히 담았으나, 출입구가 상단 중앙보다 왼쪽이고 앞다리를 크게 뻗은 동작은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성은 등을 보이며 화면 깊숙한 출입구 쪽으로 달린다. 머리도 전방을 향하지만 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 몸의 중심은 출입구 오른쪽에 있어 문으로 향하는 경로가 다소 옆으로 어긋나 보인다. 겨누는 무기나 손에 든 물체는 없다.",
        "built_space": "안쪽 복도에서 바깥을 보는 구도이며 열린 출입구 하나가 중앙 왼쪽에 있다. 인물은 아직 실내이고 문턱과 양발 아래 바닥이 보인다. 왼쪽에는 걸린 수건 하나와 검은 물품 하나, 잘린 의자 하나가 있고, 오른쪽에는 냉장고 하나, 조리용 선반 하나, 바닥의 상자형 기기 하나, 일부 보이는 탁자와 스툴이 있다. 낡은 금속 벽과 천장, 수건 및 냉장고의 좌우 관계는 참고 사진과 대체로 맞는다. 인물이 중앙보다 오른쪽이고 출입구도 지정된 상단 중앙에서 벗어난다. 불가능한 거울 반사는 없다.",
        "entities": "인물은 한 명뿐이며, 짧고 헝클어진 검은 머리와 마른 남성 체격, 때와 얼룩이 있는 어두운 셔츠, 올리브색 바지와 낡은 부츠가 참고 인물과 대체로 일치한다. 뒷모습이므로 한국계 미국인이라는 정체성, 정확한 얼굴 나이와 멍, 인이어 착용 여부는 판별할 수 없다. 다리 부상도 명확하지 않다. 바깥에는 낮의 컨테이너 정착촌이 보이고 다른 사람이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "두 발 모두 바닥에서 떨어져 있다. 뒤쪽 다리는 아래뒤로 뻗고 반대쪽 무릎은 접혀 있으며, 상체의 전경과 팔 동작은 바닥을 밀어낸 뒤의 짧은 달리기 비행 구간으로 설명 가능하다. 현재 접촉 지지점은 없지만 이륙한 실내 바닥과 전방 착지 공간이 있으므로 근거 없는 부유는 아니다. 다만 앞다리를 크게 전방으로 뻗었다기보다는 무릎을 접어 회수하는 순간에 가깝다. 집기들은 바닥이나 선반에 놓이고 수건은 벽에 걸려 있다."
       },
       {
        "label": "B",
        "direction": "등을 보인 남성의 머리와 몸은 열린 출입구 및 그 너머 골목 방향을 향한다. 눈 자체는 가려져 있다. 출입구가 몸 중심보다 왼쪽에 있지만 A보다 몸과 목적지의 화면상 간격이 작아 퇴실 동선이 더 명료하다. 손에 든 물건이나 별도의 조준 대상은 없다.",
        "built_space": "실내에서 출입구를 향한 낮은 추격 시점으로, 전신이 거의 중앙에 놓이고 발 아래 바닥도 충분히 보인다. 열린 출입구는 하나이며 중앙 왼쪽 배경에 있고, 인물은 문턱을 넘지 않았다. 왼쪽에는 수건 하나와 검은 물품 하나, 일부 잘린 의자가 보인다. 오른쪽에는 냉장고 하나, 조리 선반 하나, 아래쪽 상자들, 물통 하나와 스툴 하나가 보인다. 금속 벽·천장과 주요 집기의 좌우 배치는 참고 장소와 대체로 이어진다. A보다 인물 크기와 중앙 배치가 지정된 와이드 구도에 가깝지만 출입구는 여전히 왼쪽으로 치우친다. 바닥의 흐릿한 반사에는 명백한 광학적 모순이 없다.",
        "entities": "한 명의 마른 남성이 있으며, 헝클어진 검은 머리, 낡고 얼룩진 어두운 셔츠, 올리브색 바지와 부츠가 참고 이미지의 외형을 따른다. 얼굴을 노출하지 않아 뒷모습 지시를 지켰으며, 얼굴의 멍·정확한 나이·민족적 정체성과 인이어는 이 각도에서 확인할 수 없다. 다리 부상은 뚜렷하지 않다. 열린 문 너머 낮의 컨테이너 골목이 보이며 다른 인물이나 그래픽 자막은 없다.",
        "hard_violations": [],
        "physics": "양발이 바닥에서 떨어져 있고, 뒤로 남은 다리와 접힌 반대쪽 무릎, 앞으로 기운 몸통과 굽힌 팔이 달리기의 비행 구간을 형성한다. 직접적인 접촉 지지는 없지만 직전 발차기의 출발면과 다음 발을 디딜 실내 바닥이 명확하여 물리적으로 가능한 순간이다. 앞다리의 큰 전방 신전은 보이지 않아 요청한 보폭과는 차이가 있다. 냉장고·선반·상자·물통·스툴에는 각각 바닥 지지가 있고 조리 도구는 선반 위에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "문턱 이전의 공중 질주와 전신 후면은 구현했지만, 인물이 더 크게 오른쪽에 치우치고 출입구가 왼쪽으로 밀려 지정된 중앙 추격 구도에서 더 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "중앙의 전신과 양발 아래 바닥, 문턱 이전의 질주를 더 충실히 담았으나, 출입구가 상단 중앙보다 왼쪽이고 앞다리를 크게 뻗은 동작은 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남성은 등을 보이며 화면 깊숙한 출입구 쪽으로 달린다. 머리도 전방을 향하지만 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 몸의 중심은 출입구 오른쪽에 있어 문으로 향하는 경로가 다소 옆으로 어긋나 보인다. 겨누는 무기나 손에 든 물체는 없다.",
        "built_space": "안쪽 복도에서 바깥을 보는 구도이며 열린 출입구 하나가 중앙 왼쪽에 있다. 인물은 아직 실내이고 문턱과 양발 아래 바닥이 보인다. 왼쪽에는 걸린 수건 하나와 검은 물품 하나, 잘린 의자 하나가 있고, 오른쪽에는 냉장고 하나, 조리용 선반 하나, 바닥의 상자형 기기 하나, 일부 보이는 탁자와 스툴이 있다. 낡은 금속 벽과 천장, 수건 및 냉장고의 좌우 관계는 참고 사진과 대체로 맞는다. 인물이 중앙보다 오른쪽이고 출입구도 지정된 상단 중앙에서 벗어난다. 불가능한 거울 반사는 없다.",
        "entities": "인물은 한 명뿐이며, 짧고 헝클어진 검은 머리와 마른 남성 체격, 때와 얼룩이 있는 어두운 셔츠, 올리브색 바지와 낡은 부츠가 참고 인물과 대체로 일치한다. 뒷모습이므로 한국계 미국인이라는 정체성, 정확한 얼굴 나이와 멍, 인이어 착용 여부는 판별할 수 없다. 다리 부상도 명확하지 않다. 바깥에는 낮의 컨테이너 정착촌이 보이고 다른 사람이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "두 발 모두 바닥에서 떨어져 있다. 뒤쪽 다리는 아래뒤로 뻗고 반대쪽 무릎은 접혀 있으며, 상체의 전경과 팔 동작은 바닥을 밀어낸 뒤의 짧은 달리기 비행 구간으로 설명 가능하다. 현재 접촉 지지점은 없지만 이륙한 실내 바닥과 전방 착지 공간이 있으므로 근거 없는 부유는 아니다. 다만 앞다리를 크게 전방으로 뻗었다기보다는 무릎을 접어 회수하는 순간에 가깝다. 집기들은 바닥이나 선반에 놓이고 수건은 벽에 걸려 있다."
       },
       {
        "label": "A",
        "direction": "등을 보인 남성의 머리와 몸은 열린 출입구 및 그 너머 골목 방향을 향한다. 눈 자체는 가려져 있다. 출입구가 몸 중심보다 왼쪽에 있지만 A보다 몸과 목적지의 화면상 간격이 작아 퇴실 동선이 더 명료하다. 손에 든 물건이나 별도의 조준 대상은 없다.",
        "built_space": "실내에서 출입구를 향한 낮은 추격 시점으로, 전신이 거의 중앙에 놓이고 발 아래 바닥도 충분히 보인다. 열린 출입구는 하나이며 중앙 왼쪽 배경에 있고, 인물은 문턱을 넘지 않았다. 왼쪽에는 수건 하나와 검은 물품 하나, 일부 잘린 의자가 보인다. 오른쪽에는 냉장고 하나, 조리 선반 하나, 아래쪽 상자들, 물통 하나와 스툴 하나가 보인다. 금속 벽·천장과 주요 집기의 좌우 배치는 참고 장소와 대체로 이어진다. A보다 인물 크기와 중앙 배치가 지정된 와이드 구도에 가깝지만 출입구는 여전히 왼쪽으로 치우친다. 바닥의 흐릿한 반사에는 명백한 광학적 모순이 없다.",
        "entities": "한 명의 마른 남성이 있으며, 헝클어진 검은 머리, 낡고 얼룩진 어두운 셔츠, 올리브색 바지와 부츠가 참고 이미지의 외형을 따른다. 얼굴을 노출하지 않아 뒷모습 지시를 지켰으며, 얼굴의 멍·정확한 나이·민족적 정체성과 인이어는 이 각도에서 확인할 수 없다. 다리 부상은 뚜렷하지 않다. 열린 문 너머 낮의 컨테이너 골목이 보이며 다른 인물이나 그래픽 자막은 없다.",
        "hard_violations": [],
        "physics": "양발이 바닥에서 떨어져 있고, 뒤로 남은 다리와 접힌 반대쪽 무릎, 앞으로 기운 몸통과 굽힌 팔이 달리기의 비행 구간을 형성한다. 직접적인 접촉 지지는 없지만 직전 발차기의 출발면과 다음 발을 디딜 실내 바닥이 명확하여 물리적으로 가능한 순간이다. 앞다리의 큰 전방 신전은 보이지 않아 요청한 보폭과는 차이가 있다. 냉장고·선반·상자·물통·스툴에는 각각 바닥 지지가 있고 조리 도구는 선반 위에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.589
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.589
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1589
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 배경 요소(수건, 선반 위 기기)를 완벽히 유지하면서 지시된 공중 질주 동작과 프레이밍을 정확히 구현함."
   },
   {
    "label": "B",
    "score": 1589,
    "verdict_ko": "지시된 도약 동작은 잘 표현했으나 레퍼런스에 없는 테이블과 가방이 추가되어 공간 일관성이 어긋남."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S24sh4_sel.png",
    "asset_id": "4a6ca536-1857-46fc-b004-fac332356499",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab095f-57c9-72ab-9682-69a394a0621c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh4"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S25sh11::signage": {
  "fp": "de1560a0a81d285f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::reception_clearing": {
  "input_fingerprint": "b2df4204e23a5ddb",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "reception_clearing",
    "tags": [
     "S25sh11",
     "S25sh17",
     "S25sh28",
     "S26sh9"
    ]
   },
   "context_sig": "3cbf067e4cc307fc"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사) / 인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 어쩌다...앰버와 라울이 피로연장까지 들어선다. (*피로연장이라고 해봤자 그냥 빈 공터에 천막 치는 정도)\n- /피로연장 -N\n\nTIME OF DAY (lock): sunset to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사) / 인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 어쩌다...앰버와 라울이 피로연장까지 들어선다. (*피로연장이라고 해봤자 그냥 빈 공터에 천막 치는 정도)\n- /피로연장 -N\n\nTIME OF DAY (lock): sunset to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reception_clearing_3488b1.png",
  "asset_id": "296e6be3-728c-4aba-9a0e-c68a9359e450",
  "input_asset_ids": [
   "dfa4f22a-7333-4964-9636-0525b7191ce0"
  ],
  "origin_tag": "S25sh11",
  "place_text": "In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.",
  "origin_inputs": {
   "place_text": "In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.",
   "time_of_day_en": "sunset to night",
   "conti_asset_id": "dfa4f22a-7333-4964-9636-0525b7191ce0"
  }
 },
 "S25sh11::bgfirst_bg": {
  "input_fingerprint": "f0ec8fa52661e359",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 환한 표정으로 찰리와 함께 한쪽 발을 바닥에서 떼고 경쾌한 춤동작을 취한 mid-action 전신 구도.\n\nLOCATION (lock): In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.\n\nTIME OF DAY (lock): sunset to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the settled lateral section of the dance path at 앰버's waist height, looking slightly upward from the oblique side and leaving clear space above both heads and beneath both feet. 앰버 occupies left-center in three-quarter view, smiling with one foot lifted, while 찰리 turns toward her at right-center; both glance down toward their joined hands as their bodies complete different phases of the turn. Keep the surrounding dancers peripheral, with unequal step phases, varied shoulder angles and irregular gaps rather than a synchronized ring.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사람들이 음악에 맞춰 춤추고 있음); used as Clear ground around the pair preserves footwork; peripheral dancers have staggered steps, varied arm angles and naturally uneven spacing; 피로연 천막 (공터에 설치되어 있음) — A partial side view sits behind the shared dance space; used as Identifies the improvised reception without enclosing the pair in a symmetrical backdrop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued dusk illumination with controlled contrast, allowing the warmth of the moment to come from expressions and contact rather than an imposed warm color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 앰버가 환한 표정으로 찰리와 함께 한쪽 발을 바닥에서 떼고 경쾌한 춤동작을 취한 mid-action 전신 구도.\n\nLOCATION (lock): In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk.\n\nTIME OF DAY (lock): sunset to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the settled lateral section of the dance path at 앰버's waist height, looking slightly upward from the oblique side and leaving clear space above both heads and beneath both feet. 앰버 occupies left-center in three-quarter view, smiling with one foot lifted, while 찰리 turns toward her at right-center; both glance down toward their joined hands as their bodies complete different phases of the turn. Keep the surrounding dancers peripheral, with unequal step phases, varied shoulder angles and irregular gaps rather than a synchronized ring.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사람들이 음악에 맞춰 춤추고 있음); used as Clear ground around the pair preserves footwork; peripheral dancers have staggered steps, varied arm angles and naturally uneven spacing; 피로연 천막 (공터에 설치되어 있음) — A partial side view sits behind the shared dance space; used as Identifies the improvised reception without enclosing the pair in a symmetrical backdrop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued dusk illumination with controlled contrast, allowing the warmth of the moment to come from expressions and contact rather than an imposed warm color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh11__bgfirst_bg.png",
  "asset_id": "01187f03-8a03-4f0b-81e3-25b6ad120734",
  "input_asset_ids": [
   "dfa4f22a-7333-4964-9636-0525b7191ce0",
   "296e6be3-728c-4aba-9a0e-c68a9359e450"
  ]
 },
 "S25sh11": {
  "input_fingerprint": "4c32c70f3ffab954",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 환한 표정으로 찰리와 함께 한쪽 발을 바닥에서 떼고 경쾌한 춤동작을 취한 mid-action 전신 구도.\n\nLOCATION (lock): In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the settled lateral section of the dance path at 앰버's waist height, looking slightly upward from the oblique side and leaving clear space above both heads and beneath both feet. 앰버 occupies left-center in three-quarter view, smiling with one foot lifted, while 찰리 turns toward her at right-center; both glance down toward their joined hands as their bodies complete different phases of the turn. Keep the surrounding dancers peripheral, with unequal step phases, varied shoulder angles and irregular gaps rather than a synchronized ring.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사람들이 음악에 맞춰 춤추고 있음); used as Clear ground around the pair preserves footwork; peripheral dancers have staggered steps, varied arm angles and naturally uneven spacing; 피로연 천막 (공터에 설치되어 있음) — A partial side view sits behind the shared dance space; used as Identifies the improvised reception without enclosing the pair in a symmetrical backdrop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued dusk illumination with controlled contrast, allowing the warmth of the moment to come from expressions and contact rather than an imposed warm color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the repaired jukebox playing music at dusk. Charlie retains his old coat and hat disguise, blue-lit eyes and worn UBIQ chest logo, with the earlier surface dirt washed away. 앰버: She is dancing, with her previously established mask and waist tool pouch retained; her recurring cough has not been resolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 환한 표정으로 찰리와 함께 한쪽 발을 바닥에서 떼고 경쾌한 춤동작을 취한 mid-action 전신 구도.\n\nLOCATION (lock): In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the settled lateral section of the dance path at 앰버's waist height, looking slightly upward from the oblique side and leaving clear space above both heads and beneath both feet. 앰버 occupies left-center in three-quarter view, smiling with one foot lifted, while 찰리 turns toward her at right-center; both glance down toward their joined hands as their bodies complete different phases of the turn. Keep the surrounding dancers peripheral, with unequal step phases, varied shoulder angles and irregular gaps rather than a synchronized ring.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사람들이 음악에 맞춰 춤추고 있음); used as Clear ground around the pair preserves footwork; peripheral dancers have staggered steps, varied arm angles and naturally uneven spacing; 피로연 천막 (공터에 설치되어 있음) — A partial side view sits behind the shared dance space; used as Identifies the improvised reception without enclosing the pair in a symmetrical backdrop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued dusk illumination with controlled contrast, allowing the warmth of the moment to come from expressions and contact rather than an imposed warm color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the repaired jukebox playing music at dusk. Charlie retains his old coat and hat disguise, blue-lit eyes and worn UBIQ chest logo, with the earlier surface dirt washed away. 앰버: She is dancing, with her previously established mask and waist tool pouch retained; her recurring cough has not been resolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 앰버가 환한 표정으로 찰리와 함께 한쪽 발을 바닥에서 떼고 경쾌한 춤동작을 취한 mid-action 전신 구도.\n\nLOCATION (lock): In the central dancing area of an outdoor wedding reception, held under a simple canopy in a refugee-settlement clearing at dusk. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track along the settled lateral section of the dance path at 앰버's waist height, looking slightly upward from the oblique side and leaving clear space above both heads and beneath both feet. 앰버 occupies left-center in three-quarter view, smiling with one foot lifted, while 찰리 turns toward her at right-center; both glance down toward their joined hands as their bodies complete different phases of the turn. Keep the surrounding dancers peripheral, with unequal step phases, varied shoulder angles and irregular gaps rather than a synchronized ring.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사람들이 음악에 맞춰 춤추고 있음); used as Clear ground around the pair preserves footwork; peripheral dancers have staggered steps, varied arm angles and naturally uneven spacing; 피로연 천막 (공터에 설치되어 있음) — A partial side view sits behind the shared dance space; used as Identifies the improvised reception without enclosing the pair in a symmetrical backdrop.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued dusk illumination with controlled contrast, allowing the warmth of the moment to come from expressions and contact rather than an imposed warm color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception occupies a tented vacant lot, with the repaired jukebox playing music at dusk. Charlie retains his old coat and hat disguise, blue-lit eyes and worn UBIQ chest logo, with the earlier surface dirt washed away. 앰버: She is dancing, with her previously established mask and waist tool pouch retained; her recurring cough has not been resolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh11__bgfirst_bg.png",
     "asset_id": "01187f03-8a03-4f0b-81e3-25b6ad120734",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S25sh11.png",
     "asset_id": "dfa4f22a-7333-4964-9636-0525b7191ce0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reception_clearing_3488b1.png",
     "asset_id": "296e6be3-728c-4aba-9a0e-c68a9359e450",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "앰버는 오른쪽 아래의 맞잡은 손 쪽으로 얼굴과 눈을 향한다. 찰리는 왼쪽의 앰버와 손 쪽을 향하지만, 손이 눈높이 가까이 올라와 있어 시선이 거의 수평이다. 두 사람이 함께 손을 내려다보는 구도는 아니다. 주변 사람들은 서로 다른 방향으로 몸을 돌리고 있다.",
    "built_space": "왼쪽에 주크박스 한 대와 음향 장비가 있고, 뒤에는 여러 지주에 걸친 천막, 전구 줄, 삼각 깃발, 컨테이너 주거지가 보인다. 참조 장소의 주요 재료와 배치를 유지한다. 앰버는 왼쪽 중앙, 찰리는 오른쪽 중앙이며 머리 위와 발 아래에 여백이 있다. 주변 무용수들의 간격과 팔 각도는 불규칙하고 두 주인공의 발밑 공간은 확보되어 있다. 천막은 부분적인 측면보다는 화면 상부 전반을 덮는다.",
    "entities": "앰버는 금발의 어린 여자아이로, 참조와 유사한 얼굴과 카키 작업복, 방진 마스크, 가죽 공구 주머니를 갖췄다. 눈매에서 즐거운 표정이 읽힌다. 찰리는 짧고 육중한 베이지색 기계 몸체, 흰 얼굴판, 두 점 형태의 눈과 선 형태의 입, 파란 가슴 원자로, 낡은 코트와 모자를 갖췄다. 다만 눈은 요구된 파란색이 아니라 주황색이고, 요구된 가슴 상표는 식별되지 않는다. 배경 인물들은 명시된 주변 무용수로 읽힌다. 전체 조명은 요구보다 강한 주황색 기운을 띤다.",
    "hard_violations": [],
    "physics": "앰버는 화면 오른쪽 부츠로 바닥을 딛고 다른 다리를 접어 들었다. 찰리도 화면 오른쪽 발로 지지하며 반대쪽 발을 앞으로 들어 올렸다. 두 인물의 손은 실제로 맞닿아 있고, 자유로운 팔은 균형을 잡는 방향으로 벌어져 있다. 공중에 든 다리는 몸통과 지지 다리에 연결되어 있으며 지지 없는 부유는 없다. 코트와 주머니도 몸에 걸리거나 벨트에 부착되어 있다."
   },
   {
    "label": "A",
    "direction": "앰버는 오른쪽 아래의 맞잡은 손을 바라보고, 찰리 역시 머리를 왼쪽 아래로 숙여 같은 손 접촉점을 향한다. 찰리의 몸통도 앰버 쪽으로 돌아 있다. 두 인물의 시선 목표가 명확하며, 손을 내려다보면서 서로 다른 회전 단계에 있다는 지시에 더 가깝다.",
    "built_space": "왼쪽에 주크박스 한 대, 그 주변에 두 개의 높은 스피커와 장비 상자가 보인다. 천막 지주, 전구 줄, 깃발, 뒤편 컨테이너 건물과 오른쪽의 작은 천막 공간이 참조 장소와 대응한다. 두 인물은 지정된 좌우 중앙에 있고 전신과 위아래 여백이 확보되어 있다. 비스듬한 낮은 시점도 읽힌다. 그러나 공터가 완전히 비어 있어 명시된 주변 무용수와 불규칙한 군중 배치는 누락되었다.",
    "entities": "앰버는 금발의 어린 여자아이이며 참조와 비슷한 얼굴, 카키 작업복, 방진 마스크와 허리 공구 주머니를 유지한다. 찰리는 흰 얼굴판과 단순한 눈·입, 베이지 장갑판, 파란 원자로, 코트와 모자를 갖춘 기계다. 다만 앰버보다 훨씬 크게 표현되어 키 작고 땅딸막한 성인 남성 정도라는 체격 설정에서 멀어진다. 눈은 파란색이 아닌 주황색이며 가슴 상표도 식별되지 않는다. 주크박스는 보이지만 음악 재생 여부는 정지 화면으로 확인할 수 없다. 조명은 절제된 황혼광보다 따뜻한 주황색이 강하다.",
    "hard_violations": [],
    "physics": "앰버는 화면 오른쪽 부츠에 체중을 싣고 반대쪽 무릎을 접어 발을 들었다. 찰리는 화면 오른쪽의 넓은 기계 발을 바닥에 붙이고 반대쪽 무릎과 발을 들어 올렸다. 서로 맞잡은 손과 바깥으로 뻗은 팔이 회전 중 균형을 보조한다. 두 인물 모두 지지 발이 분명하므로 떠 있는 몸은 아니다. 모자는 머리에 얹혀 있고 코트와 공구 주머니도 정상적으로 지지된다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "주변 무용수와 작은 찰리의 체격은 더 충실하지만, 찰리가 높이 든 맞잡은 손을 거의 수평으로 바라보아 두 인물 모두 손을 내려다보라는 핵심 시선 지시에서 벗어난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "전신 여백과 비스듬한 저각도에서 두 인물이 맞잡은 손을 내려다보는 동작은 더 정확하지만, 주변 무용수가 없고 찰리의 체격이 지나치게 크다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 오른쪽 아래의 맞잡은 손 쪽으로 얼굴과 눈을 향한다. 찰리는 왼쪽의 앰버와 손 쪽을 향하지만, 손이 눈높이 가까이 올라와 있어 시선이 거의 수평이다. 두 사람이 함께 손을 내려다보는 구도는 아니다. 주변 사람들은 서로 다른 방향으로 몸을 돌리고 있다.",
        "built_space": "왼쪽에 주크박스 한 대와 음향 장비가 있고, 뒤에는 여러 지주에 걸친 천막, 전구 줄, 삼각 깃발, 컨테이너 주거지가 보인다. 참조 장소의 주요 재료와 배치를 유지한다. 앰버는 왼쪽 중앙, 찰리는 오른쪽 중앙이며 머리 위와 발 아래에 여백이 있다. 주변 무용수들의 간격과 팔 각도는 불규칙하고 두 주인공의 발밑 공간은 확보되어 있다. 천막은 부분적인 측면보다는 화면 상부 전반을 덮는다.",
        "entities": "앰버는 금발의 어린 여자아이로, 참조와 유사한 얼굴과 카키 작업복, 방진 마스크, 가죽 공구 주머니를 갖췄다. 눈매에서 즐거운 표정이 읽힌다. 찰리는 짧고 육중한 베이지색 기계 몸체, 흰 얼굴판, 두 점 형태의 눈과 선 형태의 입, 파란 가슴 원자로, 낡은 코트와 모자를 갖췄다. 다만 눈은 요구된 파란색이 아니라 주황색이고, 요구된 가슴 상표는 식별되지 않는다. 배경 인물들은 명시된 주변 무용수로 읽힌다. 전체 조명은 요구보다 강한 주황색 기운을 띤다.",
        "hard_violations": [],
        "physics": "앰버는 화면 오른쪽 부츠로 바닥을 딛고 다른 다리를 접어 들었다. 찰리도 화면 오른쪽 발로 지지하며 반대쪽 발을 앞으로 들어 올렸다. 두 인물의 손은 실제로 맞닿아 있고, 자유로운 팔은 균형을 잡는 방향으로 벌어져 있다. 공중에 든 다리는 몸통과 지지 다리에 연결되어 있으며 지지 없는 부유는 없다. 코트와 주머니도 몸에 걸리거나 벨트에 부착되어 있다."
       },
       {
        "label": "B",
        "direction": "앰버는 오른쪽 아래의 맞잡은 손을 바라보고, 찰리 역시 머리를 왼쪽 아래로 숙여 같은 손 접촉점을 향한다. 찰리의 몸통도 앰버 쪽으로 돌아 있다. 두 인물의 시선 목표가 명확하며, 손을 내려다보면서 서로 다른 회전 단계에 있다는 지시에 더 가깝다.",
        "built_space": "왼쪽에 주크박스 한 대, 그 주변에 두 개의 높은 스피커와 장비 상자가 보인다. 천막 지주, 전구 줄, 깃발, 뒤편 컨테이너 건물과 오른쪽의 작은 천막 공간이 참조 장소와 대응한다. 두 인물은 지정된 좌우 중앙에 있고 전신과 위아래 여백이 확보되어 있다. 비스듬한 낮은 시점도 읽힌다. 그러나 공터가 완전히 비어 있어 명시된 주변 무용수와 불규칙한 군중 배치는 누락되었다.",
        "entities": "앰버는 금발의 어린 여자아이이며 참조와 비슷한 얼굴, 카키 작업복, 방진 마스크와 허리 공구 주머니를 유지한다. 찰리는 흰 얼굴판과 단순한 눈·입, 베이지 장갑판, 파란 원자로, 코트와 모자를 갖춘 기계다. 다만 앰버보다 훨씬 크게 표현되어 키 작고 땅딸막한 성인 남성 정도라는 체격 설정에서 멀어진다. 눈은 파란색이 아닌 주황색이며 가슴 상표도 식별되지 않는다. 주크박스는 보이지만 음악 재생 여부는 정지 화면으로 확인할 수 없다. 조명은 절제된 황혼광보다 따뜻한 주황색이 강하다.",
        "hard_violations": [],
        "physics": "앰버는 화면 오른쪽 부츠에 체중을 싣고 반대쪽 무릎을 접어 발을 들었다. 찰리는 화면 오른쪽의 넓은 기계 발을 바닥에 붙이고 반대쪽 무릎과 발을 들어 올렸다. 서로 맞잡은 손과 바깥으로 뻗은 팔이 회전 중 균형을 보조한다. 두 인물 모두 지지 발이 분명하므로 떠 있는 몸은 아니다. 모자는 머리에 얹혀 있고 코트와 공구 주머니도 정상적으로 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "주변 무용수와 작은 찰리의 체격은 더 충실하지만, 찰리가 높이 든 맞잡은 손을 거의 수평으로 바라보아 두 인물 모두 손을 내려다보라는 핵심 시선 지시에서 벗어난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "전신 여백과 비스듬한 저각도에서 두 인물이 맞잡은 손을 내려다보는 동작은 더 정확하지만, 주변 무용수가 없고 찰리의 체격이 지나치게 크다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 오른쪽 아래의 맞잡은 손 쪽으로 얼굴과 눈을 향한다. 찰리는 왼쪽의 앰버와 손 쪽을 향하지만, 손이 눈높이 가까이 올라와 있어 시선이 거의 수평이다. 두 사람이 함께 손을 내려다보는 구도는 아니다. 주변 사람들은 서로 다른 방향으로 몸을 돌리고 있다.",
        "built_space": "왼쪽에 주크박스 한 대와 음향 장비가 있고, 뒤에는 여러 지주에 걸친 천막, 전구 줄, 삼각 깃발, 컨테이너 주거지가 보인다. 참조 장소의 주요 재료와 배치를 유지한다. 앰버는 왼쪽 중앙, 찰리는 오른쪽 중앙이며 머리 위와 발 아래에 여백이 있다. 주변 무용수들의 간격과 팔 각도는 불규칙하고 두 주인공의 발밑 공간은 확보되어 있다. 천막은 부분적인 측면보다는 화면 상부 전반을 덮는다.",
        "entities": "앰버는 금발의 어린 여자아이로, 참조와 유사한 얼굴과 카키 작업복, 방진 마스크, 가죽 공구 주머니를 갖췄다. 눈매에서 즐거운 표정이 읽힌다. 찰리는 짧고 육중한 베이지색 기계 몸체, 흰 얼굴판, 두 점 형태의 눈과 선 형태의 입, 파란 가슴 원자로, 낡은 코트와 모자를 갖췄다. 다만 눈은 요구된 파란색이 아니라 주황색이고, 요구된 가슴 상표는 식별되지 않는다. 배경 인물들은 명시된 주변 무용수로 읽힌다. 전체 조명은 요구보다 강한 주황색 기운을 띤다.",
        "hard_violations": [],
        "physics": "앰버는 화면 오른쪽 부츠로 바닥을 딛고 다른 다리를 접어 들었다. 찰리도 화면 오른쪽 발로 지지하며 반대쪽 발을 앞으로 들어 올렸다. 두 인물의 손은 실제로 맞닿아 있고, 자유로운 팔은 균형을 잡는 방향으로 벌어져 있다. 공중에 든 다리는 몸통과 지지 다리에 연결되어 있으며 지지 없는 부유는 없다. 코트와 주머니도 몸에 걸리거나 벨트에 부착되어 있다."
       },
       {
        "label": "A",
        "direction": "앰버는 오른쪽 아래의 맞잡은 손을 바라보고, 찰리 역시 머리를 왼쪽 아래로 숙여 같은 손 접촉점을 향한다. 찰리의 몸통도 앰버 쪽으로 돌아 있다. 두 인물의 시선 목표가 명확하며, 손을 내려다보면서 서로 다른 회전 단계에 있다는 지시에 더 가깝다.",
        "built_space": "왼쪽에 주크박스 한 대, 그 주변에 두 개의 높은 스피커와 장비 상자가 보인다. 천막 지주, 전구 줄, 깃발, 뒤편 컨테이너 건물과 오른쪽의 작은 천막 공간이 참조 장소와 대응한다. 두 인물은 지정된 좌우 중앙에 있고 전신과 위아래 여백이 확보되어 있다. 비스듬한 낮은 시점도 읽힌다. 그러나 공터가 완전히 비어 있어 명시된 주변 무용수와 불규칙한 군중 배치는 누락되었다.",
        "entities": "앰버는 금발의 어린 여자아이이며 참조와 비슷한 얼굴, 카키 작업복, 방진 마스크와 허리 공구 주머니를 유지한다. 찰리는 흰 얼굴판과 단순한 눈·입, 베이지 장갑판, 파란 원자로, 코트와 모자를 갖춘 기계다. 다만 앰버보다 훨씬 크게 표현되어 키 작고 땅딸막한 성인 남성 정도라는 체격 설정에서 멀어진다. 눈은 파란색이 아닌 주황색이며 가슴 상표도 식별되지 않는다. 주크박스는 보이지만 음악 재생 여부는 정지 화면으로 확인할 수 없다. 조명은 절제된 황혼광보다 따뜻한 주황색이 강하다.",
        "hard_violations": [],
        "physics": "앰버는 화면 오른쪽 부츠에 체중을 싣고 반대쪽 무릎을 접어 발을 들었다. 찰리는 화면 오른쪽의 넓은 기계 발을 바닥에 붙이고 반대쪽 무릎과 발을 들어 올렸다. 서로 맞잡은 손과 바깥으로 뻗은 팔이 회전 중 균형을 보조한다. 두 인물 모두 지지 발이 분명하므로 떠 있는 몸은 아니다. 모자는 머리에 얹혀 있고 코트와 공구 주머니도 정상적으로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 6,
   "A": 7
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "주변 무용수와 작은 찰리의 체격은 더 충실하지만, 찰리가 높이 든 맞잡은 손을 거의 수평으로 바라보아 두 인물 모두 손을 내려다보라는 핵심 시선 지시에서 벗어난다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "전신 여백과 비스듬한 저각도에서 두 인물이 맞잡은 손을 내려다보는 동작은 더 정확하지만, 주변 무용수가 없고 찰리의 체격이 지나치게 크다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reception_clearing_3488b1.png",
    "asset_id": "296e6be3-728c-4aba-9a0e-c68a9359e450",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0964-c7f3-7ce6-aa15-de80aa58cd0d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh11__bgfirst_bg.png",
   "bg_asset_id": "01187f03-8a03-4f0b-81e3-25b6ad120734",
   "bg_record_key": "S25sh11::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "reception_clearing",
   "groupbg_asset_id": "296e6be3-728c-4aba-9a0e-c68a9359e450"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S25sh17::signage": {
  "fp": "5a6426b05a90dc30",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S25sh17": {
  "input_fingerprint": "73c0a328c60b8137",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 찰리의 낡은 금속 가슴 중앙에서 둥근 링 모양의 강렬한 빛이 뿜어져 나오는 근접 찰나.\n\nLOCATION (lock): In the darkened outdoor reception clearing, close to the silent jukebox after the settlement's lights have gone out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inherited diagonal approach at 찰리's chest height, keeping the oblique angle and a level view rather than squaring up to his torso. Crop out his face and 앰버, placing the emerging ring at center within the worn metal chest, with adjoining torso contours providing scale and a narrow margin of the darkened clearing still visible. 찰리 remains motionless with his unseen head oriented toward the off-screen jukebox; the emphasis is the final reduction in camera distance, not an additional pose change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사이렌 이후 사람들이 춤을 멈춘 장소); used as Only a narrow, unresolved margin remains around Charlie's torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The emerging ring emits intense light into the blackout, with controlled highlight bloom preserving its circular shape and the adjacent metal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The streetlights have gone out, leaving the reception dark, and the jukebox has stopped. Charlie's chest ring begins emitting light; his old coat and hat, worn chest logo and blue-lit eyes remain established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 찰리의 낡은 금속 가슴 중앙에서 둥근 링 모양의 강렬한 빛이 뿜어져 나오는 근접 찰나.\n\nLOCATION (lock): In the darkened outdoor reception clearing, close to the silent jukebox after the settlement's lights have gone out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inherited diagonal approach at 찰리's chest height, keeping the oblique angle and a level view rather than squaring up to his torso. Crop out his face and 앰버, placing the emerging ring at center within the worn metal chest, with adjoining torso contours providing scale and a narrow margin of the darkened clearing still visible. 찰리 remains motionless with his unseen head oriented toward the off-screen jukebox; the emphasis is the final reduction in camera distance, not an additional pose change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사이렌 이후 사람들이 춤을 멈춘 장소); used as Only a narrow, unresolved margin remains around Charlie's torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The emerging ring emits intense light into the blackout, with controlled highlight bloom preserving its circular shape and the adjacent metal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The streetlights have gone out, leaving the reception dark, and the jukebox has stopped. Charlie's chest ring begins emitting light; his old coat and hat, worn chest logo and blue-lit eyes remain established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 찰리의 낡은 금속 가슴 중앙에서 둥근 링 모양의 강렬한 빛이 뿜어져 나오는 근접 찰나.\n\nLOCATION (lock): In the darkened outdoor reception clearing, close to the silent jukebox after the settlement's lights have gone out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inherited diagonal approach at 찰리's chest height, keeping the oblique angle and a level view rather than squaring up to his torso. Crop out his face and 앰버, placing the emerging ring at center within the worn metal chest, with adjoining torso contours providing scale and a narrow margin of the darkened clearing still visible. 찰리 remains motionless with his unseen head oriented toward the off-screen jukebox; the emphasis is the final reduction in camera distance, not an additional pose change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 피로연 공터 (사이렌 이후 사람들이 춤을 멈춘 장소); used as Only a narrow, unresolved margin remains around Charlie's torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The emerging ring emits intense light into the blackout, with controlled highlight bloom preserving its circular shape and the adjacent metal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The streetlights have gone out, leaving the reception dark, and the jukebox has stopped. Charlie's chest ring begins emitting light; his old coat and hat, worn chest logo and blue-lit eyes remain established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
    "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스와 텐트 기둥들이 배치되어 있음.",
    "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱 끝이 미세하게 보임.",
    "hard_violations": [],
    "physics": "어깨의 형태가 갈색 코트를 자연스럽게 지탱하고 있으며 직립한 자세를 유지함."
   },
   {
    "label": "B",
    "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
    "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스가 배치되어 있음.",
    "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱과 하관이 확연히 드러남.",
    "hard_violations": [],
    "physics": "금속 어깨 위로 코트가 걸쳐져 있으며 몸통이 안정적으로 자세를 지탱함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴을 완전히 크롭하지 못했고 화면 밖에 있어야 할 쥬크박스가 켜진 채로 배경에 나타나지만, B에 비해 레퍼런스의 흉부 장갑 형태에 조금 더 가깝습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "얼굴의 하단이 크게 노출되어 크롭 지시를 어겼고, 꺼져 있어야 할 쥬크박스가 불이 켜진 채 화면에 등장하며 이전 샷의 팔 포즈가 유지되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
        "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스와 텐트 기둥들이 배치되어 있음.",
        "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱 끝이 미세하게 보임.",
        "hard_violations": [],
        "physics": "어깨의 형태가 갈색 코트를 자연스럽게 지탱하고 있으며 직립한 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
        "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스가 배치되어 있음.",
        "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱과 하관이 확연히 드러남.",
        "hard_violations": [],
        "physics": "금속 어깨 위로 코트가 걸쳐져 있으며 몸통이 안정적으로 자세를 지탱함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴을 완전히 크롭하지 못했고 화면 밖에 있어야 할 쥬크박스가 켜진 채로 배경에 나타나지만, B에 비해 레퍼런스의 흉부 장갑 형태에 조금 더 가깝습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "얼굴의 하단이 크게 노출되어 크롭 지시를 어겼고, 꺼져 있어야 할 쥬크박스가 불이 켜진 채 화면에 등장하며 이전 샷의 팔 포즈가 유지되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
        "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스와 텐트 기둥들이 배치되어 있음.",
        "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱 끝이 미세하게 보임.",
        "hard_violations": [],
        "physics": "어깨의 형태가 갈색 코트를 자연스럽게 지탱하고 있으며 직립한 자세를 유지함."
       },
       {
        "label": "B",
        "direction": "카메라가 찰리의 가슴을 사선으로 바라보며 가슴 중앙의 푸른 빛을 발하는 링을 향함.",
        "built_space": "어두워진 야외 공터로, 왼쪽 배경에 조명이 켜진 쥬크박스가 배치되어 있음.",
        "entities": "샌드 베이지색 금속 흉부와 갈색 코트를 입은 찰리. 중앙의 푸른 원자로 링이 있으며 화면 상단에 흰색 마스크의 턱과 하관이 확연히 드러남.",
        "hard_violations": [],
        "physics": "금속 어깨 위로 코트가 걸쳐져 있으며 몸통이 안정적으로 자세를 지탱함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낡은 가슴과 코트의 재질은 충실하지만, 링이 중앙에서 더 벗어나고 공터가 넓고 선명하게 드러나며 정전된 배경도 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "비스듬한 가슴 근접 시점과 중앙에 가까운 선명한 발광 링이 더 적합하지만, 얼굴 하단 노출과 넓고 불이 남은 배경은 지시와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "가슴을 수평에 가까운 비스듬한 방향에서 본다. 링은 화면 중앙보다 오른쪽 위에 있으며 가슴 밖으로 푸른 빛을 낸다. 눈은 보이지 않아 화면 왼쪽 뒤의 주크박스를 향한 시선은 확인할 수 없다. 무기나 겨냥하는 물체는 없다.",
        "built_space": "왼쪽 배경에 주크박스 한 대, 겹쳐 보이는 천막 지붕과 지지 기둥, 삼각 깃발, 장비 상자와 젖은 바닥이 보인다. 이전 장면의 공터 재료와 배치는 대체로 이어지지만, 배경이 좁고 미해결된 여백이 아니라 화면 왼쪽의 상당 부분을 차지한다. 주크박스의 주황색 조명이 켜져 있어 정전 분위기가 불완전하다. 바닥의 빛 반사는 보이는 광원으로 설명된다.",
        "entities": "찰리의 샌드 베이지 금속 흉갑, 마모와 긁힘, 갈색 낡은 코트, 푸른 원자로가 보이며 인간 피부나 다른 사람은 없다. 흉갑의 세부 분할은 참고와 조금 다르다. 원자로에는 밝은 외곽 링과 내부 소용돌이가 함께 보인다. 상단에 흰 얼굴판의 아래쪽 일부가 남아 있어 얼굴을 완전히 제외하지는 않았다. 모자와 눈은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "원자로는 금속 테두리와 흉갑 안에 고정되어 있고 팔은 관절로 몸통에 연결되어 있다. 코트는 어깨와 몸통에 걸쳐 자연스럽게 내려온다. 푸른 빛이 인접 금속과 천에 반사되며 떠 있는 물체는 없다. 발과 지면 접촉은 근접 구도 밖이므로 지지 상태나 자세 변화를 단정할 수 없다."
       },
       {
        "label": "B",
        "direction": "흉갑의 좌우 면과 링의 원근 차이가 드러나는 비스듬한 가슴 높이 시점이다. 발광 링은 화면 중앙에서 약간 오른쪽 위에 있지만 A보다 중앙에 가깝다. 얼굴 하단은 보이나 눈은 잘려 있어 왼쪽 뒤 주크박스를 향하는 시선은 검증할 수 없다. 겨냥하거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤에 주크박스 한 대, 천막 지붕과 기둥들, 삼각 깃발, 검은 장비와 젖은 공터 바닥이 보인다. 참고 장소의 주요 재료와 시설은 이어지며 중복 주크박스는 없다. 다만 배경 공간과 시설이 상당히 식별되어 좁고 흐릿한 여백이라는 요구에는 못 미친다. 주크박스와 뒤쪽의 따뜻한 점광원이 켜져 있어 정전 조건과 다르다. 바닥 반사는 광원과 모순되지 않는다.",
        "entities": "찰리 한 개체의 낡은 베이지 장갑판, 육중한 팔 일부, 갈색 코트와 푸른 가슴 원자로가 보인다. 링의 경계와 주변 금속 디테일이 강한 발광 속에서도 유지된다. 흉갑 분할과 원자로 테두리의 세부는 참고와 완전히 같지는 않다. 흰 마스크 하단이 A보다 더 드러나 얼굴 제외 지시를 완전히 지키지 못했다. 앰버나 다른 사람, 새 문구는 없다.",
        "hard_violations": [],
        "physics": "발광 장치는 볼트와 테두리가 있는 가슴 소켓에 장착되어 있고, 팔의 금속 부품은 관절에 연결되어 있다. 코트는 몸통에 지지되어 접히며 발광에 따른 푸른 반사가 주변 표면에 나타난다. 지지 없이 떠 있는 물체나 불가능한 신체 연결은 보이지 않는다. 하체가 잘려 있어 발의 접촉이나 전신 정지 자세는 평가할 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낡은 가슴과 코트의 재질은 충실하지만, 링이 중앙에서 더 벗어나고 공터가 넓고 선명하게 드러나며 정전된 배경도 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "비스듬한 가슴 근접 시점과 중앙에 가까운 선명한 발광 링이 더 적합하지만, 얼굴 하단 노출과 넓고 불이 남은 배경은 지시와 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "가슴을 수평에 가까운 비스듬한 방향에서 본다. 링은 화면 중앙보다 오른쪽 위에 있으며 가슴 밖으로 푸른 빛을 낸다. 눈은 보이지 않아 화면 왼쪽 뒤의 주크박스를 향한 시선은 확인할 수 없다. 무기나 겨냥하는 물체는 없다.",
        "built_space": "왼쪽 배경에 주크박스 한 대, 겹쳐 보이는 천막 지붕과 지지 기둥, 삼각 깃발, 장비 상자와 젖은 바닥이 보인다. 이전 장면의 공터 재료와 배치는 대체로 이어지지만, 배경이 좁고 미해결된 여백이 아니라 화면 왼쪽의 상당 부분을 차지한다. 주크박스의 주황색 조명이 켜져 있어 정전 분위기가 불완전하다. 바닥의 빛 반사는 보이는 광원으로 설명된다.",
        "entities": "찰리의 샌드 베이지 금속 흉갑, 마모와 긁힘, 갈색 낡은 코트, 푸른 원자로가 보이며 인간 피부나 다른 사람은 없다. 흉갑의 세부 분할은 참고와 조금 다르다. 원자로에는 밝은 외곽 링과 내부 소용돌이가 함께 보인다. 상단에 흰 얼굴판의 아래쪽 일부가 남아 있어 얼굴을 완전히 제외하지는 않았다. 모자와 눈은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "원자로는 금속 테두리와 흉갑 안에 고정되어 있고 팔은 관절로 몸통에 연결되어 있다. 코트는 어깨와 몸통에 걸쳐 자연스럽게 내려온다. 푸른 빛이 인접 금속과 천에 반사되며 떠 있는 물체는 없다. 발과 지면 접촉은 근접 구도 밖이므로 지지 상태나 자세 변화를 단정할 수 없다."
       },
       {
        "label": "A",
        "direction": "흉갑의 좌우 면과 링의 원근 차이가 드러나는 비스듬한 가슴 높이 시점이다. 발광 링은 화면 중앙에서 약간 오른쪽 위에 있지만 A보다 중앙에 가깝다. 얼굴 하단은 보이나 눈은 잘려 있어 왼쪽 뒤 주크박스를 향하는 시선은 검증할 수 없다. 겨냥하거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤에 주크박스 한 대, 천막 지붕과 기둥들, 삼각 깃발, 검은 장비와 젖은 공터 바닥이 보인다. 참고 장소의 주요 재료와 시설은 이어지며 중복 주크박스는 없다. 다만 배경 공간과 시설이 상당히 식별되어 좁고 흐릿한 여백이라는 요구에는 못 미친다. 주크박스와 뒤쪽의 따뜻한 점광원이 켜져 있어 정전 조건과 다르다. 바닥 반사는 광원과 모순되지 않는다.",
        "entities": "찰리 한 개체의 낡은 베이지 장갑판, 육중한 팔 일부, 갈색 코트와 푸른 가슴 원자로가 보인다. 링의 경계와 주변 금속 디테일이 강한 발광 속에서도 유지된다. 흉갑 분할과 원자로 테두리의 세부는 참고와 완전히 같지는 않다. 흰 마스크 하단이 A보다 더 드러나 얼굴 제외 지시를 완전히 지키지 못했다. 앰버나 다른 사람, 새 문구는 없다.",
        "hard_violations": [],
        "physics": "발광 장치는 볼트와 테두리가 있는 가슴 소켓에 장착되어 있고, 팔의 금속 부품은 관절에 연결되어 있다. 코트는 몸통에 지지되어 접히며 발광에 따른 푸른 반사가 주변 표면에 나타난다. 지지 없이 떠 있는 물체나 불가능한 신체 연결은 보이지 않는다. 하체가 잘려 있어 발의 접촉이나 전신 정지 자세는 평가할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.657
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.657
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1657
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "얼굴을 완전히 크롭하지 못했고 화면 밖에 있어야 할 쥬크박스가 켜진 채로 배경에 나타나지만, B에 비해 레퍼런스의 흉부 장갑 형태에 조금 더 가깝습니다."
   },
   {
    "label": "B",
    "score": 1657,
    "verdict_ko": "얼굴의 하단이 크게 노출되어 크롭 지시를 어겼고, 꺼져 있어야 할 쥬크박스가 불이 켜진 채 화면에 등장하며 이전 샷의 팔 포즈가 유지되지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh11_sel.png",
    "asset_id": "f4ac48ab-2343-4e98-aed8-83f95f53aaf0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab096d-be15-7c85-9a86-83ca860d1604",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh11"
  }
 },
 "S25sh28::signage": {
  "fp": "530c8aee2188d1d1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S25sh28": {
  "input_fingerprint": "267cb3c588f41076",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 혼비백산하여 입을 벌린 군중들 틈바구니로 이현우가 앰버와 찰리의 팔을 꽉 잡은 채 몸을 앞으로 기울이고 전력 질주하는 mid-stride 역동적인 전신.\n\nLOCATION (lock): Among the scattering guests in the outdoor reception clearing, on the escape route away from the militia trucks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track backward at the runners' speed from ahead and to the side of their escape path, maintaining the inherited waist-height, three-quarter view with all three bodies and feet visible. 이현우 leads diagonally toward the lower-left opening, leaning into his stride at center-left while gripping 앰버's arm to his left and 찰리's arm to his right; keep both hand-to-arm contacts separated in silhouette, with all three looking along their escape route rather than toward the lens. Frightened, open-mouthed people break around the frame edges in mismatched strides and head turns, leaving the trio's route and grips unobscured.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: clear escape lane between scattering people in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 피로연 공터의 탈출 틈 (놀란 사람들이 흩어지며 일행이 빠져나갈 틈이 생김); used as An unobstructed diagonal lane leads toward the lower left; edge figures differ in footfall, lean, spacing and gaze rather than repeating one panic pose.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The restored reception lighting keeps the escape clearly readable, with restrained highlights and controlled contrast rather than an additional action-lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding streets are brightly relit, some lamps have burst, the jukebox has resumed, and a militia truck has arrived. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is escaping through the reception in his outer top, still carrying facial bruises and the leg injury. 앰버: She is leaving in the escape, retaining her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 혼비백산하여 입을 벌린 군중들 틈바구니로 이현우가 앰버와 찰리의 팔을 꽉 잡은 채 몸을 앞으로 기울이고 전력 질주하는 mid-stride 역동적인 전신.\n\nLOCATION (lock): Among the scattering guests in the outdoor reception clearing, on the escape route away from the militia trucks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track backward at the runners' speed from ahead and to the side of their escape path, maintaining the inherited waist-height, three-quarter view with all three bodies and feet visible. 이현우 leads diagonally toward the lower-left opening, leaning into his stride at center-left while gripping 앰버's arm to his left and 찰리's arm to his right; keep both hand-to-arm contacts separated in silhouette, with all three looking along their escape route rather than toward the lens. Frightened, open-mouthed people break around the frame edges in mismatched strides and head turns, leaving the trio's route and grips unobscured.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: clear escape lane between scattering people in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 피로연 공터의 탈출 틈 (놀란 사람들이 흩어지며 일행이 빠져나갈 틈이 생김); used as An unobstructed diagonal lane leads toward the lower left; edge figures differ in footfall, lean, spacing and gaze rather than repeating one panic pose.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The restored reception lighting keeps the escape clearly readable, with restrained highlights and controlled contrast rather than an additional action-lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding streets are brightly relit, some lamps have burst, the jukebox has resumed, and a militia truck has arrived. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is escaping through the reception in his outer top, still carrying facial bruises and the leg injury. 앰버: She is leaving in the escape, retaining her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night.\n\nSHOT TEXT (authoritative, Korean): 혼비백산하여 입을 벌린 군중들 틈바구니로 이현우가 앰버와 찰리의 팔을 꽉 잡은 채 몸을 앞으로 기울이고 전력 질주하는 mid-stride 역동적인 전신.\n\nLOCATION (lock): Among the scattering guests in the outdoor reception clearing, on the escape route away from the militia trucks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track backward at the runners' speed from ahead and to the side of their escape path, maintaining the inherited waist-height, three-quarter view with all three bodies and feet visible. 이현우 leads diagonally toward the lower-left opening, leaning into his stride at center-left while gripping 앰버's arm to his left and 찰리's arm to his right; keep both hand-to-arm contacts separated in silhouette, with all three looking along their escape route rather than toward the lens. Frightened, open-mouthed people break around the frame edges in mismatched strides and head turns, leaving the trio's route and grips unobscured.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: clear escape lane between scattering people in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 피로연 공터의 탈출 틈 (놀란 사람들이 흩어지며 일행이 빠져나갈 틈이 생김); used as An unobstructed diagonal lane leads toward the lower left; edge figures differ in footfall, lean, spacing and gaze rather than repeating one panic pose.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The restored reception lighting keeps the escape clearly readable, with restrained highlights and controlled contrast rather than an additional action-lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception and surrounding streets are brightly relit, some lamps have burst, the jukebox has resumed, and a militia truck has arrived. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is escaping through the reception in his outer top, still carrying facial bruises and the leg injury. 앰버: She is leaving in the escape, retaining her mask and waist tool pouch.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "현우와 앰버는 화면 왼쪽 전방을 바라보고 달리며, 찰리의 얼굴도 왼쪽으로 돌아가 있다. 렌즈를 직접 바라보는 주역은 없다. 다만 몸과 발의 진행은 좌하단 대각선보다 카메라 정면에 가깝다. 가장자리 군중은 좌우로 흩어지며 서로 다른 방향으로 고개를 돌린다.",
    "built_space": "젖은 공터 뒤로 천막과 지지 기둥들, 전구 줄, 삼각 깃발, 여러 피로연 탁자와 의자가 이어진다. 왼쪽 배경에 주크박스 한 대, 오른쪽 배경에 천막을 씌운 트럭 한 대가 보인다. 이전 장면의 젖은 바닥과 따뜻한 실용 조명을 대체로 유지한다. 앰버는 왼쪽, 현우는 중앙 왼쪽, 찰리는 오른쪽에 있고 세 몸과 발이 모두 들어온다. 좌하단 모서리에 흐릿한 군중 일부가 있으나 그 안쪽 탈출 공간은 비교적 열려 있다.",
    "entities": "현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 어두운 오염 셔츠와 바지, 얼굴 상처 및 찢어진 무릎이 보인다. 인이어는 명확하지 않다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니를 유지한다. 찰리는 짧고 육중한 베이지 기계 몸체와 흰 마스크 얼굴, 갈색 외투와 모자, 푸른 가슴 원자로를 갖췄지만 눈은 지시된 파란색이 아니라 주황색이다. 현우와 앰버의 손은 서로 겹쳐 잡혀 있고, 현우의 반대 손은 찰리의 기계 손을 잡아 두 팔을 붙드는 지시와 다르다. 입을 벌린 군중 외에 트럭 주변의 무장·헬멧 인물도 보인다.",
    "hard_violations": [
     "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 주변에 총기를 든 헬멧 착용 무장 인물을 추가했다."
    ],
    "physics": "현우와 앰버는 앞쪽 발을 바닥에 내딛고 뒷다리를 접었으며, 찰리도 한 발을 딛고 반대 발을 들어 달리는 형태다. 발과 지면, 굽힌 무릎이 체중 지지와 다음 보폭을 설명한다. 양쪽 손 접촉은 서로 떨어져 보이고 물리적으로 연결되지만, 팔을 꽉 잡는 동작은 아니다. 외투와 머리카락의 움직임도 달리기로 설명되며 근거 없이 떠 있는 몸은 없다."
   },
   {
    "label": "A",
    "direction": "현우와 앰버는 왼쪽 전방의 탈출 방향을 바라보고, 찰리도 얼굴을 왼쪽으로 돌린다. 현우는 A보다 상체를 더 깊게 숙여 전방 추진이 강하다. 그러나 세 인물의 보폭은 여전히 정면 성분이 크다. 왼쪽 전경 인물의 다리가 좌하단 탈출 구역을 가로지르고, 오른쪽 전경 인물은 화면 오른쪽 바깥을 바라본다.",
    "built_space": "젖은 야외 공터, 천막과 여러 기둥, 전구 줄과 삼각 깃발이 이어진다. 왼쪽에 주크박스 한 대와 검은 스피커 한 대, 오른쪽 뒤에 적재함을 보이는 트럭 한 대가 있다. 노을과 따뜻한 피로연 조명은 장소의 연속성을 유지한다. 세 주역의 전신과 발은 보이지만 발끝이 아래 테두리에 더 가까우며, 좌하단을 가로지르는 군중의 다리가 지정된 비어 있는 탈출로를 훼손한다. 양쪽 손 접촉 자체는 군중에 가려지지 않는다.",
    "entities": "현우는 젊은 동아시아계 남성의 얼굴과 검은 머리, 더러운 어두운 셔츠와 바지, 뺨 상처 및 인이어를 갖췄다. 다리 부상은 A보다 분명하지 않다. 앰버의 어린 얼굴, 금발, 착용한 방진 마스크, 카키 작업복과 가죽 공구 주머니는 유지된다. 찰리는 흰 기계 얼굴, 베이지 장갑판, 갈색 외투와 챙 모자, 푸른 가슴 원자로를 유지하지만 눈은 주황색이다. 현우는 앰버 및 찰리와 손을 맞잡은 형태이며 양쪽 팔을 붙들지 않는다. 군중 외에 트럭 적재함의 헬멧 착용 무장 인물들이 뚜렷하다.",
    "hard_violations": [
     "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 적재함 등에 헬멧 착용 무장 인물들을 추가했다."
    ],
    "physics": "현우와 앰버는 앞쪽 발의 착지와 뒤로 접힌 다리로 질주를 표현한다. 찰리는 한쪽 기계 발을 앞으로 내밀고 반대쪽을 뒤로 들어, 도약과 착지 사이의 짧은 달리기 국면으로 읽힌다. 무릎과 몸의 전진 자세가 그 움직임을 뒷받침하므로 무근거한 부유는 아니다. 두 손 연결도 물리적으로 가능하며, 마스크와 공구 주머니는 얼굴과 허리 끈에 고정되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "세 인물의 발까지 담고 좌하단 탈출 공간을 비교적 잘 확보했지만, 허용되지 않은 무장 인물과 팔 대신 손을 잡는 동작, 찰리의 주황색 눈이 지시를 어긴다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "현우의 전경사와 질주 자세는 더 강하지만, 좌하단을 가로지르는 군중이 지정 탈출로를 막으며 무장 인물 추가와 손잡기·눈 색상 오류도 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우와 앰버는 화면 왼쪽 전방을 바라보고 달리며, 찰리의 얼굴도 왼쪽으로 돌아가 있다. 렌즈를 직접 바라보는 주역은 없다. 다만 몸과 발의 진행은 좌하단 대각선보다 카메라 정면에 가깝다. 가장자리 군중은 좌우로 흩어지며 서로 다른 방향으로 고개를 돌린다.",
        "built_space": "젖은 공터 뒤로 천막과 지지 기둥들, 전구 줄, 삼각 깃발, 여러 피로연 탁자와 의자가 이어진다. 왼쪽 배경에 주크박스 한 대, 오른쪽 배경에 천막을 씌운 트럭 한 대가 보인다. 이전 장면의 젖은 바닥과 따뜻한 실용 조명을 대체로 유지한다. 앰버는 왼쪽, 현우는 중앙 왼쪽, 찰리는 오른쪽에 있고 세 몸과 발이 모두 들어온다. 좌하단 모서리에 흐릿한 군중 일부가 있으나 그 안쪽 탈출 공간은 비교적 열려 있다.",
        "entities": "현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 어두운 오염 셔츠와 바지, 얼굴 상처 및 찢어진 무릎이 보인다. 인이어는 명확하지 않다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니를 유지한다. 찰리는 짧고 육중한 베이지 기계 몸체와 흰 마스크 얼굴, 갈색 외투와 모자, 푸른 가슴 원자로를 갖췄지만 눈은 지시된 파란색이 아니라 주황색이다. 현우와 앰버의 손은 서로 겹쳐 잡혀 있고, 현우의 반대 손은 찰리의 기계 손을 잡아 두 팔을 붙드는 지시와 다르다. 입을 벌린 군중 외에 트럭 주변의 무장·헬멧 인물도 보인다.",
        "hard_violations": [
         "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 주변에 총기를 든 헬멧 착용 무장 인물을 추가했다."
        ],
        "physics": "현우와 앰버는 앞쪽 발을 바닥에 내딛고 뒷다리를 접었으며, 찰리도 한 발을 딛고 반대 발을 들어 달리는 형태다. 발과 지면, 굽힌 무릎이 체중 지지와 다음 보폭을 설명한다. 양쪽 손 접촉은 서로 떨어져 보이고 물리적으로 연결되지만, 팔을 꽉 잡는 동작은 아니다. 외투와 머리카락의 움직임도 달리기로 설명되며 근거 없이 떠 있는 몸은 없다."
       },
       {
        "label": "B",
        "direction": "현우와 앰버는 왼쪽 전방의 탈출 방향을 바라보고, 찰리도 얼굴을 왼쪽으로 돌린다. 현우는 A보다 상체를 더 깊게 숙여 전방 추진이 강하다. 그러나 세 인물의 보폭은 여전히 정면 성분이 크다. 왼쪽 전경 인물의 다리가 좌하단 탈출 구역을 가로지르고, 오른쪽 전경 인물은 화면 오른쪽 바깥을 바라본다.",
        "built_space": "젖은 야외 공터, 천막과 여러 기둥, 전구 줄과 삼각 깃발이 이어진다. 왼쪽에 주크박스 한 대와 검은 스피커 한 대, 오른쪽 뒤에 적재함을 보이는 트럭 한 대가 있다. 노을과 따뜻한 피로연 조명은 장소의 연속성을 유지한다. 세 주역의 전신과 발은 보이지만 발끝이 아래 테두리에 더 가까우며, 좌하단을 가로지르는 군중의 다리가 지정된 비어 있는 탈출로를 훼손한다. 양쪽 손 접촉 자체는 군중에 가려지지 않는다.",
        "entities": "현우는 젊은 동아시아계 남성의 얼굴과 검은 머리, 더러운 어두운 셔츠와 바지, 뺨 상처 및 인이어를 갖췄다. 다리 부상은 A보다 분명하지 않다. 앰버의 어린 얼굴, 금발, 착용한 방진 마스크, 카키 작업복과 가죽 공구 주머니는 유지된다. 찰리는 흰 기계 얼굴, 베이지 장갑판, 갈색 외투와 챙 모자, 푸른 가슴 원자로를 유지하지만 눈은 주황색이다. 현우는 앰버 및 찰리와 손을 맞잡은 형태이며 양쪽 팔을 붙들지 않는다. 군중 외에 트럭 적재함의 헬멧 착용 무장 인물들이 뚜렷하다.",
        "hard_violations": [
         "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 적재함 등에 헬멧 착용 무장 인물들을 추가했다."
        ],
        "physics": "현우와 앰버는 앞쪽 발의 착지와 뒤로 접힌 다리로 질주를 표현한다. 찰리는 한쪽 기계 발을 앞으로 내밀고 반대쪽을 뒤로 들어, 도약과 착지 사이의 짧은 달리기 국면으로 읽힌다. 무릎과 몸의 전진 자세가 그 움직임을 뒷받침하므로 무근거한 부유는 아니다. 두 손 연결도 물리적으로 가능하며, 마스크와 공구 주머니는 얼굴과 허리 끈에 고정되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "세 인물의 발까지 담고 좌하단 탈출 공간을 비교적 잘 확보했지만, 허용되지 않은 무장 인물과 팔 대신 손을 잡는 동작, 찰리의 주황색 눈이 지시를 어긴다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "현우의 전경사와 질주 자세는 더 강하지만, 좌하단을 가로지르는 군중이 지정 탈출로를 막으며 무장 인물 추가와 손잡기·눈 색상 오류도 남는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우와 앰버는 화면 왼쪽 전방을 바라보고 달리며, 찰리의 얼굴도 왼쪽으로 돌아가 있다. 렌즈를 직접 바라보는 주역은 없다. 다만 몸과 발의 진행은 좌하단 대각선보다 카메라 정면에 가깝다. 가장자리 군중은 좌우로 흩어지며 서로 다른 방향으로 고개를 돌린다.",
        "built_space": "젖은 공터 뒤로 천막과 지지 기둥들, 전구 줄, 삼각 깃발, 여러 피로연 탁자와 의자가 이어진다. 왼쪽 배경에 주크박스 한 대, 오른쪽 배경에 천막을 씌운 트럭 한 대가 보인다. 이전 장면의 젖은 바닥과 따뜻한 실용 조명을 대체로 유지한다. 앰버는 왼쪽, 현우는 중앙 왼쪽, 찰리는 오른쪽에 있고 세 몸과 발이 모두 들어온다. 좌하단 모서리에 흐릿한 군중 일부가 있으나 그 안쪽 탈출 공간은 비교적 열려 있다.",
        "entities": "현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 어두운 오염 셔츠와 바지, 얼굴 상처 및 찢어진 무릎이 보인다. 인이어는 명확하지 않다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니를 유지한다. 찰리는 짧고 육중한 베이지 기계 몸체와 흰 마스크 얼굴, 갈색 외투와 모자, 푸른 가슴 원자로를 갖췄지만 눈은 지시된 파란색이 아니라 주황색이다. 현우와 앰버의 손은 서로 겹쳐 잡혀 있고, 현우의 반대 손은 찰리의 기계 손을 잡아 두 팔을 붙드는 지시와 다르다. 입을 벌린 군중 외에 트럭 주변의 무장·헬멧 인물도 보인다.",
        "hard_violations": [
         "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 주변에 총기를 든 헬멧 착용 무장 인물을 추가했다."
        ],
        "physics": "현우와 앰버는 앞쪽 발을 바닥에 내딛고 뒷다리를 접었으며, 찰리도 한 발을 딛고 반대 발을 들어 달리는 형태다. 발과 지면, 굽힌 무릎이 체중 지지와 다음 보폭을 설명한다. 양쪽 손 접촉은 서로 떨어져 보이고 물리적으로 연결되지만, 팔을 꽉 잡는 동작은 아니다. 외투와 머리카락의 움직임도 달리기로 설명되며 근거 없이 떠 있는 몸은 없다."
       },
       {
        "label": "A",
        "direction": "현우와 앰버는 왼쪽 전방의 탈출 방향을 바라보고, 찰리도 얼굴을 왼쪽으로 돌린다. 현우는 A보다 상체를 더 깊게 숙여 전방 추진이 강하다. 그러나 세 인물의 보폭은 여전히 정면 성분이 크다. 왼쪽 전경 인물의 다리가 좌하단 탈출 구역을 가로지르고, 오른쪽 전경 인물은 화면 오른쪽 바깥을 바라본다.",
        "built_space": "젖은 야외 공터, 천막과 여러 기둥, 전구 줄과 삼각 깃발이 이어진다. 왼쪽에 주크박스 한 대와 검은 스피커 한 대, 오른쪽 뒤에 적재함을 보이는 트럭 한 대가 있다. 노을과 따뜻한 피로연 조명은 장소의 연속성을 유지한다. 세 주역의 전신과 발은 보이지만 발끝이 아래 테두리에 더 가까우며, 좌하단을 가로지르는 군중의 다리가 지정된 비어 있는 탈출로를 훼손한다. 양쪽 손 접촉 자체는 군중에 가려지지 않는다.",
        "entities": "현우는 젊은 동아시아계 남성의 얼굴과 검은 머리, 더러운 어두운 셔츠와 바지, 뺨 상처 및 인이어를 갖췄다. 다리 부상은 A보다 분명하지 않다. 앰버의 어린 얼굴, 금발, 착용한 방진 마스크, 카키 작업복과 가죽 공구 주머니는 유지된다. 찰리는 흰 기계 얼굴, 베이지 장갑판, 갈색 외투와 챙 모자, 푸른 가슴 원자로를 유지하지만 눈은 주황색이다. 현우는 앰버 및 찰리와 손을 맞잡은 형태이며 양쪽 팔을 붙들지 않는다. 군중 외에 트럭 적재함의 헬멧 착용 무장 인물들이 뚜렷하다.",
        "hard_violations": [
         "숏에 허용된 세 주역과 피로연 군중 외에, 트럭 적재함 등에 헬멧 착용 무장 인물들을 추가했다."
        ],
        "physics": "현우와 앰버는 앞쪽 발의 착지와 뒤로 접힌 다리로 질주를 표현한다. 찰리는 한쪽 기계 발을 앞으로 내밀고 반대쪽을 뒤로 들어, 도약과 착지 사이의 짧은 달리기 국면으로 읽힌다. 무릎과 몸의 전진 자세가 그 움직임을 뒷받침하므로 무근거한 부유는 아니다. 두 손 연결도 물리적으로 가능하며, 마스크와 공구 주머니는 얼굴과 허리 끈에 고정되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 4,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "세 인물의 발까지 담고 좌하단 탈출 공간을 비교적 잘 확보했지만, 허용되지 않은 무장 인물과 팔 대신 손을 잡는 동작, 찰리의 주황색 눈이 지시를 어긴다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "현우의 전경사와 질주 자세는 더 강하지만, 좌하단을 가로지르는 군중이 지정 탈출로를 막으며 무장 인물 추가와 손잡기·눈 색상 오류도 남는다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh17_sel.png",
    "asset_id": "51f7a09e-ae54-4847-a177-3ebd0907f551",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0972-c4b0-7292-888c-6df0d9928984",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh17"
  }
 },
 "S26sh2::signage": {
  "fp": "6f65f5bd9d0facdc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::fc790b8df36b11f8": {
  "subjects": [],
  "subject_text": "인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중\n맨홀 아래로 길게 이어지는 어두운 지하 하수도. 수직 사다리와 두 갈래로 나뉘는 통로가 있으며, 외부로 이어지는 배수 출구가 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L176",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::escape_sewer_tunnel": {
  "input_fingerprint": "e1fec7a4efeafb54",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "escape_sewer_tunnel",
    "tags": [
     "S26sh2"
    ]
   },
   "context_sig": "3ceb20b3c968047e"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우일행, 사다리 타고 지하 하수도로 내려와 달린다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우일행, 사다리 타고 지하 하수도로 내려와 달린다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_escape_sewer_tunnel_92cbdc.png",
  "asset_id": "ceae8e6f-7f8b-4fad-8462-dcd7e12700fd",
  "input_asset_ids": [
   "7edf71e3-a57e-4e4f-a6d0-b67517785059"
  ],
  "origin_tag": "S26sh2",
  "place_text": "Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.",
  "origin_inputs": {
   "place_text": "Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.",
   "time_of_day_en": "night",
   "conti_asset_id": "7edf71e3-a57e-4e4f-a6d0-b67517785059"
  }
 },
 "S26sh2::bgfirst_bg": {
  "input_fingerprint": "15a41c4e0bd67013",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 겉옷으로 앰버의 입과 코를 단단히 감싸 쥔 채 하수도 물살을 가르며 달리는 mid-stride 찰나의 이현우, 좁은 터널 구도.\n\nLOCATION (lock): Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside the pair from 이현우's left-front quarter at lower-chest height, looking slightly upward along the narrow sewer passage. Keep 이현우 at left-center and 앰버 immediately beside him at right-center in a thigh-up moving composition: one hand holds his removed outer garment over her mouth and nose while his other hand keeps hers, with the cloth confined to her lower face. 이현우 watches the route beyond the camera's side and 앰버 lowers her eyes toward her footing, their different running phases and a small strip of disturbed water at the bottom conveying motion without separating them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 좁은 하수도 통로 (현우와 앰버가 달리고 있는 통로) — The passage recedes behind the pair along their running axis; used as Close lateral boundaries compress the composition without adding unsupported pipes or fixtures; 하수도 물살 (달리는 움직임으로 갈라지고 있음); used as A limited lower-frame strip supplies movement context without making reflected imagery the subject; 현우의 겉옷 (벗어서 앰버의 입과 코를 감싸고 있음) — The folded cloth is seen obliquely against Amber's lower face beneath Hyunwoo's hand; used as Makes his protective action readable at a natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the sewer, preserving the protective hand-and-cloth gesture without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 겉옷으로 앰버의 입과 코를 단단히 감싸 쥔 채 하수도 물살을 가르며 달리는 mid-stride 찰나의 이현우, 좁은 터널 구도.\n\nLOCATION (lock): Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside the pair from 이현우's left-front quarter at lower-chest height, looking slightly upward along the narrow sewer passage. Keep 이현우 at left-center and 앰버 immediately beside him at right-center in a thigh-up moving composition: one hand holds his removed outer garment over her mouth and nose while his other hand keeps hers, with the cloth confined to her lower face. 이현우 watches the route beyond the camera's side and 앰버 lowers her eyes toward her footing, their different running phases and a small strip of disturbed water at the bottom conveying motion without separating them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 좁은 하수도 통로 (현우와 앰버가 달리고 있는 통로) — The passage recedes behind the pair along their running axis; used as Close lateral boundaries compress the composition without adding unsupported pipes or fixtures; 하수도 물살 (달리는 움직임으로 갈라지고 있음); used as A limited lower-frame strip supplies movement context without making reflected imagery the subject; 현우의 겉옷 (벗어서 앰버의 입과 코를 감싸고 있음) — The folded cloth is seen obliquely against Amber's lower face beneath Hyunwoo's hand; used as Makes his protective action readable at a natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the sewer, preserving the protective hand-and-cloth gesture without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh2__bgfirst_bg.png",
  "asset_id": "5d2ace75-61f8-4777-b4db-b6de6cabd6d1",
  "input_asset_ids": [
   "7edf71e3-a57e-4e4f-a6d0-b67517785059",
   "ceae8e6f-7f8b-4fad-8462-dcd7e12700fd"
  ]
 },
 "S26sh2": {
  "input_fingerprint": "8c08e1c7a412119e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신의 겉옷으로 앰버의 입과 코를 단단히 감싸 쥔 채 하수도 물살을 가르며 달리는 mid-stride 찰나의 이현우, 좁은 터널 구도.\n\nLOCATION (lock): Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside the pair from 이현우's left-front quarter at lower-chest height, looking slightly upward along the narrow sewer passage. Keep 이현우 at left-center and 앰버 immediately beside him at right-center in a thigh-up moving composition: one hand holds his removed outer garment over her mouth and nose while his other hand keeps hers, with the cloth confined to her lower face. 이현우 watches the route beyond the camera's side and 앰버 lowers her eyes toward her footing, their different running phases and a small strip of disturbed water at the bottom conveying motion without separating them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 좁은 하수도 통로 (현우와 앰버가 달리고 있는 통로) — The passage recedes behind the pair along their running axis; used as Close lateral boundaries compress the composition without adding unsupported pipes or fixtures; 하수도 물살 (달리는 움직임으로 갈라지고 있음); used as A limited lower-frame strip supplies movement context without making reflected imagery the subject; 현우의 겉옷 (벗어서 앰버의 입과 코를 감싸고 있음) — The folded cloth is seen obliquely against Amber's lower face beneath Hyunwoo's hand; used as Makes his protective action readable at a natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the sewer, preserving the protective hand-and-cloth gesture without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish-yard manhole cover has been pushed aside, leaving ladder access to the sewer open. Charlie retains his old coat and hat disguise and blue-lit eyes; at the reception, the restored lights and burst lamps remain. 이현우: He is running through the sewer with his outer garment removed. His facial bruises and leg injury remain. 앰버: Her mouth is covered with a removed outer garment as she runs; her mask and waist tool pouch remain established.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신의 겉옷으로 앰버의 입과 코를 단단히 감싸 쥔 채 하수도 물살을 가르며 달리는 mid-stride 찰나의 이현우, 좁은 터널 구도.\n\nLOCATION (lock): Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside the pair from 이현우's left-front quarter at lower-chest height, looking slightly upward along the narrow sewer passage. Keep 이현우 at left-center and 앰버 immediately beside him at right-center in a thigh-up moving composition: one hand holds his removed outer garment over her mouth and nose while his other hand keeps hers, with the cloth confined to her lower face. 이현우 watches the route beyond the camera's side and 앰버 lowers her eyes toward her footing, their different running phases and a small strip of disturbed water at the bottom conveying motion without separating them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 좁은 하수도 통로 (현우와 앰버가 달리고 있는 통로) — The passage recedes behind the pair along their running axis; used as Close lateral boundaries compress the composition without adding unsupported pipes or fixtures; 하수도 물살 (달리는 움직임으로 갈라지고 있음); used as A limited lower-frame strip supplies movement context without making reflected imagery the subject; 현우의 겉옷 (벗어서 앰버의 입과 코를 감싸고 있음) — The folded cloth is seen obliquely against Amber's lower face beneath Hyunwoo's hand; used as Makes his protective action readable at a natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the sewer, preserving the protective hand-and-cloth gesture without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish-yard manhole cover has been pushed aside, leaving ladder access to the sewer open. Charlie retains his old coat and hat disguise and blue-lit eyes; at the reception, the restored lights and burst lamps remain. 이현우: He is running through the sewer with his outer garment removed. His facial bruises and leg injury remain. 앰버: Her mouth is covered with a removed outer garment as she runs; her mask and waist tool pouch remain established.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신의 겉옷으로 앰버의 입과 코를 단단히 감싸 쥔 채 하수도 물살을 가르며 달리는 mid-stride 찰나의 이현우, 좁은 터널 구도.\n\nLOCATION (lock): Inside a narrow underground sewer passage beneath the refugee settlement, with shallow water along the route. Only minimal visibility is specified; no fixed lighting source is established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside the pair from 이현우's left-front quarter at lower-chest height, looking slightly upward along the narrow sewer passage. Keep 이현우 at left-center and 앰버 immediately beside him at right-center in a thigh-up moving composition: one hand holds his removed outer garment over her mouth and nose while his other hand keeps hers, with the cloth confined to her lower face. 이현우 watches the route beyond the camera's side and 앰버 lowers her eyes toward her footing, their different running phases and a small strip of disturbed water at the bottom conveying motion without separating them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 좁은 하수도 통로 (현우와 앰버가 달리고 있는 통로) — The passage recedes behind the pair along their running axis; used as Close lateral boundaries compress the composition without adding unsupported pipes or fixtures; 하수도 물살 (달리는 움직임으로 갈라지고 있음); used as A limited lower-frame strip supplies movement context without making reflected imagery the subject; 현우의 겉옷 (벗어서 앰버의 입과 코를 감싸고 있음) — The folded cloth is seen obliquely against Amber's lower face beneath Hyunwoo's hand; used as Makes his protective action readable at a natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the sewer, preserving the protective hand-and-cloth gesture without specifying an unsupported fixture or light color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rubbish-yard manhole cover has been pushed aside, leaving ladder access to the sewer open. Charlie retains his old coat and hat disguise and blue-lit eyes; at the reception, the restored lights and burst lamps remain. 이현우: He is running through the sewer with his outer garment removed. His facial bruises and leg injury remain. 앰버: Her mouth is covered with a removed outer garment as she runs; her mask and waist tool pouch remain established.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh2__bgfirst_bg.png",
     "asset_id": "5d2ace75-61f8-4777-b4db-b6de6cabd6d1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S26sh2.png",
     "asset_id": "7edf71e3-a57e-4e4f-a6d0-b67517785059",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_escape_sewer_tunnel_92cbdc.png",
     "asset_id": "ceae8e6f-7f8b-4fad-8462-dcd7e12700fd",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "현우는 화면 왼쪽 카메라 밖의 진행 경로를 보고, 앰버는 앞쪽 바닥을 향해 눈을 내린다. 두 사람은 통로 안쪽에서 카메라 쪽으로 나란히 이동한다. 시선은 요구에 가깝지만 현우의 손은 앰버의 입과 코가 아니라 어깨에 놓여 있다.",
    "built_space": "젖고 얼룩진 아치형 통로 하나, 양쪽의 낮은 턱, 중앙 수로가 보인다. 왼쪽 벽에는 가는 배관 한 줄, 오른쪽 상부에는 굵은 배관 여러 줄과 가는 벽 배관이 있으며, 뒤쪽 왼편에는 밝은 수직 개구부 일부가 보인다. 참고 장소의 낡은 재질과 배치에 가깝다. 두 사람은 턱 사이 수로에 있다. 거의 정면인 구도이며 무릎 부근까지 보여 지정된 허벅지 위 구도보다 조금 넓다.",
    "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 젊은 동아시아계 남성 외형, 짧은 검은 머리, 마른 체격, 어두운 긴소매 셔츠와 인이어 장치는 참고에 가깝다. 얼굴의 멍은 뚜렷하지 않다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 주머니, 허리 앞쪽에 매달린 방진 마스크가 보인다. 혼혈 여부는 외관만으로 확정할 수 없다. 겉옷은 어깨와 등 쪽에 걸쳐져 있고 앰버의 입과 코를 덮지 않는다. 서로 손도 잡지 않는다.",
    "hard_violations": [],
    "physics": "두 사람의 다리는 프레임 아래 수로 바닥 방향으로 이어지며 발 접촉은 잘려 있다. 공중에 떠 있다고 볼 근거는 없다. 옷은 어깨와 등에 걸려 지지되고 현우의 손은 앰버의 어깨에 닿는다. 자세는 느린 보행에 가까우며, 달리는 보폭과 강하게 갈라지는 물살은 뚜렷하지 않다."
   },
   {
    "label": "A",
    "direction": "현우는 화면 왼쪽 카메라 밖을 보며 앞으로 달린다. 앰버는 발밑이 아니라 전방의 화면 왼쪽을 본다. 현우가 뻗은 손은 앰버의 입과 코에 있는 마스크와 천 부위에 닿아 보호 행동의 목표는 맞는다. 다른 손은 달리기 동작으로 떨어져 있어 앰버의 손을 잡지 않는다.",
    "built_space": "아치형 통로 하나와 양쪽 낮은 턱, 중앙의 얕은 수로가 보이고 통로는 두 사람 뒤로 이어진다. 노출 배관과 사다리는 보이지 않는다. 벽은 비교적 균일한 밝은 콘크리트 구획으로, 참고의 거칠고 심하게 오염된 벽 및 배관이 있는 장소와 차이가 크다. 현우는 왼쪽 중앙, 앰버는 바로 옆 오른쪽 중앙에 배치된다. 거의 정면의 낮은 시점이며 무릎 아래까지 일부 보여 지정된 허벅지 위 구도보다 넓다.",
    "entities": "인물은 두 명뿐이다. 현우의 검은 머리, 젊은 동아시아계 남성 외형, 마른 체격과 인이어 장치는 맞지만 참고의 칼라 있는 긴소매 셔츠 대신 반소매 티셔츠를 입었다. 바지 무릎은 찢어져 있으나 부상 자체는 확인하기 어렵다. 앰버는 금발의 어린 여자아이이고 하이테크 마스크가 코와 입에 보인다. 카키색 옷은 참고의 긴소매 작업복과 다르며 허리 공구 주머니는 확인되지 않는다. 현우의 손이 아래 얼굴을 누르지만 마스크 전면이 드러나고, 겉옷은 얼굴 아래에서 가슴과 어깨까지 크게 늘어진다.",
    "hard_violations": [],
    "physics": "현우는 한쪽 무릎을 앞으로 들고 반대쪽 다리를 뒤로 접었으며, 앰버는 다른 보폭으로 달린다. 나머지 발의 접지는 프레임 밖이지만 다리의 굽힘과 전진 자세가 달리기로 설명되며 무근거한 부유는 아니다. 겉옷은 앰버의 어깨와 현우의 손으로 지지되고 천이 아래로 처진다. 하단 수면의 물보라가 이동을 뒷받침한다. 다만 앰버의 두 손과 현우의 자유로운 손이 각각 떨어져 있어 손을 잡고 달리는 동작은 아니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "하수도 장소와 인물 외형은 가깝지만, 앰버의 입과 코가 완전히 드러나고 달리기 대신 걷는 모습이라 핵심 보호 행동을 재현하지 못했다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "현우가 앰버의 아래 얼굴을 누르며 함께 달리는 순간은 더 충실하지만, 손잡기·앰버의 하향 시선·정확한 하수도 구조와 의상은 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 화면 왼쪽 카메라 밖의 진행 경로를 보고, 앰버는 앞쪽 바닥을 향해 눈을 내린다. 두 사람은 통로 안쪽에서 카메라 쪽으로 나란히 이동한다. 시선은 요구에 가깝지만 현우의 손은 앰버의 입과 코가 아니라 어깨에 놓여 있다.",
        "built_space": "젖고 얼룩진 아치형 통로 하나, 양쪽의 낮은 턱, 중앙 수로가 보인다. 왼쪽 벽에는 가는 배관 한 줄, 오른쪽 상부에는 굵은 배관 여러 줄과 가는 벽 배관이 있으며, 뒤쪽 왼편에는 밝은 수직 개구부 일부가 보인다. 참고 장소의 낡은 재질과 배치에 가깝다. 두 사람은 턱 사이 수로에 있다. 거의 정면인 구도이며 무릎 부근까지 보여 지정된 허벅지 위 구도보다 조금 넓다.",
        "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 젊은 동아시아계 남성 외형, 짧은 검은 머리, 마른 체격, 어두운 긴소매 셔츠와 인이어 장치는 참고에 가깝다. 얼굴의 멍은 뚜렷하지 않다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 주머니, 허리 앞쪽에 매달린 방진 마스크가 보인다. 혼혈 여부는 외관만으로 확정할 수 없다. 겉옷은 어깨와 등 쪽에 걸쳐져 있고 앰버의 입과 코를 덮지 않는다. 서로 손도 잡지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 다리는 프레임 아래 수로 바닥 방향으로 이어지며 발 접촉은 잘려 있다. 공중에 떠 있다고 볼 근거는 없다. 옷은 어깨와 등에 걸려 지지되고 현우의 손은 앰버의 어깨에 닿는다. 자세는 느린 보행에 가까우며, 달리는 보폭과 강하게 갈라지는 물살은 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "현우는 화면 왼쪽 카메라 밖을 보며 앞으로 달린다. 앰버는 발밑이 아니라 전방의 화면 왼쪽을 본다. 현우가 뻗은 손은 앰버의 입과 코에 있는 마스크와 천 부위에 닿아 보호 행동의 목표는 맞는다. 다른 손은 달리기 동작으로 떨어져 있어 앰버의 손을 잡지 않는다.",
        "built_space": "아치형 통로 하나와 양쪽 낮은 턱, 중앙의 얕은 수로가 보이고 통로는 두 사람 뒤로 이어진다. 노출 배관과 사다리는 보이지 않는다. 벽은 비교적 균일한 밝은 콘크리트 구획으로, 참고의 거칠고 심하게 오염된 벽 및 배관이 있는 장소와 차이가 크다. 현우는 왼쪽 중앙, 앰버는 바로 옆 오른쪽 중앙에 배치된다. 거의 정면의 낮은 시점이며 무릎 아래까지 일부 보여 지정된 허벅지 위 구도보다 넓다.",
        "entities": "인물은 두 명뿐이다. 현우의 검은 머리, 젊은 동아시아계 남성 외형, 마른 체격과 인이어 장치는 맞지만 참고의 칼라 있는 긴소매 셔츠 대신 반소매 티셔츠를 입었다. 바지 무릎은 찢어져 있으나 부상 자체는 확인하기 어렵다. 앰버는 금발의 어린 여자아이이고 하이테크 마스크가 코와 입에 보인다. 카키색 옷은 참고의 긴소매 작업복과 다르며 허리 공구 주머니는 확인되지 않는다. 현우의 손이 아래 얼굴을 누르지만 마스크 전면이 드러나고, 겉옷은 얼굴 아래에서 가슴과 어깨까지 크게 늘어진다.",
        "hard_violations": [],
        "physics": "현우는 한쪽 무릎을 앞으로 들고 반대쪽 다리를 뒤로 접었으며, 앰버는 다른 보폭으로 달린다. 나머지 발의 접지는 프레임 밖이지만 다리의 굽힘과 전진 자세가 달리기로 설명되며 무근거한 부유는 아니다. 겉옷은 앰버의 어깨와 현우의 손으로 지지되고 천이 아래로 처진다. 하단 수면의 물보라가 이동을 뒷받침한다. 다만 앰버의 두 손과 현우의 자유로운 손이 각각 떨어져 있어 손을 잡고 달리는 동작은 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "하수도 장소와 인물 외형은 가깝지만, 앰버의 입과 코가 완전히 드러나고 달리기 대신 걷는 모습이라 핵심 보호 행동을 재현하지 못했다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "현우가 앰버의 아래 얼굴을 누르며 함께 달리는 순간은 더 충실하지만, 손잡기·앰버의 하향 시선·정확한 하수도 구조와 의상은 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 화면 왼쪽 카메라 밖의 진행 경로를 보고, 앰버는 앞쪽 바닥을 향해 눈을 내린다. 두 사람은 통로 안쪽에서 카메라 쪽으로 나란히 이동한다. 시선은 요구에 가깝지만 현우의 손은 앰버의 입과 코가 아니라 어깨에 놓여 있다.",
        "built_space": "젖고 얼룩진 아치형 통로 하나, 양쪽의 낮은 턱, 중앙 수로가 보인다. 왼쪽 벽에는 가는 배관 한 줄, 오른쪽 상부에는 굵은 배관 여러 줄과 가는 벽 배관이 있으며, 뒤쪽 왼편에는 밝은 수직 개구부 일부가 보인다. 참고 장소의 낡은 재질과 배치에 가깝다. 두 사람은 턱 사이 수로에 있다. 거의 정면인 구도이며 무릎 부근까지 보여 지정된 허벅지 위 구도보다 조금 넓다.",
        "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 젊은 동아시아계 남성 외형, 짧은 검은 머리, 마른 체격, 어두운 긴소매 셔츠와 인이어 장치는 참고에 가깝다. 얼굴의 멍은 뚜렷하지 않다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 주머니, 허리 앞쪽에 매달린 방진 마스크가 보인다. 혼혈 여부는 외관만으로 확정할 수 없다. 겉옷은 어깨와 등 쪽에 걸쳐져 있고 앰버의 입과 코를 덮지 않는다. 서로 손도 잡지 않는다.",
        "hard_violations": [],
        "physics": "두 사람의 다리는 프레임 아래 수로 바닥 방향으로 이어지며 발 접촉은 잘려 있다. 공중에 떠 있다고 볼 근거는 없다. 옷은 어깨와 등에 걸려 지지되고 현우의 손은 앰버의 어깨에 닿는다. 자세는 느린 보행에 가까우며, 달리는 보폭과 강하게 갈라지는 물살은 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "현우는 화면 왼쪽 카메라 밖을 보며 앞으로 달린다. 앰버는 발밑이 아니라 전방의 화면 왼쪽을 본다. 현우가 뻗은 손은 앰버의 입과 코에 있는 마스크와 천 부위에 닿아 보호 행동의 목표는 맞는다. 다른 손은 달리기 동작으로 떨어져 있어 앰버의 손을 잡지 않는다.",
        "built_space": "아치형 통로 하나와 양쪽 낮은 턱, 중앙의 얕은 수로가 보이고 통로는 두 사람 뒤로 이어진다. 노출 배관과 사다리는 보이지 않는다. 벽은 비교적 균일한 밝은 콘크리트 구획으로, 참고의 거칠고 심하게 오염된 벽 및 배관이 있는 장소와 차이가 크다. 현우는 왼쪽 중앙, 앰버는 바로 옆 오른쪽 중앙에 배치된다. 거의 정면의 낮은 시점이며 무릎 아래까지 일부 보여 지정된 허벅지 위 구도보다 넓다.",
        "entities": "인물은 두 명뿐이다. 현우의 검은 머리, 젊은 동아시아계 남성 외형, 마른 체격과 인이어 장치는 맞지만 참고의 칼라 있는 긴소매 셔츠 대신 반소매 티셔츠를 입었다. 바지 무릎은 찢어져 있으나 부상 자체는 확인하기 어렵다. 앰버는 금발의 어린 여자아이이고 하이테크 마스크가 코와 입에 보인다. 카키색 옷은 참고의 긴소매 작업복과 다르며 허리 공구 주머니는 확인되지 않는다. 현우의 손이 아래 얼굴을 누르지만 마스크 전면이 드러나고, 겉옷은 얼굴 아래에서 가슴과 어깨까지 크게 늘어진다.",
        "hard_violations": [],
        "physics": "현우는 한쪽 무릎을 앞으로 들고 반대쪽 다리를 뒤로 접었으며, 앰버는 다른 보폭으로 달린다. 나머지 발의 접지는 프레임 밖이지만 다리의 굽힘과 전진 자세가 달리기로 설명되며 무근거한 부유는 아니다. 겉옷은 앰버의 어깨와 현우의 손으로 지지되고 천이 아래로 처진다. 하단 수면의 물보라가 이동을 뒷받침한다. 다만 앰버의 두 손과 현우의 자유로운 손이 각각 떨어져 있어 손을 잡고 달리는 동작은 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 3,
   "A": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "하수도 장소와 인물 외형은 가깝지만, 앰버의 입과 코가 완전히 드러나고 달리기 대신 걷는 모습이라 핵심 보호 행동을 재현하지 못했다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "현우가 앰버의 아래 얼굴을 누르며 함께 달리는 순간은 더 충실하지만, 손잡기·앰버의 하향 시선·정확한 하수도 구조와 의상은 어긋난다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_escape_sewer_tunnel_92cbdc.png",
    "asset_id": "ceae8e6f-7f8b-4fad-8462-dcd7e12700fd",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab097b-458e-7325-b530-b48379bb45c4",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh2__bgfirst_bg.png",
   "bg_asset_id": "5d2ace75-61f8-4777-b4db-b6de6cabd6d1",
   "bg_record_key": "S26sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "escape_sewer_tunnel",
   "groupbg_asset_id": "ceae8e6f-7f8b-4fad-8462-dcd7e12700fd"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S26sh9::signage": {
  "fp": "c00c4ee61579b20b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S26sh9": {
  "input_fingerprint": "502e1ac380ef86b8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연의 굳은 표정과 미세한 떨림을 놓치지 않고 비스듬히 노려보는 박철진의 잔인한 눈빛.\n\nLOCATION (lock): At the outdoor reception clearing, beside the group of wedding guests being held under guard at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inherited lateral move behind and beside 미연 at shoulder height, looking past her near shoulder toward 박철진's oblique three-quarter face. Keep 미연's tense lower profile and slightly trembling shoulder at the left foreground edge, with her eyes just beyond the crop and lowered, while 박철진 occupies right-center and fixes his gaze toward her off-screen eyes rather than the lens. Favor his eyes through selective focus while retaining her rigid jaw as readable foreground evidence, making the eyeline shift the shot's sole accent.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 피로연장 중앙의 하객 무리 (위협받으며 한가운데 모여 있음); used as A softly resolved background cluster situates Miyeon among the detained guests; uneven head heights, small shoulder turns and varied spacing avoid a duplicated lineup.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use controlled nighttime illumination appropriate to the reception, retaining detail in both faces without adding a new source, color cue or threatening lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and burst lamps, while the escape route remains the underground sewer. Charlie retains his old coat and hat disguise and blue-lit eyes. 미연: She stands in the front portion of the gathered guests, visibly alarmed. 박철진: He remains at the reception interrogation, watching intently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연의 굳은 표정과 미세한 떨림을 놓치지 않고 비스듬히 노려보는 박철진의 잔인한 눈빛.\n\nLOCATION (lock): At the outdoor reception clearing, beside the group of wedding guests being held under guard at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inherited lateral move behind and beside 미연 at shoulder height, looking past her near shoulder toward 박철진's oblique three-quarter face. Keep 미연's tense lower profile and slightly trembling shoulder at the left foreground edge, with her eyes just beyond the crop and lowered, while 박철진 occupies right-center and fixes his gaze toward her off-screen eyes rather than the lens. Favor his eyes through selective focus while retaining her rigid jaw as readable foreground evidence, making the eyeline shift the shot's sole accent.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 피로연장 중앙의 하객 무리 (위협받으며 한가운데 모여 있음); used as A softly resolved background cluster situates Miyeon among the detained guests; uneven head heights, small shoulder turns and varied spacing avoid a duplicated lineup.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use controlled nighttime illumination appropriate to the reception, retaining detail in both faces without adding a new source, color cue or threatening lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and burst lamps, while the escape route remains the underground sewer. Charlie retains his old coat and hat disguise and blue-lit eyes. 미연: She stands in the front portion of the gathered guests, visibly alarmed. 박철진: He remains at the reception interrogation, watching intently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연의 굳은 표정과 미세한 떨림을 놓치지 않고 비스듬히 노려보는 박철진의 잔인한 눈빛.\n\nLOCATION (lock): At the outdoor reception clearing, beside the group of wedding guests being held under guard at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inherited lateral move behind and beside 미연 at shoulder height, looking past her near shoulder toward 박철진's oblique three-quarter face. Keep 미연's tense lower profile and slightly trembling shoulder at the left foreground edge, with her eyes just beyond the crop and lowered, while 박철진 occupies right-center and fixes his gaze toward her off-screen eyes rather than the lens. Favor his eyes through selective focus while retaining her rigid jaw as readable foreground evidence, making the eyeline shift the shot's sole accent.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 피로연장 중앙의 하객 무리 (위협받으며 한가운데 모여 있음); used as A softly resolved background cluster situates Miyeon among the detained guests; uneven head heights, small shoulder turns and varied spacing avoid a duplicated lineup.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use controlled nighttime illumination appropriate to the reception, retaining detail in both faces without adding a new source, color cue or threatening lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The reception retains its restored lighting and burst lamps, while the escape route remains the underground sewer. Charlie retains his old coat and hat disguise and blue-lit eyes. 미연: She stands in the front portion of the gathered guests, visibly alarmed. 박철진: He remains at the reception interrogation, watching intently.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진의 시선이 화면 밖 미연의 눈높이를 향해 비스듬히 꽂혀 있으며, 미연은 시선을 아래로 향한 채 굳어 있습니다.",
    "built_space": "야외 피로연장의 야간 조명, 뒤편의 군용 트럭 등 레퍼런스의 환경이 유지되었으며 배경의 하객들이 바닥에 주저앉아 통제받는 모습이 보입니다.",
    "entities": "박철진과 미연의 얼굴, 체형, 복장(어두운 전투복, 붉은 완장, 모자 등)이 레퍼런스와 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며 어색한 신체 왜곡은 관찰되지 않습니다."
   },
   {
    "label": "B",
    "direction": "박철진이 미연 쪽을 쳐다보고 있으나 시선이 다소 엇나간 느낌이 있으며, 미연은 시선을 아래로 내리고 있습니다.",
    "built_space": "야간 야외 피로연장의 조명과 시설물이 존재하나, 배경의 하객들이 위협받거나 억류된 모습 없이 평범하게 서 있습니다.",
    "entities": "지정된 두 인물의 얼굴, 머리, 복장, 붉은 완장 등 외형 조건이 레퍼런스와 잘 일치합니다.",
    "hard_violations": [],
    "physics": "중력에 맞게 자연스럽게 서 있으며 지지 기반이나 자세에 물리적 오류가 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 피사계 심도가 프롬프트의 지시사항에 매우 근접하며, 위협받는 배경 하객들의 모습이 잘 표현되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물의 외형은 훌륭하게 구현되었으나 미연의 눈이 프레임에 너무 많이 포함되었고, 배경의 하객들이 억류된 상태로 보이지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선이 화면 밖 미연의 눈높이를 향해 비스듬히 꽂혀 있으며, 미연은 시선을 아래로 향한 채 굳어 있습니다.",
        "built_space": "야외 피로연장의 야간 조명, 뒤편의 군용 트럭 등 레퍼런스의 환경이 유지되었으며 배경의 하객들이 바닥에 주저앉아 통제받는 모습이 보입니다.",
        "entities": "박철진과 미연의 얼굴, 체형, 복장(어두운 전투복, 붉은 완장, 모자 등)이 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며 어색한 신체 왜곡은 관찰되지 않습니다."
       },
       {
        "label": "B",
        "direction": "박철진이 미연 쪽을 쳐다보고 있으나 시선이 다소 엇나간 느낌이 있으며, 미연은 시선을 아래로 내리고 있습니다.",
        "built_space": "야간 야외 피로연장의 조명과 시설물이 존재하나, 배경의 하객들이 위협받거나 억류된 모습 없이 평범하게 서 있습니다.",
        "entities": "지정된 두 인물의 얼굴, 머리, 복장, 붉은 완장 등 외형 조건이 레퍼런스와 잘 일치합니다.",
        "hard_violations": [],
        "physics": "중력에 맞게 자연스럽게 서 있으며 지지 기반이나 자세에 물리적 오류가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 피사계 심도가 프롬프트의 지시사항에 매우 근접하며, 위협받는 배경 하객들의 모습이 잘 표현되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물의 외형은 훌륭하게 구현되었으나 미연의 눈이 프레임에 너무 많이 포함되었고, 배경의 하객들이 억류된 상태로 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선이 화면 밖 미연의 눈높이를 향해 비스듬히 꽂혀 있으며, 미연은 시선을 아래로 향한 채 굳어 있습니다.",
        "built_space": "야외 피로연장의 야간 조명, 뒤편의 군용 트럭 등 레퍼런스의 환경이 유지되었으며 배경의 하객들이 바닥에 주저앉아 통제받는 모습이 보입니다.",
        "entities": "박철진과 미연의 얼굴, 체형, 복장(어두운 전투복, 붉은 완장, 모자 등)이 레퍼런스와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며 어색한 신체 왜곡은 관찰되지 않습니다."
       },
       {
        "label": "B",
        "direction": "박철진이 미연 쪽을 쳐다보고 있으나 시선이 다소 엇나간 느낌이 있으며, 미연은 시선을 아래로 내리고 있습니다.",
        "built_space": "야간 야외 피로연장의 조명과 시설물이 존재하나, 배경의 하객들이 위협받거나 억류된 모습 없이 평범하게 서 있습니다.",
        "entities": "지정된 두 인물의 얼굴, 머리, 복장, 붉은 완장 등 외형 조건이 레퍼런스와 잘 일치합니다.",
        "hard_violations": [],
        "physics": "중력에 맞게 자연스럽게 서 있으며 지지 기반이나 자세에 물리적 오류가 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "박철진이 미연을 곁눈질하는 방향과 장소는 맞지만, 카메라가 미연의 앞옆에 있어 요구된 어깨 너머 시점과 눈을 제외한 하단 옆얼굴 크롭을 놓쳤다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미연의 뒤쪽 어깨를 넘는 시점과 박철진의 집요한 눈빛·선택적 초점이 더 정확하지만, 두 후보 모두 미연의 눈을 화면 밖으로 잘라야 한다는 지시에는 실패했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 얼굴과 눈동자는 화면 왼쪽 미연의 눈 부근을 향하며 렌즈를 보지 않는다. 미연은 오른쪽 앞쪽을 약간 낮춰 보고 있어 박철진과 눈을 맞추지 않는다. 오른쪽 경비 인물들의 총기는 대체로 아래로 향하며 특정 하객을 직접 조준하는 모습은 분명하지 않다.",
        "built_space": "왼쪽과 중앙 뒤에 천막 지붕과 여러 지지대, 삼각 깃발 및 전구 줄이 있고 오른쪽 뒤에는 덮개 달린 트럭 한 대가 보인다. 젖은 바닥은 기존 피로연장의 재질과 조명을 따른다. 중앙 하객들은 서로 다른 높이와 간격으로 서 있고 경비 인물들은 오른쪽에 있다. 미연은 왼쪽 전경, 박철진은 오른쪽 중앙에 허리까지 잡히지만, 미연의 가슴 앞면까지 보여 카메라가 지정된 뒤옆보다는 앞옆에 있다. 바닥의 빛 반사는 가능한 위치다.",
        "entities": "박철진은 중년 한국인 남성 설정에 부합하는 외형이며 참고의 얼굴, 짙은 모자, 남색 전투복, 붉은 완장과 흰 표식이 대체로 유지된다. 미연도 참고의 중년 여성 얼굴, 검은 단발과 남색 니트에 가깝다. 다만 미연의 눈과 얼굴 대부분이 보여 명시된 하단 옆얼굴 크롭과 다르다. 배경에는 구체적인 사람으로 표현된 하객과 경비 인물들이 있으며 이전 스틸의 주요 인물은 재등장하지 않는다. 추가 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "두 주인공의 몸통은 화면 아래로 자연스럽게 이어지는 선 자세이고 발은 크롭 밖이다. 배경 하객과 경비 인물은 지면에 서 있으며 공중에 뜬 몸은 없다. 보이는 총기는 손과 몸 앞쪽에 걸쳐 지지된다. 천막과 전구 줄은 지지대에 매달리고 트럭은 지면에 놓인다. 미연의 굳은 턱과 어깨 긴장은 읽히지만 미세한 떨림 자체는 정지 화면에서 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진은 고개를 미연 쪽으로 기울이고 눈동자를 왼쪽 미연의 눈 부근에 고정한다. 렌즈를 향한 시선이 아니다. 미연은 고개와 눈을 낮춰 오른쪽 아래를 보며 시선을 피한다. 오른쪽 뒤 경비 인물들의 무기는 몸 앞에서 아래쪽으로 향하고, 선명한 특정 조준 대상은 없다.",
        "built_space": "중앙 뒤에 지지대가 있는 천막과 전구 줄, 삼각 깃발이 이어지고 오른쪽 뒤에 덮개 달린 트럭 한 대가 있다. 젖은 바닥과 따뜻한 실용등은 참고 장소에 부합하며 불가능한 거울 반사나 시설 중복은 없다. 하객들은 중앙에 모여 있고 앞쪽 일부는 몸을 낮춰 머리 높이가 달라진다. 미연의 뒤쪽 어깨가 왼쪽 전경을 이루고 박철진이 오른쪽 중앙을 차지해 지정된 뒤옆 시점에 더 가깝다. 다만 미연의 눈까지 화면 안에 남아 있다.",
        "entities": "박철진의 중년 남성 얼굴과 체격, 짙은 모자, 낡은 남색 전투복 및 붉은 완장은 참고와 대체로 일치한다. 미연의 검은 단발, 중년 여성 옆얼굴과 남색 니트도 참고를 따른다. 박철진의 눈에 초점이 더 집중되고 미연의 굳은 턱은 전경에서 읽힌다. 배경 하객은 얼굴과 옷을 갖춘 실제 사람으로 표현되며 반복 복제된 줄처럼 보이지 않는다. 새 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "박철진의 약간 기울어진 상체와 미연의 돌아선 어깨는 서 있는 몸에서 가능한 자세이고 하체는 크롭 밖으로 이어진다. 배경의 낮은 인물들은 몸을 숙이거나 웅크린 모습이며, 가려진 접촉점만으로 부유한다고 볼 근거는 없다. 경비 인물은 바닥에 서 있고 무기는 몸 앞에서 지지된다. 천막·전구 줄·트럭도 정상적인 지지를 갖는다. 어깨의 긴장은 보이지만 실제 떨림은 확인하기 어렵다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "박철진이 미연을 곁눈질하는 방향과 장소는 맞지만, 카메라가 미연의 앞옆에 있어 요구된 어깨 너머 시점과 눈을 제외한 하단 옆얼굴 크롭을 놓쳤다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미연의 뒤쪽 어깨를 넘는 시점과 박철진의 집요한 눈빛·선택적 초점이 더 정확하지만, 두 후보 모두 미연의 눈을 화면 밖으로 잘라야 한다는 지시에는 실패했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 얼굴과 눈동자는 화면 왼쪽 미연의 눈 부근을 향하며 렌즈를 보지 않는다. 미연은 오른쪽 앞쪽을 약간 낮춰 보고 있어 박철진과 눈을 맞추지 않는다. 오른쪽 경비 인물들의 총기는 대체로 아래로 향하며 특정 하객을 직접 조준하는 모습은 분명하지 않다.",
        "built_space": "왼쪽과 중앙 뒤에 천막 지붕과 여러 지지대, 삼각 깃발 및 전구 줄이 있고 오른쪽 뒤에는 덮개 달린 트럭 한 대가 보인다. 젖은 바닥은 기존 피로연장의 재질과 조명을 따른다. 중앙 하객들은 서로 다른 높이와 간격으로 서 있고 경비 인물들은 오른쪽에 있다. 미연은 왼쪽 전경, 박철진은 오른쪽 중앙에 허리까지 잡히지만, 미연의 가슴 앞면까지 보여 카메라가 지정된 뒤옆보다는 앞옆에 있다. 바닥의 빛 반사는 가능한 위치다.",
        "entities": "박철진은 중년 한국인 남성 설정에 부합하는 외형이며 참고의 얼굴, 짙은 모자, 남색 전투복, 붉은 완장과 흰 표식이 대체로 유지된다. 미연도 참고의 중년 여성 얼굴, 검은 단발과 남색 니트에 가깝다. 다만 미연의 눈과 얼굴 대부분이 보여 명시된 하단 옆얼굴 크롭과 다르다. 배경에는 구체적인 사람으로 표현된 하객과 경비 인물들이 있으며 이전 스틸의 주요 인물은 재등장하지 않는다. 추가 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "두 주인공의 몸통은 화면 아래로 자연스럽게 이어지는 선 자세이고 발은 크롭 밖이다. 배경 하객과 경비 인물은 지면에 서 있으며 공중에 뜬 몸은 없다. 보이는 총기는 손과 몸 앞쪽에 걸쳐 지지된다. 천막과 전구 줄은 지지대에 매달리고 트럭은 지면에 놓인다. 미연의 굳은 턱과 어깨 긴장은 읽히지만 미세한 떨림 자체는 정지 화면에서 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진은 고개를 미연 쪽으로 기울이고 눈동자를 왼쪽 미연의 눈 부근에 고정한다. 렌즈를 향한 시선이 아니다. 미연은 고개와 눈을 낮춰 오른쪽 아래를 보며 시선을 피한다. 오른쪽 뒤 경비 인물들의 무기는 몸 앞에서 아래쪽으로 향하고, 선명한 특정 조준 대상은 없다.",
        "built_space": "중앙 뒤에 지지대가 있는 천막과 전구 줄, 삼각 깃발이 이어지고 오른쪽 뒤에 덮개 달린 트럭 한 대가 있다. 젖은 바닥과 따뜻한 실용등은 참고 장소에 부합하며 불가능한 거울 반사나 시설 중복은 없다. 하객들은 중앙에 모여 있고 앞쪽 일부는 몸을 낮춰 머리 높이가 달라진다. 미연의 뒤쪽 어깨가 왼쪽 전경을 이루고 박철진이 오른쪽 중앙을 차지해 지정된 뒤옆 시점에 더 가깝다. 다만 미연의 눈까지 화면 안에 남아 있다.",
        "entities": "박철진의 중년 남성 얼굴과 체격, 짙은 모자, 낡은 남색 전투복 및 붉은 완장은 참고와 대체로 일치한다. 미연의 검은 단발, 중년 여성 옆얼굴과 남색 니트도 참고를 따른다. 박철진의 눈에 초점이 더 집중되고 미연의 굳은 턱은 전경에서 읽힌다. 배경 하객은 얼굴과 옷을 갖춘 실제 사람으로 표현되며 반복 복제된 줄처럼 보이지 않는다. 새 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "박철진의 약간 기울어진 상체와 미연의 돌아선 어깨는 서 있는 몸에서 가능한 자세이고 하체는 크롭 밖으로 이어진다. 배경의 낮은 인물들은 몸을 숙이거나 웅크린 모습이며, 가려진 접촉점만으로 부유한다고 볼 근거는 없다. 경비 인물은 바닥에 서 있고 무기는 몸 앞에서 지지된다. 천막·전구 줄·트럭도 정상적인 지지를 갖는다. 어깨의 긴장은 보이지만 실제 떨림은 확인하기 어렵다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.69
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.69
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1690
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "구도와 피사계 심도가 프롬프트의 지시사항에 매우 근접하며, 위협받는 배경 하객들의 모습이 잘 표현되었습니다."
   },
   {
    "label": "B",
    "score": 1690,
    "verdict_ko": "인물의 외형은 훌륭하게 구현되었으나 미연의 눈이 프레임에 너무 많이 포함되었고, 배경의 하객들이 억류된 상태로 보이지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S25sh28_sel.png",
    "asset_id": "fa61214b-c9fe-43d5-b438-ed6d5133bf64",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1296917>",
    "asset_id": "4355f93d-0fde-4b21-a98a-8202637de522",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab098e-112c-7c92-806a-5cdd161e1f7f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh28"
  },
  "staged_characters_added": [
   "C08"
  ]
 },
 "S26sh12::signage": {
  "fp": "6ef63a2bf61d014c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::sewer_junction": {
  "input_fingerprint": "ffebc5ba7160c224",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "sewer_junction",
    "tags": [
     "S26sh12"
    ]
   },
   "context_sig": "e418a732a34b977b"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우일행, 두 갈래 길로 나눠지는.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우일행, 두 갈래 길로 나눠지는.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_sewer_junction_75d7f3.png",
  "asset_id": "4d4807b6-ee2e-48eb-91c9-6bd1ad304eb5",
  "input_asset_ids": [
   "bbca828d-8e8e-429a-af43-6126dd370473"
  ],
  "origin_tag": "S26sh12",
  "place_text": "At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.",
  "origin_inputs": {
   "place_text": "At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.",
   "time_of_day_en": "night",
   "conti_asset_id": "bbca828d-8e8e-429a-af43-6126dd370473"
  }
 },
 "S26sh12::bgfirst_bg": {
  "input_fingerprint": "47baa15387522c03",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 결연하고 진지한 표정으로 페드로의 양어깨를 두 손으로 꽉 쥔 상반신.\n\nLOCATION (lock): At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the gentle push from behind and outside 페드로's near shoulder, at 이현우's upper-chest height with a slight upward tilt, keeping the lens offset from their face-to-face axis. 페드로's shoulder occupies the lower-left edge while 이현우's serious three-quarter face sits right of center; both forearms and both hands remain visible gripping 페드로's shoulders. Their attention stays on each other's faces, with the tightening distance—not a lighting or blocking change—carrying the urgency of their parting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지하 하수도 갈림길 (The route divides into two passages) — The branching space recedes obliquely behind 이현우; used as A subdued background reminder of the impending separation, subordinate to the shoulder grip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable faces and hands without introducing an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 결연하고 진지한 표정으로 페드로의 양어깨를 두 손으로 꽉 쥔 상반신.\n\nLOCATION (lock): At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the gentle push from behind and outside 페드로's near shoulder, at 이현우's upper-chest height with a slight upward tilt, keeping the lens offset from their face-to-face axis. 페드로's shoulder occupies the lower-left edge while 이현우's serious three-quarter face sits right of center; both forearms and both hands remain visible gripping 페드로's shoulders. Their attention stays on each other's faces, with the tightening distance—not a lighting or blocking change—carrying the urgency of their parting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지하 하수도 갈림길 (The route divides into two passages) — The branching space recedes obliquely behind 이현우; used as A subdued background reminder of the impending separation, subordinate to the shoulder grip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable faces and hands without introducing an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh12__bgfirst_bg.png",
  "asset_id": "56eea23e-4f38-4674-bd99-72ea4b90ee0d",
  "input_asset_ids": [
   "bbca828d-8e8e-429a-af43-6126dd370473",
   "4d4807b6-ee2e-48eb-91c9-6bd1ad304eb5"
  ]
 },
 "S26sh12": {
  "input_fingerprint": "111d7f242993538f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 결연하고 진지한 표정으로 페드로의 양어깨를 두 손으로 꽉 쥔 상반신.\n\nLOCATION (lock): At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the gentle push from behind and outside 페드로's near shoulder, at 이현우's upper-chest height with a slight upward tilt, keeping the lens offset from their face-to-face axis. 페드로's shoulder occupies the lower-left edge while 이현우's serious three-quarter face sits right of center; both forearms and both hands remain visible gripping 페드로's shoulders. Their attention stays on each other's faces, with the tightening distance—not a lighting or blocking change—carrying the urgency of their parting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지하 하수도 갈림길 (The route divides into two passages) — The branching space recedes obliquely behind 이현우; used as A subdued background reminder of the impending separation, subordinate to the shoulder grip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable faces and hands without introducing an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route reaches a fork. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is at the sewer fork without his outer garment, still bearing facial bruises and the leg injury. 페드로: He is at the fork, preparing to take a separate route.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 결연하고 진지한 표정으로 페드로의 양어깨를 두 손으로 꽉 쥔 상반신.\n\nLOCATION (lock): At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the gentle push from behind and outside 페드로's near shoulder, at 이현우's upper-chest height with a slight upward tilt, keeping the lens offset from their face-to-face axis. 페드로's shoulder occupies the lower-left edge while 이현우's serious three-quarter face sits right of center; both forearms and both hands remain visible gripping 페드로's shoulders. Their attention stays on each other's faces, with the tightening distance—not a lighting or blocking change—carrying the urgency of their parting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지하 하수도 갈림길 (The route divides into two passages) — The branching space recedes obliquely behind 이현우; used as A subdued background reminder of the impending separation, subordinate to the shoulder grip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable faces and hands without introducing an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route reaches a fork. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is at the sewer fork without his outer garment, still bearing facial bruises and the leg injury. 페드로: He is at the fork, preparing to take a separate route.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 결연하고 진지한 표정으로 페드로의 양어깨를 두 손으로 꽉 쥔 상반신.\n\nLOCATION (lock): At the fork of the underground sewer passage beneath the refugee settlement. The nighttime tunnel has no specified fixed light source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the gentle push from behind and outside 페드로's near shoulder, at 이현우's upper-chest height with a slight upward tilt, keeping the lens offset from their face-to-face axis. 페드로's shoulder occupies the lower-left edge while 이현우's serious three-quarter face sits right of center; both forearms and both hands remain visible gripping 페드로's shoulders. Their attention stays on each other's faces, with the tightening distance—not a lighting or blocking change—carrying the urgency of their parting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지하 하수도 갈림길 (The route divides into two passages) — The branching space recedes obliquely behind 이현우; used as A subdued background reminder of the impending separation, subordinate to the shoulder grip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable faces and hands without introducing an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer route reaches a fork. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes. 이현우: He is at the sewer fork without his outer garment, still bearing facial bruises and the leg injury. 페드로: He is at the fork, preparing to take a separate route.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh12__bgfirst_bg.png",
     "asset_id": "56eea23e-4f38-4674-bd99-72ea4b90ee0d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S26sh12.png",
     "asset_id": "bbca828d-8e8e-429a-af43-6126dd370473",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_sewer_junction_75d7f3.png",
     "asset_id": "4d4807b6-ee2e-48eb-91c9-6bd1ad304eb5",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 페드로가 서로의 얼굴을 마주보고 있음.",
    "built_space": "하수도 갈림길의 중앙 기둥과 사다리가 보이나 한쪽 통로가 가려짐.",
    "entities": "이현우의 외양(상처, 셔츠, 무전기)과 페드로의 복장(비니, 재킷)이 일치함.",
    "hard_violations": [
     "[gemini-pro] 화면 좌측에 위치한 이현우의 오른손 엄지손가락이 바깥쪽을 향하고 있어 해부학적으로 불가능함"
    ],
    "physics": "손이 어깨에 닿아 있으나, 오른손의 구조가 해부학적으로 꺾일 수 없는 형태임."
   },
   {
    "label": "B",
    "direction": "이현우가 결연한 표정으로 페드로를 응시하며 서로 마주봄.",
    "built_space": "중앙의 사다리 기둥과 갈라지는 두 하수도 통로가 정확히 배치됨.",
    "entities": "인물들의 외양과 복장, 양어깨를 쥔 두 손과 팔뚝의 위치가 완벽히 일치함.",
    "hard_violations": [],
    "physics": "어깨를 꽉 쥔 두 손과 몸의 자세가 중력과 구조에 맞게 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "해부학적으로 불가능한 손 구조(오른손의 엄지 방향 오류)로 인해 치명적 위반이 발생했으며, 프레이밍 지시도 완벽히 따르지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "요청된 숄더뷰 앵글과 인물의 배치, 배경의 하수도 갈림길 등 모든 세부 사항을 사실적이고 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로가 서로의 얼굴을 마주보고 있음.",
        "built_space": "하수도 갈림길의 중앙 기둥과 사다리가 보이나 한쪽 통로가 가려짐.",
        "entities": "이현우의 외양(상처, 셔츠, 무전기)과 페드로의 복장(비니, 재킷)이 일치함.",
        "hard_violations": [
         "화면 좌측에 위치한 이현우의 오른손 엄지손가락이 바깥쪽을 향하고 있어 해부학적으로 불가능함"
        ],
        "physics": "손이 어깨에 닿아 있으나, 오른손의 구조가 해부학적으로 꺾일 수 없는 형태임."
       },
       {
        "label": "B",
        "direction": "이현우가 결연한 표정으로 페드로를 응시하며 서로 마주봄.",
        "built_space": "중앙의 사다리 기둥과 갈라지는 두 하수도 통로가 정확히 배치됨.",
        "entities": "인물들의 외양과 복장, 양어깨를 쥔 두 손과 팔뚝의 위치가 완벽히 일치함.",
        "hard_violations": [],
        "physics": "어깨를 꽉 쥔 두 손과 몸의 자세가 중력과 구조에 맞게 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "해부학적으로 불가능한 손 구조(오른손의 엄지 방향 오류)로 인해 치명적 위반이 발생했으며, 프레이밍 지시도 완벽히 따르지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "요청된 숄더뷰 앵글과 인물의 배치, 배경의 하수도 갈림길 등 모든 세부 사항을 사실적이고 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 페드로가 서로의 얼굴을 마주보고 있음.",
        "built_space": "하수도 갈림길의 중앙 기둥과 사다리가 보이나 한쪽 통로가 가려짐.",
        "entities": "이현우의 외양(상처, 셔츠, 무전기)과 페드로의 복장(비니, 재킷)이 일치함.",
        "hard_violations": [
         "화면 좌측에 위치한 이현우의 오른손 엄지손가락이 바깥쪽을 향하고 있어 해부학적으로 불가능함"
        ],
        "physics": "손이 어깨에 닿아 있으나, 오른손의 구조가 해부학적으로 꺾일 수 없는 형태임."
       },
       {
        "label": "B",
        "direction": "이현우가 결연한 표정으로 페드로를 응시하며 서로 마주봄.",
        "built_space": "중앙의 사다리 기둥과 갈라지는 두 하수도 통로가 정확히 배치됨.",
        "entities": "인물들의 외양과 복장, 양어깨를 쥔 두 손과 팔뚝의 위치가 완벽히 일치함.",
        "hard_violations": [],
        "physics": "어깨를 꽉 쥔 두 손과 몸의 자세가 중력과 구조에 맞게 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "진지한 삼사분면 얼굴과 양어깨를 쥔 두 손, 상반신 미디엄 구도가 더 충실하지만 먼 쪽 전완은 페드로에게 가려져 있다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "서로를 보는 시선과 어깨를 붙잡는 행동은 맞지만 페드로의 등이 전경을 과점유하고 먼 쪽 손까지 화면 끝에서 잘려, 지정된 양손·양쪽 전완 가시성이 더 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 눈과 얼굴은 왼쪽 앞의 페드로 얼굴을 향한다. 페드로도 고개를 이현우 쪽으로 돌리고 있으나 눈 자체는 보이지 않는다. 두 손은 각각 페드로의 양어깨에 놓여 있어 행동의 대상이 정확하다.",
        "built_space": "배경에 좌우 아치형 통로 두 개, 그 사이의 벽체, 중앙 뒤쪽의 수직 사다리 한 개와 상부 개구부 한 개가 보인다. 오른쪽 상단의 굵은 배관과 벽면의 가는 배관, 젖은 벽과 물바닥도 장소 참조에 부합한다. 두 사람은 갈림길 앞에 있으며 구조물과 충돌하지 않는다. 페드로의 뒤쪽에서 비껴 보는 상반신 구도이고 이현우의 얼굴은 중앙보다 조금 오른쪽에 있다. 다만 배경 분기가 비스듬히 물러나기보다는 비교적 정면으로 펼쳐진다.",
        "entities": "인물은 두 명뿐이다. 이현우는 참조와 유사한 앳된 동아시아계 남성 얼굴, 헝클어진 짧은 검은 머리, 마른 체격이며 얼굴 상처와 검은 인이어가 보인다. 겉옷 없이 어둡고 오염된 셔츠를 입고 있다. 페드로는 검은 비니와 낡은 올리브색 재킷을 착용해 참조의 복장과 맞지만 얼굴 대부분이 가려져 정확한 얼굴 정체성과 라틴계 혼혈 외형은 확인하기 어렵다. 하의와 다리 부상은 구도 밖이므로 판단하지 않는다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "가까운 손은 손가락으로 재킷 어깨를 감싸고, 먼 손도 반대편 어깨 위에 접촉한다. 가까운 전완은 몸까지 자연스럽게 이어진다. 먼 쪽 전완은 페드로의 몸에 가려져 보이지 않지만 손 위치와 팔의 진행 방향에 물리적 모순은 없다. 발은 화면 밖이며 상체에 부유를 암시하는 자세는 없다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽 위에 있는 페드로의 얼굴을 똑바로 바라본다. 페드로의 머리도 이현우를 향하고 있으나 눈은 가려져 있다. 가까운 손은 페드로의 가까운 어깨를 잡고, 왼쪽 끝의 다른 손은 반대편 어깨에 닿아 있다.",
        "built_space": "오른쪽 아치 통로 한 개가 뚜렷하고 왼쪽 통로 한 개는 페드로 뒤로 일부만 보인다. 중앙 뒤의 사다리 한 개와 천장 개구부 한 개, 상단의 굵은 배관 및 벽면 배관들이 장소 참조와 대응한다. 물바닥과 젖은 구조물의 규모도 자연스럽다. 다만 페드로의 머리와 등이 왼쪽 절반 이상을 크게 차지하여 어깨를 하단 왼쪽 가장자리에 두라는 지정보다 전경 비중이 크다. 이현우를 살짝 올려다보는 카메라 각도도 뚜렷하지 않다.",
        "entities": "이현우와 페드로 두 명만 보인다. 이현우의 앳된 동아시아계 남성 외형, 검은 머리, 마른 체격, 얼굴 상처, 인이어와 피·먼지가 묻은 어두운 셔츠는 요구와 대체로 맞는다. 페드로의 비니, 짙은 머리와 낡은 올리브색 재킷은 참조와 맞지만 뒷모습 위주라 얼굴과 정확한 정체성은 확인하기 어렵다. 하의와 다리 부상은 화면 밖이다. 불필요한 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "가까운 손과 전완은 자연스럽게 연결되고 손가락이 어깨 천을 눌러 잡고 있다. 먼 손은 화면 왼쪽 경계에서 일부 잘리지만 반대편 어깨와 접촉한 부분은 보인다. 그쪽 전완은 페드로 뒤에 가려진다. 이는 부유하거나 분리된 손의 증거는 아니며, 보이는 상체 자세도 가능한 동작이다. 다만 두 전완과 두 손을 모두 보여 달라는 명시적 조건은 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "진지한 삼사분면 얼굴과 양어깨를 쥔 두 손, 상반신 미디엄 구도가 더 충실하지만 먼 쪽 전완은 페드로에게 가려져 있다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "서로를 보는 시선과 어깨를 붙잡는 행동은 맞지만 페드로의 등이 전경을 과점유하고 먼 쪽 손까지 화면 끝에서 잘려, 지정된 양손·양쪽 전완 가시성이 더 떨어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 눈과 얼굴은 왼쪽 앞의 페드로 얼굴을 향한다. 페드로도 고개를 이현우 쪽으로 돌리고 있으나 눈 자체는 보이지 않는다. 두 손은 각각 페드로의 양어깨에 놓여 있어 행동의 대상이 정확하다.",
        "built_space": "배경에 좌우 아치형 통로 두 개, 그 사이의 벽체, 중앙 뒤쪽의 수직 사다리 한 개와 상부 개구부 한 개가 보인다. 오른쪽 상단의 굵은 배관과 벽면의 가는 배관, 젖은 벽과 물바닥도 장소 참조에 부합한다. 두 사람은 갈림길 앞에 있으며 구조물과 충돌하지 않는다. 페드로의 뒤쪽에서 비껴 보는 상반신 구도이고 이현우의 얼굴은 중앙보다 조금 오른쪽에 있다. 다만 배경 분기가 비스듬히 물러나기보다는 비교적 정면으로 펼쳐진다.",
        "entities": "인물은 두 명뿐이다. 이현우는 참조와 유사한 앳된 동아시아계 남성 얼굴, 헝클어진 짧은 검은 머리, 마른 체격이며 얼굴 상처와 검은 인이어가 보인다. 겉옷 없이 어둡고 오염된 셔츠를 입고 있다. 페드로는 검은 비니와 낡은 올리브색 재킷을 착용해 참조의 복장과 맞지만 얼굴 대부분이 가려져 정확한 얼굴 정체성과 라틴계 혼혈 외형은 확인하기 어렵다. 하의와 다리 부상은 구도 밖이므로 판단하지 않는다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "가까운 손은 손가락으로 재킷 어깨를 감싸고, 먼 손도 반대편 어깨 위에 접촉한다. 가까운 전완은 몸까지 자연스럽게 이어진다. 먼 쪽 전완은 페드로의 몸에 가려져 보이지 않지만 손 위치와 팔의 진행 방향에 물리적 모순은 없다. 발은 화면 밖이며 상체에 부유를 암시하는 자세는 없다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽 위에 있는 페드로의 얼굴을 똑바로 바라본다. 페드로의 머리도 이현우를 향하고 있으나 눈은 가려져 있다. 가까운 손은 페드로의 가까운 어깨를 잡고, 왼쪽 끝의 다른 손은 반대편 어깨에 닿아 있다.",
        "built_space": "오른쪽 아치 통로 한 개가 뚜렷하고 왼쪽 통로 한 개는 페드로 뒤로 일부만 보인다. 중앙 뒤의 사다리 한 개와 천장 개구부 한 개, 상단의 굵은 배관 및 벽면 배관들이 장소 참조와 대응한다. 물바닥과 젖은 구조물의 규모도 자연스럽다. 다만 페드로의 머리와 등이 왼쪽 절반 이상을 크게 차지하여 어깨를 하단 왼쪽 가장자리에 두라는 지정보다 전경 비중이 크다. 이현우를 살짝 올려다보는 카메라 각도도 뚜렷하지 않다.",
        "entities": "이현우와 페드로 두 명만 보인다. 이현우의 앳된 동아시아계 남성 외형, 검은 머리, 마른 체격, 얼굴 상처, 인이어와 피·먼지가 묻은 어두운 셔츠는 요구와 대체로 맞는다. 페드로의 비니, 짙은 머리와 낡은 올리브색 재킷은 참조와 맞지만 뒷모습 위주라 얼굴과 정확한 정체성은 확인하기 어렵다. 하의와 다리 부상은 화면 밖이다. 불필요한 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "가까운 손과 전완은 자연스럽게 연결되고 손가락이 어깨 천을 눌러 잡고 있다. 먼 손은 화면 왼쪽 경계에서 일부 잘리지만 반대편 어깨와 접촉한 부분은 보인다. 그쪽 전완은 페드로 뒤에 가려진다. 이는 부유하거나 분리된 손의 증거는 아니며, 보이는 상체 자세도 가능한 동작이다. 다만 두 전완과 두 손을 모두 보여 달라는 명시적 조건은 충족하지 못한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.208,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.958,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 화면 좌측에 위치한 이현우의 오른손 엄지손가락이 바깥쪽을 향하고 있어 해부학적으로 불가능함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 958,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 958,
    "verdict_ko": "해부학적으로 불가능한 손 구조(오른손의 엄지 방향 오류)로 인해 치명적 위반이 발생했으며, 프레이밍 지시도 완벽히 따르지 않았습니다.  ★위반: [gemini-pro] 화면 좌측에 위치한 이현우의 오른손 엄지손가락이 바깥쪽을 향하고 있어 해부학적으로 불가능함"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요청된 숄더뷰 앵글과 인물의 배치, 배경의 하수도 갈림길 등 모든 세부 사항을 사실적이고 정확하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_sewer_junction_75d7f3.png",
    "asset_id": "4d4807b6-ee2e-48eb-91c9-6bd1ad304eb5",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0993-3140-75e5-b6ea-4aeb0b3e4149",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S26sh12__bgfirst_bg.png",
   "bg_asset_id": "56eea23e-4f38-4674-bd99-72ea4b90ee0d",
   "bg_record_key": "S26sh12::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "sewer_junction",
   "groupbg_asset_id": "4d4807b6-ee2e-48eb-91c9-6bd1ad304eb5"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S27sh11::signage": {
  "fp": "7a43da9d48a5fdb4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::van_pickup_street": {
  "input_fingerprint": "f0195c76de6f0315",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "van_pickup_street",
    "tags": [
     "S27sh11",
     "S27sh17",
     "S27sh19"
    ]
   },
   "context_sig": "7306bb12dfdb6980"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 도로를 따라 빠르게 걷는 현우 일행.\n- 황급히 찰리 앞에 멈춰 서는 자동차.\n- 차에 오르는 찰리를 눈여겨보는 누군가의 시선(이하 구도환)\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 거리·컨테이너 골목, 시장통, 전경, 비탈길, 입구·진입로, 공터 피로연장, 골목 쓰레기 집하장·맨홀 입구, 해일이 덮치는 거리: 겹겹이 쌓인 컨테이너 주택과 노점상, 좁은 골목이 얽힌 근미래 빈민가. 밤이 되면 전력 통제로 일제히 어두워진다. (특징: '난민 거주지역' 표지판이 걸린 철조망 입구; 거미줄처럼 얽힌 전선과 녹슨 컨테이너 박스들; 노점상(바나나 등 과일)과 통행하는 다양한 복장의 난민들; 자체 제작 제복과 완장을 착용한 박철진과 민병대원들; 밤이 되어 가로등과 불빛이 일제히 꺼지는 정전 연출; 찰리의 가슴 링이 푸르게 발광하며 주변 가로등과 전구들이 터질 듯이 환해지는 광경; 음악에 맞춰 문워크를 추는 찰리와 하객들의 피로연 공터 춤판; 쓰레기 더미 아래 숨겨진 육중한 금속 맨홀 뚜껑; 인공제방 붕괴 후 거대한 바닷물 해일이 컨테이너와 사람들을 덮치는 시각적 재난 묘사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 도로를 따라 빠르게 걷는 현우 일행.\n- 황급히 찰리 앞에 멈춰 서는 자동차.\n- 차에 오르는 찰리를 눈여겨보는 누군가의 시선(이하 구도환)\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_van_pickup_street_41dfb3.png",
  "asset_id": "5a81c29f-f19e-40c7-87af-19982c6fea88",
  "input_asset_ids": [
   "8f5f9ec5-20ce-4a35-9f05-3aa7f9c47f3a"
  ],
  "origin_tag": "S27sh11",
  "place_text": "On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.",
  "origin_inputs": {
   "place_text": "On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.",
   "time_of_day_en": "night",
   "conti_asset_id": "8f5f9ec5-20ce-4a35-9f05-3aa7f9c47f3a"
  }
 },
 "S27sh11::bgfirst_bg": {
  "input_fingerprint": "b7fe47c931acfea9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 다리 바로 앞에서 거친 마찰을 일으키며 멈춰 선 자동차의 앞범퍼 근접 찰나.\n\nLOCATION (lock): On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral tracking move outside the vehicle's path, looking diagonally across the narrow gap between 찰리's lower legs at left and the stopped front bumper at right, never along the vehicle's frontal axis. Tighten only the viewing distance at this endpoint: the bumper occupies less than a third of the image, with road visible beneath the gap and 찰리's uneven foot placement retaining the interrupted running action. His head and eyeline remain above the crop as his attention turns toward the vehicle's headlights.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 찰리의 하퇴와 발 in the middle-left of the frame, midground; Stopped front bumper with a visible gap before Charlie's legs in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 자동차 앞범퍼 (The vehicle has stopped immediately before 찰리 without contact) — The front corner and a short portion of the bumper's side are visible obliquely; used as Establishes the near collision through the remaining gap, without exaggerated foreground scale; 도로 (Visible underneath the stationary bumper and 찰리's feet); used as Maintains a continuous ground plane and a credible scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established headlights give the stopping moment sharp local contrast while retaining precise detail in 찰리's hard-surface legs.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 다리 바로 앞에서 거친 마찰을 일으키며 멈춰 선 자동차의 앞범퍼 근접 찰나.\n\nLOCATION (lock): On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral tracking move outside the vehicle's path, looking diagonally across the narrow gap between 찰리's lower legs at left and the stopped front bumper at right, never along the vehicle's frontal axis. Tighten only the viewing distance at this endpoint: the bumper occupies less than a third of the image, with road visible beneath the gap and 찰리's uneven foot placement retaining the interrupted running action. His head and eyeline remain above the crop as his attention turns toward the vehicle's headlights.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 찰리의 하퇴와 발 in the middle-left of the frame, midground; Stopped front bumper with a visible gap before Charlie's legs in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 자동차 앞범퍼 (The vehicle has stopped immediately before 찰리 without contact) — The front corner and a short portion of the bumper's side are visible obliquely; used as Establishes the near collision through the remaining gap, without exaggerated foreground scale; 도로 (Visible underneath the stationary bumper and 찰리's feet); used as Maintains a continuous ground plane and a credible scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established headlights give the stopping moment sharp local contrast while retaining precise detail in 찰리's hard-surface legs.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S27sh11__bgfirst_bg.png",
  "asset_id": "551e6bb7-a5ab-40f4-8d76-0e7157383116",
  "input_asset_ids": [
   "8f5f9ec5-20ce-4a35-9f05-3aa7f9c47f3a",
   "5a81c29f-f19e-40c7-87af-19982c6fea88"
  ]
 },
 "S27sh11": {
  "input_fingerprint": "e2f0312416e2ae40",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 다리 바로 앞에서 거친 마찰을 일으키며 멈춰 선 자동차의 앞범퍼 근접 찰나.\n\nLOCATION (lock): On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral tracking move outside the vehicle's path, looking diagonally across the narrow gap between 찰리's lower legs at left and the stopped front bumper at right, never along the vehicle's frontal axis. Tighten only the viewing distance at this endpoint: the bumper occupies less than a third of the image, with road visible beneath the gap and 찰리's uneven foot placement retaining the interrupted running action. His head and eyeline remain above the crop as his attention turns toward the vehicle's headlights.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 찰리의 하퇴와 발 in the middle-left of the frame, midground; Stopped front bumper with a visible gap before Charlie's legs in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 자동차 앞범퍼 (The vehicle has stopped immediately before 찰리 without contact) — The front corner and a short portion of the bumper's side are visible obliquely; used as Establishes the near collision through the remaining gap, without exaggerated foreground scale; 도로 (Visible underneath the stationary bumper and 찰리's feet); used as Maintains a continuous ground plane and a credible scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established headlights give the stopping moment sharp local contrast while retaining precise detail in 찰리's hard-surface legs.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has stopped immediately before Charlie with its headlights on; streetlights continue switching on and off along his route. Charlie retains his coat and hat disguise and blue-lit eyes, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 다리 바로 앞에서 거친 마찰을 일으키며 멈춰 선 자동차의 앞범퍼 근접 찰나.\n\nLOCATION (lock): On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral tracking move outside the vehicle's path, looking diagonally across the narrow gap between 찰리's lower legs at left and the stopped front bumper at right, never along the vehicle's frontal axis. Tighten only the viewing distance at this endpoint: the bumper occupies less than a third of the image, with road visible beneath the gap and 찰리's uneven foot placement retaining the interrupted running action. His head and eyeline remain above the crop as his attention turns toward the vehicle's headlights.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 찰리의 하퇴와 발 in the middle-left of the frame, midground; Stopped front bumper with a visible gap before Charlie's legs in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 자동차 앞범퍼 (The vehicle has stopped immediately before 찰리 without contact) — The front corner and a short portion of the bumper's side are visible obliquely; used as Establishes the near collision through the remaining gap, without exaggerated foreground scale; 도로 (Visible underneath the stationary bumper and 찰리's feet); used as Maintains a continuous ground plane and a credible scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established headlights give the stopping moment sharp local contrast while retaining precise detail in 찰리's hard-surface legs.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has stopped immediately before Charlie with its headlights on; streetlights continue switching on and off along his route. Charlie retains his coat and hat disguise and blue-lit eyes, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 다리 바로 앞에서 거친 마찰을 일으키며 멈춰 선 자동차의 앞범퍼 근접 찰나.\n\nLOCATION (lock): On the refugee-settlement roadway immediately in front of a braking van, within its headlight beams. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral tracking move outside the vehicle's path, looking diagonally across the narrow gap between 찰리's lower legs at left and the stopped front bumper at right, never along the vehicle's frontal axis. Tighten only the viewing distance at this endpoint: the bumper occupies less than a third of the image, with road visible beneath the gap and 찰리's uneven foot placement retaining the interrupted running action. His head and eyeline remain above the crop as his attention turns toward the vehicle's headlights.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 찰리의 하퇴와 발 in the middle-left of the frame, midground; Stopped front bumper with a visible gap before Charlie's legs in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 자동차 앞범퍼 (The vehicle has stopped immediately before 찰리 without contact) — The front corner and a short portion of the bumper's side are visible obliquely; used as Establishes the near collision through the remaining gap, without exaggerated foreground scale; 도로 (Visible underneath the stationary bumper and 찰리's feet); used as Maintains a continuous ground plane and a credible scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established headlights give the stopping moment sharp local contrast while retaining precise detail in 찰리's hard-surface legs.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has stopped immediately before Charlie with its headlights on; streetlights continue switching on and off along his route. Charlie retains his coat and hat disguise and blue-lit eyes, and the small bird has flown away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S27sh11__bgfirst_bg.png",
     "asset_id": "551e6bb7-a5ab-40f4-8d76-0e7157383116",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S27sh11.png",
     "asset_id": "8f5f9ec5-20ce-4a35-9f05-3aa7f9c47f3a",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_van_pickup_street_41dfb3.png",
     "asset_id": "5a81c29f-f19e-40c7-87af-19982c6fea88",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 찰리의 다리와 자동차 앞범퍼 사이의 간격을 대각선으로 바라보고 있음.",
    "built_space": "야간의 난민 거주지 도로. 젖은 노면과 배경의 컨테이너, 전신주 등이 로케이션 레퍼런스와 정확히 일치함.",
    "entities": "샌드 베이지색 기계 장갑판으로 이루어진 찰리의 하퇴와 발, 코트 자락, 그리고 자동차의 앞범퍼가 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "찰리의 한쪽 발은 평평하게 딛고 다른 쪽 발은 뒤꿈치가 들려 있어 달리다 멈춘 물리적 관성과 체중 이동이 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 찰리의 하반신 및 늘어진 팔과 자동차 앞범퍼 사이의 틈을 대각선으로 바라보고 있음.",
    "built_space": "야간의 젖은 도로와 컨테이너 배경 등 로케이션 레퍼런스의 요소를 충실히 반영함.",
    "entities": "찰리의 기계 다리와 팔, 코트 자락, 자동차 앞범퍼가 묘사됨.",
    "hard_violations": [],
    "physics": "두 발이 모두 땅에 평평하게 닿아 있어 정지 상태로 지지되고 있으며, 달리다 멈춘 찰나의 관성이 느껴지지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "달려오다 멈춘 듯한 역동적인 발의 자세(들린 뒤꿈치)를 잘 표현하였으며, 명시된 하퇴와 발 위주의 프레이밍을 매우 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍과 배경 및 캐릭터의 재질 구현은 훌륭하나, 두 발이 평평하게 땅에 닿아 있어 지시문이 요구한 '달리다 멈춘 동작'의 역동성이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 찰리의 다리와 자동차 앞범퍼 사이의 간격을 대각선으로 바라보고 있음.",
        "built_space": "야간의 난민 거주지 도로. 젖은 노면과 배경의 컨테이너, 전신주 등이 로케이션 레퍼런스와 정확히 일치함.",
        "entities": "샌드 베이지색 기계 장갑판으로 이루어진 찰리의 하퇴와 발, 코트 자락, 그리고 자동차의 앞범퍼가 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 한쪽 발은 평평하게 딛고 다른 쪽 발은 뒤꿈치가 들려 있어 달리다 멈춘 물리적 관성과 체중 이동이 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 찰리의 하반신 및 늘어진 팔과 자동차 앞범퍼 사이의 틈을 대각선으로 바라보고 있음.",
        "built_space": "야간의 젖은 도로와 컨테이너 배경 등 로케이션 레퍼런스의 요소를 충실히 반영함.",
        "entities": "찰리의 기계 다리와 팔, 코트 자락, 자동차 앞범퍼가 묘사됨.",
        "hard_violations": [],
        "physics": "두 발이 모두 땅에 평평하게 닿아 있어 정지 상태로 지지되고 있으며, 달리다 멈춘 찰나의 관성이 느껴지지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "달려오다 멈춘 듯한 역동적인 발의 자세(들린 뒤꿈치)를 잘 표현하였으며, 명시된 하퇴와 발 위주의 프레이밍을 매우 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍과 배경 및 캐릭터의 재질 구현은 훌륭하나, 두 발이 평평하게 땅에 닿아 있어 지시문이 요구한 '달리다 멈춘 동작'의 역동성이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 찰리의 다리와 자동차 앞범퍼 사이의 간격을 대각선으로 바라보고 있음.",
        "built_space": "야간의 난민 거주지 도로. 젖은 노면과 배경의 컨테이너, 전신주 등이 로케이션 레퍼런스와 정확히 일치함.",
        "entities": "샌드 베이지색 기계 장갑판으로 이루어진 찰리의 하퇴와 발, 코트 자락, 그리고 자동차의 앞범퍼가 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 한쪽 발은 평평하게 딛고 다른 쪽 발은 뒤꿈치가 들려 있어 달리다 멈춘 물리적 관성과 체중 이동이 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 찰리의 하반신 및 늘어진 팔과 자동차 앞범퍼 사이의 틈을 대각선으로 바라보고 있음.",
        "built_space": "야간의 젖은 도로와 컨테이너 배경 등 로케이션 레퍼런스의 요소를 충실히 반영함.",
        "entities": "찰리의 기계 다리와 팔, 코트 자락, 자동차 앞범퍼가 묘사됨.",
        "hard_violations": [],
        "physics": "두 발이 모두 땅에 평평하게 닿아 있어 정지 상태로 지지되고 있으며, 달리다 멈춘 찰나의 관성이 느껴지지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "비접촉 간격과 짧고 육중한 기계 다리는 잘 구현했지만, 손과 허벅지까지 크게 들어오고 가까운 발이 강조되어 지정된 중경 하퇴 중심 구도에서 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "하퇴 중심의 크롭, 앞뒤로 어긋난 발, 우측 범퍼와의 간격 및 장소 재현이 더 충실하나, 다리가 참조보다 길고 차량 측면 노출은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 시선은 화면 밖이라 전조등을 바라보는지는 확인할 수 없다. 두 발끝은 화면 오른쪽 앞쪽을 향하고, 차량 전면은 왼쪽 전방의 찰리 쪽을 향한다. 전조등이 다리와 그 앞 노면을 비춘다. 카메라는 차량 정면축에서 비껴나 있지만 범퍼 측면보다는 그릴과 정면이 많이 보인다.",
        "built_space": "왼쪽에는 철망과 임시 점포, 오른쪽에는 적층 컨테이너와 외부 계단 한 곳, 거리에는 가로등과 전선이 보인다. 차량에는 번호판 한 장, 원형 안개등 하나, 그릴 하나와 전조등 두 곳의 일부가 보이며 부품 중복은 없다. 왼쪽 다리와 오른쪽 범퍼 사이에 도로가 이어지고 접촉하지 않는다. 범퍼는 대략 우측 3분의 1 안에 있으나 가까운 발과 손이 전경처럼 크게 보인다.",
        "entities": "찰리 한 개체의 두 다리와 한 손, 외투 자락이 보인다. 샌드 베이지 장갑판, 노출 관절, 짧고 굵은 다리와 분절된 발은 기계 참조에 가깝다. 얼굴·눈·모자는 크롭 밖이라 평가하지 않는다. 낡은 밴과 젖은 도로, 야간 전조등은 요청과 맞으며 새나 다른 사람은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발의 발바닥과 발가락이 젖은 도로에 닿아 몸을 지탱한다. 무릎은 굽혀져 있고 양발의 깊이가 달라 정지 직전 체중을 받는 자세로 가능하지만, 달리기의 중단보다는 낮게 버티는 자세가 강하다. 차량은 오른쪽 아래 보이는 타이어로 지지되며 범퍼는 차체에 연결되어 있다. 거친 제동 마찰의 직접적인 흔적은 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "머리와 시선은 위로 잘려 보이지 않는다. 범퍼에 가까운 발은 오른쪽을 향하고 다른 발은 왼쪽 뒤에 놓여 비대칭 보폭을 만든다. 차량 전면과 켜진 전조등은 찰리 쪽을 향한다. 카메라는 낮은 사선 위치이나 차량의 짧은 측면보다 정면 그릴이 우세하게 보인다.",
        "built_space": "왼쪽 철조망 울타리와 현수막 한 장, 그 뒤 컨테이너, 차양 아래 점포와 진열대, 오른쪽 컨테이너와 외부 계단 한 곳이 장소 참조와 가깝다. 전선과 가로등이 도로 깊이 방향으로 이어진다. 차량에는 번호판 한 장, 원형 안개등 하나, 그릴 하나와 전조등 두 곳의 일부가 보인다. 하퇴는 중간 왼쪽, 범퍼는 오른쪽에 놓이고 그 사이와 아래의 도로가 끊기지 않는다. 젖은 노면의 광원 반사도 가능한 배치다.",
        "entities": "외투 아래 찰리의 기계 다리 두 개와 분절된 발이 보인다. 베이지색 장갑판과 원형 발목 관절은 참조에 부합하지만, 하퇴가 참조보다 길고 날씬하다. 얼굴·푸른 눈·모자는 프레임 밖이다. 밴, 전조등, 도로가 있으며 다른 사람과 새는 없다. 현수막의 '난민 거주구역'은 장소 참조에 있는 문구다.",
        "hard_violations": [],
        "physics": "앞쪽 발과 뒤쪽 발 모두 도로에 닿아 있으며, 서로 다른 깊이와 각도로 몸을 지지한다. 달리다가 보폭을 벌린 채 멈추는 자세로 가능하고 공중에 뜬 신체는 없다. 차량은 보이는 앞바퀴로 도로에 지지되고 범퍼는 차체에 붙어 있다. 범퍼와 가까운 발 사이에는 명확한 비접촉 간격이 있다. 낮게 퍼진 옅은 안개성 흔적은 있지만 거친 제동 마찰 자체는 명확하지 않다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "비접촉 간격과 짧고 육중한 기계 다리는 잘 구현했지만, 손과 허벅지까지 크게 들어오고 가까운 발이 강조되어 지정된 중경 하퇴 중심 구도에서 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "하퇴 중심의 크롭, 앞뒤로 어긋난 발, 우측 범퍼와의 간격 및 장소 재현이 더 충실하나, 다리가 참조보다 길고 차량 측면 노출은 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 시선은 화면 밖이라 전조등을 바라보는지는 확인할 수 없다. 두 발끝은 화면 오른쪽 앞쪽을 향하고, 차량 전면은 왼쪽 전방의 찰리 쪽을 향한다. 전조등이 다리와 그 앞 노면을 비춘다. 카메라는 차량 정면축에서 비껴나 있지만 범퍼 측면보다는 그릴과 정면이 많이 보인다.",
        "built_space": "왼쪽에는 철망과 임시 점포, 오른쪽에는 적층 컨테이너와 외부 계단 한 곳, 거리에는 가로등과 전선이 보인다. 차량에는 번호판 한 장, 원형 안개등 하나, 그릴 하나와 전조등 두 곳의 일부가 보이며 부품 중복은 없다. 왼쪽 다리와 오른쪽 범퍼 사이에 도로가 이어지고 접촉하지 않는다. 범퍼는 대략 우측 3분의 1 안에 있으나 가까운 발과 손이 전경처럼 크게 보인다.",
        "entities": "찰리 한 개체의 두 다리와 한 손, 외투 자락이 보인다. 샌드 베이지 장갑판, 노출 관절, 짧고 굵은 다리와 분절된 발은 기계 참조에 가깝다. 얼굴·눈·모자는 크롭 밖이라 평가하지 않는다. 낡은 밴과 젖은 도로, 야간 전조등은 요청과 맞으며 새나 다른 사람은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발의 발바닥과 발가락이 젖은 도로에 닿아 몸을 지탱한다. 무릎은 굽혀져 있고 양발의 깊이가 달라 정지 직전 체중을 받는 자세로 가능하지만, 달리기의 중단보다는 낮게 버티는 자세가 강하다. 차량은 오른쪽 아래 보이는 타이어로 지지되며 범퍼는 차체에 연결되어 있다. 거친 제동 마찰의 직접적인 흔적은 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "머리와 시선은 위로 잘려 보이지 않는다. 범퍼에 가까운 발은 오른쪽을 향하고 다른 발은 왼쪽 뒤에 놓여 비대칭 보폭을 만든다. 차량 전면과 켜진 전조등은 찰리 쪽을 향한다. 카메라는 낮은 사선 위치이나 차량의 짧은 측면보다 정면 그릴이 우세하게 보인다.",
        "built_space": "왼쪽 철조망 울타리와 현수막 한 장, 그 뒤 컨테이너, 차양 아래 점포와 진열대, 오른쪽 컨테이너와 외부 계단 한 곳이 장소 참조와 가깝다. 전선과 가로등이 도로 깊이 방향으로 이어진다. 차량에는 번호판 한 장, 원형 안개등 하나, 그릴 하나와 전조등 두 곳의 일부가 보인다. 하퇴는 중간 왼쪽, 범퍼는 오른쪽에 놓이고 그 사이와 아래의 도로가 끊기지 않는다. 젖은 노면의 광원 반사도 가능한 배치다.",
        "entities": "외투 아래 찰리의 기계 다리 두 개와 분절된 발이 보인다. 베이지색 장갑판과 원형 발목 관절은 참조에 부합하지만, 하퇴가 참조보다 길고 날씬하다. 얼굴·푸른 눈·모자는 프레임 밖이다. 밴, 전조등, 도로가 있으며 다른 사람과 새는 없다. 현수막의 '난민 거주구역'은 장소 참조에 있는 문구다.",
        "hard_violations": [],
        "physics": "앞쪽 발과 뒤쪽 발 모두 도로에 닿아 있으며, 서로 다른 깊이와 각도로 몸을 지지한다. 달리다가 보폭을 벌린 채 멈추는 자세로 가능하고 공중에 뜬 신체는 없다. 차량은 보이는 앞바퀴로 도로에 지지되고 범퍼는 차체에 붙어 있다. 범퍼와 가까운 발 사이에는 명확한 비접촉 간격이 있다. 낮게 퍼진 옅은 안개성 흔적은 있지만 거친 제동 마찰 자체는 명확하지 않다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.653
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.653
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1653
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "달려오다 멈춘 듯한 역동적인 발의 자세(들린 뒤꿈치)를 잘 표현하였으며, 명시된 하퇴와 발 위주의 프레이밍을 매우 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1653,
    "verdict_ko": "프레이밍과 배경 및 캐릭터의 재질 구현은 훌륭하나, 두 발이 평평하게 땅에 닿아 있어 지시문이 요구한 '달리다 멈춘 동작'의 역동성이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_van_pickup_street_41dfb3.png",
    "asset_id": "5a81c29f-f19e-40c7-87af-19982c6fea88",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab099d-81ed-76dc-b998-31cc0436dff1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S27sh11__bgfirst_bg.png",
   "bg_asset_id": "551e6bb7-a5ab-40f4-8d76-0e7157383116",
   "bg_record_key": "S27sh11::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "van_pickup_street",
   "groupbg_asset_id": "5a81c29f-f19e-40c7-87af-19982c6fea88"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S27sh17::signage": {
  "fp": "2637e64821ef0281",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh17": {
  "input_fingerprint": "adf339cafef71d47",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 불빛 아래 얼굴이 드러난 신부가 찰리를 바라보며 믿기지 않는 듯 당황한 표정을 띤 구도.\n\nLOCATION (lock): On the road in front of the stopped van, where its headlights reveal the man who has stepped out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the rising camera beside and slightly behind 찰리, below 신부's eye level, looking gently upward at the priest's three-quarter face without crossing the established vehicle–찰리 axis. 찰리's head and near shoulder occupy the lower-left foreground as he looks up at 신부, whose halted approach and bewildered expression occupy the right midground. Hold their positions and let the priest's newly readable face become the single emphasis of the reveal, his gaze directed down toward 찰리 rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 멈춰 선 차량 (Stationary behind the priest's approach) — Only an oblique portion of the vehicle's front remains at the rear edge of the composition; used as Preserves the geography of the stop without competing with the priest's face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established headlight contrast while allowing 신부's now-revealed facial features to read clearly, without adding a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights shining, and nearby streetlights retain their intermittent switching. Charlie remains before the vehicle in his old coat and hat disguise; the bird is no longer perched on him. 신부: He has stepped out of the van and stands near its front, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 불빛 아래 얼굴이 드러난 신부가 찰리를 바라보며 믿기지 않는 듯 당황한 표정을 띤 구도.\n\nLOCATION (lock): On the road in front of the stopped van, where its headlights reveal the man who has stepped out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the rising camera beside and slightly behind 찰리, below 신부's eye level, looking gently upward at the priest's three-quarter face without crossing the established vehicle–찰리 axis. 찰리's head and near shoulder occupy the lower-left foreground as he looks up at 신부, whose halted approach and bewildered expression occupy the right midground. Hold their positions and let the priest's newly readable face become the single emphasis of the reveal, his gaze directed down toward 찰리 rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 멈춰 선 차량 (Stationary behind the priest's approach) — Only an oblique portion of the vehicle's front remains at the rear edge of the composition; used as Preserves the geography of the stop without competing with the priest's face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established headlight contrast while allowing 신부's now-revealed facial features to read clearly, without adding a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights shining, and nearby streetlights retain their intermittent switching. Charlie remains before the vehicle in his old coat and hat disguise; the bird is no longer perched on him. 신부: He has stepped out of the van and stands near its front, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 불빛 아래 얼굴이 드러난 신부가 찰리를 바라보며 믿기지 않는 듯 당황한 표정을 띤 구도.\n\nLOCATION (lock): On the road in front of the stopped van, where its headlights reveal the man who has stepped out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the rising camera beside and slightly behind 찰리, below 신부's eye level, looking gently upward at the priest's three-quarter face without crossing the established vehicle–찰리 axis. 찰리's head and near shoulder occupy the lower-left foreground as he looks up at 신부, whose halted approach and bewildered expression occupy the right midground. Hold their positions and let the priest's newly readable face become the single emphasis of the reveal, his gaze directed down toward 찰리 rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 멈춰 선 차량 (Stationary behind the priest's approach) — Only an oblique portion of the vehicle's front remains at the rear edge of the composition; used as Preserves the geography of the stop without competing with the priest's face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the established headlight contrast while allowing 신부's now-revealed facial features to read clearly, without adding a new source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van remains stopped with its headlights shining, and nearby streetlights retain their intermittent switching. Charlie remains before the vehicle in his old coat and hat disguise; the bird is no longer perched on him. 신부: He has stepped out of the van and stands near its front, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "신부는 아래로 찰리를, 찰리는 위로 신부를 향해 시선을 둠.",
    "built_space": "젖은 도로 위, 배경에 컨테이너가 있으며 우측에 정차된 차량의 전면부가 보임.",
    "entities": "찰리는 코트와 모자를 쓴 기계 몸체. 신부는 60대 남성이나 레퍼런스와 달리 라펠이 있는 양복 재킷을 입음.",
    "hard_violations": [],
    "physics": "두 사람 모두 지면에 서 있으며 지지 상태에 무리가 없음."
   },
   {
    "label": "B",
    "direction": "신부는 시선을 아래로 향해 찰리를 바라보고, 찰리는 고개를 들어 신부를 향함.",
    "built_space": "젖은 도로, 배경의 컨테이너, 우측에 멈춰 선 밴과 헤드라이트가 위치함.",
    "entities": "찰리는 모자와 코트를 착용한 기계 형태. 신부는 레퍼런스와 일치하는 라펠 없는 사제복을 착용함.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 서서 안정적인 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 앵글과 샷 크기를 훌륭하게 구현했으며, 신부의 당황한 표정과 사제복 의상 디테일까지 레퍼런스를 정확히 반영했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "카메라 구도와 배경은 지시문과 일치하나, 신부의 의상이 레퍼런스 이미지와 달리 일반적인 양복 재킷으로 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "신부는 시선을 아래로 향해 찰리를 바라보고, 찰리는 고개를 들어 신부를 향함.",
        "built_space": "젖은 도로, 배경의 컨테이너, 우측에 멈춰 선 밴과 헤드라이트가 위치함.",
        "entities": "찰리는 모자와 코트를 착용한 기계 형태. 신부는 레퍼런스와 일치하는 라펠 없는 사제복을 착용함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서서 안정적인 자세를 유지함."
       },
       {
        "label": "A",
        "direction": "신부는 아래로 찰리를, 찰리는 위로 신부를 향해 시선을 둠.",
        "built_space": "젖은 도로 위, 배경에 컨테이너가 있으며 우측에 정차된 차량의 전면부가 보임.",
        "entities": "찰리는 코트와 모자를 쓴 기계 몸체. 신부는 60대 남성이나 레퍼런스와 달리 라펠이 있는 양복 재킷을 입음.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 서 있으며 지지 상태에 무리가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 앵글과 샷 크기를 훌륭하게 구현했으며, 신부의 당황한 표정과 사제복 의상 디테일까지 레퍼런스를 정확히 반영했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "카메라 구도와 배경은 지시문과 일치하나, 신부의 의상이 레퍼런스 이미지와 달리 일반적인 양복 재킷으로 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 시선을 아래로 향해 찰리를 바라보고, 찰리는 고개를 들어 신부를 향함.",
        "built_space": "젖은 도로, 배경의 컨테이너, 우측에 멈춰 선 밴과 헤드라이트가 위치함.",
        "entities": "찰리는 모자와 코트를 착용한 기계 형태. 신부는 레퍼런스와 일치하는 라펠 없는 사제복을 착용함.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서서 안정적인 자세를 유지함."
       },
       {
        "label": "A",
        "direction": "신부는 아래로 찰리를, 찰리는 위로 신부를 향해 시선을 둠.",
        "built_space": "젖은 도로 위, 배경에 컨테이너가 있으며 우측에 정차된 차량의 전면부가 보임.",
        "entities": "찰리는 코트와 모자를 쓴 기계 몸체. 신부는 60대 남성이나 레퍼런스와 달리 라펠이 있는 양복 재킷을 입음.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 서 있으며 지지 상태에 무리가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "찰리 뒤의 완만한 로우앵글 미디엄 숏, 찰리를 내려다보는 당황한 얼굴, 참조와 가까운 사제복을 충실히 구현한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "시선과 당황한 표정은 맞지만, 더 기울어진 앙각과 넓게 드러난 차량, 참조에 없는 라펠 재킷이 A보다 덜 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부는 고개를 왼쪽 아래로 기울여 전경의 찰리 얼굴을 바라보며, 렌즈를 응시하지 않는다. 찰리의 뒤통수와 옆면은 오른쪽 신부 쪽으로 향한다. 찰리의 눈은 가려져 정확한 눈동자 방향은 확인할 수 없다. 차량 전조등은 차량 앞 도로와 카메라 쪽으로 켜져 있다.",
        "built_space": "왼쪽 전경에 찰리의 머리와 어깨, 오른쪽 중경에 신부가 배치되어 있다. 오른쪽 가장자리에는 밴 한 대의 비스듬한 전면 일부와 켜진 전조등 한 개, 그릴 일부, 번호판 한 개가 보인다. 젖은 도로 양옆의 컨테이너 구조물, 왼쪽 철망, 상부 전선과 여러 가로등이 이전 장소의 재질과 야간 환경을 잇는다. 신부 눈높이 아래에서 올려다보는 미디엄 숏이며, 찰리의 머리는 지정된 왼쪽 아래보다 위쪽까지 크게 차지한다.",
        "entities": "등장 개체는 신부와 찰리뿐이다. 신부는 참조와 가까운 짧은 회색 머리와 얼굴을 가진 60대 한국인 남성으로 보이며, 흰 로만칼라와 라펠 없는 검은 단추식 사제복이 참조에 가깝다. 찰리는 낡고 젖은 모자와 외투를 착용하고, 노출된 머리에는 샌드 베이지 장갑판과 원형 기계 관절이 보인다. 가려진 얼굴의 흰 마스크와 눈은 검증할 수 없다. 새나 추가 인물, 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 인물의 상체는 화면 아래로 자연스럽게 이어지며, 발이 프레임 밖이라는 이유로 부유한다고 볼 근거는 없다. 신부의 손은 몸 옆에 내려와 있고 얼굴과 상체를 찰리 쪽으로 조금 기울인 자세가 가능하다. 모자는 찰리의 머리에 얹혀 있고 외투는 어깨를 감싼다. 차량과 인물이 공중에 떠 있는 징후는 없으며, 도로의 불빛 반사도 젖은 노면과 부합한다."
       },
       {
        "label": "B",
        "direction": "신부의 시선은 왼쪽 아래 찰리의 머리로 향하고, 살짝 벌어진 입과 모인 눈썹이 믿기 어려워하는 반응을 드러낸다. 찰리의 머리는 오른쪽 위 신부를 향하지만 눈 자체는 보이지 않는다. 밴의 켜진 전조등은 전방 도로와 카메라 방향을 비춘다.",
        "built_space": "찰리는 왼쪽 전경, 신부는 오른쪽 중경에 서 있으며 오른쪽 뒤에 밴 한 대가 있다. 전조등 한 개, 그릴 일부와 비교적 넓은 앞유리가 보인다. 젖은 도로, 컨테이너 건물, 오른쪽 외부 계단 한 줄, 상부 전선과 여러 가로등이 참조 장소의 성격을 유지한다. 미디엄 숏이지만 A보다 앙각과 화면 기울기가 강하고, 차량 전면의 노출 면적도 더 커서 가장자리의 보조 배경이라는 지시에는 덜 가깝다.",
        "entities": "신부와 찰리만 보인다. 신부의 나이, 한국인 남성으로 보이는 외모, 짧은 회색 머리와 흰 로만칼라는 참조에 가깝다. 다만 검은 사제복 위에 넓은 라펠이 있는 재킷이 더해져 참조의 단정한 라펠 없는 상의와 다르다. 찰리는 젖은 낡은 모자와 외투, 베이지색 기계 머리와 원형 관절을 갖추고 있다. 흰 얼굴판과 눈은 가려져 있으며 새나 다른 인물, 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "신부는 상체를 찰리 쪽으로 약간 숙이고 양손을 몸 옆에 둔다. 발은 잘렸지만 몸이 화면 아래의 다리로 이어지는 정상적인 서 있는 자세이며, 지지 없는 부유는 보이지 않는다. 찰리의 머리와 어깨도 연결되어 있고 모자와 외투는 각각 머리와 몸에 지지된다. 차체는 도로 위 차량의 일부로 자연스럽게 이어지고, 젖은 표면의 반사도 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "찰리 뒤의 완만한 로우앵글 미디엄 숏, 찰리를 내려다보는 당황한 얼굴, 참조와 가까운 사제복을 충실히 구현한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "시선과 당황한 표정은 맞지만, 더 기울어진 앙각과 넓게 드러난 차량, 참조에 없는 라펠 재킷이 A보다 덜 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 고개를 왼쪽 아래로 기울여 전경의 찰리 얼굴을 바라보며, 렌즈를 응시하지 않는다. 찰리의 뒤통수와 옆면은 오른쪽 신부 쪽으로 향한다. 찰리의 눈은 가려져 정확한 눈동자 방향은 확인할 수 없다. 차량 전조등은 차량 앞 도로와 카메라 쪽으로 켜져 있다.",
        "built_space": "왼쪽 전경에 찰리의 머리와 어깨, 오른쪽 중경에 신부가 배치되어 있다. 오른쪽 가장자리에는 밴 한 대의 비스듬한 전면 일부와 켜진 전조등 한 개, 그릴 일부, 번호판 한 개가 보인다. 젖은 도로 양옆의 컨테이너 구조물, 왼쪽 철망, 상부 전선과 여러 가로등이 이전 장소의 재질과 야간 환경을 잇는다. 신부 눈높이 아래에서 올려다보는 미디엄 숏이며, 찰리의 머리는 지정된 왼쪽 아래보다 위쪽까지 크게 차지한다.",
        "entities": "등장 개체는 신부와 찰리뿐이다. 신부는 참조와 가까운 짧은 회색 머리와 얼굴을 가진 60대 한국인 남성으로 보이며, 흰 로만칼라와 라펠 없는 검은 단추식 사제복이 참조에 가깝다. 찰리는 낡고 젖은 모자와 외투를 착용하고, 노출된 머리에는 샌드 베이지 장갑판과 원형 기계 관절이 보인다. 가려진 얼굴의 흰 마스크와 눈은 검증할 수 없다. 새나 추가 인물, 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 인물의 상체는 화면 아래로 자연스럽게 이어지며, 발이 프레임 밖이라는 이유로 부유한다고 볼 근거는 없다. 신부의 손은 몸 옆에 내려와 있고 얼굴과 상체를 찰리 쪽으로 조금 기울인 자세가 가능하다. 모자는 찰리의 머리에 얹혀 있고 외투는 어깨를 감싼다. 차량과 인물이 공중에 떠 있는 징후는 없으며, 도로의 불빛 반사도 젖은 노면과 부합한다."
       },
       {
        "label": "A",
        "direction": "신부의 시선은 왼쪽 아래 찰리의 머리로 향하고, 살짝 벌어진 입과 모인 눈썹이 믿기 어려워하는 반응을 드러낸다. 찰리의 머리는 오른쪽 위 신부를 향하지만 눈 자체는 보이지 않는다. 밴의 켜진 전조등은 전방 도로와 카메라 방향을 비춘다.",
        "built_space": "찰리는 왼쪽 전경, 신부는 오른쪽 중경에 서 있으며 오른쪽 뒤에 밴 한 대가 있다. 전조등 한 개, 그릴 일부와 비교적 넓은 앞유리가 보인다. 젖은 도로, 컨테이너 건물, 오른쪽 외부 계단 한 줄, 상부 전선과 여러 가로등이 참조 장소의 성격을 유지한다. 미디엄 숏이지만 A보다 앙각과 화면 기울기가 강하고, 차량 전면의 노출 면적도 더 커서 가장자리의 보조 배경이라는 지시에는 덜 가깝다.",
        "entities": "신부와 찰리만 보인다. 신부의 나이, 한국인 남성으로 보이는 외모, 짧은 회색 머리와 흰 로만칼라는 참조에 가깝다. 다만 검은 사제복 위에 넓은 라펠이 있는 재킷이 더해져 참조의 단정한 라펠 없는 상의와 다르다. 찰리는 젖은 낡은 모자와 외투, 베이지색 기계 머리와 원형 관절을 갖추고 있다. 흰 얼굴판과 눈은 가려져 있으며 새나 다른 인물, 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "신부는 상체를 찰리 쪽으로 약간 숙이고 양손을 몸 옆에 둔다. 발은 잘렸지만 몸이 화면 아래의 다리로 이어지는 정상적인 서 있는 자세이며, 지지 없는 부유는 보이지 않는다. 찰리의 머리와 어깨도 연결되어 있고 모자와 외투는 각각 머리와 몸에 지지된다. 차체는 도로 위 차량의 일부로 자연스럽게 이어지고, 젖은 표면의 반사도 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.603,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.603,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1603
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 앵글과 샷 크기를 훌륭하게 구현했으며, 신부의 당황한 표정과 사제복 의상 디테일까지 레퍼런스를 정확히 반영했습니다."
   },
   {
    "label": "A",
    "score": 1603,
    "verdict_ko": "카메라 구도와 배경은 지시문과 일치하나, 신부의 의상이 레퍼런스 이미지와 달리 일반적인 양복 재킷으로 묘사되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S27sh11_sel.png",
    "asset_id": "85ef7a79-6beb-4a24-8f5e-840c5e2ad2b2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1305657>",
    "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09a5-e365-7e56-a267-8dd7e6ee373e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh11"
  },
  "staged_characters_added": [
   "C01"
  ]
 },
 "S27sh19::signage": {
  "fp": "f98a84d38c3172c8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh19": {
  "input_fingerprint": "e60c597fa6368800",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차량 뒷좌석에 한 발을 딛고 오르는 중인 찰리를 어두운 골목 구석에서 몰래 예의 주시하는 구도환의 어깨 위 측면.\n\nLOCATION (lock): In a dark corner of a settlement alley overlooking the stopped van's rear boarding doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut to the alley, hold at 구도환's shoulder height just behind and to one side of him, looking obliquely toward the rear passenger doorway. His near shoulder and a sliver of his intent profile occupy the left foreground edge, while 찰리 appears nearly full-body at right in the background, bending into the vehicle with one foot already inside and his attention lowered to his footing. Keep the camera static so 구도환's fixed attention on 찰리 and the open doorway connecting their sightline establish covert observation rather than a shared interaction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 구도환 in the middle-left of the frame, foreground, looks toward 찰리 boarding the rear passenger doorway; 찰리 in the middle-right of the frame, background, moves toward rear passenger compartment; Open rear passenger doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 골목 구석 (구도환 remains concealed in the dark corner) — The corner borders the observer's side of the frame; used as Provides foreground concealment without blocking the view of boarding; 차량 뒷좌석 출입구 (Open while 찰리 boards) — The rear passenger opening is seen obliquely from the alley; used as Anchors the watched action in the background; the visible vehicle section remains below two-fifths of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the alley observer subdued against the more readable vehicle area, using restrained nighttime contrast without inventing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van has a crude halogen cross ornament on its roof and scattered cargo boxes in the rear compartment. Charlie boards in his old coat and hat disguise, retaining his blue-lit eyes and worn chest logo. 구도환: He remains nearby, watching the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차량 뒷좌석에 한 발을 딛고 오르는 중인 찰리를 어두운 골목 구석에서 몰래 예의 주시하는 구도환의 어깨 위 측면.\n\nLOCATION (lock): In a dark corner of a settlement alley overlooking the stopped van's rear boarding doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut to the alley, hold at 구도환's shoulder height just behind and to one side of him, looking obliquely toward the rear passenger doorway. His near shoulder and a sliver of his intent profile occupy the left foreground edge, while 찰리 appears nearly full-body at right in the background, bending into the vehicle with one foot already inside and his attention lowered to his footing. Keep the camera static so 구도환's fixed attention on 찰리 and the open doorway connecting their sightline establish covert observation rather than a shared interaction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 구도환 in the middle-left of the frame, foreground, looks toward 찰리 boarding the rear passenger doorway; 찰리 in the middle-right of the frame, background, moves toward rear passenger compartment; Open rear passenger doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 골목 구석 (구도환 remains concealed in the dark corner) — The corner borders the observer's side of the frame; used as Provides foreground concealment without blocking the view of boarding; 차량 뒷좌석 출입구 (Open while 찰리 boards) — The rear passenger opening is seen obliquely from the alley; used as Anchors the watched action in the background; the visible vehicle section remains below two-fifths of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the alley observer subdued against the more readable vehicle area, using restrained nighttime contrast without inventing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van has a crude halogen cross ornament on its roof and scattered cargo boxes in the rear compartment. Charlie boards in his old coat and hat disguise, retaining his blue-lit eyes and worn chest logo. 구도환: He remains nearby, watching the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차량 뒷좌석에 한 발을 딛고 오르는 중인 찰리를 어두운 골목 구석에서 몰래 예의 주시하는 구도환의 어깨 위 측면.\n\nLOCATION (lock): In a dark corner of a settlement alley overlooking the stopped van's rear boarding doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut to the alley, hold at 구도환's shoulder height just behind and to one side of him, looking obliquely toward the rear passenger doorway. His near shoulder and a sliver of his intent profile occupy the left foreground edge, while 찰리 appears nearly full-body at right in the background, bending into the vehicle with one foot already inside and his attention lowered to his footing. Keep the camera static so 구도환's fixed attention on 찰리 and the open doorway connecting their sightline establish covert observation rather than a shared interaction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 구도환 in the middle-left of the frame, foreground, looks toward 찰리 boarding the rear passenger doorway; 찰리 in the middle-right of the frame, background, moves toward rear passenger compartment; Open rear passenger doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 골목 구석 (구도환 remains concealed in the dark corner) — The corner borders the observer's side of the frame; used as Provides foreground concealment without blocking the view of boarding; 차량 뒷좌석 출입구 (Open while 찰리 boards) — The rear passenger opening is seen obliquely from the alley; used as Anchors the watched action in the background; the visible vehicle section remains below two-fifths of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the alley observer subdued against the more readable vehicle area, using restrained nighttime contrast without inventing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The stopped van has a crude halogen cross ornament on its roof and scattered cargo boxes in the rear compartment. Charlie boards in his old coat and hat disguise, retaining his blue-lit eyes and worn chest logo. 구도환: He remains nearby, watching the boarding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "구도환은 왼쪽 전경에서 화면 오른쪽의 밴에 오르고 있는 찰리를 예의 주시하고 있습니다. 찰리는 밴에 오르며 발을 딛기 위해 바닥 쪽으로 시선을 내리고 있습니다.",
    "built_space": "왼쪽 전경에 구도환을 가려주는 어두운 골목 구석 벽면이 자리하고, 오른쪽 배경에는 측면 출입문이 열린 어두운 색상의 밴이 주차되어 있습니다. 밴 지붕에는 빛나는 할로겐 십자가가 있고, 차량 내부에는 화물 상자들이 적재되어 있으며, 배경은 이전 샷과 일치하는 젖은 노면과 가로등이 있는 거리입니다.",
    "entities": "구도환은 50대 한국인 남성의 측면 모습으로 짧은 흑발과 낡은 갈색 점퍼를 입고 있어 참조와 일치합니다. 찰리는 이전 샷과 동일한 젖은 질감의 어두운 코트와 모자를 쓰고 있으며, 샌드 베이지색 기계 다리와 팔, 흰색 마스크 얼굴(눈 구멍과 입 선)이 찰리의 캐릭터 참조와 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "구도환은 바닥에 안정적으로 서 있습니다. 찰리는 왼발을 땅에 지지하고 오른발을 밴 내부 바닥에 디딘 채, 오른팔로 밴 내부 프레임을 짚어 탑승하는 체중을 물리적으로 자연스럽게 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "구도환은 왼쪽 전경에서 밴에 탑승하려는 찰리를 향해 시선을 고정하고 있습니다. 찰리는 밴 내부를 향해 고개를 숙이고 진입하고 있습니다.",
    "built_space": "왼쪽 전경에 골목 벽면이 있고, 오른쪽에 측면 문이 열린 밴이 정차해 있습니다. 밴 내부에는 종이 상자들이 쌓여 있고 지붕에는 빛나는 십자가 장식이 있으며, 비에 젖은 밤거리의 배경이 묘사되었습니다.",
    "entities": "구도환의 측면 실루엣과 의상(갈색 점퍼)은 참조와 잘 맞습니다. 찰리는 코트와 모자를 착용하고 베이지색 기계 다리를 보이고 있으나, 머리 측면에 캐릭터 참조 시트에 없는 검은 원형 속 기하학적 무늬가 임의로 생성되었습니다. 밴 지붕의 십자가는 텍스트가 흐르는 LED 전광판처럼 묘사되었습니다.",
    "hard_violations": [],
    "physics": "구도환은 두 발로 바닥에 서서 안정적인 자세를 유지합니다. 찰리는 왼발을 지면에 두고 오른발을 밴 발판에 올린 채 탑승을 진행하고 있어 물리적 지지가 자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "오버더숄더 앵글과 찰리가 밴에 탑승하는 동작을 완벽하게 담아냈으며, 찰리의 기계 마스크와 베이지색 장갑판, 할로겐 십자가 등 세부 묘사가 프롬프트와 참조 이미지에 매우 충실합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "전체적인 구도와 상황은 잘 구현되었으나, 찰리의 머리 측면에 참조에 없는 임의의 문양이 추가되었고 지붕의 십자가가 할로겐보다는 LED 간판에 가깝게 표현되어 세부 디테일이 다소 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 왼쪽 전경에서 화면 오른쪽의 밴에 오르고 있는 찰리를 예의 주시하고 있습니다. 찰리는 밴에 오르며 발을 딛기 위해 바닥 쪽으로 시선을 내리고 있습니다.",
        "built_space": "왼쪽 전경에 구도환을 가려주는 어두운 골목 구석 벽면이 자리하고, 오른쪽 배경에는 측면 출입문이 열린 어두운 색상의 밴이 주차되어 있습니다. 밴 지붕에는 빛나는 할로겐 십자가가 있고, 차량 내부에는 화물 상자들이 적재되어 있으며, 배경은 이전 샷과 일치하는 젖은 노면과 가로등이 있는 거리입니다.",
        "entities": "구도환은 50대 한국인 남성의 측면 모습으로 짧은 흑발과 낡은 갈색 점퍼를 입고 있어 참조와 일치합니다. 찰리는 이전 샷과 동일한 젖은 질감의 어두운 코트와 모자를 쓰고 있으며, 샌드 베이지색 기계 다리와 팔, 흰색 마스크 얼굴(눈 구멍과 입 선)이 찰리의 캐릭터 참조와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "구도환은 바닥에 안정적으로 서 있습니다. 찰리는 왼발을 땅에 지지하고 오른발을 밴 내부 바닥에 디딘 채, 오른팔로 밴 내부 프레임을 짚어 탑승하는 체중을 물리적으로 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "구도환은 왼쪽 전경에서 밴에 탑승하려는 찰리를 향해 시선을 고정하고 있습니다. 찰리는 밴 내부를 향해 고개를 숙이고 진입하고 있습니다.",
        "built_space": "왼쪽 전경에 골목 벽면이 있고, 오른쪽에 측면 문이 열린 밴이 정차해 있습니다. 밴 내부에는 종이 상자들이 쌓여 있고 지붕에는 빛나는 십자가 장식이 있으며, 비에 젖은 밤거리의 배경이 묘사되었습니다.",
        "entities": "구도환의 측면 실루엣과 의상(갈색 점퍼)은 참조와 잘 맞습니다. 찰리는 코트와 모자를 착용하고 베이지색 기계 다리를 보이고 있으나, 머리 측면에 캐릭터 참조 시트에 없는 검은 원형 속 기하학적 무늬가 임의로 생성되었습니다. 밴 지붕의 십자가는 텍스트가 흐르는 LED 전광판처럼 묘사되었습니다.",
        "hard_violations": [],
        "physics": "구도환은 두 발로 바닥에 서서 안정적인 자세를 유지합니다. 찰리는 왼발을 지면에 두고 오른발을 밴 발판에 올린 채 탑승을 진행하고 있어 물리적 지지가 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "오버더숄더 앵글과 찰리가 밴에 탑승하는 동작을 완벽하게 담아냈으며, 찰리의 기계 마스크와 베이지색 장갑판, 할로겐 십자가 등 세부 묘사가 프롬프트와 참조 이미지에 매우 충실합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "전체적인 구도와 상황은 잘 구현되었으나, 찰리의 머리 측면에 참조에 없는 임의의 문양이 추가되었고 지붕의 십자가가 할로겐보다는 LED 간판에 가깝게 표현되어 세부 디테일이 다소 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 왼쪽 전경에서 화면 오른쪽의 밴에 오르고 있는 찰리를 예의 주시하고 있습니다. 찰리는 밴에 오르며 발을 딛기 위해 바닥 쪽으로 시선을 내리고 있습니다.",
        "built_space": "왼쪽 전경에 구도환을 가려주는 어두운 골목 구석 벽면이 자리하고, 오른쪽 배경에는 측면 출입문이 열린 어두운 색상의 밴이 주차되어 있습니다. 밴 지붕에는 빛나는 할로겐 십자가가 있고, 차량 내부에는 화물 상자들이 적재되어 있으며, 배경은 이전 샷과 일치하는 젖은 노면과 가로등이 있는 거리입니다.",
        "entities": "구도환은 50대 한국인 남성의 측면 모습으로 짧은 흑발과 낡은 갈색 점퍼를 입고 있어 참조와 일치합니다. 찰리는 이전 샷과 동일한 젖은 질감의 어두운 코트와 모자를 쓰고 있으며, 샌드 베이지색 기계 다리와 팔, 흰색 마스크 얼굴(눈 구멍과 입 선)이 찰리의 캐릭터 참조와 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "구도환은 바닥에 안정적으로 서 있습니다. 찰리는 왼발을 땅에 지지하고 오른발을 밴 내부 바닥에 디딘 채, 오른팔로 밴 내부 프레임을 짚어 탑승하는 체중을 물리적으로 자연스럽게 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "구도환은 왼쪽 전경에서 밴에 탑승하려는 찰리를 향해 시선을 고정하고 있습니다. 찰리는 밴 내부를 향해 고개를 숙이고 진입하고 있습니다.",
        "built_space": "왼쪽 전경에 골목 벽면이 있고, 오른쪽에 측면 문이 열린 밴이 정차해 있습니다. 밴 내부에는 종이 상자들이 쌓여 있고 지붕에는 빛나는 십자가 장식이 있으며, 비에 젖은 밤거리의 배경이 묘사되었습니다.",
        "entities": "구도환의 측면 실루엣과 의상(갈색 점퍼)은 참조와 잘 맞습니다. 찰리는 코트와 모자를 착용하고 베이지색 기계 다리를 보이고 있으나, 머리 측면에 캐릭터 참조 시트에 없는 검은 원형 속 기하학적 무늬가 임의로 생성되었습니다. 밴 지붕의 십자가는 텍스트가 흐르는 LED 전광판처럼 묘사되었습니다.",
        "hard_violations": [],
        "physics": "구도환은 두 발로 바닥에 서서 안정적인 자세를 유지합니다. 찰리는 왼발을 지면에 두고 오른발을 밴 발판에 올린 채 탑승을 진행하고 있어 물리적 지지가 자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "외투의 연속성과 한 발을 걸친 승차는 잘 맞지만, 찰리가 발밑보다 실내를 향하고 상체를 덜 숙여 지정된 승차 순간은 B보다 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 은폐 관찰자의 시선과 오른쪽 찰리의 숙인 고개·굽힌 몸·문턱에 디딘 발이 지시된 순간에 더 충실하나, 외투 소매의 연속성은 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 왼쪽 전경에서 얼굴을 오른쪽으로 돌려 열린 출입구의 찰리를 보고 있다. 찰리의 몸과 들어 올린 다리는 차량 안으로 향하지만, 고개는 발밑보다 실내 앞쪽을 향한다. 서로 마주 보는 상호작용은 아니다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 벽 모서리와 배수관이 관찰자를 가리고, 오른쪽에는 열린 측면 승객 출입구 하나와 지붕 십자가 하나가 보인다. 문 안에는 등받이와 머리받침, 여러 화물 상자가 있으며 승차 문턱은 찰리의 발과 연결된다. 젖은 도로, 철망, 컨테이너와 전선은 참조 장소의 재질과 야간 분위기를 따른다. 차량은 화면 약 3분의 1을 차지하지만, 구도환의 머리와 상체는 요구된 가장자리의 좁은 어깨·옆얼굴보다 크게 들어온다.",
        "entities": "등장 개체는 구도환과 찰리뿐이다. 구도환은 짧은 검은 머리와 낡은 갈색 점퍼를 입은 중년 동아시아계 남성으로 보이며, 얼굴 대부분이 가려져 정확한 동일성은 확인하기 어렵다. 찰리는 작은 기계 몸체, 베이지 장갑판, 원형 귀 부품, 낡은 모자와 긴 소매 외투를 유지한다. 흰 얼굴은 일부만 보이고 눈 색과 가슴 표식은 가려져 판단할 수 없다. 검은 밴, 지붕의 발광 십자가, 실내 상자가 있다.",
        "hard_violations": [],
        "physics": "찰리의 한 발은 젖은 노면에, 굽힌 반대쪽 다리의 발은 차량 문턱에 놓여 있다. 앞으로 뻗은 손도 출입구 안쪽에 닿아 승차를 보조하는 자세다. 몸이 떠 있거나 지지 없는 물체는 보이지 않는다. 구도환의 하체는 화면 밖이며, 보이는 상체에 물리적 모순은 없다. 십자가는 지붕 받침에 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "구도환의 얼굴 방향은 오른쪽 출입구에서 승차하는 찰리에게 향한다. 찰리는 고개를 아래로 숙여 문턱과 들어 올린 발 쪽을 보고, 상체와 팔은 실내로 향한다. 관찰자에게 돌아보지 않아 몰래 지켜보는 관계가 유지된다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 벽 모서리와 배관 뒤에 구도환이 있고, 오른쪽 밴에는 열린 측면 승객 출입구 하나와 지붕 십자가 하나가 있다. 실내 좌석 일부와 적재물이 보이며 찰리가 들어갈 문턱 공간이 확보되어 있다. 차량은 화면의 5분의 2 미만으로 읽히고, 찰리는 오른쪽에 거의 전신으로 들어온다. 철망과 컨테이너, 젖은 포장 및 가로등이 참조의 장소 성격을 잇는다. 다만 구도환은 어깨와 옆얼굴의 얇은 일부만 보이라는 지시보다 크게 잡혔다.",
        "entities": "구도환과 찰리 외의 사람은 없다. 구도환의 짧은 검은 머리, 중년 동아시아계 남성의 옆모습과 갈색 점퍼는 설정에 부합한다. 찰리의 흰 기계 얼굴, 원형 귀 부품, 베이지 장갑과 짧은 다리, 낡은 모자가 보인다. 다만 팔의 장갑이 크게 노출되어 이전 장면의 긴 소매 외투보다는 소매 없는 덮개처럼 보인다. 눈의 발광색과 가슴 표식은 이 각도에서 확인되지 않는다. 밴과 지붕 십자가, 뒤쪽 적재물은 존재한다.",
        "hard_violations": [],
        "physics": "찰리는 바깥쪽 발바닥으로 노면을 지지하고 반대쪽 발을 실내 문턱에 올린 채 무릎과 허리를 굽힌다. 팔은 출입구 안으로 뻗어 있으며 손끝은 일부 가려지지만, 두 발의 접촉만으로도 승차 자세의 지지가 성립한다. 차량은 바퀴로 지면에 서 있고 십자가는 지붕 거치대에 고정되어 있다. 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "외투의 연속성과 한 발을 걸친 승차는 잘 맞지만, 찰리가 발밑보다 실내를 향하고 상체를 덜 숙여 지정된 승차 순간은 B보다 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 은폐 관찰자의 시선과 오른쪽 찰리의 숙인 고개·굽힌 몸·문턱에 디딘 발이 지시된 순간에 더 충실하나, 외투 소매의 연속성은 떨어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 왼쪽 전경에서 얼굴을 오른쪽으로 돌려 열린 출입구의 찰리를 보고 있다. 찰리의 몸과 들어 올린 다리는 차량 안으로 향하지만, 고개는 발밑보다 실내 앞쪽을 향한다. 서로 마주 보는 상호작용은 아니다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 벽 모서리와 배수관이 관찰자를 가리고, 오른쪽에는 열린 측면 승객 출입구 하나와 지붕 십자가 하나가 보인다. 문 안에는 등받이와 머리받침, 여러 화물 상자가 있으며 승차 문턱은 찰리의 발과 연결된다. 젖은 도로, 철망, 컨테이너와 전선은 참조 장소의 재질과 야간 분위기를 따른다. 차량은 화면 약 3분의 1을 차지하지만, 구도환의 머리와 상체는 요구된 가장자리의 좁은 어깨·옆얼굴보다 크게 들어온다.",
        "entities": "등장 개체는 구도환과 찰리뿐이다. 구도환은 짧은 검은 머리와 낡은 갈색 점퍼를 입은 중년 동아시아계 남성으로 보이며, 얼굴 대부분이 가려져 정확한 동일성은 확인하기 어렵다. 찰리는 작은 기계 몸체, 베이지 장갑판, 원형 귀 부품, 낡은 모자와 긴 소매 외투를 유지한다. 흰 얼굴은 일부만 보이고 눈 색과 가슴 표식은 가려져 판단할 수 없다. 검은 밴, 지붕의 발광 십자가, 실내 상자가 있다.",
        "hard_violations": [],
        "physics": "찰리의 한 발은 젖은 노면에, 굽힌 반대쪽 다리의 발은 차량 문턱에 놓여 있다. 앞으로 뻗은 손도 출입구 안쪽에 닿아 승차를 보조하는 자세다. 몸이 떠 있거나 지지 없는 물체는 보이지 않는다. 구도환의 하체는 화면 밖이며, 보이는 상체에 물리적 모순은 없다. 십자가는 지붕 받침에 고정되어 있다."
       },
       {
        "label": "A",
        "direction": "구도환의 얼굴 방향은 오른쪽 출입구에서 승차하는 찰리에게 향한다. 찰리는 고개를 아래로 숙여 문턱과 들어 올린 발 쪽을 보고, 상체와 팔은 실내로 향한다. 관찰자에게 돌아보지 않아 몰래 지켜보는 관계가 유지된다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 벽 모서리와 배관 뒤에 구도환이 있고, 오른쪽 밴에는 열린 측면 승객 출입구 하나와 지붕 십자가 하나가 있다. 실내 좌석 일부와 적재물이 보이며 찰리가 들어갈 문턱 공간이 확보되어 있다. 차량은 화면의 5분의 2 미만으로 읽히고, 찰리는 오른쪽에 거의 전신으로 들어온다. 철망과 컨테이너, 젖은 포장 및 가로등이 참조의 장소 성격을 잇는다. 다만 구도환은 어깨와 옆얼굴의 얇은 일부만 보이라는 지시보다 크게 잡혔다.",
        "entities": "구도환과 찰리 외의 사람은 없다. 구도환의 짧은 검은 머리, 중년 동아시아계 남성의 옆모습과 갈색 점퍼는 설정에 부합한다. 찰리의 흰 기계 얼굴, 원형 귀 부품, 베이지 장갑과 짧은 다리, 낡은 모자가 보인다. 다만 팔의 장갑이 크게 노출되어 이전 장면의 긴 소매 외투보다는 소매 없는 덮개처럼 보인다. 눈의 발광색과 가슴 표식은 이 각도에서 확인되지 않는다. 밴과 지붕 십자가, 뒤쪽 적재물은 존재한다.",
        "hard_violations": [],
        "physics": "찰리는 바깥쪽 발바닥으로 노면을 지지하고 반대쪽 발을 실내 문턱에 올린 채 무릎과 허리를 굽힌다. 팔은 출입구 안으로 뻗어 있으며 손끝은 일부 가려지지만, 두 발의 접촉만으로도 승차 자세의 지지가 성립한다. 차량은 바퀴로 지면에 서 있고 십자가는 지붕 거치대에 고정되어 있다. 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.575
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.575
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1575
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "오버더숄더 앵글과 찰리가 밴에 탑승하는 동작을 완벽하게 담아냈으며, 찰리의 기계 마스크와 베이지색 장갑판, 할로겐 십자가 등 세부 묘사가 프롬프트와 참조 이미지에 매우 충실합니다."
   },
   {
    "label": "B",
    "score": 1575,
    "verdict_ko": "전체적인 구도와 상황은 잘 구현되었으나, 찰리의 머리 측면에 참조에 없는 임의의 문양이 추가되었고 지붕의 십자가가 할로겐보다는 LED 간판에 가깝게 표현되어 세부 디테일이 다소 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S27sh17_sel.png",
    "asset_id": "39f2ada4-ab6b-425d-8291-157826b080e5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09aa-56d1-76eb-8da9-840bc8bab332",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh17"
  }
 },
 "S28sh2::signage": {
  "fp": "152e7dd5f3f05b35",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9ecfbfd758317e06": {
  "subjects": [],
  "subject_text": "신부의 밴 내부\n앞쪽 운전 공간 뒤로 긴 뒷좌석 공간이 이어지는 실내. 물품 상자들이 듬성듬성 놓여 좌석 사이에 좁은 틈을 만든다.",
  "identity": "canonical",
  "scope_id": "L39",
  "scope_role": "location_interior",
  "scope_sha": "ad2cc6875eed2105"
 },
 "groupbg::utility_van_interior": {
  "input_fingerprint": "edc8baf33386142a",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "utility_van_interior",
    "tags": [
     "S28sh2",
     "S28sh5"
    ]
   },
   "context_sig": "157ad5cc103230cb"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n신부의 밴 내부: 지붕에 십자가 할로겐 전등이 달린 구급대 형태의 낡은 승합차 내부다. (특징: 차량 지붕에 달린 조악한 십자가 형태의 점등 장치; 뒷좌석 화물칸에 쌓인 물품 상자들과 그 사이에 웅크린 현우, 앰버, 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 마치 구급대 차량처럼, 뒷좌석에는 물건 박스가 듬성듬성. 그 사이사이에 현우와 앰버, 찰리가 숨어있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n신부의 밴 내부: 지붕에 십자가 할로겐 전등이 달린 구급대 형태의 낡은 승합차 내부다. (특징: 차량 지붕에 달린 조악한 십자가 형태의 점등 장치; 뒷좌석 화물칸에 쌓인 물품 상자들과 그 사이에 웅크린 현우, 앰버, 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 마치 구급대 차량처럼, 뒷좌석에는 물건 박스가 듬성듬성. 그 사이사이에 현우와 앰버, 찰리가 숨어있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_utility_van_interior_df21e0.png",
  "asset_id": "f960bc7f-bc99-4b62-ace2-488d54c9aa93",
  "input_asset_ids": [
   "550b93f5-f7c2-436b-a456-385435066af9"
  ],
  "origin_tag": "S28sh2",
  "place_text": "Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.",
  "origin_inputs": {
   "place_text": "Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.",
   "time_of_day_en": "night",
   "conti_asset_id": "550b93f5-f7c2-436b-a456-385435066af9"
  }
 },
 "S28sh2::bgfirst_bg": {
  "input_fingerprint": "9820be486abed713",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 밴 차량 뒷좌석의 어지러운 화물 박스들 사이에 웅크리고 숨어 있는 이현우, 앰버, 찰리의 구도.\n\nLOCATION (lock): Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just behind the front seats, offset to the passenger side and above the group's crouched heads, looking diagonally downward through the spaces between the boxes. 이현우 folds forward at left with his gaze lowered into the space around his knees, 앰버 tucks herself into the center and looks toward 찰리, and 찰리 crouches at right with his head inclined toward her. Capture the already-concealed entry pose before the warning or forced downward movement, keeping the scattered boxes below shoulder height in the composition and no single box larger than a quarter of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 물건 박스들 (Scattered through the rear seating area with gaps occupied by the concealed group) — Different top and side faces appear along the diagonal view; used as Create irregular partial occlusions and separate the three crouched silhouettes without forming a solid wall; 앞좌석 등받이 (Between the camera's position and the front cabin) — A cropped rear face borders the near edge of the image; used as Locates the camera inside the van without obstructing the group; 밴 뒷좌석 (Partially visible around the boxes and crouched occupants) — Seen diagonally from above and in front; used as Supplies human-scale spatial context for the hiding positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued cabin illumination and controlled contrast keep the three faces legible without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 밴 차량 뒷좌석의 어지러운 화물 박스들 사이에 웅크리고 숨어 있는 이현우, 앰버, 찰리의 구도.\n\nLOCATION (lock): Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just behind the front seats, offset to the passenger side and above the group's crouched heads, looking diagonally downward through the spaces between the boxes. 이현우 folds forward at left with his gaze lowered into the space around his knees, 앰버 tucks herself into the center and looks toward 찰리, and 찰리 crouches at right with his head inclined toward her. Capture the already-concealed entry pose before the warning or forced downward movement, keeping the scattered boxes below shoulder height in the composition and no single box larger than a quarter of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 물건 박스들 (Scattered through the rear seating area with gaps occupied by the concealed group) — Different top and side faces appear along the diagonal view; used as Create irregular partial occlusions and separate the three crouched silhouettes without forming a solid wall; 앞좌석 등받이 (Between the camera's position and the front cabin) — A cropped rear face borders the near edge of the image; used as Locates the camera inside the van without obstructing the group; 밴 뒷좌석 (Partially visible around the boxes and crouched occupants) — Seen diagonally from above and in front; used as Supplies human-scale spatial context for the hiding positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued cabin illumination and controlled contrast keep the three faces legible without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh2__bgfirst_bg.png",
  "asset_id": "9fab6aaa-d8ac-4731-b287-1ff70e160480",
  "input_asset_ids": [
   "550b93f5-f7c2-436b-a456-385435066af9",
   "f960bc7f-bc99-4b62-ace2-488d54c9aa93"
  ]
 },
 "S28sh2": {
  "input_fingerprint": "e5d859843a135afc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 밴 차량 뒷좌석의 어지러운 화물 박스들 사이에 웅크리고 숨어 있는 이현우, 앰버, 찰리의 구도.\n\nLOCATION (lock): Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just behind the front seats, offset to the passenger side and above the group's crouched heads, looking diagonally downward through the spaces between the boxes. 이현우 folds forward at left with his gaze lowered into the space around his knees, 앰버 tucks herself into the center and looks toward 찰리, and 찰리 crouches at right with his head inclined toward her. Capture the already-concealed entry pose before the warning or forced downward movement, keeping the scattered boxes below shoulder height in the composition and no single box larger than a quarter of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 물건 박스들 (Scattered through the rear seating area with gaps occupied by the concealed group) — Different top and side faces appear along the diagonal view; used as Create irregular partial occlusions and separate the three crouched silhouettes without forming a solid wall; 앞좌석 등받이 (Between the camera's position and the front cabin) — A cropped rear face borders the near edge of the image; used as Locates the camera inside the van without obstructing the group; 밴 뒷좌석 (Partially visible around the boxes and crouched occupants) — Seen diagonally from above and in front; used as Supplies human-scale spatial context for the hiding positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued cabin illumination and controlled contrast keep the three faces legible without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has a crude roof-mounted halogen cross and scattered cargo boxes in its rear compartment. Charlie is concealed among the boxes, retaining his old coat and hat disguise and blue-lit eyes. 이현우: He is concealed among the rear cargo boxes, with facial bruises and the persistent leg injury. His outer garment remains removed. 앰버: She is concealed among the rear cargo boxes, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 밴 차량 뒷좌석의 어지러운 화물 박스들 사이에 웅크리고 숨어 있는 이현우, 앰버, 찰리의 구도.\n\nLOCATION (lock): Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just behind the front seats, offset to the passenger side and above the group's crouched heads, looking diagonally downward through the spaces between the boxes. 이현우 folds forward at left with his gaze lowered into the space around his knees, 앰버 tucks herself into the center and looks toward 찰리, and 찰리 crouches at right with his head inclined toward her. Capture the already-concealed entry pose before the warning or forced downward movement, keeping the scattered boxes below shoulder height in the composition and no single box larger than a quarter of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 물건 박스들 (Scattered through the rear seating area with gaps occupied by the concealed group) — Different top and side faces appear along the diagonal view; used as Create irregular partial occlusions and separate the three crouched silhouettes without forming a solid wall; 앞좌석 등받이 (Between the camera's position and the front cabin) — A cropped rear face borders the near edge of the image; used as Locates the camera inside the van without obstructing the group; 밴 뒷좌석 (Partially visible around the boxes and crouched occupants) — Seen diagonally from above and in front; used as Supplies human-scale spatial context for the hiding positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued cabin illumination and controlled contrast keep the three faces legible without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has a crude roof-mounted halogen cross and scattered cargo boxes in its rear compartment. Charlie is concealed among the boxes, retaining his old coat and hat disguise and blue-lit eyes. 이현우: He is concealed among the rear cargo boxes, with facial bruises and the persistent leg injury. His outer garment remains removed. 앰버: She is concealed among the rear cargo boxes, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 밴 차량 뒷좌석의 어지러운 화물 박스들 사이에 웅크리고 숨어 있는 이현우, 앰버, 찰리의 구도.\n\nLOCATION (lock): Inside the van's rear passenger area, in the gaps between sparsely placed cargo boxes. Nighttime light is limited to what reaches the compartment through the vehicle openings. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just behind the front seats, offset to the passenger side and above the group's crouched heads, looking diagonally downward through the spaces between the boxes. 이현우 folds forward at left with his gaze lowered into the space around his knees, 앰버 tucks herself into the center and looks toward 찰리, and 찰리 crouches at right with his head inclined toward her. Capture the already-concealed entry pose before the warning or forced downward movement, keeping the scattered boxes below shoulder height in the composition and no single box larger than a quarter of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 물건 박스들 (Scattered through the rear seating area with gaps occupied by the concealed group) — Different top and side faces appear along the diagonal view; used as Create irregular partial occlusions and separate the three crouched silhouettes without forming a solid wall; 앞좌석 등받이 (Between the camera's position and the front cabin) — A cropped rear face borders the near edge of the image; used as Locates the camera inside the van without obstructing the group; 밴 뒷좌석 (Partially visible around the boxes and crouched occupants) — Seen diagonally from above and in front; used as Supplies human-scale spatial context for the hiding positions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued cabin illumination and controlled contrast keep the three faces legible without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van has a crude roof-mounted halogen cross and scattered cargo boxes in its rear compartment. Charlie is concealed among the boxes, retaining his old coat and hat disguise and blue-lit eyes. 이현우: He is concealed among the rear cargo boxes, with facial bruises and the persistent leg injury. His outer garment remains removed. 앰버: She is concealed among the rear cargo boxes, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh2__bgfirst_bg.png",
     "asset_id": "9fab6aaa-d8ac-4731-b287-1ff70e160480",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S28sh2.png",
     "asset_id": "550b93f5-f7c2-436b-a456-385435066af9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_utility_van_interior_df21e0.png",
     "asset_id": "f960bc7f-bc99-4b62-ace2-488d54c9aa93",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 자신의 무릎 쪽으로 시선을 내리깔고 있으며, 앰버는 오른쪽의 찰리를 향해 시선을 던지고 있습니다. 찰리 역시 고개를 숙여 앰버를 바라봅니다. 카메라는 앞좌석 뒤에서 대각선 아래로 인물들을 내려다봅니다.",
    "built_space": "밴 내부 공간으로, 전경에 앞좌석 등받이가 프레임을 분할하며 배치되어 있습니다. 화물 박스들이 어깨 높이 아래로 흩어져 있고, 인물들은 그 사이의 빈 공간에 자리 잡고 있습니다. 공간의 비례와 구조는 로케이션 참조와 일치합니다.",
    "entities": "이현우는 낡은 셔츠를 입고 있으나 무전기는 보이지 않습니다. 앰버는 마스크와 작업복을 착용했으나 머리를 묶은 형태입니다. 찰리는 기계 몸체에 코트와 모자로 변장했으나 프롬프트가 요구한 '푸른색 눈' 대신 참조 이미지의 주황색 눈을 유지하고 있습니다.",
    "hard_violations": [],
    "physics": "세 인물 모두 밴의 바닥에 안정적으로 웅크린 자세를 취하고 있으며, 찰리의 기계 팔 등 모든 신체 부위가 물리적으로 자연스럽게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우는 고개를 숙여 무릎 아래 공간을 응시하고, 앰버는 중앙에 웅크린 채 찰리를 바라봅니다. 찰리는 고개를 앰버 쪽으로 기울이고 마주 봅니다. 카메라 앵글은 지시된 대로 앞좌석 너머에서 대각선 아래로 향하고 있습니다.",
    "built_space": "위치 참조와 일치하는 밴 내부입니다. 카메라 앞쪽에 앞좌석 등받이가 일부 보이며, 인물들은 흩어진 종이상자와 금속 화물 박스들 사이에 몸을 숨기고 있습니다. 상자들의 크기와 배치가 인물들을 적절히 가려줍니다.",
    "entities": "이현우의 왼쪽 귀에 소형 인이어 무전기가 명확하게 보이며 의상과 외모가 참조와 일치합니다. 앰버는 참조 이미지대로 풀어내린 금발과 방진 마스크를 정확히 착용했습니다. 찰리는 코트와 벙거지 모자를 썼으며, 눈은 프롬프트의 푸른색 대신 주황색으로 표현되었습니다.",
    "hard_violations": [],
    "physics": "이현우와 앰버는 바닥에 발을 딛고 웅크린 채 체중을 안정적으로 지탱하고 있으며, 찰리 역시 바닥과 상자 위에 기계 팔을 얹어 자연스러운 은폐 자세를 유지하고 있습니다. 떠 있거나 지지되지 않은 객체는 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "앰버의 머리스타일(풀어내린 금발)을 참조 이미지와 동일하게 유지하고 이현우의 인이어 무전기를 정확히 묘사하여 캐릭터 디테일 구현도가 더 높습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "전반적인 구도와 인물들의 시선 처리는 훌륭하나, 앰버의 머리가 묶여 있어 참조와 다르고 이현우의 인이어 무전기가 보이지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 자신의 무릎 쪽으로 시선을 내리깔고 있으며, 앰버는 오른쪽의 찰리를 향해 시선을 던지고 있습니다. 찰리 역시 고개를 숙여 앰버를 바라봅니다. 카메라는 앞좌석 뒤에서 대각선 아래로 인물들을 내려다봅니다.",
        "built_space": "밴 내부 공간으로, 전경에 앞좌석 등받이가 프레임을 분할하며 배치되어 있습니다. 화물 박스들이 어깨 높이 아래로 흩어져 있고, 인물들은 그 사이의 빈 공간에 자리 잡고 있습니다. 공간의 비례와 구조는 로케이션 참조와 일치합니다.",
        "entities": "이현우는 낡은 셔츠를 입고 있으나 무전기는 보이지 않습니다. 앰버는 마스크와 작업복을 착용했으나 머리를 묶은 형태입니다. 찰리는 기계 몸체에 코트와 모자로 변장했으나 프롬프트가 요구한 '푸른색 눈' 대신 참조 이미지의 주황색 눈을 유지하고 있습니다.",
        "hard_violations": [],
        "physics": "세 인물 모두 밴의 바닥에 안정적으로 웅크린 자세를 취하고 있으며, 찰리의 기계 팔 등 모든 신체 부위가 물리적으로 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 고개를 숙여 무릎 아래 공간을 응시하고, 앰버는 중앙에 웅크린 채 찰리를 바라봅니다. 찰리는 고개를 앰버 쪽으로 기울이고 마주 봅니다. 카메라 앵글은 지시된 대로 앞좌석 너머에서 대각선 아래로 향하고 있습니다.",
        "built_space": "위치 참조와 일치하는 밴 내부입니다. 카메라 앞쪽에 앞좌석 등받이가 일부 보이며, 인물들은 흩어진 종이상자와 금속 화물 박스들 사이에 몸을 숨기고 있습니다. 상자들의 크기와 배치가 인물들을 적절히 가려줍니다.",
        "entities": "이현우의 왼쪽 귀에 소형 인이어 무전기가 명확하게 보이며 의상과 외모가 참조와 일치합니다. 앰버는 참조 이미지대로 풀어내린 금발과 방진 마스크를 정확히 착용했습니다. 찰리는 코트와 벙거지 모자를 썼으며, 눈은 프롬프트의 푸른색 대신 주황색으로 표현되었습니다.",
        "hard_violations": [],
        "physics": "이현우와 앰버는 바닥에 발을 딛고 웅크린 채 체중을 안정적으로 지탱하고 있으며, 찰리 역시 바닥과 상자 위에 기계 팔을 얹어 자연스러운 은폐 자세를 유지하고 있습니다. 떠 있거나 지지되지 않은 객체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "앰버의 머리스타일(풀어내린 금발)을 참조 이미지와 동일하게 유지하고 이현우의 인이어 무전기를 정확히 묘사하여 캐릭터 디테일 구현도가 더 높습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "전반적인 구도와 인물들의 시선 처리는 훌륭하나, 앰버의 머리가 묶여 있어 참조와 다르고 이현우의 인이어 무전기가 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 자신의 무릎 쪽으로 시선을 내리깔고 있으며, 앰버는 오른쪽의 찰리를 향해 시선을 던지고 있습니다. 찰리 역시 고개를 숙여 앰버를 바라봅니다. 카메라는 앞좌석 뒤에서 대각선 아래로 인물들을 내려다봅니다.",
        "built_space": "밴 내부 공간으로, 전경에 앞좌석 등받이가 프레임을 분할하며 배치되어 있습니다. 화물 박스들이 어깨 높이 아래로 흩어져 있고, 인물들은 그 사이의 빈 공간에 자리 잡고 있습니다. 공간의 비례와 구조는 로케이션 참조와 일치합니다.",
        "entities": "이현우는 낡은 셔츠를 입고 있으나 무전기는 보이지 않습니다. 앰버는 마스크와 작업복을 착용했으나 머리를 묶은 형태입니다. 찰리는 기계 몸체에 코트와 모자로 변장했으나 프롬프트가 요구한 '푸른색 눈' 대신 참조 이미지의 주황색 눈을 유지하고 있습니다.",
        "hard_violations": [],
        "physics": "세 인물 모두 밴의 바닥에 안정적으로 웅크린 자세를 취하고 있으며, 찰리의 기계 팔 등 모든 신체 부위가 물리적으로 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 고개를 숙여 무릎 아래 공간을 응시하고, 앰버는 중앙에 웅크린 채 찰리를 바라봅니다. 찰리는 고개를 앰버 쪽으로 기울이고 마주 봅니다. 카메라 앵글은 지시된 대로 앞좌석 너머에서 대각선 아래로 향하고 있습니다.",
        "built_space": "위치 참조와 일치하는 밴 내부입니다. 카메라 앞쪽에 앞좌석 등받이가 일부 보이며, 인물들은 흩어진 종이상자와 금속 화물 박스들 사이에 몸을 숨기고 있습니다. 상자들의 크기와 배치가 인물들을 적절히 가려줍니다.",
        "entities": "이현우의 왼쪽 귀에 소형 인이어 무전기가 명확하게 보이며 의상과 외모가 참조와 일치합니다. 앰버는 참조 이미지대로 풀어내린 금발과 방진 마스크를 정확히 착용했습니다. 찰리는 코트와 벙거지 모자를 썼으며, 눈은 프롬프트의 푸른색 대신 주황색으로 표현되었습니다.",
        "hard_violations": [],
        "physics": "이현우와 앰버는 바닥에 발을 딛고 웅크린 채 체중을 안정적으로 지탱하고 있으며, 찰리 역시 바닥과 상자 위에 기계 팔을 얹어 자연스러운 은폐 자세를 유지하고 있습니다. 떠 있거나 지지되지 않은 객체는 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "머리 위에서 내려다보는 각도와 상자 사이에 몸을 접어 숨은 구도가 B보다 가깝지만, 조수석 쪽 사선 시점은 약하고 찰리의 눈이 지정된 파란색이 아니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "세 인물의 시선 관계와 찰리의 외투 위장은 충실하지만, 더 낮고 정면적인 시점이 지정된 하향 사선 구도에서 멀어지며 찰리의 눈도 주황색이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 이현우는 상체와 고개를 접고 자기 무릎 주변을 내려다본다. 중앙 앰버의 눈은 오른쪽 찰리의 얼굴을 향한다. 찰리도 얼굴을 왼쪽 아래 앰버 쪽으로 기울인다. 요구된 세 인물의 시선 관계가 성립하며, 무기나 별도로 조준하는 물건은 없다.",
        "built_space": "뒤쪽에는 창이 하나씩 달린 양개문 두 짝, 좌우에는 각각 측창 두 면이 보인다. 가까운 양쪽 가장자리에 앞좌석 등받이 두 개가 잘려 들어오며, 낡은 밝은색 내벽과 어두운 바닥은 장소 참조와 부합한다. 십여 개의 종이 상자와 여러 하드케이스가 전경·인물 사이·후방에 나뉘어 있고, 어느 단일 상자도 화면의 사분의 일을 넘지 않는다. 세 인물은 왼쪽·중앙·오른쪽 틈을 차지한다. 상자의 윗면과 옆면이 보이는 하향 와이드 구도이지만 후문 중심을 향한 축이 강해 조수석 쪽에서 가로지르는 사선은 충분하지 않다. 뒷좌석 자체는 상자와 몸에 가려 식별하기 어렵고 천장 십자가 위치는 프레임 밖이다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며, 볼의 상처, 인이어 장치, 겉옷 없는 오염된 어두운 셔츠가 확인된다. 다리 부상의 구체적 상태는 가림 때문에 확정하기 어렵다. 앰버는 참조처럼 금발의 어린 여자아이이며 방진 마스크와 카키 작업복을 착용한다. 허리 도구 주머니와 기침 상태는 명확히 확인되지 않는다. 찰리는 샌드 베이지 장갑, 흰 기계 얼굴, 점 형태의 눈과 선 형태의 입, 큰 팔과 푸른 가슴 장치를 지닌 로봇이다. 낡은 모자와 외투도 있지만 외투는 어깨에 걸친 형태로 장갑 노출이 많다. 눈은 파란색 유지 지시와 달리 주황색이다. 등장 개체는 지정된 셋뿐이다.",
        "hard_violations": [],
        "physics": "이현우와 앰버는 무릎을 몸 가까이 접고 바닥 쪽으로 낮게 웅크린 자세이며, 하체 접점 일부는 전경 상자에 가려진다. 찰리의 팔과 손은 굽힌 하체 및 낮은 화물 옆으로 내려와 있다. 세 몸 모두 바닥까지 이어지는 공간 안에 자리하며 공중에 뜬 모습은 없다. 상자와 케이스는 바닥 또는 아래 화물에 놓여 있고, 모자는 머리에, 외투는 어깨에 걸려 있어 지지 관계가 자연스럽다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽에서 무릎과 손이 있는 아래 공간을 본다. 앰버는 중앙에서 오른쪽 위 찰리의 얼굴을 바라보고, 찰리는 고개를 왼쪽 아래로 숙여 앰버를 향한다. 앰버의 손은 자기 마스크 아래에 올라와 있으며 타인을 가리키거나 별도 물건을 조작하지 않는다. 세 인물의 방향 관계는 지시에 맞는다.",
        "built_space": "뒤 양개문 두 짝과 각 문의 창 하나, 좌우 측창 각각 두 면, 전경 양쪽의 잘린 앞좌석 등받이 두 개가 보인다. 내벽·창틀·문 손잡이와 야간 외부 풍경은 장소 참조에 가깝다. 십여 개의 종이 상자와 여러 하드케이스가 인물 앞뒤에 놓여 있고, 인물들은 지정된 좌·중·우의 틈에 들어가 있다. 전경 상자가 하체를 부분적으로 가리지만 단일 상자가 화면의 사분의 일을 차지하지는 않는다. 다만 카메라가 차량 중앙축에 가깝고 A보다 하향 각도가 약해, 머리 위 조수석 쪽에서 상자 틈을 대각선으로 보는 구도가 덜 구현됐다. 뒷좌석은 뚜렷하게 식별되지 않으며 천장 일부에는 십자가가 보이지 않지만 그 설치 지점까지 포함됐는지는 확정하기 어렵다.",
        "entities": "이현우는 젊은 동아시아계 남성으로 검은 머리, 얼굴 상처, 오염된 어두운 셔츠와 바지를 갖추고 겉옷은 없다. 손으로 다리 아래쪽을 감싼 자세가 부상 지속 상태와 양립한다. 앰버는 금발의 어린 여자아이로 마스크, 카키 작업복, 허리의 가죽 도구 주머니를 착용한다. 참조의 풀어 내린 머리와 달리 머리를 묶었으며 마스크 아래 손을 올린 모습은 기침 상태와 양립하지만 기침 자체를 확정할 수는 없다. 찰리는 흰 기계 얼굴과 베이지 장갑, 육중한 팔, 푸른 가슴 장치를 유지하며 낡은 외투와 챙 달린 모자로 위장한다. 눈은 요구된 파란색이 아니라 주황색이다. 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 낮게 앉아 한쪽 무릎을 세우고 두 손으로 다리 아래쪽을 잡는다. 앰버도 하체를 접어 바닥에 앉아 있으며 한쪽 팔꿈치와 팔을 굽힌 몸 가까이 둔다. 찰리의 굵은 팔은 몸 앞의 낮은 케이스와 접힌 하체 부근에 걸쳐 있고 손도 팔에 자연스럽게 연결된다. 화물은 바닥이나 다른 화물 위에 놓여 있다. 가려진 발의 접점을 모두 확인할 수는 없지만 지지 없는 공중 부유나 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "머리 위에서 내려다보는 각도와 상자 사이에 몸을 접어 숨은 구도가 B보다 가깝지만, 조수석 쪽 사선 시점은 약하고 찰리의 눈이 지정된 파란색이 아니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "세 인물의 시선 관계와 찰리의 외투 위장은 충실하지만, 더 낮고 정면적인 시점이 지정된 하향 사선 구도에서 멀어지며 찰리의 눈도 주황색이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 이현우는 상체와 고개를 접고 자기 무릎 주변을 내려다본다. 중앙 앰버의 눈은 오른쪽 찰리의 얼굴을 향한다. 찰리도 얼굴을 왼쪽 아래 앰버 쪽으로 기울인다. 요구된 세 인물의 시선 관계가 성립하며, 무기나 별도로 조준하는 물건은 없다.",
        "built_space": "뒤쪽에는 창이 하나씩 달린 양개문 두 짝, 좌우에는 각각 측창 두 면이 보인다. 가까운 양쪽 가장자리에 앞좌석 등받이 두 개가 잘려 들어오며, 낡은 밝은색 내벽과 어두운 바닥은 장소 참조와 부합한다. 십여 개의 종이 상자와 여러 하드케이스가 전경·인물 사이·후방에 나뉘어 있고, 어느 단일 상자도 화면의 사분의 일을 넘지 않는다. 세 인물은 왼쪽·중앙·오른쪽 틈을 차지한다. 상자의 윗면과 옆면이 보이는 하향 와이드 구도이지만 후문 중심을 향한 축이 강해 조수석 쪽에서 가로지르는 사선은 충분하지 않다. 뒷좌석 자체는 상자와 몸에 가려 식별하기 어렵고 천장 십자가 위치는 프레임 밖이다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며, 볼의 상처, 인이어 장치, 겉옷 없는 오염된 어두운 셔츠가 확인된다. 다리 부상의 구체적 상태는 가림 때문에 확정하기 어렵다. 앰버는 참조처럼 금발의 어린 여자아이이며 방진 마스크와 카키 작업복을 착용한다. 허리 도구 주머니와 기침 상태는 명확히 확인되지 않는다. 찰리는 샌드 베이지 장갑, 흰 기계 얼굴, 점 형태의 눈과 선 형태의 입, 큰 팔과 푸른 가슴 장치를 지닌 로봇이다. 낡은 모자와 외투도 있지만 외투는 어깨에 걸친 형태로 장갑 노출이 많다. 눈은 파란색 유지 지시와 달리 주황색이다. 등장 개체는 지정된 셋뿐이다.",
        "hard_violations": [],
        "physics": "이현우와 앰버는 무릎을 몸 가까이 접고 바닥 쪽으로 낮게 웅크린 자세이며, 하체 접점 일부는 전경 상자에 가려진다. 찰리의 팔과 손은 굽힌 하체 및 낮은 화물 옆으로 내려와 있다. 세 몸 모두 바닥까지 이어지는 공간 안에 자리하며 공중에 뜬 모습은 없다. 상자와 케이스는 바닥 또는 아래 화물에 놓여 있고, 모자는 머리에, 외투는 어깨에 걸려 있어 지지 관계가 자연스럽다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽에서 무릎과 손이 있는 아래 공간을 본다. 앰버는 중앙에서 오른쪽 위 찰리의 얼굴을 바라보고, 찰리는 고개를 왼쪽 아래로 숙여 앰버를 향한다. 앰버의 손은 자기 마스크 아래에 올라와 있으며 타인을 가리키거나 별도 물건을 조작하지 않는다. 세 인물의 방향 관계는 지시에 맞는다.",
        "built_space": "뒤 양개문 두 짝과 각 문의 창 하나, 좌우 측창 각각 두 면, 전경 양쪽의 잘린 앞좌석 등받이 두 개가 보인다. 내벽·창틀·문 손잡이와 야간 외부 풍경은 장소 참조에 가깝다. 십여 개의 종이 상자와 여러 하드케이스가 인물 앞뒤에 놓여 있고, 인물들은 지정된 좌·중·우의 틈에 들어가 있다. 전경 상자가 하체를 부분적으로 가리지만 단일 상자가 화면의 사분의 일을 차지하지는 않는다. 다만 카메라가 차량 중앙축에 가깝고 A보다 하향 각도가 약해, 머리 위 조수석 쪽에서 상자 틈을 대각선으로 보는 구도가 덜 구현됐다. 뒷좌석은 뚜렷하게 식별되지 않으며 천장 일부에는 십자가가 보이지 않지만 그 설치 지점까지 포함됐는지는 확정하기 어렵다.",
        "entities": "이현우는 젊은 동아시아계 남성으로 검은 머리, 얼굴 상처, 오염된 어두운 셔츠와 바지를 갖추고 겉옷은 없다. 손으로 다리 아래쪽을 감싼 자세가 부상 지속 상태와 양립한다. 앰버는 금발의 어린 여자아이로 마스크, 카키 작업복, 허리의 가죽 도구 주머니를 착용한다. 참조의 풀어 내린 머리와 달리 머리를 묶었으며 마스크 아래 손을 올린 모습은 기침 상태와 양립하지만 기침 자체를 확정할 수는 없다. 찰리는 흰 기계 얼굴과 베이지 장갑, 육중한 팔, 푸른 가슴 장치를 유지하며 낡은 외투와 챙 달린 모자로 위장한다. 눈은 요구된 파란색이 아니라 주황색이다. 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 낮게 앉아 한쪽 무릎을 세우고 두 손으로 다리 아래쪽을 잡는다. 앰버도 하체를 접어 바닥에 앉아 있으며 한쪽 팔꿈치와 팔을 굽힌 몸 가까이 둔다. 찰리의 굵은 팔은 몸 앞의 낮은 케이스와 접힌 하체 부근에 걸쳐 있고 손도 팔에 자연스럽게 연결된다. 화물은 바닥이나 다른 화물 위에 놓여 있다. 가려진 발의 접점을 모두 확인할 수는 없지만 지지 없는 공중 부유나 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.635,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.635,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1635
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "앰버의 머리스타일(풀어내린 금발)을 참조 이미지와 동일하게 유지하고 이현우의 인이어 무전기를 정확히 묘사하여 캐릭터 디테일 구현도가 더 높습니다."
   },
   {
    "label": "A",
    "score": 1635,
    "verdict_ko": "전반적인 구도와 인물들의 시선 처리는 훌륭하나, 앰버의 머리가 묶여 있어 참조와 다르고 이현우의 인이어 무전기가 보이지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_utility_van_interior_df21e0.png",
    "asset_id": "f960bc7f-bc99-4b62-ace2-488d54c9aa93",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09af-3f5f-784c-9396-8b32c8bcc552",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh2__bgfirst_bg.png",
   "bg_asset_id": "9fab6aaa-d8ac-4731-b287-1ff70e160480",
   "bg_record_key": "S28sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "utility_van_interior",
   "groupbg_asset_id": "f960bc7f-bc99-4b62-ace2-488d54c9aa93"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S28sh5::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 샷은 승합차 운전석에 앉은 신부가 창문을 내리고 외부를 향해 미소를 짓는 장면입니다. 차량 내부의 운전석 위치와 창문의 방향, 그리고 인물의 위치 관계가 정확히 일치해야만 이야기의 흐름이 깨지지 않으므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "d43979d3ec8fb3cf"
 },
 "S28sh5::signage": {
  "fp": "92906644ddc5b42e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::141a31ed2201": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_141a31ed2201.png",
  "place_text": "At the driver's seat inside the van, beside the lowered side window at a nighttime checkpoint. Exterior checkpoint light reaches the driver's face.",
  "input_fingerprint": "71ddf7015b9475bf"
 },
 "S28sh5::confined_fp": {
  "reads": {
   "controls": "The steering wheel is positioned at the front of the driver's seat along the bottom edge of the diagram.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "The camera is located in the front passenger seat on the top edge, pointing diagonally forward and left toward the driver.",
   "occupants": "The driver's seat on the bottom edge is occupied by the priest (신부)."
  },
  "mismatches": [],
  "scene_description_en": "The camera is stationed in the front passenger seat, looking diagonally forward and left toward the driver. The priest occupies the driver's seat, appearing in the center-left of the frame in profile as he looks outward. The lowered driver's side window sits on the left edge of the screen, serving as the background beyond his face. A section of the driver's seatback remains in view behind his shoulder. No mirrors or additional occupants appear from this vantage point.",
  "fixed": false,
  "input_fingerprint": "60dfe3c56121fb8d"
 },
 "S28sh5": {
  "input_fingerprint": "6f74e85c712f6fff",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차창을 내린 채 창밖을 향해 눈웃음을 지으며 여유롭게 미소 짓는 신부의 얼굴.\n\nLOCATION (lock): At the driver's seat inside the van, beside the lowered side window at a nighttime checkpoint. Exterior checkpoint light reaches the driver's face. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle inside the front cabin diagonally behind and to the passenger side of 신부, at his seated shoulder height with a slight upward view across his profile. His face occupies the center-left, with the lowered driver's window opening on the left edge and a narrow portion of the seat behind him; he turns toward the militia member outside the frame, smiling with relaxed, narrowed eyes. Hold the established position and let his outward eyeline and easy smile carry the change in emphasis, never redirecting his attention toward our lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 운전석 차창 개구부 (The window is lowered) — The open aperture lies beyond the priest's outward-facing profile; no glass intervenes in his sightline; used as Provides look room toward the unseen militia member; 운전석 등받이 (Supports the seated priest) — A narrow oblique section remains behind his shoulder; used as Retains cabin context within the face-centered framing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime cabin light gives the priest's smile gentle facial separation without importing the roof ornament's later illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van retains its roof-mounted halogen cross and rear cargo boxes; its driver's window is lowered for the checkpoint. Charlie is pushed down into concealment among the boxes, still in his old coat and hat. 신부: He remains at the wheel, wearing his clerical collar and smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the front passenger seat, looking diagonally forward and left toward the driver. The priest occupies the driver's seat, appearing in the center-left of the frame in profile as he looks outward. The lowered driver's side window sits on the left edge of the screen, serving as the background beyond his face. A section of the driver's seatback remains in view behind his shoulder. No mirrors or additional occupants appear from this vantage point.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차창을 내린 채 창밖을 향해 눈웃음을 지으며 여유롭게 미소 짓는 신부의 얼굴.\n\nLOCATION (lock): At the driver's seat inside the van, beside the lowered side window at a nighttime checkpoint. Exterior checkpoint light reaches the driver's face. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime cabin light gives the priest's smile gentle facial separation without importing the roof ornament's later illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van retains its roof-mounted halogen cross and rear cargo boxes; its driver's window is lowered for the checkpoint. Charlie is pushed down into concealment among the boxes, still in his old coat and hat. 신부: He remains at the wheel, wearing his clerical collar and smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed in the front passenger seat, looking diagonally forward and left toward the driver. The priest occupies the driver's seat, appearing in the center-left of the frame in profile as he looks outward. The lowered driver's side window sits on the left edge of the screen, serving as the background beyond his face. A section of the driver's seatback remains in view behind his shoulder. No mirrors or additional occupants appear from this vantage point.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차창을 내린 채 창밖을 향해 눈웃음을 지으며 여유롭게 미소 짓는 신부의 얼굴.\n\nLOCATION (lock): At the driver's seat inside the van, beside the lowered side window at a nighttime checkpoint. Exterior checkpoint light reaches the driver's face. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime cabin light gives the priest's smile gentle facial separation without importing the roof ornament's later illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The van retains its roof-mounted halogen cross and rear cargo boxes; its driver's window is lowered for the checkpoint. Charlie is pushed down into concealment among the boxes, still in his old coat and hat. 신부: He remains at the wheel, wearing his clerical collar and smiling.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh5_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "신부",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh5_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "신부",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
    "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이의 일부가 보이고 창밖으로는 야간 검문소가 위치합니다.",
    "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
    "hard_violations": [
     "[gpt-high] 창밖 중앙의 검문 인물 한 명과 왼쪽 가장자리의 부분 인물 한 명을 추가하여, 신부만 등장해야 한다는 조건을 위반했다."
    ],
    "physics": "인물은 운전석에 자연스럽게 앉아 운전대를 잡고 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
    "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이가 보이고 창밖으로는 야간 검문소가 위치합니다.",
    "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
    "hard_violations": [
     "[gpt-high] 화면 밖에 있어야 할 검문 인물 두 명을 창밖에 추가하여, 신부 외에는 누구도 등장시키지 말라는 조건을 위반했다."
    ],
    "physics": "인물은 운전석에 앉아 있으며 손은 프레임 하단의 운전대에 올려져 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 앵글, 인물의 위치 및 시선 처리, 그리고 외부 체크포인트의 조명 등 프롬프트의 지시사항을 정확하게 반영한 우수한 결과물입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프롬프트의 전반적인 지시사항을 잘 따랐으나, 창틀 부분의 렌더링이 약간 부자연스럽고 A에 비해 감정 표현의 디테일이 미세하게 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
        "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이의 일부가 보이고 창밖으로는 야간 검문소가 위치합니다.",
        "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "인물은 운전석에 자연스럽게 앉아 운전대를 잡고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
        "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이가 보이고 창밖으로는 야간 검문소가 위치합니다.",
        "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "인물은 운전석에 앉아 있으며 손은 프레임 하단의 운전대에 올려져 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 앵글, 인물의 위치 및 시선 처리, 그리고 외부 체크포인트의 조명 등 프롬프트의 지시사항을 정확하게 반영한 우수한 결과물입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프롬프트의 전반적인 지시사항을 잘 따랐으나, 창틀 부분의 렌더링이 약간 부자연스럽고 A에 비해 감정 표현의 디테일이 미세하게 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
        "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이의 일부가 보이고 창밖으로는 야간 검문소가 위치합니다.",
        "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "인물은 운전석에 자연스럽게 앉아 운전대를 잡고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 열린 왼쪽 창문 너머 프레임 밖을 향하고 있습니다.",
        "built_space": "밴 내부의 조수석 쪽에서 운전석을 바라보는 구도이며, 뒤편으로 등받이가 보이고 창밖으로는 야간 검문소가 위치합니다.",
        "entities": "레퍼런스와 일치하는 60대 한국인 남성이 로만칼라가 있는 검은색 사제복을 입고 있습니다.",
        "hard_violations": [],
        "physics": "인물은 운전석에 앉아 있으며 손은 프레임 하단의 운전대에 올려져 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "창밖을 향한 옆얼굴과 편안한 눈웃음은 더 충실하지만, 금지된 외부 인물 두 명을 추가했고 얼굴 중심 클로즈업보다 넓게 잡았다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "창밖 시선과 사제 복장은 맞지만, 외부 인물 두 명이 노출되고 상반신·운전대 중심의 넓은 구도와 덜 좁혀진 눈이 지정된 얼굴 클로즈업에서 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부의 얼굴과 눈은 화면 왼쪽의 열린 운전석 창밖을 향하며 렌즈를 보지 않는다. 눈을 편안하게 좁히고 입을 다문 채 미소 짓는다. 시선은 창밖 화면 밖으로 이어져 특정 경비 인물과 맞닿는지는 확인할 수 없다. 배경 인물이 든 발광 봉은 아래쪽을 향한다.",
        "built_space": "왼쪽에 운전석 창 개구부 하나, 오른쪽 어깨 뒤에 머리받침과 등받이 한 조, 왼쪽 아래에 운전대 일부가 보인다. 창은 내려가 있어 얼굴과 외부 사이에 유리가 보이지 않는다. 바깥에는 초소 하나와 조명 기둥 두 개, 차단 시설이 있다. 실내에서 비스듬히 바라보지만 지정된 뒤쪽 시점은 뚜렷하지 않고, 얼굴이 중앙보다 오른쪽에 있으며 상반신과 등받이를 너무 넓게 포함한다. 실제 이전 장소 사진은 제공되지 않아 마감과 배치의 연속성은 확인할 수 없다.",
        "entities": "짧은 회색 머리의 고령 동아시아계 남성 한 명이 검은 사제복과 흰 로만칼라를 착용하여 한국인 60대 신부라는 설정 및 인물 참조와 대체로 부합한다. 다만 창밖에 반사 조끼를 입은 인물 두 명이 추가되어 신부만 보여야 한다는 조건을 어긴다. 안전벨트가 보이며, 찰리·화물 상자·지붕 십자가는 이 구도 밖이므로 미노출 자체는 결함이 아니다. 글자나 도식 오버레이는 없다.",
        "hard_violations": [
         "화면 밖에 있어야 할 검문 인물 두 명을 창밖에 추가하여, 신부 외에는 누구도 등장시키지 말라는 조건을 위반했다."
        ],
        "physics": "신부는 운전석 등받이 앞에 자연스럽게 앉아 있고 안전벨트가 몸통을 가로지른다. 좌판과 골반은 화면 밖이지만 상체에 부유나 비정상적인 접촉은 없다. 팔은 운전대 쪽으로 뻗어 있으나 손의 접촉은 크롭 때문에 명확하지 않다. 배경 인물들은 지면에 서 있으며 발광 봉은 인물의 손 아래로 이어져 보인다."
       },
       {
        "label": "B",
        "direction": "신부는 화면 왼쪽 위의 창밖을 바라보며 카메라와 눈을 맞추지 않는다. 입꼬리가 올라가고 치아가 조금 보이지만 A보다 눈이 더 열려 있어 느긋한 눈웃음은 약하다. 시선의 대상인 화면 밖 민병대원은 확인되지 않으며, 보이는 배경 인물에게 정확히 시선이 닿는다고 단정할 수 없다.",
        "built_space": "왼쪽에 열린 운전석 창 하나, 신부 뒤에 머리받침과 등받이 한 조, 아래쪽 전경에 운전대 하나가 보인다. 창 개구부를 가로막는 유리는 없다. 외부에는 초소 하나, 두 개의 발광 면이 달린 조명 기둥 하나와 차단 시설이 보인다. 신부의 운전석 위치는 자연스럽지만, 얼굴은 중앙 오른쪽에 놓이고 손·팔·몸통·넓은 등받이까지 포함해 지정된 얼굴 중심 클로즈업보다 넓다. 뒤쪽에서 옆얼굴을 가로지르는 시점보다 앞쪽으로 열린 사선 구도에 가깝다.",
        "entities": "짧은 회색 머리와 주름진 얼굴의 고령 동아시아계 남성이 검은 사제복, 흰 로만칼라를 착용해 신부의 연령대·외형·복장 조건에 대체로 맞는다. 외부에는 반사 조끼를 입은 인물 한 명과 왼쪽 가장자리에 잘린 또 다른 인물이 있어 허용 인원 조건을 어긴다. 찰리·상자·지붕 십자가는 클로즈업에서 요구할 가시 요소가 아니다. 추가 문구나 도식은 없다.",
        "hard_violations": [
         "창밖 중앙의 검문 인물 한 명과 왼쪽 가장자리의 부분 인물 한 명을 추가하여, 신부만 등장해야 한다는 조건을 위반했다."
        ],
        "physics": "신부의 등과 어깨는 운전석 등받이 앞에 위치하고 안전벨트가 몸통에 밀착된다. 왼쪽 전경의 손은 운전대 윗부분을 잡고 있어 손과 사용 물체의 접촉이 성립한다. 하체는 화면 밖이며 보이는 상체에는 부유나 불가능한 관절 자세가 없다. 외부 인물들은 노면에 서 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "창밖을 향한 옆얼굴과 편안한 눈웃음은 더 충실하지만, 금지된 외부 인물 두 명을 추가했고 얼굴 중심 클로즈업보다 넓게 잡았다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "창밖 시선과 사제 복장은 맞지만, 외부 인물 두 명이 노출되고 상반신·운전대 중심의 넓은 구도와 덜 좁혀진 눈이 지정된 얼굴 클로즈업에서 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "신부의 얼굴과 눈은 화면 왼쪽의 열린 운전석 창밖을 향하며 렌즈를 보지 않는다. 눈을 편안하게 좁히고 입을 다문 채 미소 짓는다. 시선은 창밖 화면 밖으로 이어져 특정 경비 인물과 맞닿는지는 확인할 수 없다. 배경 인물이 든 발광 봉은 아래쪽을 향한다.",
        "built_space": "왼쪽에 운전석 창 개구부 하나, 오른쪽 어깨 뒤에 머리받침과 등받이 한 조, 왼쪽 아래에 운전대 일부가 보인다. 창은 내려가 있어 얼굴과 외부 사이에 유리가 보이지 않는다. 바깥에는 초소 하나와 조명 기둥 두 개, 차단 시설이 있다. 실내에서 비스듬히 바라보지만 지정된 뒤쪽 시점은 뚜렷하지 않고, 얼굴이 중앙보다 오른쪽에 있으며 상반신과 등받이를 너무 넓게 포함한다. 실제 이전 장소 사진은 제공되지 않아 마감과 배치의 연속성은 확인할 수 없다.",
        "entities": "짧은 회색 머리의 고령 동아시아계 남성 한 명이 검은 사제복과 흰 로만칼라를 착용하여 한국인 60대 신부라는 설정 및 인물 참조와 대체로 부합한다. 다만 창밖에 반사 조끼를 입은 인물 두 명이 추가되어 신부만 보여야 한다는 조건을 어긴다. 안전벨트가 보이며, 찰리·화물 상자·지붕 십자가는 이 구도 밖이므로 미노출 자체는 결함이 아니다. 글자나 도식 오버레이는 없다.",
        "hard_violations": [
         "화면 밖에 있어야 할 검문 인물 두 명을 창밖에 추가하여, 신부 외에는 누구도 등장시키지 말라는 조건을 위반했다."
        ],
        "physics": "신부는 운전석 등받이 앞에 자연스럽게 앉아 있고 안전벨트가 몸통을 가로지른다. 좌판과 골반은 화면 밖이지만 상체에 부유나 비정상적인 접촉은 없다. 팔은 운전대 쪽으로 뻗어 있으나 손의 접촉은 크롭 때문에 명확하지 않다. 배경 인물들은 지면에 서 있으며 발광 봉은 인물의 손 아래로 이어져 보인다."
       },
       {
        "label": "A",
        "direction": "신부는 화면 왼쪽 위의 창밖을 바라보며 카메라와 눈을 맞추지 않는다. 입꼬리가 올라가고 치아가 조금 보이지만 A보다 눈이 더 열려 있어 느긋한 눈웃음은 약하다. 시선의 대상인 화면 밖 민병대원은 확인되지 않으며, 보이는 배경 인물에게 정확히 시선이 닿는다고 단정할 수 없다.",
        "built_space": "왼쪽에 열린 운전석 창 하나, 신부 뒤에 머리받침과 등받이 한 조, 아래쪽 전경에 운전대 하나가 보인다. 창 개구부를 가로막는 유리는 없다. 외부에는 초소 하나, 두 개의 발광 면이 달린 조명 기둥 하나와 차단 시설이 보인다. 신부의 운전석 위치는 자연스럽지만, 얼굴은 중앙 오른쪽에 놓이고 손·팔·몸통·넓은 등받이까지 포함해 지정된 얼굴 중심 클로즈업보다 넓다. 뒤쪽에서 옆얼굴을 가로지르는 시점보다 앞쪽으로 열린 사선 구도에 가깝다.",
        "entities": "짧은 회색 머리와 주름진 얼굴의 고령 동아시아계 남성이 검은 사제복, 흰 로만칼라를 착용해 신부의 연령대·외형·복장 조건에 대체로 맞는다. 외부에는 반사 조끼를 입은 인물 한 명과 왼쪽 가장자리에 잘린 또 다른 인물이 있어 허용 인원 조건을 어긴다. 찰리·상자·지붕 십자가는 클로즈업에서 요구할 가시 요소가 아니다. 추가 문구나 도식은 없다.",
        "hard_violations": [
         "창밖 중앙의 검문 인물 한 명과 왼쪽 가장자리의 부분 인물 한 명을 추가하여, 신부만 등장해야 한다는 조건을 위반했다."
        ],
        "physics": "신부의 등과 어깨는 운전석 등받이 앞에 위치하고 안전벨트가 몸통에 밀착된다. 왼쪽 전경의 손은 운전대 윗부분을 잡고 있어 손과 사용 물체의 접촉이 성립한다. 하체는 화면 밖이며 보이는 상체에는 부유나 불가능한 관절 자세가 없다. 외부 인물들은 노면에 서 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.607
   },
   "violations": {
    "B": [
     "[gpt-high] 화면 밖에 있어야 할 검문 인물 두 명을 창밖에 추가하여, 신부 외에는 누구도 등장시키지 말라는 조건을 위반했다."
    ],
    "A": [
     "[gpt-high] 창밖 중앙의 검문 인물 한 명과 왼쪽 가장자리의 부분 인물 한 명을 추가하여, 신부만 등장해야 한다는 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1417,
   "B": 1607
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "카메라 앵글, 인물의 위치 및 시선 처리, 그리고 외부 체크포인트의 조명 등 프롬프트의 지시사항을 정확하게 반영한 우수한 결과물입니다.  ★위반: [gpt-high] 창밖 중앙의 검문 인물 한 명과 왼쪽 가장자리의 부분 인물 한 명을 추가하여, 신부만 등장해야 한다는 조건을 위반했다."
   },
   {
    "label": "B",
    "score": 1607,
    "verdict_ko": "프롬프트의 전반적인 지시사항을 잘 따랐으나, 창틀 부분의 렌더링이 약간 부자연스럽고 A에 비해 감정 표현의 디테일이 미세하게 부족합니다.  ★위반: [gpt-high] 화면 밖에 있어야 할 검문 인물 두 명을 창밖에 추가하여, 신부 외에는 누구도 등장시키지 말라는 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S28sh5_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "신부",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09b8-e17d-79a9-a507-1f16e1dcfec0",
  "confined_fp": {
   "base_key": "confinedfp::141a31ed2201",
   "apt_reason": "이 샷은 승합차 운전석에 앉은 신부가 창문을 내리고 외부를 향해 미소를 짓는 장면입니다. 차량 내부의 운전석 위치와 창문의 방향, 그리고 인물의 위치 관계가 정확히 일치해야만 이야기의 흐름이 깨지지 않으므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh2"
  }
 },
 "S29sh5::signage": {
  "fp": "1c03b66a4ae557b2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::25a2329362ef654b": {
  "subjects": [],
  "subject_text": "인천 성당 지하 기도소, 지하실 연결 계단\n계단 아래의 단출한 지하 기도실. 십자가와 작은 책상, 의자, 풍금이 놓여 있으며, 구석의 낡은 라디오와 촛불·랜턴이 보인다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L174",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::basement_prayer_room": {
  "input_fingerprint": "8c0926c963c5515a",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "basement_prayer_room",
    "tags": [
     "S29sh5",
     "S30sh10",
     "S30sh14",
     "S30sh6",
     "S32sh1",
     "S32sh11",
     "S32sh5"
    ]
   },
   "context_sig": "05c3c3f68bd5cd00"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 성당 지하 기도소, 지하실 연결 계단: 촛불과 소형 랜턴에 의지하는 좁고 은밀한 지하 제단이다. (특징: 목재 십자가, 작은 책상, 낡은 풍금 라디오와 의자; 찰리의 가슴 발광에 맞춰 사방의 랜턴들이 켜지는 모습; 낡은 라디오의 주파수 다이얼이 저절로 움직이며 기계적 음성 조합을 내는 장비 작동; 계단을 뛰어 내려오는 신부와 방독/연구 장비 흔적이 없는 일반 복장의 구도환)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 십자가와 작은 책상, 의자, 작은 풍금이 덩그러니 놓여 있는 단출한 공간.\n- 작은 풍금을 연주하고 있는 찰리, ♬ \"Country Road, Take Me Home\" 곁에서 신기한 눈으로 바라보는 앰버.\n- 32. 성당. 지하실 - N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 성당 지하 기도소, 지하실 연결 계단: 촛불과 소형 랜턴에 의지하는 좁고 은밀한 지하 제단이다. (특징: 목재 십자가, 작은 책상, 낡은 풍금 라디오와 의자; 찰리의 가슴 발광에 맞춰 사방의 랜턴들이 켜지는 모습; 낡은 라디오의 주파수 다이얼이 저절로 움직이며 기계적 음성 조합을 내는 장비 작동; 계단을 뛰어 내려오는 신부와 방독/연구 장비 흔적이 없는 일반 복장의 구도환)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 십자가와 작은 책상, 의자, 작은 풍금이 덩그러니 놓여 있는 단출한 공간.\n- 작은 풍금을 연주하고 있는 찰리, ♬ \"Country Road, Take Me Home\" 곁에서 신기한 눈으로 바라보는 앰버.\n- 32. 성당. 지하실 - N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_basement_prayer_room_5d61e9.png",
  "asset_id": "2aaacaed-7cb6-454b-8aea-3633d16bda64",
  "input_asset_ids": [
   "ab6cbe3a-e4fd-4ee3-8c9d-4b81f3a0cc3c"
  ],
  "origin_tag": "S29sh5",
  "place_text": "Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.",
  "origin_inputs": {
   "place_text": "Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.",
   "time_of_day_en": "night",
   "conti_asset_id": "ab6cbe3a-e4fd-4ee3-8c9d-4b81f3a0cc3c"
  }
 },
 "S29sh5::bgfirst_bg": {
  "input_fingerprint": "447120bbd7134390",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 위층의 둔탁한 소리에 일제히 숨을 죽인 채 굳은 얼굴로 천장을 올려다보는 신부, 앰버, 찰리, 이현우.\n\nLOCATION (lock): Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at a low oblique position inside the prayer room, looking slightly upward across the four figures with open space above their heads and a strip of ceiling visible. 신부 at left arrests his movement, 앰버 draws her shoulders inward near center-left, 찰리 pauses with his previously raised hand still suspended at center-right, and 이현우 leans back slightly at right; all direct naturally inclined upward gazes toward the ceiling after the knocking. Keep their spacing and illumination stable, emphasizing the shared upward eyeline while varying head angles and bodily tension so their frightened stillness does not become a synchronized lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 기도소 천장 (Above the group as knocking comes from the floor upstairs) — A narrow underside section is visible above the figures; used as Connects their upward attention to the source direction without requiring extreme neck angles; 작은 책상과 의자 (Sparsely placed in the prayer room) — Oblique sides remain visible behind the group; used as Give the lower background scale and preserve the room's sparse layout; 작은 풍금 (Present in the prayer room and not being played) — A partial side view sits at the background edge; used as Maintains environmental continuity without competing with the collective reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established candlelight gives restrained, localized illumination, leaving the surrounding prayer room subdued while preserving each startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 위층의 둔탁한 소리에 일제히 숨을 죽인 채 굳은 얼굴로 천장을 올려다보는 신부, 앰버, 찰리, 이현우.\n\nLOCATION (lock): Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at a low oblique position inside the prayer room, looking slightly upward across the four figures with open space above their heads and a strip of ceiling visible. 신부 at left arrests his movement, 앰버 draws her shoulders inward near center-left, 찰리 pauses with his previously raised hand still suspended at center-right, and 이현우 leans back slightly at right; all direct naturally inclined upward gazes toward the ceiling after the knocking. Keep their spacing and illumination stable, emphasizing the shared upward eyeline while varying head angles and bodily tension so their frightened stillness does not become a synchronized lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 기도소 천장 (Above the group as knocking comes from the floor upstairs) — A narrow underside section is visible above the figures; used as Connects their upward attention to the source direction without requiring extreme neck angles; 작은 책상과 의자 (Sparsely placed in the prayer room) — Oblique sides remain visible behind the group; used as Give the lower background scale and preserve the room's sparse layout; 작은 풍금 (Present in the prayer room and not being played) — A partial side view sits at the background edge; used as Maintains environmental continuity without competing with the collective reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established candlelight gives restrained, localized illumination, leaving the surrounding prayer room subdued while preserving each startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh5__bgfirst_bg.png",
  "asset_id": "69138a0b-9ed7-46b9-b365-6866080ed7c6",
  "input_asset_ids": [
   "ab6cbe3a-e4fd-4ee3-8c9d-4b81f3a0cc3c",
   "2aaacaed-7cb6-454b-8aea-3633d16bda64"
  ]
 },
 "S29sh5": {
  "input_fingerprint": "77ece651cc446112",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위층의 둔탁한 소리에 일제히 숨을 죽인 채 굳은 얼굴로 천장을 올려다보는 신부, 앰버, 찰리, 이현우.\n\nLOCATION (lock): Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at a low oblique position inside the prayer room, looking slightly upward across the four figures with open space above their heads and a strip of ceiling visible. 신부 at left arrests his movement, 앰버 draws her shoulders inward near center-left, 찰리 pauses with his previously raised hand still suspended at center-right, and 이현우 leans back slightly at right; all direct naturally inclined upward gazes toward the ceiling after the knocking. Keep their spacing and illumination stable, emphasizing the shared upward eyeline while varying head angles and bodily tension so their frightened stillness does not become a synchronized lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 기도소 천장 (Above the group as knocking comes from the floor upstairs) — A narrow underside section is visible above the figures; used as Connects their upward attention to the source direction without requiring extreme neck angles; 작은 책상과 의자 (Sparsely placed in the prayer room) — Oblique sides remain visible behind the group; used as Give the lower background scale and preserve the room's sparse layout; 작은 풍금 (Present in the prayer room and not being played) — A partial side view sits at the background edge; used as Maintains environmental continuity without competing with the collective reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established candlelight gives restrained, localized illumination, leaving the surrounding prayer room subdued while preserving each startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, small desk, chairs and a small organ, with candle illumination; the church's electric bulbs remain burst. Charlie retains his old coat and hat disguise and blue-lit eyes. 신부: He remains in the basement, wearing his clerical collar and carrying the candle used on the stairs. 이현우: He is in the basement with facial bruises and the injured leg, still without his outer garment. 앰버: She is in the basement, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위층의 둔탁한 소리에 일제히 숨을 죽인 채 굳은 얼굴로 천장을 올려다보는 신부, 앰버, 찰리, 이현우.\n\nLOCATION (lock): Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at a low oblique position inside the prayer room, looking slightly upward across the four figures with open space above their heads and a strip of ceiling visible. 신부 at left arrests his movement, 앰버 draws her shoulders inward near center-left, 찰리 pauses with his previously raised hand still suspended at center-right, and 이현우 leans back slightly at right; all direct naturally inclined upward gazes toward the ceiling after the knocking. Keep their spacing and illumination stable, emphasizing the shared upward eyeline while varying head angles and bodily tension so their frightened stillness does not become a synchronized lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 기도소 천장 (Above the group as knocking comes from the floor upstairs) — A narrow underside section is visible above the figures; used as Connects their upward attention to the source direction without requiring extreme neck angles; 작은 책상과 의자 (Sparsely placed in the prayer room) — Oblique sides remain visible behind the group; used as Give the lower background scale and preserve the room's sparse layout; 작은 풍금 (Present in the prayer room and not being played) — A partial side view sits at the background edge; used as Maintains environmental continuity without competing with the collective reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established candlelight gives restrained, localized illumination, leaving the surrounding prayer room subdued while preserving each startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, small desk, chairs and a small organ, with candle illumination; the church's electric bulbs remain burst. Charlie retains his old coat and hat disguise and blue-lit eyes. 신부: He remains in the basement, wearing his clerical collar and carrying the candle used on the stairs. 이현우: He is in the basement with facial bruises and the injured leg, still without his outer garment. 앰버: She is in the basement, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위층의 둔탁한 소리에 일제히 숨을 죽인 채 굳은 얼굴로 천장을 올려다보는 신부, 앰버, 찰리, 이현우.\n\nLOCATION (lock): Inside the church's sparsely furnished basement prayer room, below the stairs to the ground floor. A carried candle lights the room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at a low oblique position inside the prayer room, looking slightly upward across the four figures with open space above their heads and a strip of ceiling visible. 신부 at left arrests his movement, 앰버 draws her shoulders inward near center-left, 찰리 pauses with his previously raised hand still suspended at center-right, and 이현우 leans back slightly at right; all direct naturally inclined upward gazes toward the ceiling after the knocking. Keep their spacing and illumination stable, emphasizing the shared upward eyeline while varying head angles and bodily tension so their frightened stillness does not become a synchronized lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 기도소 천장 (Above the group as knocking comes from the floor upstairs) — A narrow underside section is visible above the figures; used as Connects their upward attention to the source direction without requiring extreme neck angles; 작은 책상과 의자 (Sparsely placed in the prayer room) — Oblique sides remain visible behind the group; used as Give the lower background scale and preserve the room's sparse layout; 작은 풍금 (Present in the prayer room and not being played) — A partial side view sits at the background edge; used as Maintains environmental continuity without competing with the collective reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established candlelight gives restrained, localized illumination, leaving the surrounding prayer room subdued while preserving each startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement prayer room contains a cross, small desk, chairs and a small organ, with candle illumination; the church's electric bulbs remain burst. Charlie retains his old coat and hat disguise and blue-lit eyes. 신부: He remains in the basement, wearing his clerical collar and carrying the candle used on the stairs. 이현우: He is in the basement with facial bruises and the injured leg, still without his outer garment. 앰버: She is in the basement, retaining her mask and waist tool pouch and her unresolved coughing condition.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh5__bgfirst_bg.png",
     "asset_id": "69138a0b-9ed7-46b9-b365-6866080ed7c6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S29sh5.png",
     "asset_id": "ab6cbe3a-e4fd-4ee3-8c9d-4b81f3a0cc3c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_basement_prayer_room_5d61e9.png",
     "asset_id": "2aaacaed-7cb6-454b-8aea-3633d16bda64",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "4명의 인물 모두 시선이 위쪽 천장을 향하고 있음.",
    "built_space": "지하실 기도실 내부. 지정된 가구(풍금, 책상, 의자, 십자가)가 배치되어 있으며 로우 앵글 구도가 적용됨.",
    "entities": "신부, 앰버, 이현우의 묘사는 양호하나, 찰리가 레퍼런스(키 작고 뚱뚱한 체형)와 달리 그룹 내에서 가장 키가 큰 비율로 왜곡되어 나타남.",
    "hard_violations": [],
    "physics": "모든 인물이 지면에 온전히 서서 자세를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "신부, 앰버, 찰리, 이현우 모두 위층 소리의 근원지인 천장 쪽으로 시선을 고정하고 있음.",
    "built_space": "지하실 기도실 내부. 로우 앵글 구도를 통해 천장 일부가 보이며, 좌측 계단과 풍금, 중앙의 책상 및 배경 십자가가 알맞게 배치됨.",
    "entities": "신부(수단, 양초), 앰버(마스크, 작업복), 이현우(상처, 붕대, 인이어), 찰리(고릴라형 체형, 코트, 모자) 모두 레퍼런스의 외형과 체형 크기를 정확히 반영함.",
    "hard_violations": [],
    "physics": "4명의 인물 모두 바닥에 안정적으로 발을 딛고 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 구도와 천장을 향한 시선을 완벽히 구현했으며, 찰리의 작고 뚱뚱한 체형과 이현우의 인이어 등 세부 설정까지 정확히 묘사했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 작고 뚱뚱해야 할 찰리의 체형이 지나치게 크고 길게 묘사되어 캐릭터 레퍼런스에 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "신부, 앰버, 찰리, 이현우 모두 위층 소리의 근원지인 천장 쪽으로 시선을 고정하고 있음.",
        "built_space": "지하실 기도실 내부. 로우 앵글 구도를 통해 천장 일부가 보이며, 좌측 계단과 풍금, 중앙의 책상 및 배경 십자가가 알맞게 배치됨.",
        "entities": "신부(수단, 양초), 앰버(마스크, 작업복), 이현우(상처, 붕대, 인이어), 찰리(고릴라형 체형, 코트, 모자) 모두 레퍼런스의 외형과 체형 크기를 정확히 반영함.",
        "hard_violations": [],
        "physics": "4명의 인물 모두 바닥에 안정적으로 발을 딛고 서 있음."
       },
       {
        "label": "A",
        "direction": "4명의 인물 모두 시선이 위쪽 천장을 향하고 있음.",
        "built_space": "지하실 기도실 내부. 지정된 가구(풍금, 책상, 의자, 십자가)가 배치되어 있으며 로우 앵글 구도가 적용됨.",
        "entities": "신부, 앰버, 이현우의 묘사는 양호하나, 찰리가 레퍼런스(키 작고 뚱뚱한 체형)와 달리 그룹 내에서 가장 키가 큰 비율로 왜곡되어 나타남.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 온전히 서서 자세를 유지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 구도와 천장을 향한 시선을 완벽히 구현했으며, 찰리의 작고 뚱뚱한 체형과 이현우의 인이어 등 세부 설정까지 정확히 묘사했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 작고 뚱뚱해야 할 찰리의 체형이 지나치게 크고 길게 묘사되어 캐릭터 레퍼런스에 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부, 앰버, 찰리, 이현우 모두 위층 소리의 근원지인 천장 쪽으로 시선을 고정하고 있음.",
        "built_space": "지하실 기도실 내부. 로우 앵글 구도를 통해 천장 일부가 보이며, 좌측 계단과 풍금, 중앙의 책상 및 배경 십자가가 알맞게 배치됨.",
        "entities": "신부(수단, 양초), 앰버(마스크, 작업복), 이현우(상처, 붕대, 인이어), 찰리(고릴라형 체형, 코트, 모자) 모두 레퍼런스의 외형과 체형 크기를 정확히 반영함.",
        "hard_violations": [],
        "physics": "4명의 인물 모두 바닥에 안정적으로 발을 딛고 서 있음."
       },
       {
        "label": "A",
        "direction": "4명의 인물 모두 시선이 위쪽 천장을 향하고 있음.",
        "built_space": "지하실 기도실 내부. 지정된 가구(풍금, 책상, 의자, 십자가)가 배치되어 있으며 로우 앵글 구도가 적용됨.",
        "entities": "신부, 앰버, 이현우의 묘사는 양호하나, 찰리가 레퍼런스(키 작고 뚱뚱한 체형)와 달리 그룹 내에서 가장 키가 큰 비율로 왜곡되어 나타남.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 온전히 서서 자세를 유지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낮은 와이드 구도와 찰리의 짧고 육중한 체형은 잘 맞지만, 찰리가 천장보다 정면을 바라보고 이현우도 앞으로 기울어 핵심 반응 연출이 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "네 인물의 천장 방향 시선과 서로 다른 긴장 자세를 더 충실히 구현했으나, 찰리의 큰 키와 길어진 체형, 주황색 눈은 지정과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부는 왼쪽 위 천장, 앰버는 자기 머리 위쪽, 이현우는 오른쪽 위 천장을 바라본다. 찰리의 얼굴과 두 눈은 거의 카메라 쪽 정면을 향해 있어 천장의 소리 근원을 함께 올려다보는 반응이 분명하지 않다. 찰리의 든 손은 가슴 앞에서 멈춰 있으며 특정 대상을 가리키지는 않는다.",
        "built_space": "낮은 실내 시점에서 네 인물이 신부–앰버–찰리–이현우 순으로 배치된다. 위쪽에 보와 배관이 있는 콘크리트 천장, 왼쪽에 위층으로 올라가는 계단, 뒤쪽에 십자가 하나와 작은 책상 하나가 보인다. 왼쪽 풍금 하나는 인물에 상당 부분 가려져 있다. 왼쪽 의자 두 개와 풍금 앞 낮은 걸상 하나, 책상 오른쪽 의자 하나, 오른쪽 벽의 의자들이 부분적으로 보이며 후자의 정확한 수는 가림 때문에 확정하기 어렵다. 오른쪽 책장 하나와 높은 창도 장소 참조와 연결된다. 천장 면적은 요구한 좁은 띠보다 넓고 인물들의 하단은 프레임에 잘린다.",
        "entities": "등장 인물은 지정된 네 명뿐이다. 신부는 짧은 회색 머리의 동아시아계 노년 남성으로 검은 사제복과 흰 로만칼라를 착용하고 촛대 하나를 들고 있다. 앰버는 금발의 어린 소녀로 방진 마스크, 카키 작업복, 허리 공구 주머니가 보이며 참조보다 약간 성숙해 보인다. 찰리는 흰 기계 얼굴, 베이지 장갑판, 긴 팔과 짧은 다리, 낡은 외투와 모자, 푸른 가슴 원자로를 갖췄지만 눈은 지정된 파란색이 아니라 주황색이다. 이현우는 헝클어진 검은 머리의 동아시아계 청소년 남성으로 얼굴 상처, 인이어 장치, 더러운 셔츠와 바지, 다리 붕대가 보이고 겉옷은 없다. 실내는 촛불 중심의 어두운 환경이다.",
        "hard_violations": [],
        "physics": "신부의 손이 촛대 손잡이를 잡고 있어 초가 지지된다. 앰버의 두 손은 가슴 앞에서 서로 맞닿고, 찰리의 든 손은 굽힌 팔과 어깨 관절로 지지된다. 인물들의 발은 대부분 프레임 밖이지만 하체는 바닥 쪽으로 자연스럽게 이어지며 공중에 떠 있다는 징후는 없다. 이현우는 몸을 약간 앞으로 숙여 지정된 뒤로 물러나는 긴장과 다르지만 물리적으로 불가능한 자세는 아니다. 배경 가구와 촛불은 바닥이나 가구 상판에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "신부는 왼쪽 위 천장, 앰버는 중앙 위 천장, 찰리는 고개를 들어 오른쪽 위 천장, 이현우는 더 오른쪽 위 천장으로 주의를 향한다. 찰리는 동공 대신 발광 눈을 갖지만 얼굴 면 자체가 위로 기울어 소리의 근원을 향한 반응이 읽힌다. 네 인물의 고개 각도는 서로 다르며 찰리의 손은 가슴 높이에 든 채 멈춰 있다.",
        "built_space": "실내의 낮은 사선 시점과 신부–앰버–찰리–이현우의 좌우 순서를 지킨다. 콘크리트 천장과 보·배관, 왼쪽 상승 계단, 오른쪽 높은 창이 참조 장소와 맞는다. 뒤쪽에 십자가 하나, 작은 책상 하나, 왼쪽 가장자리의 부분적으로 가려진 풍금 하나, 오른쪽 책장 하나가 있다. 왼쪽 의자 두 개와 낮은 걸상 하나, 책상 오른쪽 의자 하나, 오른쪽 벽의 여러 의자 일부가 보인다. 인물들이 가구 앞에 서 있어 자리 충돌은 없다. 머리 위 여백은 확보됐지만 천장은 좁은 띠보다 넓게 드러나고, 찰리는 요구보다 크게 화면을 차지한다.",
        "entities": "지정된 네 인물만 보인다. 신부는 노년의 동아시아계 남성으로 로만칼라와 검은 사제복을 착용했지만 참조의 상하의보다 긴 수단 형태다. 금발 어린 소녀 앰버는 마스크, 오염된 카키 작업복, 공구 벨트와 주머니를 유지한다. 찰리는 베이지 기계 몸체와 흰 마스크형 얼굴, 모자와 낡은 외투, 푸른 가슴 원자로를 갖췄으나 짧고 뚱뚱한 고릴라형보다 키 크고 길쭉하며 눈도 파란색이 아닌 주황색이다. 이현우는 검은 머리의 동아시아계 청소년 남성으로 얼굴의 멍과 상처, 낡고 더러운 셔츠, 붕대 감긴 다리가 보이고 겉옷은 없다. 인이어 장치는 이 각도에서 명확히 식별되지 않는다.",
        "hard_violations": [],
        "physics": "신부는 초의 아래쪽을 손으로 잡고 있다. 앰버는 어깨를 안으로 모으고 손을 가슴에 붙여 실제로 가능한 움츠림을 보인다. 찰리의 정지한 손은 팔꿈치와 어깨로 지지된다. 이현우는 한쪽 다리를 내밀고 다른 다리에 체중을 둔 채 상체를 뒤로 기울이는 자세로 읽히며, 발 접점은 프레임 밖이지만 무지지 부유의 징후는 없다. 네 인물의 몸과 의복은 중력에 맞게 이어지고 배경 물건들도 상판이나 바닥의 지지를 받는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낮은 와이드 구도와 찰리의 짧고 육중한 체형은 잘 맞지만, 찰리가 천장보다 정면을 바라보고 이현우도 앞으로 기울어 핵심 반응 연출이 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "네 인물의 천장 방향 시선과 서로 다른 긴장 자세를 더 충실히 구현했으나, 찰리의 큰 키와 길어진 체형, 주황색 눈은 지정과 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 왼쪽 위 천장, 앰버는 자기 머리 위쪽, 이현우는 오른쪽 위 천장을 바라본다. 찰리의 얼굴과 두 눈은 거의 카메라 쪽 정면을 향해 있어 천장의 소리 근원을 함께 올려다보는 반응이 분명하지 않다. 찰리의 든 손은 가슴 앞에서 멈춰 있으며 특정 대상을 가리키지는 않는다.",
        "built_space": "낮은 실내 시점에서 네 인물이 신부–앰버–찰리–이현우 순으로 배치된다. 위쪽에 보와 배관이 있는 콘크리트 천장, 왼쪽에 위층으로 올라가는 계단, 뒤쪽에 십자가 하나와 작은 책상 하나가 보인다. 왼쪽 풍금 하나는 인물에 상당 부분 가려져 있다. 왼쪽 의자 두 개와 풍금 앞 낮은 걸상 하나, 책상 오른쪽 의자 하나, 오른쪽 벽의 의자들이 부분적으로 보이며 후자의 정확한 수는 가림 때문에 확정하기 어렵다. 오른쪽 책장 하나와 높은 창도 장소 참조와 연결된다. 천장 면적은 요구한 좁은 띠보다 넓고 인물들의 하단은 프레임에 잘린다.",
        "entities": "등장 인물은 지정된 네 명뿐이다. 신부는 짧은 회색 머리의 동아시아계 노년 남성으로 검은 사제복과 흰 로만칼라를 착용하고 촛대 하나를 들고 있다. 앰버는 금발의 어린 소녀로 방진 마스크, 카키 작업복, 허리 공구 주머니가 보이며 참조보다 약간 성숙해 보인다. 찰리는 흰 기계 얼굴, 베이지 장갑판, 긴 팔과 짧은 다리, 낡은 외투와 모자, 푸른 가슴 원자로를 갖췄지만 눈은 지정된 파란색이 아니라 주황색이다. 이현우는 헝클어진 검은 머리의 동아시아계 청소년 남성으로 얼굴 상처, 인이어 장치, 더러운 셔츠와 바지, 다리 붕대가 보이고 겉옷은 없다. 실내는 촛불 중심의 어두운 환경이다.",
        "hard_violations": [],
        "physics": "신부의 손이 촛대 손잡이를 잡고 있어 초가 지지된다. 앰버의 두 손은 가슴 앞에서 서로 맞닿고, 찰리의 든 손은 굽힌 팔과 어깨 관절로 지지된다. 인물들의 발은 대부분 프레임 밖이지만 하체는 바닥 쪽으로 자연스럽게 이어지며 공중에 떠 있다는 징후는 없다. 이현우는 몸을 약간 앞으로 숙여 지정된 뒤로 물러나는 긴장과 다르지만 물리적으로 불가능한 자세는 아니다. 배경 가구와 촛불은 바닥이나 가구 상판에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "신부는 왼쪽 위 천장, 앰버는 중앙 위 천장, 찰리는 고개를 들어 오른쪽 위 천장, 이현우는 더 오른쪽 위 천장으로 주의를 향한다. 찰리는 동공 대신 발광 눈을 갖지만 얼굴 면 자체가 위로 기울어 소리의 근원을 향한 반응이 읽힌다. 네 인물의 고개 각도는 서로 다르며 찰리의 손은 가슴 높이에 든 채 멈춰 있다.",
        "built_space": "실내의 낮은 사선 시점과 신부–앰버–찰리–이현우의 좌우 순서를 지킨다. 콘크리트 천장과 보·배관, 왼쪽 상승 계단, 오른쪽 높은 창이 참조 장소와 맞는다. 뒤쪽에 십자가 하나, 작은 책상 하나, 왼쪽 가장자리의 부분적으로 가려진 풍금 하나, 오른쪽 책장 하나가 있다. 왼쪽 의자 두 개와 낮은 걸상 하나, 책상 오른쪽 의자 하나, 오른쪽 벽의 여러 의자 일부가 보인다. 인물들이 가구 앞에 서 있어 자리 충돌은 없다. 머리 위 여백은 확보됐지만 천장은 좁은 띠보다 넓게 드러나고, 찰리는 요구보다 크게 화면을 차지한다.",
        "entities": "지정된 네 인물만 보인다. 신부는 노년의 동아시아계 남성으로 로만칼라와 검은 사제복을 착용했지만 참조의 상하의보다 긴 수단 형태다. 금발 어린 소녀 앰버는 마스크, 오염된 카키 작업복, 공구 벨트와 주머니를 유지한다. 찰리는 베이지 기계 몸체와 흰 마스크형 얼굴, 모자와 낡은 외투, 푸른 가슴 원자로를 갖췄으나 짧고 뚱뚱한 고릴라형보다 키 크고 길쭉하며 눈도 파란색이 아닌 주황색이다. 이현우는 검은 머리의 동아시아계 청소년 남성으로 얼굴의 멍과 상처, 낡고 더러운 셔츠, 붕대 감긴 다리가 보이고 겉옷은 없다. 인이어 장치는 이 각도에서 명확히 식별되지 않는다.",
        "hard_violations": [],
        "physics": "신부는 초의 아래쪽을 손으로 잡고 있다. 앰버는 어깨를 안으로 모으고 손을 가슴에 붙여 실제로 가능한 움츠림을 보인다. 찰리의 정지한 손은 팔꿈치와 어깨로 지지된다. 이현우는 한쪽 다리를 내밀고 다른 다리에 체중을 둔 채 상체를 뒤로 기울이는 자세로 읽히며, 발 접점은 프레임 밖이지만 무지지 부유의 징후는 없다. 네 인물의 몸과 의복은 중력에 맞게 이어지고 배경 물건들도 상판이나 바닥의 지지를 받는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 로우 앵글 구도와 천장을 향한 시선을 완벽히 구현했으며, 찰리의 작고 뚱뚱한 체형과 이현우의 인이어 등 세부 설정까지 정확히 묘사했습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "구도와 배경은 적절하나, 작고 뚱뚱해야 할 찰리의 체형이 지나치게 크고 길게 묘사되어 캐릭터 레퍼런스에 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_basement_prayer_room_5d61e9.png",
    "asset_id": "2aaacaed-7cb6-454b-8aea-3633d16bda64",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09c4-0f60-7dba-af10-30ceb8650c0c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh5__bgfirst_bg.png",
   "bg_asset_id": "69138a0b-9ed7-46b9-b365-6866080ed7c6",
   "bg_record_key": "S29sh5::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "basement_prayer_room",
   "groupbg_asset_id": "2aaacaed-7cb6-454b-8aea-3633d16bda64"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh11::signage": {
  "fp": "676bc408fd72be80",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::church_doorstep": {
  "input_fingerprint": "cdab17cdf8620ae7",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "church_doorstep",
    "tags": [
     "S29sh11"
    ]
   },
   "context_sig": "fed842722bc486b3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 성당 지하 기도소, 지하실 연결 계단: 촛불과 소형 랜턴에 의지하는 좁고 은밀한 지하 제단이다. (특징: 목재 십자가, 작은 책상, 낡은 풍금 라디오와 의자; 찰리의 가슴 발광에 맞춰 사방의 랜턴들이 켜지는 모습; 낡은 라디오의 주파수 다이얼이 저절로 움직이며 기계적 음성 조합을 내는 장비 작동; 계단을 뛰어 내려오는 신부와 방독/연구 장비 흔적이 없는 일반 복장의 구도환)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우도 작은 창으로 빼꼼 보는데... 어떤 50대 남자(씬 37)와 신부가 얘길 나누는 모습이 보인다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 성당 지하 기도소, 지하실 연결 계단: 촛불과 소형 랜턴에 의지하는 좁고 은밀한 지하 제단이다. (특징: 목재 십자가, 작은 책상, 낡은 풍금 라디오와 의자; 찰리의 가슴 발광에 맞춰 사방의 랜턴들이 켜지는 모습; 낡은 라디오의 주파수 다이얼이 저절로 움직이며 기계적 음성 조합을 내는 장비 작동; 계단을 뛰어 내려오는 신부와 방독/연구 장비 흔적이 없는 일반 복장의 구도환)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우도 작은 창으로 빼꼼 보는데... 어떤 50대 남자(씬 37)와 신부가 얘길 나누는 모습이 보인다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_church_doorstep_62ce3e.png",
  "asset_id": "970ffccc-52c7-4b80-a19a-7a4f9e3dc986",
  "input_asset_ids": [
   "5d3a20c2-5883-403f-b75c-b33a53ec40e2"
  ],
  "origin_tag": "S29sh11",
  "place_text": "Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.",
  "origin_inputs": {
   "place_text": "Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.",
   "time_of_day_en": "night",
   "conti_asset_id": "5d3a20c2-5883-403f-b75c-b33a53ec40e2"
  }
 },
 "S29sh11::bgfirst_bg": {
  "input_fingerprint": "4ea180f64475c178",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 작은 유리창 너머로 문 밖에서 마주 선 채 입을 열고 대화 중인 구도환과 신부의 모습.\n\nLOCATION (lock): Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the first-floor doorway at the small window's height, just behind and beside 이현우's established position but cropping him entirely out, looking obliquely through the opened aperture. The window's near edges frame 구도환 on the left and 신부 on the right outside, both visible from the upper torso as they speak toward each other's faces rather than toward the opening. Keep the camera still and the opening in soft foreground focus, making their reciprocal eyeline the emphasis without suggesting a reflection or a mediated image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 성당문에 난 작은 창 (Opened for looking outside) — The aperture's inner edges are visible; the opened glazed section is kept out of the viewing path; used as Creates a restricted foreground frame around the conversation without adding glass distortion; 성당 문밖 공간 (Occupied by the priest and 구도환 during their conversation); used as Provides a continuous exterior plane behind both men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained exterior nighttime illumination keeps the two speaking faces readable without extending the basement candlelight into this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 작은 유리창 너머로 문 밖에서 마주 선 채 입을 열고 대화 중인 구도환과 신부의 모습.\n\nLOCATION (lock): Immediately outside the church's ground-floor entrance, seen through the small window in its closed door.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the first-floor doorway at the small window's height, just behind and beside 이현우's established position but cropping him entirely out, looking obliquely through the opened aperture. The window's near edges frame 구도환 on the left and 신부 on the right outside, both visible from the upper torso as they speak toward each other's faces rather than toward the opening. Keep the camera still and the opening in soft foreground focus, making their reciprocal eyeline the emphasis without suggesting a reflection or a mediated image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 성당문에 난 작은 창 (Opened for looking outside) — The aperture's inner edges are visible; the opened glazed section is kept out of the viewing path; used as Creates a restricted foreground frame around the conversation without adding glass distortion; 성당 문밖 공간 (Occupied by the priest and 구도환 during their conversation); used as Provides a continuous exterior plane behind both men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained exterior nighttime illumination keeps the two speaking faces readable without extending the basement candlelight into this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh11__bgfirst_bg.png",
  "asset_id": "bea8f4c6-dedb-4bc3-8d6c-6395d0b62175",
  "input_asset_ids": [
   "5d3a20c2-5883-403f-b75c-b33a53ec40e2",
   "970ffccc-52c7-4b80-a19a-7a4f9e3dc986"
  ]
 },
 "S29sh11": {
  "input_fingerprint": "de0ca8ed80548bb4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 작은 유리창 너머로 문 밖에서 마주 선 채 입을 열고 대화 중인 구도환과 신부의 모습.\n\nLOCATION (lock): Immediately outside the church's ground-floor entrance, seen through the small window in its closed door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the first-floor doorway at the small window's height, just behind and beside 이현우's established position but cropping him entirely out, looking obliquely through the opened aperture. The window's near edges frame 구도환 on the left and 신부 on the right outside, both visible from the upper torso as they speak toward each other's faces rather than toward the opening. Keep the camera still and the opening in soft foreground focus, making their reciprocal eyeline the emphasis without suggesting a reflection or a mediated image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 성당문에 난 작은 창 (Opened for looking outside) — The aperture's inner edges are visible; the opened glazed section is kept out of the viewing path; used as Creates a restricted foreground frame around the conversation without adding glass distortion; 성당 문밖 공간 (Occupied by the priest and 구도환 during their conversation); used as Provides a continuous exterior plane behind both men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained exterior nighttime illumination keeps the two speaking faces readable without extending the basement candlelight into this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened, and the small viewing hatch in its door is open. The basement retains its cross, small desk, chairs and organ, with Charlie still there in his old coat and hat. 신부: He stands outside the entrance in his clerical collar. 구도환: He stands outside the church entrance, engaged in conversation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 작은 유리창 너머로 문 밖에서 마주 선 채 입을 열고 대화 중인 구도환과 신부의 모습.\n\nLOCATION (lock): Immediately outside the church's ground-floor entrance, seen through the small window in its closed door. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the first-floor doorway at the small window's height, just behind and beside 이현우's established position but cropping him entirely out, looking obliquely through the opened aperture. The window's near edges frame 구도환 on the left and 신부 on the right outside, both visible from the upper torso as they speak toward each other's faces rather than toward the opening. Keep the camera still and the opening in soft foreground focus, making their reciprocal eyeline the emphasis without suggesting a reflection or a mediated image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 성당문에 난 작은 창 (Opened for looking outside) — The aperture's inner edges are visible; the opened glazed section is kept out of the viewing path; used as Creates a restricted foreground frame around the conversation without adding glass distortion; 성당 문밖 공간 (Occupied by the priest and 구도환 during their conversation); used as Provides a continuous exterior plane behind both men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained exterior nighttime illumination keeps the two speaking faces readable without extending the basement candlelight into this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened, and the small viewing hatch in its door is open. The basement retains its cross, small desk, chairs and organ, with Charlie still there in his old coat and hat. 신부: He stands outside the entrance in his clerical collar. 구도환: He stands outside the church entrance, engaged in conversation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 작은 유리창 너머로 문 밖에서 마주 선 채 입을 열고 대화 중인 구도환과 신부의 모습.\n\nLOCATION (lock): Immediately outside the church's ground-floor entrance, seen through the small window in its closed door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the first-floor doorway at the small window's height, just behind and beside 이현우's established position but cropping him entirely out, looking obliquely through the opened aperture. The window's near edges frame 구도환 on the left and 신부 on the right outside, both visible from the upper torso as they speak toward each other's faces rather than toward the opening. Keep the camera still and the opening in soft foreground focus, making their reciprocal eyeline the emphasis without suggesting a reflection or a mediated image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 성당문에 난 작은 창 (Opened for looking outside) — The aperture's inner edges are visible; the opened glazed section is kept out of the viewing path; used as Creates a restricted foreground frame around the conversation without adding glass distortion; 성당 문밖 공간 (Occupied by the priest and 구도환 during their conversation); used as Provides a continuous exterior plane behind both men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained exterior nighttime illumination keeps the two speaking faces readable without extending the basement candlelight into this separate space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church entrance has been opened, and the small viewing hatch in its door is open. The basement retains its cross, small desk, chairs and organ, with Charlie still there in his old coat and hat. 신부: He stands outside the entrance in his clerical collar. 구도환: He stands outside the church entrance, engaged in conversation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh11__bgfirst_bg.png",
     "asset_id": "bea8f4c6-dedb-4bc3-8d6c-6395d0b62175",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S29sh11.png",
     "asset_id": "5d3a20c2-5883-403f-b75c-b33a53ec40e2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_church_doorstep_62ce3e.png",
     "asset_id": "970ffccc-52c7-4b80-a19a-7a4f9e3dc986",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "구도환과 신부가 서로의 얼굴을 마주 보며 시선을 교환하고 있으며, 구도환의 손짓이 신부를 향함.",
    "built_space": "성당 문의 작은 창틀 너머로 외부가 보이는 구도이며, 배경의 계단, 조명, 십자가 등 위치 레퍼런스의 요소들이 정확히 배치됨.",
    "entities": "왼쪽의 구도환(갈색 점퍼)과 오른쪽의 신부(사제복) 모두 레퍼런스의 인상과 복장을 정확히 반영함.",
    "hard_violations": [],
    "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 대화에 수반되는 자연스러운 신체 중심과 손의 움직임을 보여줌."
   },
   {
    "label": "B",
    "direction": "두 사람이 서로를 마주 보고 시선을 맞추고 있음.",
    "built_space": "창틀 프레임 안으로 성당 외부의 지정된 구조물과 조명 환경이 정확히 묘사됨.",
    "entities": "구도환과 신부 모두 지정된 연령, 성별, 복장 레퍼런스와 잘 일치함.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥에 서 있는 자세가 중력에 맞게 자연스러우며 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스와 일치하는 배경 및 인물 묘사를 보여주며, 손짓과 열린 입을 통해 대화 중인 상황을 매우 자연스럽게 연출함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "배경과 인물 설정은 지시사항을 잘 따랐으나, 대화하는 역동성이 후보 A에 비해 다소 부족함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환과 신부가 서로의 얼굴을 마주 보며 시선을 교환하고 있으며, 구도환의 손짓이 신부를 향함.",
        "built_space": "성당 문의 작은 창틀 너머로 외부가 보이는 구도이며, 배경의 계단, 조명, 십자가 등 위치 레퍼런스의 요소들이 정확히 배치됨.",
        "entities": "왼쪽의 구도환(갈색 점퍼)과 오른쪽의 신부(사제복) 모두 레퍼런스의 인상과 복장을 정확히 반영함.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 대화에 수반되는 자연스러운 신체 중심과 손의 움직임을 보여줌."
       },
       {
        "label": "B",
        "direction": "두 사람이 서로를 마주 보고 시선을 맞추고 있음.",
        "built_space": "창틀 프레임 안으로 성당 외부의 지정된 구조물과 조명 환경이 정확히 묘사됨.",
        "entities": "구도환과 신부 모두 지정된 연령, 성별, 복장 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 바닥에 서 있는 자세가 중력에 맞게 자연스러우며 안정적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스와 일치하는 배경 및 인물 묘사를 보여주며, 손짓과 열린 입을 통해 대화 중인 상황을 매우 자연스럽게 연출함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "배경과 인물 설정은 지시사항을 잘 따랐으나, 대화하는 역동성이 후보 A에 비해 다소 부족함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "구도환과 신부가 서로의 얼굴을 마주 보며 시선을 교환하고 있으며, 구도환의 손짓이 신부를 향함.",
        "built_space": "성당 문의 작은 창틀 너머로 외부가 보이는 구도이며, 배경의 계단, 조명, 십자가 등 위치 레퍼런스의 요소들이 정확히 배치됨.",
        "entities": "왼쪽의 구도환(갈색 점퍼)과 오른쪽의 신부(사제복) 모두 레퍼런스의 인상과 복장을 정확히 반영함.",
        "hard_violations": [],
        "physics": "두 사람 모두 지면에 안정적으로 서 있으며, 대화에 수반되는 자연스러운 신체 중심과 손의 움직임을 보여줌."
       },
       {
        "label": "B",
        "direction": "두 사람이 서로를 마주 보고 시선을 맞추고 있음.",
        "built_space": "창틀 프레임 안으로 성당 외부의 지정된 구조물과 조명 환경이 정확히 묘사됨.",
        "entities": "구도환과 신부 모두 지정된 연령, 성별, 복장 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 바닥에 서 있는 자세가 중력에 맞게 자연스러우며 안정적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "작은 개구부 너머 좌측 구도환과 우측 신부의 상반신, 서로 맞물리는 시선과 살짝 열린 입이 지정된 대화 순간을 더 충실히 구현한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "창 너머 배치와 상호 시선은 정확하지만, 신부는 입을 다문 채 듣는 모습이어서 두 사람이 입을 열고 대화하는 지정 순간에는 A보다 덜 부합한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 구도환은 오른쪽 신부의 눈과 얼굴을 바라보고, 신부도 구도환의 얼굴로 시선을 돌린다. 두 사람 모두 창이나 카메라를 보지 않으며 입술이 조금 벌어져 있다. 방향을 확인할 무기나 휴대 물건은 없다.",
        "built_space": "목제 문의 작은 직사각형 개구부 하나와 네 변의 안쪽 테두리, 오른쪽 금속 경첩 하나가 전경에 보인다. 두 사람은 그 바깥 동일한 마당에 상반신 크기로 배치된다. 오른쪽에는 석조 성당 외벽, 아치창 하나, 벽 십자가 하나, 벽등 하나와 계단·금속 난간이 있다. 낮은 담의 등 두 개가 보이고 나머지 구간은 인물에 가려진다. 멀리 밝은 십자가 하나가 보인다. 장소의 재료와 배치는 참조에 가깝고, 개구부를 가로막는 유리나 인물의 반사는 없다. 비스듬한 관찰 각도는 다소 약하다.",
        "entities": "보이는 인물은 두 명뿐이다. 구도환은 짧은 검은 머리의 50대 한국인 남성 외관이며, 참조와 유사한 얼굴에 해진 갈색 점퍼와 올리브색 안옷을 입었다. 신부는 짧은 회색 머리의 60대 한국인 남성 외관으로, 참조와 유사한 얼굴과 검은 단추 사제복·흰 로만칼라를 갖췄다. 바지는 구도 밖이라 확인할 수 없다. 이현우나 지하실 인물·비품, 추가 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 몸통은 화면 아래로 자연스럽게 이어지는 기립 자세이며, 발과 지면 접촉은 창 아래 테두리에 가려져 있다. 공중에 뜬 자세나 지지 없는 물체는 없다. 문틀과 경첩은 문에 고정되어 있고, 젖은 마당의 빛 반사는 외부 조명과 양립한다."
       },
       {
        "label": "B",
        "direction": "왼쪽 구도환과 오른쪽 신부는 옆얼굴로 서로의 얼굴을 바라본다. 구도환은 입을 벌리고 손바닥을 신부 쪽 위로 펼쳐 말하는 모습이다. 신부는 구도환을 바라보지만 입은 닫혀 있어 듣는 순간으로 읽힌다. 카메라를 향한 시선은 없다.",
        "built_space": "목제 문의 직사각형 개구부 하나, 네 변의 전경 테두리와 오른쪽 경첩 하나가 보인다. 두 인물은 창 밖의 연속된 마당에 서 있다. 오른쪽 석조 외벽에 아치창 하나와 벽 십자가 하나가 있고, 벽등 하나는 신부 머리 뒤로 일부 가려진다. 계단과 금속 난간, 낮은 담의 등 두 개, 원경의 밝은 십자가 하나가 보인다. 참조 장소의 구성을 유지하며 유리 왜곡이나 불가능한 반사는 없다. 관찰 방향은 지정된 사선보다는 정면에 가까운 편이다.",
        "entities": "두 명 모두 지정된 한국인 중·노년 남성 외관과 짧은 머리에 부합한다. 구도환의 해진 갈색 점퍼는 맞지만 목 부분의 어두운 안옷은 참조의 올리브색 둥근목 안옷과 다르다. 신부의 회색 머리, 검은 사제복과 흰 로만칼라는 참조에 가깝다. 하의는 프레임 밖이다. 추가 인물이나 휴대 소품, 문자는 없다.",
        "hard_violations": [],
        "physics": "구도환의 펼친 손은 굽힌 팔과 손목으로 연결되어 자연스러운 대화 제스처를 이룬다. 신부는 배 앞에서 두 손을 모으고 있다. 두 사람의 하체와 발은 창틀 아래로 가려져 있지만 몸통은 정상적인 기립 자세로 이어지며, 부유나 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "작은 개구부 너머 좌측 구도환과 우측 신부의 상반신, 서로 맞물리는 시선과 살짝 열린 입이 지정된 대화 순간을 더 충실히 구현한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "창 너머 배치와 상호 시선은 정확하지만, 신부는 입을 다문 채 듣는 모습이어서 두 사람이 입을 열고 대화하는 지정 순간에는 A보다 덜 부합한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 구도환은 오른쪽 신부의 눈과 얼굴을 바라보고, 신부도 구도환의 얼굴로 시선을 돌린다. 두 사람 모두 창이나 카메라를 보지 않으며 입술이 조금 벌어져 있다. 방향을 확인할 무기나 휴대 물건은 없다.",
        "built_space": "목제 문의 작은 직사각형 개구부 하나와 네 변의 안쪽 테두리, 오른쪽 금속 경첩 하나가 전경에 보인다. 두 사람은 그 바깥 동일한 마당에 상반신 크기로 배치된다. 오른쪽에는 석조 성당 외벽, 아치창 하나, 벽 십자가 하나, 벽등 하나와 계단·금속 난간이 있다. 낮은 담의 등 두 개가 보이고 나머지 구간은 인물에 가려진다. 멀리 밝은 십자가 하나가 보인다. 장소의 재료와 배치는 참조에 가깝고, 개구부를 가로막는 유리나 인물의 반사는 없다. 비스듬한 관찰 각도는 다소 약하다.",
        "entities": "보이는 인물은 두 명뿐이다. 구도환은 짧은 검은 머리의 50대 한국인 남성 외관이며, 참조와 유사한 얼굴에 해진 갈색 점퍼와 올리브색 안옷을 입었다. 신부는 짧은 회색 머리의 60대 한국인 남성 외관으로, 참조와 유사한 얼굴과 검은 단추 사제복·흰 로만칼라를 갖췄다. 바지는 구도 밖이라 확인할 수 없다. 이현우나 지하실 인물·비품, 추가 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 몸통은 화면 아래로 자연스럽게 이어지는 기립 자세이며, 발과 지면 접촉은 창 아래 테두리에 가려져 있다. 공중에 뜬 자세나 지지 없는 물체는 없다. 문틀과 경첩은 문에 고정되어 있고, 젖은 마당의 빛 반사는 외부 조명과 양립한다."
       },
       {
        "label": "A",
        "direction": "왼쪽 구도환과 오른쪽 신부는 옆얼굴로 서로의 얼굴을 바라본다. 구도환은 입을 벌리고 손바닥을 신부 쪽 위로 펼쳐 말하는 모습이다. 신부는 구도환을 바라보지만 입은 닫혀 있어 듣는 순간으로 읽힌다. 카메라를 향한 시선은 없다.",
        "built_space": "목제 문의 직사각형 개구부 하나, 네 변의 전경 테두리와 오른쪽 경첩 하나가 보인다. 두 인물은 창 밖의 연속된 마당에 서 있다. 오른쪽 석조 외벽에 아치창 하나와 벽 십자가 하나가 있고, 벽등 하나는 신부 머리 뒤로 일부 가려진다. 계단과 금속 난간, 낮은 담의 등 두 개, 원경의 밝은 십자가 하나가 보인다. 참조 장소의 구성을 유지하며 유리 왜곡이나 불가능한 반사는 없다. 관찰 방향은 지정된 사선보다는 정면에 가까운 편이다.",
        "entities": "두 명 모두 지정된 한국인 중·노년 남성 외관과 짧은 머리에 부합한다. 구도환의 해진 갈색 점퍼는 맞지만 목 부분의 어두운 안옷은 참조의 올리브색 둥근목 안옷과 다르다. 신부의 회색 머리, 검은 사제복과 흰 로만칼라는 참조에 가깝다. 하의는 프레임 밖이다. 추가 인물이나 휴대 소품, 문자는 없다.",
        "hard_violations": [],
        "physics": "구도환의 펼친 손은 굽힌 팔과 손목으로 연결되어 자연스러운 대화 제스처를 이룬다. 신부는 배 앞에서 두 손을 모으고 있다. 두 사람의 하체와 발은 창틀 아래로 가려져 있지만 몸통은 정상적인 기립 자세로 이어지며, 부유나 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "레퍼런스와 일치하는 배경 및 인물 묘사를 보여주며, 손짓과 열린 입을 통해 대화 중인 상황을 매우 자연스럽게 연출함."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "배경과 인물 설정은 지시사항을 잘 따랐으나, 대화하는 역동성이 후보 A에 비해 다소 부족함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_church_doorstep_62ce3e.png",
    "asset_id": "970ffccc-52c7-4b80-a19a-7a4f9e3dc986",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09cc-f271-777a-9cd8-6b0c6454828e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh11__bgfirst_bg.png",
   "bg_asset_id": "bea8f4c6-dedb-4bc3-8d6c-6395d0b62175",
   "bg_record_key": "S29sh11::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "church_doorstep",
   "groupbg_asset_id": "970ffccc-52c7-4b80-a19a-7a4f9e3dc986"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh12::signage": {
  "fp": "d0eecea3ee6dccac",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh12": {
  "input_fingerprint": "f4d09c546e4ac09f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 여러 대의 헬기와 드론이 기수를 비행 방향으로 둔 채 눈부신 서치라이트 빛줄기를 쏘아 내리는 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement, crossed by helicopters, drones, and downward searchlight beams. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make an explicit flashback cut to a separate ground-level exterior viewpoint, holding a steep upward angle in a distant wide view oblique to the aircrafts' travel rather than continuing a tilt from the window. Spread the helicopters across the upper field and the smaller drones through the intervening sky, with noses aligned to their individual flight paths and unequal spacing rather than a duplicated formation. The widened distance is the principal visual change, leaving generous night sky around every aircraft as the drones follow the searchlight paths.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헬기들 (Flying during the search for 찰리) — Undersides and oblique side views are visible, with noses aligned to travel; used as Form separated upper-frame search anchors, each remaining small within the wide sky; 드론들 (Flying along the helicopter searchlight paths) — Each aircraft points along its flight direction, with varied oblique presentations; used as Create dispersed depth and movement phases without a mechanically repeated formation; 밤하늘 (Dark during the flashback search); used as Supplies negative space that separates the aircraft and clarifies their search pattern.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: In this clearly separated flashback, brilliant helicopter searchlights cut through the dark night view with controlled highlights and no invented atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopter searchlights sweep the night sky, with multiple drones flying along the beams. At the church, the entrance and viewing hatch remain open, and Charlie remains in the basement in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 신부 (한국인 남성, 60대의 얼굴, 짧은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 여러 대의 헬기와 드론이 기수를 비행 방향으로 둔 채 눈부신 서치라이트 빛줄기를 쏘아 내리는 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement, crossed by helicopters, drones, and downward searchlight beams. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make an explicit flashback cut to a separate ground-level exterior viewpoint, holding a steep upward angle in a distant wide view oblique to the aircrafts' travel rather than continuing a tilt from the window. Spread the helicopters across the upper field and the smaller drones through the intervening sky, with noses aligned to their individual flight paths and unequal spacing rather than a duplicated formation. The widened distance is the principal visual change, leaving generous night sky around every aircraft as the drones follow the searchlight paths.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헬기들 (Flying during the search for 찰리) — Undersides and oblique side views are visible, with noses aligned to travel; used as Form separated upper-frame search anchors, each remaining small within the wide sky; 드론들 (Flying along the helicopter searchlight paths) — Each aircraft points along its flight direction, with varied oblique presentations; used as Create dispersed depth and movement phases without a mechanically repeated formation; 밤하늘 (Dark during the flashback search); used as Supplies negative space that separates the aircraft and clarifies their search pattern.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: In this clearly separated flashback, brilliant helicopter searchlights cut through the dark night view with controlled highlights and no invented atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopter searchlights sweep the night sky, with multiple drones flying along the beams. At the church, the entrance and viewing hatch remain open, and Charlie remains in the basement in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 신부 (한국인 남성, 60대의 얼굴, 짧은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 위로 여러 대의 헬기와 드론이 기수를 비행 방향으로 둔 채 눈부신 서치라이트 빛줄기를 쏘아 내리는 전경.\n\nLOCATION (lock): In the open night sky above the refugee settlement, crossed by helicopters, drones, and downward searchlight beams. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make an explicit flashback cut to a separate ground-level exterior viewpoint, holding a steep upward angle in a distant wide view oblique to the aircrafts' travel rather than continuing a tilt from the window. Spread the helicopters across the upper field and the smaller drones through the intervening sky, with noses aligned to their individual flight paths and unequal spacing rather than a duplicated formation. The widened distance is the principal visual change, leaving generous night sky around every aircraft as the drones follow the searchlight paths.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헬기들 (Flying during the search for 찰리) — Undersides and oblique side views are visible, with noses aligned to travel; used as Form separated upper-frame search anchors, each remaining small within the wide sky; 드론들 (Flying along the helicopter searchlight paths) — Each aircraft points along its flight direction, with varied oblique presentations; used as Create dispersed depth and movement phases without a mechanically repeated formation; 밤하늘 (Dark during the flashback search); used as Supplies negative space that separates the aircraft and clarifies their search pattern.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: In this clearly separated flashback, brilliant helicopter searchlights cut through the dark night view with controlled highlights and no invented atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Helicopter searchlights sweep the night sky, with multiple drones flying along the beams. At the church, the entrance and viewing hatch remain open, and Charlie remains in the basement in his old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 신부 (한국인 남성, 60대의 얼굴, 짧은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "헬기와 드론들의 기수가 화면 왼쪽(비행 방향)을 향하고 있으며, 서치라이트 빛줄기는 지상의 난민촌을 향해 아래로 뻗어 있음.",
    "built_space": "지상에서 하늘을 올려다보는 구도로, 하단에 난민촌을 구성하는 텐트와 구조물들이 일부 보임.",
    "entities": "지시문에 있는 헬기, 드론, 밤하늘은 존재하나, 지시문에 전혀 언급되지 않은 다수의 인물들이 전경에 크게 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people)",
     "[gpt-high] 등장이 허용되지 않은 군중을 추가하고, 그중 세 명을 큰 전경 인물로 배치했습니다.",
     "[gpt-high] 열린 밤하늘 장면에 천막 통로, 골조, 전봇대, 전선과 건물 등 별도의 지상 공간을 만들어 넣었습니다."
    ],
    "physics": "헬기와 드론 모두 비행 중인 상태로 공중에 자연스럽게 떠 있음."
   },
   {
    "label": "B",
    "direction": "헬기와 드론들이 왼쪽으로 기수를 둔 채 비행하고 있으며, 빛줄기는 아래쪽 지상을 정확히 겨냥함.",
    "built_space": "난민촌의 천막과 건물들이 하단에 넓게 깔려 있으며, 지상 뷰포인트의 각도를 형성함.",
    "entities": "헬기와 드론은 요청된 형태로 하늘에 분산되어 있으나, 지시문에 없는 사람 무리가 전경 하단을 가득 채우고 있음.",
    "hard_violations": [
     "[gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people)",
     "[gpt-high] 사람을 등장시키지 않는 숏에 다수의 구경꾼을 새로 추가했습니다.",
     "[gpt-high] 열린 밤하늘로 한정된 장면에 천막촌의 구체적 시설, 전선, 조명 기둥, 탑과 건물군을 추가했습니다."
    ],
    "physics": "항공기들이 정상적으로 공중에 부양하여 비행 중임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시문(Shot text)에 전혀 명시되지 않은 다수의 인물들을 전경에 크게 배치하여 '없는 인물 추가 불가' 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 없는 군중 실루엣을 화면 하단에 임의로 추가하여 샷의 구도와 인물 등장 규칙을 완전히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬기와 드론들의 기수가 화면 왼쪽(비행 방향)을 향하고 있으며, 서치라이트 빛줄기는 지상의 난민촌을 향해 아래로 뻗어 있음.",
        "built_space": "지상에서 하늘을 올려다보는 구도로, 하단에 난민촌을 구성하는 텐트와 구조물들이 일부 보임.",
        "entities": "지시문에 있는 헬기, 드론, 밤하늘은 존재하나, 지시문에 전혀 언급되지 않은 다수의 인물들이 전경에 크게 묘사됨.",
        "hard_violations": [
         "지시문에 없는 인물들을 전경에 추가함 (Invented people)"
        ],
        "physics": "헬기와 드론 모두 비행 중인 상태로 공중에 자연스럽게 떠 있음."
       },
       {
        "label": "B",
        "direction": "헬기와 드론들이 왼쪽으로 기수를 둔 채 비행하고 있으며, 빛줄기는 아래쪽 지상을 정확히 겨냥함.",
        "built_space": "난민촌의 천막과 건물들이 하단에 넓게 깔려 있으며, 지상 뷰포인트의 각도를 형성함.",
        "entities": "헬기와 드론은 요청된 형태로 하늘에 분산되어 있으나, 지시문에 없는 사람 무리가 전경 하단을 가득 채우고 있음.",
        "hard_violations": [
         "지시문에 없는 인물들을 전경에 추가함 (Invented people)"
        ],
        "physics": "항공기들이 정상적으로 공중에 부양하여 비행 중임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지시문(Shot text)에 전혀 명시되지 않은 다수의 인물들을 전경에 크게 배치하여 '없는 인물 추가 불가' 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 없는 군중 실루엣을 화면 하단에 임의로 추가하여 샷의 구도와 인물 등장 규칙을 완전히 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "헬기와 드론들의 기수가 화면 왼쪽(비행 방향)을 향하고 있으며, 서치라이트 빛줄기는 지상의 난민촌을 향해 아래로 뻗어 있음.",
        "built_space": "지상에서 하늘을 올려다보는 구도로, 하단에 난민촌을 구성하는 텐트와 구조물들이 일부 보임.",
        "entities": "지시문에 있는 헬기, 드론, 밤하늘은 존재하나, 지시문에 전혀 언급되지 않은 다수의 인물들이 전경에 크게 묘사됨.",
        "hard_violations": [
         "지시문에 없는 인물들을 전경에 추가함 (Invented people)"
        ],
        "physics": "헬기와 드론 모두 비행 중인 상태로 공중에 자연스럽게 떠 있음."
       },
       {
        "label": "B",
        "direction": "헬기와 드론들이 왼쪽으로 기수를 둔 채 비행하고 있으며, 빛줄기는 아래쪽 지상을 정확히 겨냥함.",
        "built_space": "난민촌의 천막과 건물들이 하단에 넓게 깔려 있으며, 지상 뷰포인트의 각도를 형성함.",
        "entities": "헬기와 드론은 요청된 형태로 하늘에 분산되어 있으나, 지시문에 없는 사람 무리가 전경 하단을 가득 채우고 있음.",
        "hard_violations": [
         "지시문에 없는 인물들을 전경에 추가함 (Invented people)"
        ],
        "physics": "항공기들이 정상적으로 공중에 부양하여 비행 중임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 군중과 지상 시설을 추가해 탈락 대상이지만, 작은 헬기들을 상부에 분산하고 넓은 밤하늘을 확보한 점은 B보다 요청한 원경에 가깝습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "올려다보는 각도는 드러나지만, 추가된 대형 전경 인물들과 정착촌 통로가 하늘 원경을 군중 관찰 장면으로 바꾸며 헬기 하나도 상부 배치에서 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "헬기 3대 모두 꼬리가 왼쪽, 기수가 오른쪽이며 측면과 아랫면이 보입니다. 서치라이트는 오른쪽 아래의 하늘과 화면 밖 지상 방향으로 향하고, 특정 인물이나 표적에 닿는 모습은 없습니다. 드론 7대의 빛도 대체로 오른쪽 아래를 향하지만 작은 기체의 앞뒤와 실제 이동 방향은 확정하기 어렵습니다. 드론들은 헬기 사이와 아래에 흩어져 있으나 모두가 헬기 광선 경로를 따라가는지는 분명하지 않습니다. 아래 군중 일부는 항공기를 올려다봅니다.",
        "built_space": "헬기는 좌상단, 상부 중앙 오른쪽, 우상단에 서로 다른 간격으로 놓여 있습니다. 그러나 하단에는 천막형 주거지, 전선과 여러 조명 기둥, 오른쪽 격자형 탑, 먼 건물군이 추가되어 있습니다. 왼쪽 전경과 하단에 다수의 사람이 서 있어, 구조물을 보여주지 않는 가파른 상향 하늘 원경보다 정착촌을 내려다보며 하늘까지 담은 전망에 가깝습니다. 반사상은 없습니다.",
        "entities": "헬기 3대, 소형 다중회전익 드론 7대, 밤하늘과 하향 광선은 요청된 대상에 해당합니다. 별도의 외형 참조는 없습니다. 화면 아래에는 최소 14명의 사람이 보이지만 이 숏은 어떤 사람도 등장시키지 않으며, 이들을 명시된 인물로 확인할 근거도 없습니다. 천막, 가로등, 전선, 탑과 도시 건물 역시 하늘만을 구성하라는 지시 밖의 구체적 추가물입니다. 읽을 수 있는 자막이나 표식은 없습니다.",
        "hard_violations": [
         "사람을 등장시키지 않는 숏에 다수의 구경꾼을 새로 추가했습니다.",
         "열린 밤하늘로 한정된 장면에 천막촌의 구체적 시설, 전선, 조명 기둥, 탑과 건물군을 추가했습니다."
        ],
        "physics": "헬기에는 주회전익과 꼬리 구조가 보이고, 드론에는 회전익 지지 팔이 보여 비행을 유지할 물리적 수단이 있습니다. 광선은 기체에 부착된 광원에서 시작합니다. 사람들의 발은 대부분 하단에 가려지지만 지면 위에 선 자세와 배치이며, 근거 없이 공중에 떠 있는 몸은 보이지 않습니다. 천막과 조명은 기둥 및 골조에 지지됩니다."
       },
       {
        "label": "B",
        "direction": "좌상단과 우상단 헬기는 오른쪽 아래를 향한 사선 기수와 아랫면을 보이며, 중앙 아래의 작은 헬기도 기수가 오른쪽입니다. 각 서치라이트는 아래쪽 또는 오른쪽 아래의 화면 밖 지상을 향합니다. 드론 7대는 하늘 곳곳에서 주로 아래로 빛을 쏘지만, 기수 방향이나 헬기 광선을 따라 이동하는 관계는 명확하지 않습니다. 전경의 세 큰 인물은 하늘의 항공기 쪽으로 고개를 들고 있습니다.",
        "built_space": "헬기 2대는 상부 좌우에 있으나 세 번째는 화면 중앙 높이까지 내려와 있어 상부 수색 거점 배치가 약해집니다. 카메라는 천막 사이 통로에서 위를 올려다보는 위치로 읽힙니다. 좌우에는 골조와 천막, 여러 전봇대와 전선, 내부 조명, 뒤쪽 건물들이 보입니다. 가까운 인물 3명이 양쪽 전경을 크게 차지하고 나머지 군중이 통로에 서 있어, 하늘만의 먼 와이드 숏과 다릅니다. 반사상은 없습니다.",
        "entities": "헬기 3대와 소형 다중회전익 드론 7대, 어두운 밤하늘 및 강한 하향 서치라이트는 요청된 대상입니다. 다만 좌상단 헬기는 A보다 크게 보입니다. 최소 12명의 사람이 추가되어 있으며, 가까운 인물들은 성인으로 보이지만 이름이나 정확한 연령·민족적 정체성을 확인할 수 없습니다. 이 숏에는 애초에 사람이 요구되지 않습니다. 천막 통로와 각종 지상 설비도 추가되었고, 읽을 수 있는 문구나 화면 오버레이는 없습니다.",
        "hard_violations": [
         "등장이 허용되지 않은 군중을 추가하고, 그중 세 명을 큰 전경 인물로 배치했습니다.",
         "열린 밤하늘 장면에 천막 통로, 골조, 전봇대, 전선과 건물 등 별도의 지상 공간을 만들어 넣었습니다."
        ],
        "physics": "헬기에는 회전익이, 드론에는 다중회전익 구조가 보여 공중 체류를 설명합니다. 서치라이트는 기체의 광원 위치와 연결됩니다. 사람들은 통로 지면에 서 있는 자세이며 발이 잘린 인물도 공중 부유로 읽히지 않습니다. 천막은 골조와 기둥에 걸려 있어 지지가 보입니다. 명백히 지지 없이 떠 있는 물체나 불가능한 신체 자세는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금지된 군중과 지상 시설을 추가해 탈락 대상이지만, 작은 헬기들을 상부에 분산하고 넓은 밤하늘을 확보한 점은 B보다 요청한 원경에 가깝습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "올려다보는 각도는 드러나지만, 추가된 대형 전경 인물들과 정착촌 통로가 하늘 원경을 군중 관찰 장면으로 바꾸며 헬기 하나도 상부 배치에서 벗어납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "헬기 3대 모두 꼬리가 왼쪽, 기수가 오른쪽이며 측면과 아랫면이 보입니다. 서치라이트는 오른쪽 아래의 하늘과 화면 밖 지상 방향으로 향하고, 특정 인물이나 표적에 닿는 모습은 없습니다. 드론 7대의 빛도 대체로 오른쪽 아래를 향하지만 작은 기체의 앞뒤와 실제 이동 방향은 확정하기 어렵습니다. 드론들은 헬기 사이와 아래에 흩어져 있으나 모두가 헬기 광선 경로를 따라가는지는 분명하지 않습니다. 아래 군중 일부는 항공기를 올려다봅니다.",
        "built_space": "헬기는 좌상단, 상부 중앙 오른쪽, 우상단에 서로 다른 간격으로 놓여 있습니다. 그러나 하단에는 천막형 주거지, 전선과 여러 조명 기둥, 오른쪽 격자형 탑, 먼 건물군이 추가되어 있습니다. 왼쪽 전경과 하단에 다수의 사람이 서 있어, 구조물을 보여주지 않는 가파른 상향 하늘 원경보다 정착촌을 내려다보며 하늘까지 담은 전망에 가깝습니다. 반사상은 없습니다.",
        "entities": "헬기 3대, 소형 다중회전익 드론 7대, 밤하늘과 하향 광선은 요청된 대상에 해당합니다. 별도의 외형 참조는 없습니다. 화면 아래에는 최소 14명의 사람이 보이지만 이 숏은 어떤 사람도 등장시키지 않으며, 이들을 명시된 인물로 확인할 근거도 없습니다. 천막, 가로등, 전선, 탑과 도시 건물 역시 하늘만을 구성하라는 지시 밖의 구체적 추가물입니다. 읽을 수 있는 자막이나 표식은 없습니다.",
        "hard_violations": [
         "사람을 등장시키지 않는 숏에 다수의 구경꾼을 새로 추가했습니다.",
         "열린 밤하늘로 한정된 장면에 천막촌의 구체적 시설, 전선, 조명 기둥, 탑과 건물군을 추가했습니다."
        ],
        "physics": "헬기에는 주회전익과 꼬리 구조가 보이고, 드론에는 회전익 지지 팔이 보여 비행을 유지할 물리적 수단이 있습니다. 광선은 기체에 부착된 광원에서 시작합니다. 사람들의 발은 대부분 하단에 가려지지만 지면 위에 선 자세와 배치이며, 근거 없이 공중에 떠 있는 몸은 보이지 않습니다. 천막과 조명은 기둥 및 골조에 지지됩니다."
       },
       {
        "label": "A",
        "direction": "좌상단과 우상단 헬기는 오른쪽 아래를 향한 사선 기수와 아랫면을 보이며, 중앙 아래의 작은 헬기도 기수가 오른쪽입니다. 각 서치라이트는 아래쪽 또는 오른쪽 아래의 화면 밖 지상을 향합니다. 드론 7대는 하늘 곳곳에서 주로 아래로 빛을 쏘지만, 기수 방향이나 헬기 광선을 따라 이동하는 관계는 명확하지 않습니다. 전경의 세 큰 인물은 하늘의 항공기 쪽으로 고개를 들고 있습니다.",
        "built_space": "헬기 2대는 상부 좌우에 있으나 세 번째는 화면 중앙 높이까지 내려와 있어 상부 수색 거점 배치가 약해집니다. 카메라는 천막 사이 통로에서 위를 올려다보는 위치로 읽힙니다. 좌우에는 골조와 천막, 여러 전봇대와 전선, 내부 조명, 뒤쪽 건물들이 보입니다. 가까운 인물 3명이 양쪽 전경을 크게 차지하고 나머지 군중이 통로에 서 있어, 하늘만의 먼 와이드 숏과 다릅니다. 반사상은 없습니다.",
        "entities": "헬기 3대와 소형 다중회전익 드론 7대, 어두운 밤하늘 및 강한 하향 서치라이트는 요청된 대상입니다. 다만 좌상단 헬기는 A보다 크게 보입니다. 최소 12명의 사람이 추가되어 있으며, 가까운 인물들은 성인으로 보이지만 이름이나 정확한 연령·민족적 정체성을 확인할 수 없습니다. 이 숏에는 애초에 사람이 요구되지 않습니다. 천막 통로와 각종 지상 설비도 추가되었고, 읽을 수 있는 문구나 화면 오버레이는 없습니다.",
        "hard_violations": [
         "등장이 허용되지 않은 군중을 추가하고, 그중 세 명을 큰 전경 인물로 배치했습니다.",
         "열린 밤하늘 장면에 천막 통로, 골조, 전봇대, 전선과 건물 등 별도의 지상 공간을 만들어 넣었습니다."
        ],
        "physics": "헬기에는 회전익이, 드론에는 다중회전익 구조가 보여 공중 체류를 설명합니다. 서치라이트는 기체의 광원 위치와 연결됩니다. 사람들은 통로 지면에 서 있는 자세이며 발이 잘린 인물도 공중 부유로 읽히지 않습니다. 천막은 골조와 기둥에 걸려 있어 지지가 보입니다. 명백히 지지 없이 떠 있는 물체나 불가능한 신체 자세는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people)",
     "[gpt-high] 등장이 허용되지 않은 군중을 추가하고, 그중 세 명을 큰 전경 인물로 배치했습니다.",
     "[gpt-high] 열린 밤하늘 장면에 천막 통로, 골조, 전봇대, 전선과 건물 등 별도의 지상 공간을 만들어 넣었습니다."
    ],
    "B": [
     "[gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people)",
     "[gpt-high] 사람을 등장시키지 않는 숏에 다수의 구경꾼을 새로 추가했습니다.",
     "[gpt-high] 열린 밤하늘로 한정된 장면에 천막촌의 구체적 시설, 전선, 조명 기둥, 탑과 건물군을 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1417,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "지시문(Shot text)에 전혀 명시되지 않은 다수의 인물들을 전경에 크게 배치하여 '없는 인물 추가 불가' 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people) / [gpt-high] 등장이 허용되지 않은 군중을 추가하고, 그중 세 명을 큰 전경 인물로 배치했습니다. / [gpt-high] 열린 밤하늘 장면에 천막 통로, 골조, 전봇대, 전선과 건물 등 별도의 지상 공간을 만들어 넣었습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시문에 없는 군중 실루엣을 화면 하단에 임의로 추가하여 샷의 구도와 인물 등장 규칙을 완전히 위반했습니다.  ★위반: [gemini-pro] 지시문에 없는 인물들을 전경에 추가함 (Invented people) / [gpt-high] 사람을 등장시키지 않는 숏에 다수의 구경꾼을 새로 추가했습니다. / [gpt-high] 열린 밤하늘로 한정된 장면에 천막촌의 구체적 시설, 전선, 조명 기둥, 탑과 건물군을 추가했습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09d4-d187-7442-a08c-06b6bddf8ba8",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S30sh6::signage": {
  "fp": "4ef38dd69db9d12c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh6": {
  "input_fingerprint": "d85ea2f441bd332f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 몸에서 퍼져 나온 빛에 반응하여 지하실 곳곳의 랜턴들이 일제히 환하게 켜지는 중의 한 순간이 포착된 전경.\n\nLOCATION (lock): Inside the church's modest basement prayer room, around the small organ and old radio. Lanterns around the room are suddenly shining brightly. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane's upward retreat from beside the organ and settle into a high oblique view diagonally across the basement, keeping 찰리 and 앰버 below the camera in the lower-left portion and the surrounding lanterns distributed through the room. 찰리 remains at the organ with his torso turned toward the old radio in the right-hand corner, his lowered attention fixed there, while 앰버 inclines toward him and watches his chest. Hold this revealed layout as the lanterns switch on together, making illumination the single emphasized change and preserving clear spatial separation among the robot, organ, radio, and individual lanterns.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 작은 풍금 (Beside 찰리 after his playing) — The keyboard side and top are visible from the elevated diagonal viewpoint; used as Anchors the preceding musical interaction while remaining smaller than two-fifths of the image; 랜턴들 (Positioned at separate locations throughout the basement); used as Their dispersed physical positions make the room-wide response legible without creating a central oversized fixture; 낡은 라디오 (The previously broken radio remains in the corner as its activation begins) — Seen obliquely within the corner rather than enlarged into an isolated detail; used as Provides the visible destination of 찰리's lowered attention and the next narrative focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light emerging from 찰리's chest coincides with the lanterns brightening throughout the basement, retaining controlled highlights, readable robot surfaces, and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs and small organ; its lanterns brighten and the old radio in the corner activates. Charlie's chest emits light, with his coat and hat disguise, blue-lit eyes and worn chest logo retained.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 몸에서 퍼져 나온 빛에 반응하여 지하실 곳곳의 랜턴들이 일제히 환하게 켜지는 중의 한 순간이 포착된 전경.\n\nLOCATION (lock): Inside the church's modest basement prayer room, around the small organ and old radio. Lanterns around the room are suddenly shining brightly. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane's upward retreat from beside the organ and settle into a high oblique view diagonally across the basement, keeping 찰리 and 앰버 below the camera in the lower-left portion and the surrounding lanterns distributed through the room. 찰리 remains at the organ with his torso turned toward the old radio in the right-hand corner, his lowered attention fixed there, while 앰버 inclines toward him and watches his chest. Hold this revealed layout as the lanterns switch on together, making illumination the single emphasized change and preserving clear spatial separation among the robot, organ, radio, and individual lanterns.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 작은 풍금 (Beside 찰리 after his playing) — The keyboard side and top are visible from the elevated diagonal viewpoint; used as Anchors the preceding musical interaction while remaining smaller than two-fifths of the image; 랜턴들 (Positioned at separate locations throughout the basement); used as Their dispersed physical positions make the room-wide response legible without creating a central oversized fixture; 낡은 라디오 (The previously broken radio remains in the corner as its activation begins) — Seen obliquely within the corner rather than enlarged into an isolated detail; used as Provides the visible destination of 찰리's lowered attention and the next narrative focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light emerging from 찰리's chest coincides with the lanterns brightening throughout the basement, retaining controlled highlights, readable robot surfaces, and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs and small organ; its lanterns brighten and the old radio in the corner activates. Charlie's chest emits light, with his coat and hat disguise, blue-lit eyes and worn chest logo retained.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 몸에서 퍼져 나온 빛에 반응하여 지하실 곳곳의 랜턴들이 일제히 환하게 켜지는 중의 한 순간이 포착된 전경.\n\nLOCATION (lock): Inside the church's modest basement prayer room, around the small organ and old radio. Lanterns around the room are suddenly shining brightly. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane's upward retreat from beside the organ and settle into a high oblique view diagonally across the basement, keeping 찰리 and 앰버 below the camera in the lower-left portion and the surrounding lanterns distributed through the room. 찰리 remains at the organ with his torso turned toward the old radio in the right-hand corner, his lowered attention fixed there, while 앰버 inclines toward him and watches his chest. Hold this revealed layout as the lanterns switch on together, making illumination the single emphasized change and preserving clear spatial separation among the robot, organ, radio, and individual lanterns.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 작은 풍금 (Beside 찰리 after his playing) — The keyboard side and top are visible from the elevated diagonal viewpoint; used as Anchors the preceding musical interaction while remaining smaller than two-fifths of the image; 랜턴들 (Positioned at separate locations throughout the basement); used as Their dispersed physical positions make the room-wide response legible without creating a central oversized fixture; 낡은 라디오 (The previously broken radio remains in the corner as its activation begins) — Seen obliquely within the corner rather than enlarged into an isolated detail; used as Provides the visible destination of 찰리's lowered attention and the next narrative focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light emerging from 찰리's chest coincides with the lanterns brightening throughout the basement, retaining controlled highlights, readable robot surfaces, and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its cross, desk, chairs and small organ; its lanterns brighten and the old radio in the corner activates. Charlie's chest emits light, with his coat and hat disguise, blue-lit eyes and worn chest logo retained.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 다소 아래를 향해 시선을 두고 있고, 앰버는 찰리 쪽을 바라보고 있음.",
    "built_space": "교회 지하실. 부감(high angle) 구도로 계단, 십자가, 풍금, 책장 등이 기준 이미지에 맞게 배치되어 있음.",
    "entities": "빛나는 가슴의 찰리(로봇), 방독면을 쓴 앰버, 풍금, 구형 라디오, 다수의 켜진 랜턴들이 모두 일치하게 등장함.",
    "hard_violations": [
     "[gpt-high] 살아 있는 사람과 얼굴 및 서 있는 인물을 금지한 명시적 조항에도 불구하고, 얼굴 일부가 보이는 성인 여성 한 명을 전신으로 등장시켰다."
    ],
    "physics": "찰리와 앰버 모두 지하실 바닥에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "등장인물이 전혀 없어 시선이나 행동 방향을 확인할 수 없음.",
    "built_space": "교회 지하실. 부감 구도로 공간과 가구들의 배치는 기준 이미지와 유사함.",
    "entities": "풍금, 라디오, 켜진 랜턴들은 존재하지만, 명시된 찰리와 앰버가 전혀 존재하지 않음.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트와 카메라 지시문에 필수적으로 명시된 찰리와 앰버가 화면에서 완전히 누락됨."
    ],
    "physics": "모든 사물과 가구들이 바닥이나 선반 위에 지지를 받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리와 앰버의 배치, 로봇 가슴의 빛과 일제히 켜지는 랜턴 등 프롬프트가 요구한 상황과 카메라 구도를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "샷 텍스트와 카메라 지시문이 명시한 핵심 피사체(찰리와 앰버)가 완전히 누락되어 프롬프트를 실패함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 다소 아래를 향해 시선을 두고 있고, 앰버는 찰리 쪽을 바라보고 있음.",
        "built_space": "교회 지하실. 부감(high angle) 구도로 계단, 십자가, 풍금, 책장 등이 기준 이미지에 맞게 배치되어 있음.",
        "entities": "빛나는 가슴의 찰리(로봇), 방독면을 쓴 앰버, 풍금, 구형 라디오, 다수의 켜진 랜턴들이 모두 일치하게 등장함.",
        "hard_violations": [],
        "physics": "찰리와 앰버 모두 지하실 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "등장인물이 전혀 없어 시선이나 행동 방향을 확인할 수 없음.",
        "built_space": "교회 지하실. 부감 구도로 공간과 가구들의 배치는 기준 이미지와 유사함.",
        "entities": "풍금, 라디오, 켜진 랜턴들은 존재하지만, 명시된 찰리와 앰버가 전혀 존재하지 않음.",
        "hard_violations": [
         "샷 텍스트와 카메라 지시문에 필수적으로 명시된 찰리와 앰버가 화면에서 완전히 누락됨."
        ],
        "physics": "모든 사물과 가구들이 바닥이나 선반 위에 지지를 받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리와 앰버의 배치, 로봇 가슴의 빛과 일제히 켜지는 랜턴 등 프롬프트가 요구한 상황과 카메라 구도를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "샷 텍스트와 카메라 지시문이 명시한 핵심 피사체(찰리와 앰버)가 완전히 누락되어 프롬프트를 실패함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 다소 아래를 향해 시선을 두고 있고, 앰버는 찰리 쪽을 바라보고 있음.",
        "built_space": "교회 지하실. 부감(high angle) 구도로 계단, 십자가, 풍금, 책장 등이 기준 이미지에 맞게 배치되어 있음.",
        "entities": "빛나는 가슴의 찰리(로봇), 방독면을 쓴 앰버, 풍금, 구형 라디오, 다수의 켜진 랜턴들이 모두 일치하게 등장함.",
        "hard_violations": [],
        "physics": "찰리와 앰버 모두 지하실 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "등장인물이 전혀 없어 시선이나 행동 방향을 확인할 수 없음.",
        "built_space": "교회 지하실. 부감 구도로 공간과 가구들의 배치는 기준 이미지와 유사함.",
        "entities": "풍금, 라디오, 켜진 랜턴들은 존재하지만, 명시된 찰리와 앰버가 전혀 존재하지 않음.",
        "hard_violations": [
         "샷 텍스트와 카메라 지시문에 필수적으로 명시된 찰리와 앰버가 화면에서 완전히 누락됨."
        ],
        "physics": "모든 사물과 가구들이 바닥이나 선반 위에 지지를 받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물 금지와 높은 사선 와이드 구도, 분산된 랜턴 및 오른쪽 라디오를 지켰지만 찰리의 발광에 반응하는 순간은 드러나지 않는다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "하단 왼쪽 배치와 가슴 발광은 구현했으나 명시적인 인물 금지를 어기고 여성을 등장시켰으며, 찰리의 주의도 라디오보다 풍금 쪽을 향한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이나 로봇이 없어 찰리의 라디오 응시와 앰버의 가슴 응시는 없다. 오른쪽 풍금 건반은 앞쪽 연주 벤치를 향하며, 오른쪽 선반의 라디오 전면은 방 안쪽으로 비스듬히 향한다.",
        "built_space": "높은 사선 시점으로 지하실을 넓게 보여준다. 왼쪽에는 위로 이어지는 계단 하나, 뒤 벽에는 큰 십자가 하나, 중앙에는 작은 탁자와 의자, 왼쪽 벽에는 의자 세 개가 보인다. 풍금 하나와 연주 벤치는 오른쪽에 있고 라디오 하나는 그 오른쪽 별도 선반에 놓였다. 켜진 랜턴 아홉 개가 벽걸이와 가구 위에 분산되어 있다. 참고의 낡은 벽·배관·계단·기도실 분위기는 유지하지만 추가 선반과 장식으로 공간이 더 채워졌으며, 하단 왼쪽 인물 중심 배치는 없다.",
        "entities": "살아 있는 사람이나 얼굴이 없어 인물 금지 조항을 충족한다. 십자가, 탁자, 의자, 건반 악기, 낡은 라디오와 여러 랜턴은 식별된다. 악기는 작은 풍금보다 업라이트 피아노에 가까운 외형이다. 찰리 자체가 없어 코트·모자·푸른 눈·마모된 가슴 로고·가슴 발광은 확인되지 않는다. 외부 개구부는 어둡고 실내는 따뜻한 실용등으로 밝혀져 야간 설정과 양립한다.",
        "hard_violations": [],
        "physics": "벽 랜턴은 고리에 걸려 있고 나머지 랜턴과 라디오는 가구 상판에 놓여 있다. 가구는 바닥에 지지되며 공중에 뜬 물체는 없다. 랜턴의 밝은 상태는 보이지만 찰리의 빛에 반응해 동시에 켜지는 과정이나 발광원은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 고개와 몸통은 화면 오른쪽 아래 풍금 쪽으로 기울어 있다. 오른쪽 벽의 더 먼 라디오가 명확한 응시 대상으로 읽히지는 않는다. 여성은 찰리 쪽으로 몸을 기울이고 얼굴도 그를 향하지만 가슴만을 응시하는지는 불명확하다. 풍금 건반은 연주 벤치 쪽을, 라디오 전면은 방 안쪽을 향한다.",
        "built_space": "높은 사선 와이드 시점이며 찰리와 여성은 하단 왼쪽, 건반과 상판이 보이는 풍금 하나는 하단 중앙에서 오른쪽에 배치됐다. 라디오 하나는 오른쪽 벽의 별도 장 위에 있다. 왼쪽 상승 계단 하나, 뒤 벽 십자가 하나, 중앙 탁자 하나, 왼쪽 줄지은 의자 세 개와 중앙·뒤쪽 의자들이 보인다. 랜턴 열 개가 벽과 가구 위에 흩어져 있다. 참고의 재료와 주요 공간 요소는 닮았지만 풍금은 오른쪽 벽이 아닌 방 안쪽으로 나와 있다.",
        "entities": "코트와 모자를 쓴 금속 로봇 한 대와 금발의 성인 여성 한 명이 보인다. 여성의 방독면과 작업복은 이전 이미지의 여성과 유사하며, 이번 이미지에는 살아 있는 사람을 등장시키지 말라는 조항과 충돌한다. 로봇 가슴은 푸르게 빛나지만 보이는 눈은 요구된 파란색이 아니라 황색이다. 가슴의 마모된 로고는 식별되지 않는다. 십자가·탁자·의자·라디오·랜턴은 있고, 건반 악기는 풍금보다 피아노에 가까워 보인다.",
        "hard_violations": [
         "살아 있는 사람과 얼굴 및 서 있는 인물을 금지한 명시적 조항에도 불구하고, 얼굴 일부가 보이는 성인 여성 한 명을 전신으로 등장시켰다."
        ],
        "physics": "로봇과 여성 모두 발로 바닥을 딛고 있으며, 기울어진 자세도 지지 범위 안에 있다. 벽 랜턴은 걸려 있고 다른 랜턴과 라디오는 가구에 받쳐져 있다. 떠 있는 인물이나 물체는 없다. 로봇 가슴의 푸른 발광과 여러 랜턴의 점등은 동시에 보이지만, 랜턴이 막 밝아지는 변화 자체는 명확하지 않다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물 금지와 높은 사선 와이드 구도, 분산된 랜턴 및 오른쪽 라디오를 지켰지만 찰리의 발광에 반응하는 순간은 드러나지 않는다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "하단 왼쪽 배치와 가슴 발광은 구현했으나 명시적인 인물 금지를 어기고 여성을 등장시켰으며, 찰리의 주의도 라디오보다 풍금 쪽을 향한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물이나 로봇이 없어 찰리의 라디오 응시와 앰버의 가슴 응시는 없다. 오른쪽 풍금 건반은 앞쪽 연주 벤치를 향하며, 오른쪽 선반의 라디오 전면은 방 안쪽으로 비스듬히 향한다.",
        "built_space": "높은 사선 시점으로 지하실을 넓게 보여준다. 왼쪽에는 위로 이어지는 계단 하나, 뒤 벽에는 큰 십자가 하나, 중앙에는 작은 탁자와 의자, 왼쪽 벽에는 의자 세 개가 보인다. 풍금 하나와 연주 벤치는 오른쪽에 있고 라디오 하나는 그 오른쪽 별도 선반에 놓였다. 켜진 랜턴 아홉 개가 벽걸이와 가구 위에 분산되어 있다. 참고의 낡은 벽·배관·계단·기도실 분위기는 유지하지만 추가 선반과 장식으로 공간이 더 채워졌으며, 하단 왼쪽 인물 중심 배치는 없다.",
        "entities": "살아 있는 사람이나 얼굴이 없어 인물 금지 조항을 충족한다. 십자가, 탁자, 의자, 건반 악기, 낡은 라디오와 여러 랜턴은 식별된다. 악기는 작은 풍금보다 업라이트 피아노에 가까운 외형이다. 찰리 자체가 없어 코트·모자·푸른 눈·마모된 가슴 로고·가슴 발광은 확인되지 않는다. 외부 개구부는 어둡고 실내는 따뜻한 실용등으로 밝혀져 야간 설정과 양립한다.",
        "hard_violations": [],
        "physics": "벽 랜턴은 고리에 걸려 있고 나머지 랜턴과 라디오는 가구 상판에 놓여 있다. 가구는 바닥에 지지되며 공중에 뜬 물체는 없다. 랜턴의 밝은 상태는 보이지만 찰리의 빛에 반응해 동시에 켜지는 과정이나 발광원은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 고개와 몸통은 화면 오른쪽 아래 풍금 쪽으로 기울어 있다. 오른쪽 벽의 더 먼 라디오가 명확한 응시 대상으로 읽히지는 않는다. 여성은 찰리 쪽으로 몸을 기울이고 얼굴도 그를 향하지만 가슴만을 응시하는지는 불명확하다. 풍금 건반은 연주 벤치 쪽을, 라디오 전면은 방 안쪽을 향한다.",
        "built_space": "높은 사선 와이드 시점이며 찰리와 여성은 하단 왼쪽, 건반과 상판이 보이는 풍금 하나는 하단 중앙에서 오른쪽에 배치됐다. 라디오 하나는 오른쪽 벽의 별도 장 위에 있다. 왼쪽 상승 계단 하나, 뒤 벽 십자가 하나, 중앙 탁자 하나, 왼쪽 줄지은 의자 세 개와 중앙·뒤쪽 의자들이 보인다. 랜턴 열 개가 벽과 가구 위에 흩어져 있다. 참고의 재료와 주요 공간 요소는 닮았지만 풍금은 오른쪽 벽이 아닌 방 안쪽으로 나와 있다.",
        "entities": "코트와 모자를 쓴 금속 로봇 한 대와 금발의 성인 여성 한 명이 보인다. 여성의 방독면과 작업복은 이전 이미지의 여성과 유사하며, 이번 이미지에는 살아 있는 사람을 등장시키지 말라는 조항과 충돌한다. 로봇 가슴은 푸르게 빛나지만 보이는 눈은 요구된 파란색이 아니라 황색이다. 가슴의 마모된 로고는 식별되지 않는다. 십자가·탁자·의자·라디오·랜턴은 있고, 건반 악기는 풍금보다 피아노에 가까워 보인다.",
        "hard_violations": [
         "살아 있는 사람과 얼굴 및 서 있는 인물을 금지한 명시적 조항에도 불구하고, 얼굴 일부가 보이는 성인 여성 한 명을 전신으로 등장시켰다."
        ],
        "physics": "로봇과 여성 모두 발로 바닥을 딛고 있으며, 기울어진 자세도 지지 범위 안에 있다. 벽 랜턴은 걸려 있고 다른 랜턴과 라디오는 가구에 받쳐져 있다. 떠 있는 인물이나 물체는 없다. 로봇 가슴의 푸른 발광과 여러 랜턴의 점등은 동시에 보이지만, 랜턴이 막 밝아지는 변화 자체는 명확하지 않다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gemini-pro] 샷 텍스트와 카메라 지시문에 필수적으로 명시된 찰리와 앰버가 화면에서 완전히 누락됨."
    ],
    "A": [
     "[gpt-high] 살아 있는 사람과 얼굴 및 서 있는 인물을 금지한 명시적 조항에도 불구하고, 얼굴 일부가 보이는 성인 여성 한 명을 전신으로 등장시켰다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1083,
   "B": 1036
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "찰리와 앰버의 배치, 로봇 가슴의 빛과 일제히 켜지는 랜턴 등 프롬프트가 요구한 상황과 카메라 구도를 충실히 구현함.  ★위반: [gpt-high] 살아 있는 사람과 얼굴 및 서 있는 인물을 금지한 명시적 조항에도 불구하고, 얼굴 일부가 보이는 성인 여성 한 명을 전신으로 등장시켰다."
   },
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "샷 텍스트와 카메라 지시문이 명시한 핵심 피사체(찰리와 앰버)가 완전히 누락되어 프롬프트를 실패함.  ★위반: [gemini-pro] 샷 텍스트와 카메라 지시문에 필수적으로 명시된 찰리와 앰버가 화면에서 완전히 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S29sh5_sel.png",
    "asset_id": "dcf3d4b1-f66d-4e4e-ab25-4e1ecfda9162",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09d9-a026-7dac-bef2-7e822126dd9d",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh5"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S30sh10::signage": {
  "fp": "ca8a68f926045f86",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh10": {
  "input_fingerprint": "788a3919372569c0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리를 발견하고 반가운 표정으로 달려와 거대한 기계 몸을 덥석 끌어안아 밀착된 구도환의 역동적인 상체.\n\nLOCATION (lock): In the open floor area of the church basement near the small organ, now lit by the bright lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 구도환's approach at upper-body distance, slightly below his chest with a gentle upward tilt, settling on the embrace rather than widening. His three-quarter face and forward-driving torso occupy the left-center as his arms close around 찰리 on the right; 구도환 looks up toward 찰리's face, while 찰리 lowers his attention toward him, with their contact clearly visible between them. Let the closing distance carry the visual change, preserving the established side of the approach.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination retains readable faces and precise robot contours, with controlled highlights and tenderness conveyed through the embrace rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the illuminated lanterns, small organ, prayer-room furnishings, and interior materials from the reference. Exclude the transient light burst that initially switched the lanterns on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain bright, and the activated old radio has begun producing assembled speech. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes beside the small organ. 구도환: He has descended into the basement and leans forward with his arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리를 발견하고 반가운 표정으로 달려와 거대한 기계 몸을 덥석 끌어안아 밀착된 구도환의 역동적인 상체.\n\nLOCATION (lock): In the open floor area of the church basement near the small organ, now lit by the bright lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 구도환's approach at upper-body distance, slightly below his chest with a gentle upward tilt, settling on the embrace rather than widening. His three-quarter face and forward-driving torso occupy the left-center as his arms close around 찰리 on the right; 구도환 looks up toward 찰리's face, while 찰리 lowers his attention toward him, with their contact clearly visible between them. Let the closing distance carry the visual change, preserving the established side of the approach.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination retains readable faces and precise robot contours, with controlled highlights and tenderness conveyed through the embrace rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the illuminated lanterns, small organ, prayer-room furnishings, and interior materials from the reference. Exclude the transient light burst that initially switched the lanterns on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain bright, and the activated old radio has begun producing assembled speech. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes beside the small organ. 구도환: He has descended into the basement and leans forward with his arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리를 발견하고 반가운 표정으로 달려와 거대한 기계 몸을 덥석 끌어안아 밀착된 구도환의 역동적인 상체.\n\nLOCATION (lock): In the open floor area of the church basement near the small organ, now lit by the bright lanterns. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 구도환's approach at upper-body distance, slightly below his chest with a gentle upward tilt, settling on the embrace rather than widening. His three-quarter face and forward-driving torso occupy the left-center as his arms close around 찰리 on the right; 구도환 looks up toward 찰리's face, while 찰리 lowers his attention toward him, with their contact clearly visible between them. Let the closing distance carry the visual change, preserving the established side of the approach.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established lantern illumination retains readable faces and precise robot contours, with controlled highlights and tenderness conveyed through the embrace rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the illuminated lanterns, small organ, prayer-room furnishings, and interior materials from the reference. Exclude the transient light burst that initially switched the lanterns on.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain bright, and the activated old radio has begun producing assembled speech. Charlie retains his old coat and hat disguise, worn chest logo and blue-lit eyes beside the small organ. 구도환: He has descended into the basement and leans forward with his arms extended in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "구도환은 찰리를 올려다보나, 찰리의 시선은 아래로 향하지 않고 정면을 바라봄.",
    "built_space": "교회 지하실. 왼쪽에 십자가와 테이블, 오른쪽에 오르간과 라디오가 레퍼런스대로 배치됨.",
    "entities": "인물들의 복장과 외형은 레퍼런스와 일치하나, 찰리의 포즈와 표정 연출이 지시된 시선 방향을 따르지 않음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 구도환의 등을 감싸고 있는 화면 왼쪽의 찰리 기계 손 엄지손가락이 아래를 향하고 있어 오른쪽 손의 구조로서 불가능함."
    ],
    "physics": "구도환의 포옹 자세는 유지되나, 화면 좌측 찰리의 기계 손 구조가 물리적/해부학적으로 불가능하게 렌더링됨."
   },
   {
    "label": "B",
    "direction": "구도환은 찰리를 올려다보고 있으며, 찰리 역시 고개를 숙여 구도환을 내려다보고 있음.",
    "built_space": "교회 지하실. 왼쪽 배경에 십자가와 테이블, 오른쪽에 오르간과 라디오가 정확하게 배치됨.",
    "entities": "구도환과 찰리 모두 레퍼런스의 특징을 정확히 반영함. 찰리는 모자와 코트를 착용했고 거대한 기계 몸체를 가짐.",
    "hard_violations": [],
    "physics": "두 인물의 자세가 자연스럽고 지탱되어 있으며, 포옹하는 팔과 기계 손의 해부학적 구조(엄지손가락 위치 등)가 올바름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 앵글과 찰리의 하향 시선, 배경 배치와 두 인물의 밀착된 포옹이 지시사항에 맞춰 충실하게 렌더링됨."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 시선 방향이 프롬프트와 다르고 기계 손의 해부학적 오류(엄지 방향)가 있어 감점됨."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 찰리를 올려다보고 있으며, 찰리 역시 고개를 숙여 구도환을 내려다보고 있음.",
        "built_space": "교회 지하실. 왼쪽 배경에 십자가와 테이블, 오른쪽에 오르간과 라디오가 정확하게 배치됨.",
        "entities": "구도환과 찰리 모두 레퍼런스의 특징을 정확히 반영함. 찰리는 모자와 코트를 착용했고 거대한 기계 몸체를 가짐.",
        "hard_violations": [],
        "physics": "두 인물의 자세가 자연스럽고 지탱되어 있으며, 포옹하는 팔과 기계 손의 해부학적 구조(엄지손가락 위치 등)가 올바름."
       },
       {
        "label": "A",
        "direction": "구도환은 찰리를 올려다보나, 찰리의 시선은 아래로 향하지 않고 정면을 바라봄.",
        "built_space": "교회 지하실. 왼쪽에 십자가와 테이블, 오른쪽에 오르간과 라디오가 레퍼런스대로 배치됨.",
        "entities": "인물들의 복장과 외형은 레퍼런스와 일치하나, 찰리의 포즈와 표정 연출이 지시된 시선 방향을 따르지 않음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 구도환의 등을 감싸고 있는 화면 왼쪽의 찰리 기계 손 엄지손가락이 아래를 향하고 있어 오른쪽 손의 구조로서 불가능함."
        ],
        "physics": "구도환의 포옹 자세는 유지되나, 화면 좌측 찰리의 기계 손 구조가 물리적/해부학적으로 불가능하게 렌더링됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 앵글과 찰리의 하향 시선, 배경 배치와 두 인물의 밀착된 포옹이 지시사항에 맞춰 충실하게 렌더링됨."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 시선 방향이 프롬프트와 다르고 기계 손의 해부학적 오류(엄지 방향)가 있어 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 찰리를 올려다보고 있으며, 찰리 역시 고개를 숙여 구도환을 내려다보고 있음.",
        "built_space": "교회 지하실. 왼쪽 배경에 십자가와 테이블, 오른쪽에 오르간과 라디오가 정확하게 배치됨.",
        "entities": "구도환과 찰리 모두 레퍼런스의 특징을 정확히 반영함. 찰리는 모자와 코트를 착용했고 거대한 기계 몸체를 가짐.",
        "hard_violations": [],
        "physics": "두 인물의 자세가 자연스럽고 지탱되어 있으며, 포옹하는 팔과 기계 손의 해부학적 구조(엄지손가락 위치 등)가 올바름."
       },
       {
        "label": "A",
        "direction": "구도환은 찰리를 올려다보나, 찰리의 시선은 아래로 향하지 않고 정면을 바라봄.",
        "built_space": "교회 지하실. 왼쪽에 십자가와 테이블, 오른쪽에 오르간과 라디오가 레퍼런스대로 배치됨.",
        "entities": "인물들의 복장과 외형은 레퍼런스와 일치하나, 찰리의 포즈와 표정 연출이 지시된 시선 방향을 따르지 않음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 구도환의 등을 감싸고 있는 화면 왼쪽의 찰리 기계 손 엄지손가락이 아래를 향하고 있어 오른쪽 손의 구조로서 불가능함."
        ],
        "physics": "구도환의 포옹 자세는 유지되나, 화면 좌측 찰리의 기계 손 구조가 물리적/해부학적으로 불가능하게 렌더링됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽에서 밀착하는 구도환의 상체와 찰리를 올려다보는 시선, 이에 응하는 찰리의 고개 숙임이 포옹 순간을 더 정확히 구현하지만 눈빛은 지정된 파란색이 아니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미디엄 구도의 밀착 포옹과 오르간 주변 공간은 잘 구현했으나 찰리의 얼굴이 구도환보다 전방을 향해 상호 주시가 약하고 눈빛도 주황색이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 화면 왼쪽에서 오른쪽의 찰리에게 몸을 기울이고 얼굴을 올려 찰리의 마스크를 바라본다. 찰리는 고개를 왼쪽 아래로 숙여 구도환의 얼굴을 향한다. 구도환의 보이는 손은 찰리의 위팔을 붙잡고, 찰리의 팔은 구도환의 등과 허리를 감싼다.",
        "built_space": "뒤쪽 왼편에 탁자 하나와 의자들이 있고, 벽 십자가 하나와 책장 하나, 책장 위 성상 하나가 보인다. 오른쪽에는 작은 오르간 하나와 벤치 하나, 그 뒤쪽 라디오 하나가 있으며 켜진 랜턴은 부분 노출을 포함해 다섯 개 보인다. 낡은 벽과 천장 배관, 양탄자는 이전 장소의 재질과 구성을 이어간다. 두 인물은 오르간 옆 열린 공간에 있어 가구와 충돌하지 않는다. 반사상이나 명백한 설비 중복은 없다.",
        "entities": "등장자는 구도환과 찰리뿐이다. 구도환은 짧은 검은 머리의 중년 동아시아계 남성으로, 참조와 유사한 얼굴 및 해진 갈색 점퍼를 갖는다. 찰리는 흰 마스크, 점 형태의 눈과 선 형태의 입, 마모된 베이지 장갑, 육중한 기계 팔, 낡은 모자와 코트, 파란 원형 가슴 동력부를 갖춘다. 다만 눈은 상태 지시의 파란색이 아니라 주황색이며, 가슴의 별도 마모 로고는 식별되지 않는다. 하체와 바지는 구도 밖이다.",
        "hard_violations": [],
        "physics": "구도환의 손바닥과 손가락이 찰리의 코트 위 위팔에 닿고 몸통도 기계 가슴에 밀착한다. 찰리의 한 손은 구도환의 등 옆에 접촉하고 다른 팔은 아래쪽 허리를 두른다. 발은 프레임 밖이지만 몸이 떠 있다는 징후는 없으며, 앞으로 기울어 달려온 뒤 포옹하는 상체 자세로 가능하다. 모자와 코트는 몸에 걸려 있고 배경 소품은 가구 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "구도환은 왼쪽에서 찰리에게 밀착해 웃으며 오른쪽 위의 찰리 얼굴을 바라본다. 찰리의 얼굴은 약간 아래로 기울었지만 구도환 쪽으로 충분히 돌아가기보다는 전방과 화면 왼쪽을 향해, 구도환 얼굴에 주의가 내려앉는 관계가 A보다 불명확하다. 구도환의 손은 찰리 위팔을 붙잡고 찰리는 그의 등을 감싼다.",
        "built_space": "왼쪽 뒤에 탁자 하나와 의자들이 있고, 벽 십자가 하나, 책장 하나와 성상 하나가 보인다. 오른쪽 앞에는 작은 오르간 하나, 그 뒤 별도 장 위에는 라디오 하나가 있으며 켜진 랜턴 네 개가 보인다. 오르간의 레이스와 책, 뒤쪽 장의 라디오 배치는 이전 장소와 대체로 일관된다. 낡은 벽과 노출 배관도 유지된다. 인물은 오르간 옆에 있고 설비 중복이나 불가능한 반사상은 보이지 않는다.",
        "entities": "구도환과 찰리 외 인물은 없다. 구도환은 참조와 유사한 중년 동아시아계 남성의 얼굴, 짧은 검은 머리와 낡은 갈색 점퍼를 갖는다. 찰리의 흰 마스크와 단순한 눈·입, 베이지 기계 장갑, 큰 팔, 낡은 코트와 모자, 파란 가슴 동력부가 보인다. 눈은 지정된 파란색이 아닌 주황색이며 별도의 마모된 가슴 로고는 확인되지 않는다. 바지와 짧은 다리 비율은 이 상체 구도에서 판단할 수 없다.",
        "hard_violations": [],
        "physics": "구도환의 보이는 손이 찰리의 위팔을 잡고 몸통이 가슴에 닿아 있다. 찰리의 손은 구도환의 등 옆을 받치고 큰 전완은 허리 앞을 가로지른다. 팔의 연결과 접촉은 가능한 포옹 자세이며, 다리가 잘렸다는 이유로 공중 부양으로 볼 근거는 없다. 배경의 라디오, 책과 랜턴에는 받침 가구가 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽에서 밀착하는 구도환의 상체와 찰리를 올려다보는 시선, 이에 응하는 찰리의 고개 숙임이 포옹 순간을 더 정확히 구현하지만 눈빛은 지정된 파란색이 아니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 구도의 밀착 포옹과 오르간 주변 공간은 잘 구현했으나 찰리의 얼굴이 구도환보다 전방을 향해 상호 주시가 약하고 눈빛도 주황색이다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 화면 왼쪽에서 오른쪽의 찰리에게 몸을 기울이고 얼굴을 올려 찰리의 마스크를 바라본다. 찰리는 고개를 왼쪽 아래로 숙여 구도환의 얼굴을 향한다. 구도환의 보이는 손은 찰리의 위팔을 붙잡고, 찰리의 팔은 구도환의 등과 허리를 감싼다.",
        "built_space": "뒤쪽 왼편에 탁자 하나와 의자들이 있고, 벽 십자가 하나와 책장 하나, 책장 위 성상 하나가 보인다. 오른쪽에는 작은 오르간 하나와 벤치 하나, 그 뒤쪽 라디오 하나가 있으며 켜진 랜턴은 부분 노출을 포함해 다섯 개 보인다. 낡은 벽과 천장 배관, 양탄자는 이전 장소의 재질과 구성을 이어간다. 두 인물은 오르간 옆 열린 공간에 있어 가구와 충돌하지 않는다. 반사상이나 명백한 설비 중복은 없다.",
        "entities": "등장자는 구도환과 찰리뿐이다. 구도환은 짧은 검은 머리의 중년 동아시아계 남성으로, 참조와 유사한 얼굴 및 해진 갈색 점퍼를 갖는다. 찰리는 흰 마스크, 점 형태의 눈과 선 형태의 입, 마모된 베이지 장갑, 육중한 기계 팔, 낡은 모자와 코트, 파란 원형 가슴 동력부를 갖춘다. 다만 눈은 상태 지시의 파란색이 아니라 주황색이며, 가슴의 별도 마모 로고는 식별되지 않는다. 하체와 바지는 구도 밖이다.",
        "hard_violations": [],
        "physics": "구도환의 손바닥과 손가락이 찰리의 코트 위 위팔에 닿고 몸통도 기계 가슴에 밀착한다. 찰리의 한 손은 구도환의 등 옆에 접촉하고 다른 팔은 아래쪽 허리를 두른다. 발은 프레임 밖이지만 몸이 떠 있다는 징후는 없으며, 앞으로 기울어 달려온 뒤 포옹하는 상체 자세로 가능하다. 모자와 코트는 몸에 걸려 있고 배경 소품은 가구 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "구도환은 왼쪽에서 찰리에게 밀착해 웃으며 오른쪽 위의 찰리 얼굴을 바라본다. 찰리의 얼굴은 약간 아래로 기울었지만 구도환 쪽으로 충분히 돌아가기보다는 전방과 화면 왼쪽을 향해, 구도환 얼굴에 주의가 내려앉는 관계가 A보다 불명확하다. 구도환의 손은 찰리 위팔을 붙잡고 찰리는 그의 등을 감싼다.",
        "built_space": "왼쪽 뒤에 탁자 하나와 의자들이 있고, 벽 십자가 하나, 책장 하나와 성상 하나가 보인다. 오른쪽 앞에는 작은 오르간 하나, 그 뒤 별도 장 위에는 라디오 하나가 있으며 켜진 랜턴 네 개가 보인다. 오르간의 레이스와 책, 뒤쪽 장의 라디오 배치는 이전 장소와 대체로 일관된다. 낡은 벽과 노출 배관도 유지된다. 인물은 오르간 옆에 있고 설비 중복이나 불가능한 반사상은 보이지 않는다.",
        "entities": "구도환과 찰리 외 인물은 없다. 구도환은 참조와 유사한 중년 동아시아계 남성의 얼굴, 짧은 검은 머리와 낡은 갈색 점퍼를 갖는다. 찰리의 흰 마스크와 단순한 눈·입, 베이지 기계 장갑, 큰 팔, 낡은 코트와 모자, 파란 가슴 동력부가 보인다. 눈은 지정된 파란색이 아닌 주황색이며 별도의 마모된 가슴 로고는 확인되지 않는다. 바지와 짧은 다리 비율은 이 상체 구도에서 판단할 수 없다.",
        "hard_violations": [],
        "physics": "구도환의 보이는 손이 찰리의 위팔을 잡고 몸통이 가슴에 닿아 있다. 찰리의 손은 구도환의 등 옆을 받치고 큰 전완은 허리 앞을 가로지른다. 팔의 연결과 접촉은 가능한 포옹 자세이며, 다리가 잘렸다는 이유로 공중 부양으로 볼 근거는 없다. 배경의 라디오, 책과 랜턴에는 받침 가구가 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.125,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 구도환의 등을 감싸고 있는 화면 왼쪽의 찰리 기계 손 엄지손가락이 아래를 향하고 있어 오른쪽 손의 구조로서 불가능함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 앵글과 찰리의 하향 시선, 배경 배치와 두 인물의 밀착된 포옹이 지시사항에 맞춰 충실하게 렌더링됨."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "찰리의 시선 방향이 프롬프트와 다르고 기계 손의 해부학적 오류(엄지 방향)가 있어 감점됨.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 구도환의 등을 감싸고 있는 화면 왼쪽의 찰리 기계 손 엄지손가락이 아래를 향하고 있어 오른쪽 손의 구조로서 불가능함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S30sh6_sel.png",
    "asset_id": "39664a01-ac18-43ee-9d36-27bbfaa65114",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09de-e379-79c3-85e5-ada26b3b4c62",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh6"
  }
 },
 "S30sh14::signage": {
  "fp": "85a28861232925b8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh14": {
  "input_fingerprint": "ec5e533dd695bfd4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환을 향해 극도의 불신을 담아 손가락을 뻗어 삿대질한 채 입을 크게 벌려 거칠게 쏘아붙이는 표정의 이현우의 화난 상반신.\n\nLOCATION (lock): Inside the lantern-lit church basement, in the shared floor space near the small organ and table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Nearly stop the track just behind and outside 구도환's shoulder at approximately eye level, looking obliquely toward 이현우 without crossing their conversation axis. Keep 구도환's rear shoulder as a narrow strip at the left foreground edge and 이현우's angry upper body at right-center, his extended finger traveling toward 구도환 rather than the lens. 이현우 fixes his attention on 구도환's face just beyond the near crop, mouth open mid-accusation, while 구도환 holds his head toward the speaker; the pointing arm is the principal positional change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern illumination and controlled contrast without a dramatic exposure shift for the accusation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 구도환 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, organ, small desk, chairs, and prayer-room surfaces from the reference. Exclude the initial activation flash and all helicopter or street-search elements.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit, and the old radio remains activated beside the established organ, desk, chairs and cross. Charlie retains his old coat and hat disguise, blue-lit eyes and worn chest logo. 이현우: He stands in the basement, visibly distrustful, with facial bruises and the injured leg; his outer garment remains removed. 구도환: He remains in the basement after the greeting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환을 향해 극도의 불신을 담아 손가락을 뻗어 삿대질한 채 입을 크게 벌려 거칠게 쏘아붙이는 표정의 이현우의 화난 상반신.\n\nLOCATION (lock): Inside the lantern-lit church basement, in the shared floor space near the small organ and table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Nearly stop the track just behind and outside 구도환's shoulder at approximately eye level, looking obliquely toward 이현우 without crossing their conversation axis. Keep 구도환's rear shoulder as a narrow strip at the left foreground edge and 이현우's angry upper body at right-center, his extended finger traveling toward 구도환 rather than the lens. 이현우 fixes his attention on 구도환's face just beyond the near crop, mouth open mid-accusation, while 구도환 holds his head toward the speaker; the pointing arm is the principal positional change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern illumination and controlled contrast without a dramatic exposure shift for the accusation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 구도환 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, organ, small desk, chairs, and prayer-room surfaces from the reference. Exclude the initial activation flash and all helicopter or street-search elements.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit, and the old radio remains activated beside the established organ, desk, chairs and cross. Charlie retains his old coat and hat disguise, blue-lit eyes and worn chest logo. 이현우: He stands in the basement, visibly distrustful, with facial bruises and the injured leg; his outer garment remains removed. 구도환: He remains in the basement after the greeting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환을 향해 극도의 불신을 담아 손가락을 뻗어 삿대질한 채 입을 크게 벌려 거칠게 쏘아붙이는 표정의 이현우의 화난 상반신.\n\nLOCATION (lock): Inside the lantern-lit church basement, in the shared floor space near the small organ and table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Nearly stop the track just behind and outside 구도환's shoulder at approximately eye level, looking obliquely toward 이현우 without crossing their conversation axis. Keep 구도환's rear shoulder as a narrow strip at the left foreground edge and 이현우's angry upper body at right-center, his extended finger traveling toward 구도환 rather than the lens. 이현우 fixes his attention on 구도환's face just beyond the near crop, mouth open mid-accusation, while 구도환 holds his head toward the speaker; the pointing arm is the principal positional change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established lantern illumination and controlled contrast without a dramatic exposure shift for the accusation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 구도환 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, organ, small desk, chairs, and prayer-room surfaces from the reference. Exclude the initial activation flash and all helicopter or street-search elements.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement lanterns remain lit, and the old radio remains activated beside the established organ, desk, chairs and cross. Charlie retains his old coat and hat disguise, blue-lit eyes and worn chest logo. 이현우: He stands in the basement, visibly distrustful, with facial bruises and the injured leg; his outer garment remains removed. 구도환: He remains in the basement after the greeting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 뻗은 손가락이 정확히 프레임 왼쪽의 구도환을 향하고 있음.",
    "built_space": "벽면의 십자가, 테이블과 의자, 바닥의 양탄자, 우측의 오르간과 책장이 이전 샷과 일치하는 위치에 배치되어 있음.",
    "entities": "이현우는 레퍼런스와 일치하는 지저분한 셔츠, 헝클어진 머리, 인이어 무전기를 착용하고 있으며, 구도환은 이전 샷의 갈색 재킷을 입은 뒷모습으로 나타남.",
    "hard_violations": [],
    "physics": "이현우의 뻗은 팔은 자연스럽게 어깨에 지지되어 있으며, 두 인물 모두 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선과 손가락이 프레임 왼쪽의 구도환을 향하고 있음.",
    "built_space": "지하실의 주요 구조와 십자가, 오르간은 존재하나, 책장 위 소품들의 세부 형태가 이전 샷과 미세하게 다름.",
    "entities": "이현우와 구도환 모두 지정된 복장 및 인상착의(지저분한 셔츠, 인이어 무전기, 뒷모습의 재킷)를 잘 반영하고 있음.",
    "hard_violations": [],
    "physics": "인물들의 자세와 뻗은 팔의 구조가 물리적으로 자연스럽게 지지되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 오버더숄더 앵글과 이현우의 격양된 표정 및 삿대질 동작을 정확히 담아냈으며, 이전 샷의 배경 디테일을 훌륭하게 유지함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프롬프트의 주요 지시사항을 잘 수행했으나, 배경에 배치된 책장과 오르간 등의 디테일이 A에 비해 이전 샷과의 일치도가 다소 떨어짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 뻗은 손가락이 정확히 프레임 왼쪽의 구도환을 향하고 있음.",
        "built_space": "벽면의 십자가, 테이블과 의자, 바닥의 양탄자, 우측의 오르간과 책장이 이전 샷과 일치하는 위치에 배치되어 있음.",
        "entities": "이현우는 레퍼런스와 일치하는 지저분한 셔츠, 헝클어진 머리, 인이어 무전기를 착용하고 있으며, 구도환은 이전 샷의 갈색 재킷을 입은 뒷모습으로 나타남.",
        "hard_violations": [],
        "physics": "이현우의 뻗은 팔은 자연스럽게 어깨에 지지되어 있으며, 두 인물 모두 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 손가락이 프레임 왼쪽의 구도환을 향하고 있음.",
        "built_space": "지하실의 주요 구조와 십자가, 오르간은 존재하나, 책장 위 소품들의 세부 형태가 이전 샷과 미세하게 다름.",
        "entities": "이현우와 구도환 모두 지정된 복장 및 인상착의(지저분한 셔츠, 인이어 무전기, 뒷모습의 재킷)를 잘 반영하고 있음.",
        "hard_violations": [],
        "physics": "인물들의 자세와 뻗은 팔의 구조가 물리적으로 자연스럽게 지지되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 오버더숄더 앵글과 이현우의 격양된 표정 및 삿대질 동작을 정확히 담아냈으며, 이전 샷의 배경 디테일을 훌륭하게 유지함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프롬프트의 주요 지시사항을 잘 수행했으나, 배경에 배치된 책장과 오르간 등의 디테일이 A에 비해 이전 샷과의 일치도가 다소 떨어짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 뻗은 손가락이 정확히 프레임 왼쪽의 구도환을 향하고 있음.",
        "built_space": "벽면의 십자가, 테이블과 의자, 바닥의 양탄자, 우측의 오르간과 책장이 이전 샷과 일치하는 위치에 배치되어 있음.",
        "entities": "이현우는 레퍼런스와 일치하는 지저분한 셔츠, 헝클어진 머리, 인이어 무전기를 착용하고 있으며, 구도환은 이전 샷의 갈색 재킷을 입은 뒷모습으로 나타남.",
        "hard_violations": [],
        "physics": "이현우의 뻗은 팔은 자연스럽게 어깨에 지지되어 있으며, 두 인물 모두 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 손가락이 프레임 왼쪽의 구도환을 향하고 있음.",
        "built_space": "지하실의 주요 구조와 십자가, 오르간은 존재하나, 책장 위 소품들의 세부 형태가 이전 샷과 미세하게 다름.",
        "entities": "이현우와 구도환 모두 지정된 복장 및 인상착의(지저분한 셔츠, 인이어 무전기, 뒷모습의 재킷)를 잘 반영하고 있음.",
        "hard_violations": [],
        "physics": "인물들의 자세와 뻗은 팔의 구조가 물리적으로 자연스럽게 지지되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "구도환을 향한 삿대질과 거칠게 쏘아붙이는 분노가 더 명확하지만, 구도환의 머리와 어깨가 왼쪽 전경을 지나치게 넓게 차지한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장소·의상·상처와 대화 구도는 유지하지만, 손가락의 목표 방향과 격한 분노가 A보다 덜 선명하며 왼쪽 어깨도 요구된 좁은 띠보다 크다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 눈은 왼쪽 구도환의 얼굴을 향하고, 뻗은 검지도 화면 왼쪽 구도환의 아래 얼굴 쪽을 가리킨다. 렌즈를 직접 겨누는 동작보다는 상대를 지목하는 동작으로 읽힌다. 구도환 역시 고개를 이현우 쪽으로 돌리고 있다.",
        "built_space": "오른쪽에 오르간 1대와 그 앞 벤치 1개, 오르간 위 가장자리에 켜진 라디오 일부가 보인다. 뒤쪽에는 작은 탁자 1개, 분명히 식별되는 등받이 의자 1개, 책장 1개, 벽 십자가 1개가 있다. 탁자 쪽·중앙 받침대·오른쪽 벽 부근에 등불이 보인다. 낡은 벽, 배관, 융단과 따뜻한 조명은 이전 장소를 따른다. 두 사람은 가구 사이 열린 공간에서 마주 본다. 눈높이의 상반신 구도이지만 구도환의 머리와 등이 왼쪽 상당 부분을 차지해 '좁은 어깨 띠' 지시와 다르다.",
        "entities": "등장인물은 두 남성뿐이다. 이현우는 젊은 동아시아계 남성으로, 짧고 헝클어진 검은 머리와 마른 체격이 참조에 부합한다. 얼굴 멍과 찰과상, 검은 인이어, 피와 먼지가 묻은 어두운 셔츠가 보이며 겉옷은 없다. 구도환은 짧은 검은 머리와 이전 장면의 낡은 갈색 재킷을 유지한다. 뒤쪽 모습만 보여 정확한 얼굴 및 연령 일치는 제한적으로만 확인된다. 찰리나 다른 인물, 추가 문구는 없다. 다리 부상은 상반신 프레임 밖이다.",
        "hard_violations": [],
        "physics": "이현우의 삿대질 팔은 어깨에서 팔꿈치와 손목으로 자연스럽게 연결되며 손가락 자세도 가능한 범위다. 두 사람의 하체와 발은 프레임 밖이지만 상체는 서 있는 자세로 이어지고 부유를 나타내는 단서는 없다. 라디오는 오르간 위에, 등불과 책은 가구 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 왼쪽 구도환의 얼굴을 향하고 구도환도 이현우를 바라보는 방향이다. 삿대질하는 팔은 왼쪽 전경으로 뻗지만 검지가 더 정면으로 단축되어, 구도환의 아래 얼굴을 지목하는지 카메라 옆을 겨누는지 A보다 덜 명확하다. 입은 열려 있으나 눈썹과 얼굴 근육의 격한 분노는 A보다 약하다.",
        "built_space": "오른쪽에 오르간 1대, 벤치 1개와 오르간 위 켜진 라디오 1대가 보인다. 왼쪽 뒤에는 탁자 1개와 등받이가 구분되는 의자 2개, 중앙 뒤에는 책장 1개와 벽 십자가 1개가 있다. 왼쪽 탁자, 오른쪽 벽 앞, 맨 오른쪽 가장자리에 등불이 보인다. 배관과 낡은 벽, 융단은 같은 지하 기도실로 읽히며 배경 가구의 크기도 무리하지 않다. 이현우는 오른쪽 중앙의 상반신으로 잡혔지만 구도환의 머리와 어깨가 여전히 왼쪽을 넓게 가려 요구된 좁은 전경 띠를 충족하지 못한다.",
        "entities": "두 남성만 등장한다. 이현우의 젊은 동아시아계 외모, 검은 헝클어진 머리, 마른 체격, 얼굴 상처, 인이어와 피 묻은 어두운 셔츠는 설정 및 참조와 대체로 일치한다. 겉옷은 없다. 구도환의 짧은 검은 머리와 갈색 재킷은 이전 장면을 따르지만 얼굴 대부분이 가려져 정확한 인상은 확인하기 어렵다. 찰리와 불필요한 인물 또는 문자 오버레이는 없다. 다리 부상은 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "뻗은 팔과 손은 몸에 정상적으로 연결되어 있고, 상체를 앞으로 기울여 상대를 지목하는 자세는 가능하다. 발은 보이지 않지만 몸이 공중에 떠 있다는 증거는 없다. 라디오와 등불은 각각 오르간과 받침 가구에 지지되며, 벤치와 탁자도 정상적으로 놓여 있다. 지지 없는 물체나 불가능한 신체 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "구도환을 향한 삿대질과 거칠게 쏘아붙이는 분노가 더 명확하지만, 구도환의 머리와 어깨가 왼쪽 전경을 지나치게 넓게 차지한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장소·의상·상처와 대화 구도는 유지하지만, 손가락의 목표 방향과 격한 분노가 A보다 덜 선명하며 왼쪽 어깨도 요구된 좁은 띠보다 크다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 눈은 왼쪽 구도환의 얼굴을 향하고, 뻗은 검지도 화면 왼쪽 구도환의 아래 얼굴 쪽을 가리킨다. 렌즈를 직접 겨누는 동작보다는 상대를 지목하는 동작으로 읽힌다. 구도환 역시 고개를 이현우 쪽으로 돌리고 있다.",
        "built_space": "오른쪽에 오르간 1대와 그 앞 벤치 1개, 오르간 위 가장자리에 켜진 라디오 일부가 보인다. 뒤쪽에는 작은 탁자 1개, 분명히 식별되는 등받이 의자 1개, 책장 1개, 벽 십자가 1개가 있다. 탁자 쪽·중앙 받침대·오른쪽 벽 부근에 등불이 보인다. 낡은 벽, 배관, 융단과 따뜻한 조명은 이전 장소를 따른다. 두 사람은 가구 사이 열린 공간에서 마주 본다. 눈높이의 상반신 구도이지만 구도환의 머리와 등이 왼쪽 상당 부분을 차지해 '좁은 어깨 띠' 지시와 다르다.",
        "entities": "등장인물은 두 남성뿐이다. 이현우는 젊은 동아시아계 남성으로, 짧고 헝클어진 검은 머리와 마른 체격이 참조에 부합한다. 얼굴 멍과 찰과상, 검은 인이어, 피와 먼지가 묻은 어두운 셔츠가 보이며 겉옷은 없다. 구도환은 짧은 검은 머리와 이전 장면의 낡은 갈색 재킷을 유지한다. 뒤쪽 모습만 보여 정확한 얼굴 및 연령 일치는 제한적으로만 확인된다. 찰리나 다른 인물, 추가 문구는 없다. 다리 부상은 상반신 프레임 밖이다.",
        "hard_violations": [],
        "physics": "이현우의 삿대질 팔은 어깨에서 팔꿈치와 손목으로 자연스럽게 연결되며 손가락 자세도 가능한 범위다. 두 사람의 하체와 발은 프레임 밖이지만 상체는 서 있는 자세로 이어지고 부유를 나타내는 단서는 없다. 라디오는 오르간 위에, 등불과 책은 가구 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 시선은 왼쪽 구도환의 얼굴을 향하고 구도환도 이현우를 바라보는 방향이다. 삿대질하는 팔은 왼쪽 전경으로 뻗지만 검지가 더 정면으로 단축되어, 구도환의 아래 얼굴을 지목하는지 카메라 옆을 겨누는지 A보다 덜 명확하다. 입은 열려 있으나 눈썹과 얼굴 근육의 격한 분노는 A보다 약하다.",
        "built_space": "오른쪽에 오르간 1대, 벤치 1개와 오르간 위 켜진 라디오 1대가 보인다. 왼쪽 뒤에는 탁자 1개와 등받이가 구분되는 의자 2개, 중앙 뒤에는 책장 1개와 벽 십자가 1개가 있다. 왼쪽 탁자, 오른쪽 벽 앞, 맨 오른쪽 가장자리에 등불이 보인다. 배관과 낡은 벽, 융단은 같은 지하 기도실로 읽히며 배경 가구의 크기도 무리하지 않다. 이현우는 오른쪽 중앙의 상반신으로 잡혔지만 구도환의 머리와 어깨가 여전히 왼쪽을 넓게 가려 요구된 좁은 전경 띠를 충족하지 못한다.",
        "entities": "두 남성만 등장한다. 이현우의 젊은 동아시아계 외모, 검은 헝클어진 머리, 마른 체격, 얼굴 상처, 인이어와 피 묻은 어두운 셔츠는 설정 및 참조와 대체로 일치한다. 겉옷은 없다. 구도환의 짧은 검은 머리와 갈색 재킷은 이전 장면을 따르지만 얼굴 대부분이 가려져 정확한 인상은 확인하기 어렵다. 찰리와 불필요한 인물 또는 문자 오버레이는 없다. 다리 부상은 프레임 밖이므로 확인할 수 없다.",
        "hard_violations": [],
        "physics": "뻗은 팔과 손은 몸에 정상적으로 연결되어 있고, 상체를 앞으로 기울여 상대를 지목하는 자세는 가능하다. 발은 보이지 않지만 몸이 공중에 떠 있다는 증거는 없다. 라디오와 등불은 각각 오르간과 받침 가구에 지지되며, 벤치와 탁자도 정상적으로 놓여 있다. 지지 없는 물체나 불가능한 신체 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "요구된 오버더숄더 앵글과 이현우의 격양된 표정 및 삿대질 동작을 정확히 담아냈으며, 이전 샷의 배경 디테일을 훌륭하게 유지함."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "프롬프트의 주요 지시사항을 잘 수행했으나, 배경에 배치된 책장과 오르간 등의 디테일이 A에 비해 이전 샷과의 일치도가 다소 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 구도환 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S30sh10_sel.png",
    "asset_id": "c80c9cb0-5bfc-46c5-b2a5-c065c6a32755",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1302308>",
    "asset_id": "3e3943a9-b3e5-4e2c-ac87-2e35c5935244",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09e5-4267-7589-b1c5-09025984b44e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh10"
  },
  "staged_characters_added": [
   "C12"
  ]
 },
 "S31sh2::signage": {
  "fp": "adf497101939dd01",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::e8f98034a9fd2db7": {
  "subjects": [],
  "subject_text": "인천 난민촌 관리사무소 내 미연 감금실, 3층 복도·창가, 내부 계단\n문으로 닫힌 작은 방과 창문이 있는 3층 복도. 복도 구석에 소화기가 비치되어 있고, 위층으로 이어지는 내부 계단이 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L165",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::admin_holding_room": {
  "input_fingerprint": "ef316f9a06dcd9ae",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "admin_holding_room",
    "tags": [
     "S31sh10",
     "S31sh11",
     "S31sh2",
     "S34sh6"
    ]
   },
   "context_sig": "3db64cd08f9fbb6d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 관리사무소 내 미연 감금실, 3층 복도·창가, 내부 계단: 구타가 자행되는 취조실과 비상사이렌이 울리는 복도 공간이다. (특징: 얼굴이 퉁퉁 부은 미연과 심문하는 박철진; 복도 구석에 비치된 붉은 소화기; 소화기로 부서진 금속 문고리; 창문 유리를 와장창 깨부수며 복도 안으로 밀어닥치는 흙빛 해일)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그중 3층에 카메라 휙~\n- 여기저기 뒤지며 3층에 도착한 현우.\n- 황급히 문을 여는 현우. 미연을 보자마자 와락 안는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 관리사무소 내 미연 감금실, 3층 복도·창가, 내부 계단: 구타가 자행되는 취조실과 비상사이렌이 울리는 복도 공간이다. (특징: 얼굴이 퉁퉁 부은 미연과 심문하는 박철진; 복도 구석에 비치된 붉은 소화기; 소화기로 부서진 금속 문고리; 창문 유리를 와장창 깨부수며 복도 안으로 밀어닥치는 흙빛 해일)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그중 3층에 카메라 휙~\n- 여기저기 뒤지며 3층에 도착한 현우.\n- 황급히 문을 여는 현우. 미연을 보자마자 와락 안는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_holding_room_e487cb.png",
  "asset_id": "58669443-c34c-42d8-945f-aaf23c0bd9d5",
  "input_asset_ids": [
   "bc9d8c5c-56d5-4456-a26a-080f1e0b2da9"
  ],
  "origin_tag": "S31sh2",
  "place_text": "In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.",
  "origin_inputs": {
   "place_text": "In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.",
   "time_of_day_en": "night",
   "conti_asset_id": "bc9d8c5c-56d5-4456-a26a-080f1e0b2da9"
  }
 },
 "S31sh2::bgfirst_bg": {
  "input_fingerprint": "0c8af098e685bfd7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 퉁퉁 부은 얼굴을 한 미연의 맞은편 의자에 박철진이 다리를 꼬고 앉은 구도.\n\nLOCATION (lock): In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the side of the interrogation axis at seated chest height with a slight upward tilt, holding the dolly's entry composition before advancing. Place 미연 on the left and 박철진 on the right in opposing three-quarter views, showing both chairs, his newly crossed legs, and the uninterrupted gap between them. 미연 holds her swollen face toward the interrogator with guarded tension, while 박철진 settles into the chair with his attention on her rather than the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Opposing chairs (Occupied by 미연 and 박철진, facing one another) — Both chairs are seen obliquely from the same side of the confrontation axis; used as Establish the interrogation geometry while leaving the space between the occupants visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office visibly well lit as established, controlling facial highlights so 미연's swelling remains legible without adding a theatrical spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 퉁퉁 부은 얼굴을 한 미연의 맞은편 의자에 박철진이 다리를 꼬고 앉은 구도.\n\nLOCATION (lock): In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the side of the interrogation axis at seated chest height with a slight upward tilt, holding the dolly's entry composition before advancing. Place 미연 on the left and 박철진 on the right in opposing three-quarter views, showing both chairs, his newly crossed legs, and the uninterrupted gap between them. 미연 holds her swollen face toward the interrogator with guarded tension, while 박철진 settles into the chair with his attention on her rather than the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Opposing chairs (Occupied by 미연 and 박철진, facing one another) — Both chairs are seen obliquely from the same side of the confrontation axis; used as Establish the interrogation geometry while leaving the space between the occupants visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office visibly well lit as established, controlling facial highlights so 미연's swelling remains legible without adding a theatrical spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh2__bgfirst_bg.png",
  "asset_id": "3e7c3550-c6c0-4685-813b-d3a028dcdd3a",
  "input_asset_ids": [
   "bc9d8c5c-56d5-4456-a26a-080f1e0b2da9",
   "58669443-c34c-42d8-945f-aaf23c0bd9d5"
  ]
 },
 "S31sh2": {
  "input_fingerprint": "68ba50314171c43e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 퉁퉁 부은 얼굴을 한 미연의 맞은편 의자에 박철진이 다리를 꼬고 앉은 구도.\n\nLOCATION (lock): In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the side of the interrogation axis at seated chest height with a slight upward tilt, holding the dolly's entry composition before advancing. Place 미연 on the left and 박철진 on the right in opposing three-quarter views, showing both chairs, his newly crossed legs, and the uninterrupted gap between them. 미연 holds her swollen face toward the interrogator with guarded tension, while 박철진 settles into the chair with his attention on her rather than the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Opposing chairs (Occupied by 미연 and 박철진, facing one another) — Both chairs are seen obliquely from the same side of the confrontation axis; used as Establish the interrogation geometry while leaving the space between the occupants visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office visibly well lit as established, controlling facial highlights so 미연's swelling remains legible without adding a theatrical spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-office building is brightly lit, with the interrogation taking place on the third floor. 미연: Her face is badly swollen from the beating, and she remains detained in the interrogation room. 박철진: He has taken a seat for the interrogation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 퉁퉁 부은 얼굴을 한 미연의 맞은편 의자에 박철진이 다리를 꼬고 앉은 구도.\n\nLOCATION (lock): In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the side of the interrogation axis at seated chest height with a slight upward tilt, holding the dolly's entry composition before advancing. Place 미연 on the left and 박철진 on the right in opposing three-quarter views, showing both chairs, his newly crossed legs, and the uninterrupted gap between them. 미연 holds her swollen face toward the interrogator with guarded tension, while 박철진 settles into the chair with his attention on her rather than the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Opposing chairs (Occupied by 미연 and 박철진, facing one another) — Both chairs are seen obliquely from the same side of the confrontation axis; used as Establish the interrogation geometry while leaving the space between the occupants visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office visibly well lit as established, controlling facial highlights so 미연's swelling remains legible without adding a theatrical spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-office building is brightly lit, with the interrogation taking place on the third floor. 미연: Her face is badly swollen from the beating, and she remains detained in the interrogation room. 박철진: He has taken a seat for the interrogation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 퉁퉁 부은 얼굴을 한 미연의 맞은편 의자에 박철진이 다리를 꼬고 앉은 구도.\n\nLOCATION (lock): In the third-floor room of the refugee-camp administration building being used for interrogation. The building's lights are on brightly at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the side of the interrogation axis at seated chest height with a slight upward tilt, holding the dolly's entry composition before advancing. Place 미연 on the left and 박철진 on the right in opposing three-quarter views, showing both chairs, his newly crossed legs, and the uninterrupted gap between them. 미연 holds her swollen face toward the interrogator with guarded tension, while 박철진 settles into the chair with his attention on her rather than the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Opposing chairs (Occupied by 미연 and 박철진, facing one another) — Both chairs are seen obliquely from the same side of the confrontation axis; used as Establish the interrogation geometry while leaving the space between the occupants visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office visibly well lit as established, controlling facial highlights so 미연's swelling remains legible without adding a theatrical spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management-office building is brightly lit, with the interrogation taking place on the third floor. 미연: Her face is badly swollen from the beating, and she remains detained in the interrogation room. 박철진: He has taken a seat for the interrogation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh2__bgfirst_bg.png",
     "asset_id": "3e7c3550-c6c0-4685-813b-d3a028dcdd3a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S31sh2.png",
     "asset_id": "bc9d8c5c-56d5-4456-a26a-080f1e0b2da9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_holding_room_e487cb.png",
     "asset_id": "58669443-c34c-42d8-945f-aaf23c0bd9d5",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연과 박철진은 책상을 사이에 두고 서로를 정면으로 응시하고 있음.",
    "built_space": "레퍼런스와 동일한 취조실 공간으로, 중앙의 책상, 우측의 4단 파일 캐비닛, 좌측의 문, 창밖의 야경, 천장 조명 등 모든 고정 요소가 정확한 위치에 올바른 수량으로 배치됨.",
    "entities": "미연은 부은 얼굴과 낡은 회색 셔츠, 작업 바지를 착용한 40대 여성으로 정확히 묘사됨. 박철진은 남색 전투복, 붉은 민병대 완장, 모자를 착용한 40대 남성으로 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 사람 모두 의자에 안정적으로 무게를 싣고 앉아 있으며, 박철진이 꼬고 있는 다리도 해부학적으로 자연스럽게 다른 다리 위에 얹혀 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
   },
   {
    "label": "B",
    "direction": "미연과 박철진이 서로의 얼굴을 마주 보며 시선을 고정하고 있음.",
    "built_space": "위치 레퍼런스의 사무실 배경(책상, 캐비닛, 문, 창문)을 정확하게 따르고 있으며, 인물들이 그 사이의 지정된 의자에 알맞게 위치함.",
    "entities": "미연의 얼굴 붓기와 의상, 박철진의 전투복과 완장 모두 프롬프트 및 레퍼런스와 일치하나, 미연의 무릎 위에 놓인 양손의 손가락들이 부자연스럽게 융합되어 있음.",
    "hard_violations": [],
    "physics": "인물들은 의자에 정상적으로 앉아 있으나, 미연의 손가락 형태가 물리적으로 불가능하게 녹아내린 듯 겹쳐 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "제시된 프롬프트의 구도, 인물의 의상 및 상태(부은 얼굴, 꼬은 다리), 그리고 레퍼런스의 배경을 완벽하게 사실적으로 재현함."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전반적인 구도와 배경은 레퍼런스를 잘 따랐으나, 미연의 손과 박철진의 완장 십자가 형태 등 세부 묘사에서 형태가 무너진 아티팩트가 존재함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연과 박철진은 책상을 사이에 두고 서로를 정면으로 응시하고 있음.",
        "built_space": "레퍼런스와 동일한 취조실 공간으로, 중앙의 책상, 우측의 4단 파일 캐비닛, 좌측의 문, 창밖의 야경, 천장 조명 등 모든 고정 요소가 정확한 위치에 올바른 수량으로 배치됨.",
        "entities": "미연은 부은 얼굴과 낡은 회색 셔츠, 작업 바지를 착용한 40대 여성으로 정확히 묘사됨. 박철진은 남색 전투복, 붉은 민병대 완장, 모자를 착용한 40대 남성으로 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 의자에 안정적으로 무게를 싣고 앉아 있으며, 박철진이 꼬고 있는 다리도 해부학적으로 자연스럽게 다른 다리 위에 얹혀 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "미연과 박철진이 서로의 얼굴을 마주 보며 시선을 고정하고 있음.",
        "built_space": "위치 레퍼런스의 사무실 배경(책상, 캐비닛, 문, 창문)을 정확하게 따르고 있으며, 인물들이 그 사이의 지정된 의자에 알맞게 위치함.",
        "entities": "미연의 얼굴 붓기와 의상, 박철진의 전투복과 완장 모두 프롬프트 및 레퍼런스와 일치하나, 미연의 무릎 위에 놓인 양손의 손가락들이 부자연스럽게 융합되어 있음.",
        "hard_violations": [],
        "physics": "인물들은 의자에 정상적으로 앉아 있으나, 미연의 손가락 형태가 물리적으로 불가능하게 녹아내린 듯 겹쳐 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "제시된 프롬프트의 구도, 인물의 의상 및 상태(부은 얼굴, 꼬은 다리), 그리고 레퍼런스의 배경을 완벽하게 사실적으로 재현함."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전반적인 구도와 배경은 레퍼런스를 잘 따랐으나, 미연의 손과 박철진의 완장 십자가 형태 등 세부 묘사에서 형태가 무너진 아티팩트가 존재함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연과 박철진은 책상을 사이에 두고 서로를 정면으로 응시하고 있음.",
        "built_space": "레퍼런스와 동일한 취조실 공간으로, 중앙의 책상, 우측의 4단 파일 캐비닛, 좌측의 문, 창밖의 야경, 천장 조명 등 모든 고정 요소가 정확한 위치에 올바른 수량으로 배치됨.",
        "entities": "미연은 부은 얼굴과 낡은 회색 셔츠, 작업 바지를 착용한 40대 여성으로 정확히 묘사됨. 박철진은 남색 전투복, 붉은 민병대 완장, 모자를 착용한 40대 남성으로 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 의자에 안정적으로 무게를 싣고 앉아 있으며, 박철진이 꼬고 있는 다리도 해부학적으로 자연스럽게 다른 다리 위에 얹혀 있음. 중력을 거스르거나 지지대가 없는 물체는 없음."
       },
       {
        "label": "B",
        "direction": "미연과 박철진이 서로의 얼굴을 마주 보며 시선을 고정하고 있음.",
        "built_space": "위치 레퍼런스의 사무실 배경(책상, 캐비닛, 문, 창문)을 정확하게 따르고 있으며, 인물들이 그 사이의 지정된 의자에 알맞게 위치함.",
        "entities": "미연의 얼굴 붓기와 의상, 박철진의 전투복과 완장 모두 프롬프트 및 레퍼런스와 일치하나, 미연의 무릎 위에 놓인 양손의 손가락들이 부자연스럽게 융합되어 있음.",
        "hard_violations": [],
        "physics": "인물들은 의자에 정상적으로 앉아 있으나, 미연의 손가락 형태가 물리적으로 불가능하게 녹아내린 듯 겹쳐 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 의자와 빈 간격을 보이는 와이드 구도, 상호 시선과 다리 꼬기를 충족하며, B보다 미연의 부기와 멍 및 비스듬한 얼굴 각도가 요구에 가깝다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "장소와 좌우 배치, 다리를 꼰 착석은 충실하지만 미연이 더 측면으로 보이고 얼굴의 심한 부기가 약해 핵심 순간의 재현에서 A에 뒤진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 미연은 얼굴과 시선을 오른쪽 박철진에게 향하고, 오른쪽 박철진도 왼쪽 미연의 얼굴을 바라본다. 두 사람 모두 카메라를 응시하지 않는다. 의자와 몸통이 서로 마주하며 박철진의 위로 올린 신발은 중앙 빈 공간 쪽으로 뻗는다.",
        "built_space": "전경의 대향 의자 두 개, 뒤쪽 책상 한 개와 사무용 회전의자 한 개가 보인다. 왼쪽 문 한 개, 중앙의 두 짝 창 한 개, 천장 선형등 두 개와 냉방기 한 개, 감지기 한 개, 오른쪽 게시판 한 개가 장소 사진과 같은 위치에 있다. 오른쪽 서류함은 인물에 일부 가려져 두 열 전체를 확인하기 어렵다. 두 사람은 각각 전경 의자에 앉고 그 사이 전경에는 장애물이 없다. 밤의 창밖과 밝은 실내, 투톤 벽과 타일 바닥도 일치한다. 두 의자 전체를 담은 와이드 구도이며, 낮은 카메라에서 천장까지 보이지만 상향 틸트는 두드러지지 않는다.",
        "entities": "인물은 지정된 두 명뿐이다. 미연은 중년 동아시아계 여성으로 검은 단발과 참고 얼굴에 가까운 외형, 먼지 묻고 해진 회색 셔츠와 어두운 작업 바지를 갖췄다. 눈 아래와 광대의 붉은 멍 및 부기가 보이지만 얼굴 전체가 퉁퉁 부은 정도는 다소 절제됐다. 박철진은 중년 동아시아계 남성으로 참고와 유사한 얼굴, 남색 전투복, 흰 표식이 있는 붉은 완장과 참고의 모자를 착용했다. 머리카락은 모자에 대부분 가려졌다. 책상 위 서류와 필기구, 뒤쪽 파일류도 장소 참고에 대응하며 추가 인물이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 각 의자 좌판에 놓이고 등은 등받이 방향을 향한다. 미연의 두 신발은 바닥에 닿고 손은 무릎 위에 놓인다. 박철진은 한쪽 발을 바닥에 딛고 반대쪽 다리를 지지 다리 위에 걸쳐 꼬았다. 들어 올린 발은 연결된 다리와 무릎의 교차로 지지되므로 부유가 아니다. 의자와 책상 다리는 바닥에 닿으며 물리적으로 불가능한 지지 관계는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "미연은 오른쪽 박철진을, 박철진은 왼쪽 미연을 바라본다. 시선의 대상은 서로이며 카메라가 아니다. 두 의자는 안쪽을 향하고 박철진의 꼰 다리와 들린 신발은 중앙으로 뻗는다. 미연의 얼굴은 요청한 비스듬한 삼사분면보다 옆얼굴에 더 가깝다.",
        "built_space": "전경의 의자 두 개와 뒤쪽 책상 한 개, 회전의자 한 개가 보이고 두 인물 사이 공간이 비어 있다. 왼쪽 문 한 개, 중앙 두 짝 창 한 개, 천장 선형등 두 개와 냉방기 한 개, 감지기 한 개, 오른쪽 게시판 한 개가 참고 장소의 배치를 따른다. 오른쪽 서류함 일부는 박철진에게 가려진다. 투톤 벽, 바닥 타일, 야간 창밖과 밝은 실내가 유지된다. 두 의자와 전신을 담는 와이드 숏이며 A와 거의 같은 카메라 위치로 보인다.",
        "entities": "지정된 중년 여성 미연과 중년 남성 박철진만 보인다. 두 사람의 동아시아계 외형과 검은 머리, 얼굴 및 체격은 참고와 대체로 맞는다. 미연의 해진 회색 셔츠, 먼지 묻은 어두운 바지와 작업화가 일치한다. 눈 아래 붉은 상처는 있지만 볼과 턱 윤곽의 심한 부기는 A보다 덜 읽힌다. 박철진은 참고와 같은 계열의 남색 전투복과 모자, 흰 표식이 있는 붉은 완장을 착용했다. 책상 위 서류와 필기구 및 배경 파일류가 보이며 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "미연은 좌판에 앉아 두 발을 바닥에 두고 손을 무릎 위에 모았다. 박철진의 엉덩이와 등은 의자에 지지되며 한 발은 바닥에 닿아 있다. 반대 다리는 아래쪽 다리 위에 교차해 얹혀 있고 신발만 공중에 있어 정상적인 다리 꼬기 자세로 성립한다. 두 의자와 책상은 바닥에 지지되고 떠 있는 독립 물체나 불가능한 관절 연결은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "두 의자와 빈 간격을 보이는 와이드 구도, 상호 시선과 다리 꼬기를 충족하며, B보다 미연의 부기와 멍 및 비스듬한 얼굴 각도가 요구에 가깝다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "장소와 좌우 배치, 다리를 꼰 착석은 충실하지만 미연이 더 측면으로 보이고 얼굴의 심한 부기가 약해 핵심 순간의 재현에서 A에 뒤진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 미연은 얼굴과 시선을 오른쪽 박철진에게 향하고, 오른쪽 박철진도 왼쪽 미연의 얼굴을 바라본다. 두 사람 모두 카메라를 응시하지 않는다. 의자와 몸통이 서로 마주하며 박철진의 위로 올린 신발은 중앙 빈 공간 쪽으로 뻗는다.",
        "built_space": "전경의 대향 의자 두 개, 뒤쪽 책상 한 개와 사무용 회전의자 한 개가 보인다. 왼쪽 문 한 개, 중앙의 두 짝 창 한 개, 천장 선형등 두 개와 냉방기 한 개, 감지기 한 개, 오른쪽 게시판 한 개가 장소 사진과 같은 위치에 있다. 오른쪽 서류함은 인물에 일부 가려져 두 열 전체를 확인하기 어렵다. 두 사람은 각각 전경 의자에 앉고 그 사이 전경에는 장애물이 없다. 밤의 창밖과 밝은 실내, 투톤 벽과 타일 바닥도 일치한다. 두 의자 전체를 담은 와이드 구도이며, 낮은 카메라에서 천장까지 보이지만 상향 틸트는 두드러지지 않는다.",
        "entities": "인물은 지정된 두 명뿐이다. 미연은 중년 동아시아계 여성으로 검은 단발과 참고 얼굴에 가까운 외형, 먼지 묻고 해진 회색 셔츠와 어두운 작업 바지를 갖췄다. 눈 아래와 광대의 붉은 멍 및 부기가 보이지만 얼굴 전체가 퉁퉁 부은 정도는 다소 절제됐다. 박철진은 중년 동아시아계 남성으로 참고와 유사한 얼굴, 남색 전투복, 흰 표식이 있는 붉은 완장과 참고의 모자를 착용했다. 머리카락은 모자에 대부분 가려졌다. 책상 위 서류와 필기구, 뒤쪽 파일류도 장소 참고에 대응하며 추가 인물이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 엉덩이는 각 의자 좌판에 놓이고 등은 등받이 방향을 향한다. 미연의 두 신발은 바닥에 닿고 손은 무릎 위에 놓인다. 박철진은 한쪽 발을 바닥에 딛고 반대쪽 다리를 지지 다리 위에 걸쳐 꼬았다. 들어 올린 발은 연결된 다리와 무릎의 교차로 지지되므로 부유가 아니다. 의자와 책상 다리는 바닥에 닿으며 물리적으로 불가능한 지지 관계는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "미연은 오른쪽 박철진을, 박철진은 왼쪽 미연을 바라본다. 시선의 대상은 서로이며 카메라가 아니다. 두 의자는 안쪽을 향하고 박철진의 꼰 다리와 들린 신발은 중앙으로 뻗는다. 미연의 얼굴은 요청한 비스듬한 삼사분면보다 옆얼굴에 더 가깝다.",
        "built_space": "전경의 의자 두 개와 뒤쪽 책상 한 개, 회전의자 한 개가 보이고 두 인물 사이 공간이 비어 있다. 왼쪽 문 한 개, 중앙 두 짝 창 한 개, 천장 선형등 두 개와 냉방기 한 개, 감지기 한 개, 오른쪽 게시판 한 개가 참고 장소의 배치를 따른다. 오른쪽 서류함 일부는 박철진에게 가려진다. 투톤 벽, 바닥 타일, 야간 창밖과 밝은 실내가 유지된다. 두 의자와 전신을 담는 와이드 숏이며 A와 거의 같은 카메라 위치로 보인다.",
        "entities": "지정된 중년 여성 미연과 중년 남성 박철진만 보인다. 두 사람의 동아시아계 외형과 검은 머리, 얼굴 및 체격은 참고와 대체로 맞는다. 미연의 해진 회색 셔츠, 먼지 묻은 어두운 바지와 작업화가 일치한다. 눈 아래 붉은 상처는 있지만 볼과 턱 윤곽의 심한 부기는 A보다 덜 읽힌다. 박철진은 참고와 같은 계열의 남색 전투복과 모자, 흰 표식이 있는 붉은 완장을 착용했다. 책상 위 서류와 필기구 및 배경 파일류가 보이며 추가 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "미연은 좌판에 앉아 두 발을 바닥에 두고 손을 무릎 위에 모았다. 박철진의 엉덩이와 등은 의자에 지지되며 한 발은 바닥에 닿아 있다. 반대 다리는 아래쪽 다리 위에 교차해 얹혀 있고 신발만 공중에 있어 정상적인 다리 꼬기 자세로 성립한다. 두 의자와 책상은 바닥에 지지되고 떠 있는 독립 물체나 불가능한 관절 연결은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.8
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1800
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "제시된 프롬프트의 구도, 인물의 의상 및 상태(부은 얼굴, 꼬은 다리), 그리고 레퍼런스의 배경을 완벽하게 사실적으로 재현함."
   },
   {
    "label": "B",
    "score": 1800,
    "verdict_ko": "전반적인 구도와 배경은 레퍼런스를 잘 따랐으나, 미연의 손과 박철진의 완장 십자가 형태 등 세부 묘사에서 형태가 무너진 아티팩트가 존재함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_holding_room_e487cb.png",
    "asset_id": "58669443-c34c-42d8-945f-aaf23c0bd9d5",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09ea-4abf-7077-acaf-2a278d0bdc12",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh2__bgfirst_bg.png",
   "bg_asset_id": "3e7c3550-c6c0-4685-813b-d3a028dcdd3a",
   "bg_record_key": "S31sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "admin_holding_room",
   "groupbg_asset_id": "58669443-c34c-42d8-945f-aaf23c0bd9d5"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S31sh10::signage": {
  "fp": "3342490127efc6b3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh10": {
  "input_fingerprint": "4eff105a833cd374",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 박철진의 귀에 바짝 입을 댄 채 귓속말을 전하는 구도.\n\nLOCATION (lock): Beside the interrogator's chair in the brightly lit third-floor detention room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the established side into a close observation just above 박철진's seated eye line, looking slightly downward across 박철진의 수하 1's near shoulder. Place the subordinate's bent head at the left edge and 박철진's ear and oblique face at right-center, keeping the whispering mouth beside—not hidden behind—the ear. 박철진의 수하 1 concentrates on delivering the message at the ear with his gaze lowered, while 박철진 stills his head to listen and keeps his eyes directed toward 미연 beyond the crop; reduced camera distance supplies the emphasis.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the office's established bright illumination with controlled contrast and no change of source or color for the private exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the interrogation room's chairs, interior surfaces, and nighttime lighting from the reference. Exclude church furnishings and any exterior construction equipment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains inside the brightly lit management office; its door has been opened for the incoming report. 박철진: He remains seated for the interrogation. 박철진의 수하 1: He has entered and leans in to deliver a quiet report, retaining his militia uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 박철진의 귀에 바짝 입을 댄 채 귓속말을 전하는 구도.\n\nLOCATION (lock): Beside the interrogator's chair in the brightly lit third-floor detention room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the established side into a close observation just above 박철진's seated eye line, looking slightly downward across 박철진의 수하 1's near shoulder. Place the subordinate's bent head at the left edge and 박철진's ear and oblique face at right-center, keeping the whispering mouth beside—not hidden behind—the ear. 박철진의 수하 1 concentrates on delivering the message at the ear with his gaze lowered, while 박철진 stills his head to listen and keeps his eyes directed toward 미연 beyond the crop; reduced camera distance supplies the emphasis.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the office's established bright illumination with controlled contrast and no change of source or color for the private exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the interrogation room's chairs, interior surfaces, and nighttime lighting from the reference. Exclude church furnishings and any exterior construction equipment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains inside the brightly lit management office; its door has been opened for the incoming report. 박철진: He remains seated for the interrogation. 박철진의 수하 1: He has entered and leans in to deliver a quiet report, retaining his militia uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진의 수하 1이 박철진의 귀에 바짝 입을 댄 채 귓속말을 전하는 구도.\n\nLOCATION (lock): Beside the interrogator's chair in the brightly lit third-floor detention room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue along the established side into a close observation just above 박철진's seated eye line, looking slightly downward across 박철진의 수하 1's near shoulder. Place the subordinate's bent head at the left edge and 박철진's ear and oblique face at right-center, keeping the whispering mouth beside—not hidden behind—the ear. 박철진의 수하 1 concentrates on delivering the message at the ear with his gaze lowered, while 박철진 stills his head to listen and keeps his eyes directed toward 미연 beyond the crop; reduced camera distance supplies the emphasis.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the office's established bright illumination with controlled contrast and no change of source or color for the private exchange.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the interrogation room's chairs, interior surfaces, and nighttime lighting from the reference. Exclude church furnishings and any exterior construction equipment.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor interrogation room remains inside the brightly lit management office; its door has been opened for the incoming report. 박철진: He remains seated for the interrogation. 박철진의 수하 1: He has entered and leans in to deliver a quiet report, retaining his militia uniform and armband.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 박철진의 수하 1 (한국인 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수하 1은 시선을 아래로 향한 채 박철진의 귀에 대고 속삭이고 있으며, 박철진은 프레임 좌측 밖(미연의 위치)을 응시하고 있음.",
    "built_space": "카메라 앵글에 맞게 배경에 이전 샷에서 확립된 넓은 가로형 창문, 외부 조명, 책상 일부가 올바른 위치와 비례로 나타남.",
    "entities": "박철진과 수하 1 모두 제공된 레퍼런스의 외모, 복장(어두운 남색 전투복, 모자)과 일치함. 완장에 십자가 마크가 추가되었으나 '민병대 완장'이라는 설정 및 이전 샷의 묘사와 부합함.",
    "hard_violations": [],
    "physics": "수하 1은 하체의 지지를 바탕으로 상체를 안정적으로 숙이고 있으며, 박철진은 의자에 앉아 있는 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "수하 1은 고개를 숙여 박철진의 귀를 향하고 있으며, 박철진은 프레임 좌측 밖을 응시함.",
    "built_space": "배경 우측에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 잘못 생성됨.",
    "entities": "두 인물 모두 레퍼런스의 외모와 복장 요소를 대체로 잘 반영하였음.",
    "hard_violations": [
     "[gemini-pro] 이전 샷의 공간 구조에 없는 창문 구조물이 배경에 임의로 발명되어 생성됨 (공간 연속성 위반)."
    ],
    "physics": "수하 1이 몸을 굽힌 자세와 박철진이 앉아있는 자세 모두 물리적 지지 기반이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트에 명시된 밀착된 카메라 구도를 완벽히 구현했으며, 이전 샷에서 확립된 실내 구조(가로형 창문, 책상 등)를 배경에 정확하게 유지하여 매우 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물들의 배치와 행동은 프롬프트를 잘 따랐으나, 배경에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 나타나 공간의 연속성을 크게 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하 1은 시선을 아래로 향한 채 박철진의 귀에 대고 속삭이고 있으며, 박철진은 프레임 좌측 밖(미연의 위치)을 응시하고 있음.",
        "built_space": "카메라 앵글에 맞게 배경에 이전 샷에서 확립된 넓은 가로형 창문, 외부 조명, 책상 일부가 올바른 위치와 비례로 나타남.",
        "entities": "박철진과 수하 1 모두 제공된 레퍼런스의 외모, 복장(어두운 남색 전투복, 모자)과 일치함. 완장에 십자가 마크가 추가되었으나 '민병대 완장'이라는 설정 및 이전 샷의 묘사와 부합함.",
        "hard_violations": [],
        "physics": "수하 1은 하체의 지지를 바탕으로 상체를 안정적으로 숙이고 있으며, 박철진은 의자에 앉아 있는 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "수하 1은 고개를 숙여 박철진의 귀를 향하고 있으며, 박철진은 프레임 좌측 밖을 응시함.",
        "built_space": "배경 우측에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 잘못 생성됨.",
        "entities": "두 인물 모두 레퍼런스의 외모와 복장 요소를 대체로 잘 반영하였음.",
        "hard_violations": [
         "이전 샷의 공간 구조에 없는 창문 구조물이 배경에 임의로 발명되어 생성됨 (공간 연속성 위반)."
        ],
        "physics": "수하 1이 몸을 굽힌 자세와 박철진이 앉아있는 자세 모두 물리적 지지 기반이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트에 명시된 밀착된 카메라 구도를 완벽히 구현했으며, 이전 샷에서 확립된 실내 구조(가로형 창문, 책상 등)를 배경에 정확하게 유지하여 매우 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물들의 배치와 행동은 프롬프트를 잘 따랐으나, 배경에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 나타나 공간의 연속성을 크게 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수하 1은 시선을 아래로 향한 채 박철진의 귀에 대고 속삭이고 있으며, 박철진은 프레임 좌측 밖(미연의 위치)을 응시하고 있음.",
        "built_space": "카메라 앵글에 맞게 배경에 이전 샷에서 확립된 넓은 가로형 창문, 외부 조명, 책상 일부가 올바른 위치와 비례로 나타남.",
        "entities": "박철진과 수하 1 모두 제공된 레퍼런스의 외모, 복장(어두운 남색 전투복, 모자)과 일치함. 완장에 십자가 마크가 추가되었으나 '민병대 완장'이라는 설정 및 이전 샷의 묘사와 부합함.",
        "hard_violations": [],
        "physics": "수하 1은 하체의 지지를 바탕으로 상체를 안정적으로 숙이고 있으며, 박철진은 의자에 앉아 있는 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "수하 1은 고개를 숙여 박철진의 귀를 향하고 있으며, 박철진은 프레임 좌측 밖을 응시함.",
        "built_space": "배경 우측에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 잘못 생성됨.",
        "entities": "두 인물 모두 레퍼런스의 외모와 복장 요소를 대체로 잘 반영하였음.",
        "hard_violations": [
         "이전 샷의 공간 구조에 없는 창문 구조물이 배경에 임의로 발명되어 생성됨 (공간 연속성 위반)."
        ],
        "physics": "수하 1이 몸을 굽힌 자세와 박철진이 앉아있는 자세 모두 물리적 지지 기반이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 야간 실내 분위기는 맞지만, 수하의 입이 귀보다 관자놀이 앞에 있고 박철진이 수하를 올려다봐 귓속말의 위치와 청취 시선 지시를 놓칩니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "수하의 가까운 어깨를 넘는 구도와 박철진의 수평에 가까운 시선은 더 적합하지만, 입과 귀의 밀착이 불명확하고 열린 문 및 수하의 완장 형태가 맞지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수하는 고개와 눈을 아래로 향하고 박철진의 화면 왼쪽 관자놀이 쪽으로 입을 내밉니다. 입은 모자챙 앞에 드러나지만 바로 옆의 귀는 확인되지 않아 귀에 바짝 대는 관계가 성립했다고 보기 어렵습니다. 박철진의 눈은 왼쪽 위 수하의 얼굴을 향하며, 화면 밖 미연을 계속 보는 지시와 다릅니다.",
        "built_space": "오른쪽 배경에 야간 창문 일부 한 곳과 상단 블라인드, 회색 계열 벽, 아래쪽 가구 표면 일부가 보입니다. 참고 공간의 재질과 밝은 실내 조명은 대체로 이어집니다. 문과 의자는 구도 밖이므로 개방 여부와 좌석 접촉은 확인할 수 없습니다. 창의 희미한 실내등 반사에 명백한 광학적 모순은 없습니다. 수하의 어깨 너머보다는 두 사람을 앞에서 함께 보는 성격이 강합니다.",
        "entities": "성인 동아시아계 남성 두 명만 보이고 추가 인물이나 문자는 없습니다. 박철진의 중년 얼굴, 남색 챙모자와 턱끈, 낡은 남색 전투복, 붉은 완장은 참고와 대체로 맞습니다. 수하의 젊은 얼굴, 검은 비니와 짧게 드러난 검은 머리, 남색 복장도 참고에 가깝습니다. 다만 수하에게 참고의 매듭형 붉은 천 대신 흰 문양이 있는 넓은 완장을 주었고 착용 팔도 달라 보입니다.",
        "hard_violations": [],
        "physics": "수하의 목과 몸통은 연결된 상태로 앞으로 숙여져 있으며 허리 아래는 화면 밖입니다. 박철진은 몸통을 세운 청취 자세로 보이지만 좌판과 엉덩이 접촉은 잘려 있습니다. 이 크롭만으로 공중 부양이나 지지 불능을 판단할 근거는 없습니다. 모자와 완장은 각각 머리와 팔에 지지되며, 별도로 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "수하는 시선을 내리고 박철진의 화면 왼쪽 머리 옆으로 몸을 기울입니다. 입술이 모자챙 아래에 보이지만 귀와 나란히 밀착한 모습보다는 관자놀이와 볼 앞쪽에 가까워 보이며, 목표 귀는 가려져 있습니다. 박철진의 눈은 거의 수평으로 화면 왼쪽을 향해 A보다 화면 밖 상대를 보는 방향에 가깝지만, 수하 쪽 곁눈질로도 읽혀 미연을 계속 주시한다고 단정하기 어렵습니다.",
        "built_space": "왼쪽 뒤에 좁은 유리창이 달린 문 한 개, 오른쪽 뒤에 야간 창문 한 곳과 블라인드, 중앙 뒤에 의자 등받이 한 개, 오른쪽 아래에 책상 일부와 작은 사무용품, 위쪽에 천장등 일부가 보입니다. 고정물의 재질과 배경 크기는 참고 사무실에 대체로 맞고 중복된 설비는 없습니다. 다만 문은 열린 상태가 아니라 닫힌 것으로 보입니다. 수하의 가까운 어깨가 왼쪽 전경을 채워 지정한 어깨 너머 관찰 구도에는 A보다 가깝습니다.",
        "entities": "보이는 사람은 지정된 두 남성뿐입니다. 박철진의 중년 얼굴, 남색 챙모자와 턱끈, 전투복과 붉은 완장이 참고에 가깝습니다. 수하의 성인 동아시아계 얼굴, 비니, 검은 머리와 남색 전투복도 대체로 맞습니다. 수하의 완장은 참고의 매듭형 붉은 천과 달리 큰 흰 십자형 문양이 있는 넓은 띠로 바뀌었습니다. 미연이나 다른 인물, 추가 문구는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "수하의 어깨와 몸통이 자연스럽게 앞으로 기울고 목이 그 기울기를 이어받습니다. 하체와 발은 크롭 밖이라 지면 접촉은 확인되지 않지만 몸이 떠 있다는 징후는 없습니다. 박철진의 앉은 몸통 자세는 가능하며 실제 좌판 접촉은 보이지 않습니다. 모자, 턱끈과 완장은 신체에 걸리거나 밀착되어 있고 지지 없이 떠 있는 소품은 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 야간 실내 분위기는 맞지만, 수하의 입이 귀보다 관자놀이 앞에 있고 박철진이 수하를 올려다봐 귓속말의 위치와 청취 시선 지시를 놓칩니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "수하의 가까운 어깨를 넘는 구도와 박철진의 수평에 가까운 시선은 더 적합하지만, 입과 귀의 밀착이 불명확하고 열린 문 및 수하의 완장 형태가 맞지 않습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "수하는 고개와 눈을 아래로 향하고 박철진의 화면 왼쪽 관자놀이 쪽으로 입을 내밉니다. 입은 모자챙 앞에 드러나지만 바로 옆의 귀는 확인되지 않아 귀에 바짝 대는 관계가 성립했다고 보기 어렵습니다. 박철진의 눈은 왼쪽 위 수하의 얼굴을 향하며, 화면 밖 미연을 계속 보는 지시와 다릅니다.",
        "built_space": "오른쪽 배경에 야간 창문 일부 한 곳과 상단 블라인드, 회색 계열 벽, 아래쪽 가구 표면 일부가 보입니다. 참고 공간의 재질과 밝은 실내 조명은 대체로 이어집니다. 문과 의자는 구도 밖이므로 개방 여부와 좌석 접촉은 확인할 수 없습니다. 창의 희미한 실내등 반사에 명백한 광학적 모순은 없습니다. 수하의 어깨 너머보다는 두 사람을 앞에서 함께 보는 성격이 강합니다.",
        "entities": "성인 동아시아계 남성 두 명만 보이고 추가 인물이나 문자는 없습니다. 박철진의 중년 얼굴, 남색 챙모자와 턱끈, 낡은 남색 전투복, 붉은 완장은 참고와 대체로 맞습니다. 수하의 젊은 얼굴, 검은 비니와 짧게 드러난 검은 머리, 남색 복장도 참고에 가깝습니다. 다만 수하에게 참고의 매듭형 붉은 천 대신 흰 문양이 있는 넓은 완장을 주었고 착용 팔도 달라 보입니다.",
        "hard_violations": [],
        "physics": "수하의 목과 몸통은 연결된 상태로 앞으로 숙여져 있으며 허리 아래는 화면 밖입니다. 박철진은 몸통을 세운 청취 자세로 보이지만 좌판과 엉덩이 접촉은 잘려 있습니다. 이 크롭만으로 공중 부양이나 지지 불능을 판단할 근거는 없습니다. 모자와 완장은 각각 머리와 팔에 지지되며, 별도로 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "수하는 시선을 내리고 박철진의 화면 왼쪽 머리 옆으로 몸을 기울입니다. 입술이 모자챙 아래에 보이지만 귀와 나란히 밀착한 모습보다는 관자놀이와 볼 앞쪽에 가까워 보이며, 목표 귀는 가려져 있습니다. 박철진의 눈은 거의 수평으로 화면 왼쪽을 향해 A보다 화면 밖 상대를 보는 방향에 가깝지만, 수하 쪽 곁눈질로도 읽혀 미연을 계속 주시한다고 단정하기 어렵습니다.",
        "built_space": "왼쪽 뒤에 좁은 유리창이 달린 문 한 개, 오른쪽 뒤에 야간 창문 한 곳과 블라인드, 중앙 뒤에 의자 등받이 한 개, 오른쪽 아래에 책상 일부와 작은 사무용품, 위쪽에 천장등 일부가 보입니다. 고정물의 재질과 배경 크기는 참고 사무실에 대체로 맞고 중복된 설비는 없습니다. 다만 문은 열린 상태가 아니라 닫힌 것으로 보입니다. 수하의 가까운 어깨가 왼쪽 전경을 채워 지정한 어깨 너머 관찰 구도에는 A보다 가깝습니다.",
        "entities": "보이는 사람은 지정된 두 남성뿐입니다. 박철진의 중년 얼굴, 남색 챙모자와 턱끈, 전투복과 붉은 완장이 참고에 가깝습니다. 수하의 성인 동아시아계 얼굴, 비니, 검은 머리와 남색 전투복도 대체로 맞습니다. 수하의 완장은 참고의 매듭형 붉은 천과 달리 큰 흰 십자형 문양이 있는 넓은 띠로 바뀌었습니다. 미연이나 다른 인물, 추가 문구는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "수하의 어깨와 몸통이 자연스럽게 앞으로 기울고 목이 그 기울기를 이어받습니다. 하체와 발은 크롭 밖이라 지면 접촉은 확인되지 않지만 몸이 떠 있다는 징후는 없습니다. 박철진의 앉은 몸통 자세는 가능하며 실제 좌판 접촉은 보이지 않습니다. 모자, 턱끈과 완장은 신체에 걸리거나 밀착되어 있고 지지 없이 떠 있는 소품은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.5
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷의 공간 구조에 없는 창문 구조물이 배경에 임의로 발명되어 생성됨 (공간 연속성 위반)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트에 명시된 밀착된 카메라 구도를 완벽히 구현했으며, 이전 샷에서 확립된 실내 구조(가로형 창문, 책상 등)를 배경에 정확하게 유지하여 매우 훌륭한 결과물입니다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "인물들의 배치와 행동은 프롬프트를 잘 따랐으나, 배경에 이전 샷의 세트 구조에 존재하지 않는 세로형 창문이 나타나 공간의 연속성을 크게 훼손했습니다.  ★위반: [gemini-pro] 이전 샷의 공간 구조에 없는 창문 구조물이 배경에 임의로 발명되어 생성됨 (공간 연속성 위반)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh2_sel.png",
    "asset_id": "be7c227b-686c-438e-afbe-a2905c8bbcbd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진의 수하 1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:728720>",
    "asset_id": "445f3c52-b87c-457e-ae88-317bd9f1202e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09f2-8a36-72d8-a310-3033d08bf552",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh2"
  }
 },
 "S31sh11::signage": {
  "fp": "b8474450644ed200",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh11": {
  "input_fingerprint": "b07dce48d76de6b3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 크게 놀란 표정으로 굳어 있는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the questioning position inside the brightly lit third-floor room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease past the subordinate's shoulder along the existing inward path and settle slightly above 박철진's seated eye line, preserving an oblique three-quarter close-up. His face sits right of center with look room on the left toward 박철진의 수하 1 outside the frame; his eyes widen and his seated body arrests as the report registers. Keep the lighting and background treatment continuous, making the turn of his attention toward the subordinate the emphasized change.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding office illumination and exposure so the widened eyes, not a lighting effect, announce the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same detention-room materials, chairs, and lighting from the reference. Exclude the earlier assault as a repeated event and any furnishings from the church basement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management office remains brightly lit, and the third-floor interrogation-room door has been opened. 박철진: He remains at the interrogation seat, suddenly alarmed by the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 크게 놀란 표정으로 굳어 있는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the questioning position inside the brightly lit third-floor room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease past the subordinate's shoulder along the existing inward path and settle slightly above 박철진's seated eye line, preserving an oblique three-quarter close-up. His face sits right of center with look room on the left toward 박철진의 수하 1 outside the frame; his eyes widen and his seated body arrests as the report registers. Keep the lighting and background treatment continuous, making the turn of his attention toward the subordinate the emphasized change.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding office illumination and exposure so the widened eyes, not a lighting effect, announce the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same detention-room materials, chairs, and lighting from the reference. Exclude the earlier assault as a repeated event and any furnishings from the church basement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management office remains brightly lit, and the third-floor interrogation-room door has been opened. 박철진: He remains at the interrogation seat, suddenly alarmed by the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 크게 놀란 표정으로 굳어 있는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the questioning position inside the brightly lit third-floor room of the camp administration building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease past the subordinate's shoulder along the existing inward path and settle slightly above 박철진's seated eye line, preserving an oblique three-quarter close-up. His face sits right of center with look room on the left toward 박철진의 수하 1 outside the frame; his eyes widen and his seated body arrests as the report registers. Keep the lighting and background treatment continuous, making the turn of his attention toward the subordinate the emphasized change.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding office illumination and exposure so the widened eyes, not a lighting effect, announce the shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the same detention-room materials, chairs, and lighting from the reference. Exclude the earlier assault as a repeated event and any furnishings from the church basement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The management office remains brightly lit, and the third-floor interrogation-room door has been opened. 박철진: He remains at the interrogation seat, suddenly alarmed by the report.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진의 시선은 화면 왼쪽 프레임 밖을 향하고 있으며, 지시된 대로 두 눈을 번쩍 뜨고 놀란 상태로 고정되어 있다.",
    "built_space": "참조 이미지와 동일하게 회색 출입문, 야경이 보이는 창문, 사무실 책상 일부가 올바른 위치와 조명 아래에 배치되어 있다.",
    "entities": "샷 텍스트에 명시된 박철진 1인만 프레임 안에 존재하며, 남색 전투복, 벙거지 모자, 붉은 완장 등 참조 이미지의 인상착의와 완벽하게 일치한다.",
    "hard_violations": [],
    "physics": "의자에 앉아 있는 몸의 무게 중심과 어깨의 각도가 자연스럽게 유지되고 있으며, 물리적인 오류는 보이지 않는다."
   },
   {
    "label": "B",
    "direction": "박철진의 시선이 화면 왼쪽을 향하고 있으며 눈을 크게 뜨고 놀란 표정을 짓고 있다.",
    "built_space": "참조 이미지에 등장했던 회색 문과 창문, 야간의 외부 조명 등이 일관성 있게 구현되어 있다.",
    "entities": "박철진의 신원과 복장은 정확하게 묘사되었으나, 화면 좌측 전경에 샷 텍스트에 허용되지 않은 다른 인물(수하)의 어깨 실루엣이 등장한다.",
    "hard_violations": [
     "[gemini-pro] 프레임 밖에 있어야 할 수하의 어깨 실루엣이 화면 좌측에 포함되어, '허용된 인물 외 다른 인물 또는 신체 일부 추가 불가' 규칙 위반",
     "[gpt-high] 화면 밖에 있어야 하고 이번 숏에 등장하도록 허용되지 않은 수하의 어깨와 몸통 일부를 왼쪽 전경에 표시했다."
    ],
    "physics": "앉은 자세와 시선 처리는 안정적이며 물리적 충돌이나 어색한 묘사는 확인되지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 박철진의 놀란 표정과 클로즈업 구도를 완벽하게 구현했으며, 프레임 내에 다른 인물을 배제하라는 엄격한 지시를 정확히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "박철진의 표정과 배경은 훌륭하게 묘사되었으나, 프레임 밖에 있어야 할 수하의 어깨가 화면 좌측을 침범하여 인물 등장 제한 규칙을 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 화면 왼쪽 프레임 밖을 향하고 있으며, 지시된 대로 두 눈을 번쩍 뜨고 놀란 상태로 고정되어 있다.",
        "built_space": "참조 이미지와 동일하게 회색 출입문, 야경이 보이는 창문, 사무실 책상 일부가 올바른 위치와 조명 아래에 배치되어 있다.",
        "entities": "샷 텍스트에 명시된 박철진 1인만 프레임 안에 존재하며, 남색 전투복, 벙거지 모자, 붉은 완장 등 참조 이미지의 인상착의와 완벽하게 일치한다.",
        "hard_violations": [],
        "physics": "의자에 앉아 있는 몸의 무게 중심과 어깨의 각도가 자연스럽게 유지되고 있으며, 물리적인 오류는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진의 시선이 화면 왼쪽을 향하고 있으며 눈을 크게 뜨고 놀란 표정을 짓고 있다.",
        "built_space": "참조 이미지에 등장했던 회색 문과 창문, 야간의 외부 조명 등이 일관성 있게 구현되어 있다.",
        "entities": "박철진의 신원과 복장은 정확하게 묘사되었으나, 화면 좌측 전경에 샷 텍스트에 허용되지 않은 다른 인물(수하)의 어깨 실루엣이 등장한다.",
        "hard_violations": [
         "프레임 밖에 있어야 할 수하의 어깨 실루엣이 화면 좌측에 포함되어, '허용된 인물 외 다른 인물 또는 신체 일부 추가 불가' 규칙 위반"
        ],
        "physics": "앉은 자세와 시선 처리는 안정적이며 물리적 충돌이나 어색한 묘사는 확인되지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 박철진의 놀란 표정과 클로즈업 구도를 완벽하게 구현했으며, 프레임 내에 다른 인물을 배제하라는 엄격한 지시를 정확히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "박철진의 표정과 배경은 훌륭하게 묘사되었으나, 프레임 밖에 있어야 할 수하의 어깨가 화면 좌측을 침범하여 인물 등장 제한 규칙을 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 시선은 화면 왼쪽 프레임 밖을 향하고 있으며, 지시된 대로 두 눈을 번쩍 뜨고 놀란 상태로 고정되어 있다.",
        "built_space": "참조 이미지와 동일하게 회색 출입문, 야경이 보이는 창문, 사무실 책상 일부가 올바른 위치와 조명 아래에 배치되어 있다.",
        "entities": "샷 텍스트에 명시된 박철진 1인만 프레임 안에 존재하며, 남색 전투복, 벙거지 모자, 붉은 완장 등 참조 이미지의 인상착의와 완벽하게 일치한다.",
        "hard_violations": [],
        "physics": "의자에 앉아 있는 몸의 무게 중심과 어깨의 각도가 자연스럽게 유지되고 있으며, 물리적인 오류는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진의 시선이 화면 왼쪽을 향하고 있으며 눈을 크게 뜨고 놀란 표정을 짓고 있다.",
        "built_space": "참조 이미지에 등장했던 회색 문과 창문, 야간의 외부 조명 등이 일관성 있게 구현되어 있다.",
        "entities": "박철진의 신원과 복장은 정확하게 묘사되었으나, 화면 좌측 전경에 샷 텍스트에 허용되지 않은 다른 인물(수하)의 어깨 실루엣이 등장한다.",
        "hard_violations": [
         "프레임 밖에 있어야 할 수하의 어깨 실루엣이 화면 좌측에 포함되어, '허용된 인물 외 다른 인물 또는 신체 일부 추가 불가' 규칙 위반"
        ],
        "physics": "앉은 자세와 시선 처리는 안정적이며 물리적 충돌이나 어색한 묘사는 확인되지 않는다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "놀란 눈과 왼쪽을 향한 시선은 맞지만, 화면 밖에 있어야 할 수하의 어깨와 몸통을 전경에 넣어 인물 제한을 위반했고 문도 닫혀 있다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박철진만 담은 오른쪽 중심의 사선 클로즈업과 화면 밖 왼쪽을 향한 놀란 시선이 가장 충실하지만, 열려 있어야 할 문은 닫혀 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진은 두 눈을 크게 뜨고 화면 왼쪽 위를 바라본다. 시선은 왼쪽 전경에 몸 일부가 들어온 상대의 얼굴이 있을 방향으로 향하며 렌즈를 보지 않는다. 다만 상대를 완전히 화면 밖에 두라는 지시와 다르다.",
        "built_space": "왼쪽에 작은 세로 유리창이 있는 회색 문 하나, 위쪽에 천장 조명 일부 하나, 오른쪽에 금속 프레임 창과 책상 일부 하나가 보인다. 밝은 회색 벽과 야간 창밖은 이전 장면과 이어진다. 문은 닫힌 상태로 보여 열린 문이라는 지속 상태를 충족하지 않는다. 얼굴은 오른쪽에 있으나 왼쪽 전경의 상대 몸통이 시선 여백을 차지한다.",
        "entities": "박철진의 중년 한국인 남성 외관, 얼굴 윤곽, 모자 아래 짧은 검은 머리, 낡은 남색 전투복, 턱끈 달린 모자와 붉은 완장은 참조와 대체로 일치한다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠져 있다. 그러나 왼쪽에 다른 사람의 어깨와 몸통 일부가 추가되어 박철진만 허용한 인물 조건을 어긴다.",
        "hard_violations": [
         "화면 밖에 있어야 하고 이번 숏에 등장하도록 허용되지 않은 수하의 어깨와 몸통 일부를 왼쪽 전경에 표시했다."
        ],
        "physics": "박철진의 머리는 목과 몸통에 자연스럽게 연결되고, 모자는 머리에 얹혀 있으며 턱끈은 아래로 늘어진다. 하체와 좌판은 프레임 밖이라 착석 접촉은 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다. 전경 인물 역시 몸 일부만 잘렸을 뿐 떠 있는 것으로 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "박철진의 얼굴과 양쪽 눈이 화면 밖 왼쪽 위의 수하를 향한다. 렌즈를 응시하지 않으며 왼쪽에 시선 여백이 확보되어 있다. 커진 눈과 벌어진 입이 보고를 듣고 놀라 멈춘 순간으로 읽힌다.",
        "built_space": "왼쪽에 세로 유리창과 손잡이가 있는 회색 문 하나, 그 옆 벽 부착 장치 하나, 오른쪽에 금속 프레임 창과 책상 하나, 그 앞에 검은 의자 등받이 일부가 보인다. 화면 아래 왼쪽에는 박철진 좌석의 금속 테두리 일부가 보인다. 벽과 창, 책상의 재질 및 야간 배경은 참조와 연속성이 있다. 다만 문은 닫혀 있어 열린 문 조건을 어긴다. 오른쪽 중심의 얼굴과 사선 구도는 맞고, 시점은 눈높이에 가깝게 읽혀 약간 높은 시점이라는 조건은 뚜렷하지 않다.",
        "entities": "허용된 박철진 한 명만 보인다. 중년 한국인 남성의 얼굴, 짧은 검은 머리, 남색 모자와 턱끈, 낡은 남색 전투복 및 오른쪽 아래에 걸친 붉은 완장이 참조와 대체로 일치한다. 눈은 해부학적으로 정상이며 표정 연기로 놀람을 나타낸다. 추가 인물이나 글자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "상체 아래 왼쪽에 좌석의 금속 테두리가 보이며 몸은 앉은 상태로 자연스럽게 이어진다. 엉덩이와 발의 접촉은 클로즈업 밖이라 확인할 수 없다. 머리와 모자는 정상적으로 지지되고 턱끈은 중력 방향으로 늘어지며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "놀란 눈과 왼쪽을 향한 시선은 맞지만, 화면 밖에 있어야 할 수하의 어깨와 몸통을 전경에 넣어 인물 제한을 위반했고 문도 닫혀 있다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박철진만 담은 오른쪽 중심의 사선 클로즈업과 화면 밖 왼쪽을 향한 놀란 시선이 가장 충실하지만, 열려 있어야 할 문은 닫혀 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진은 두 눈을 크게 뜨고 화면 왼쪽 위를 바라본다. 시선은 왼쪽 전경에 몸 일부가 들어온 상대의 얼굴이 있을 방향으로 향하며 렌즈를 보지 않는다. 다만 상대를 완전히 화면 밖에 두라는 지시와 다르다.",
        "built_space": "왼쪽에 작은 세로 유리창이 있는 회색 문 하나, 위쪽에 천장 조명 일부 하나, 오른쪽에 금속 프레임 창과 책상 일부 하나가 보인다. 밝은 회색 벽과 야간 창밖은 이전 장면과 이어진다. 문은 닫힌 상태로 보여 열린 문이라는 지속 상태를 충족하지 않는다. 얼굴은 오른쪽에 있으나 왼쪽 전경의 상대 몸통이 시선 여백을 차지한다.",
        "entities": "박철진의 중년 한국인 남성 외관, 얼굴 윤곽, 모자 아래 짧은 검은 머리, 낡은 남색 전투복, 턱끈 달린 모자와 붉은 완장은 참조와 대체로 일치한다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠져 있다. 그러나 왼쪽에 다른 사람의 어깨와 몸통 일부가 추가되어 박철진만 허용한 인물 조건을 어긴다.",
        "hard_violations": [
         "화면 밖에 있어야 하고 이번 숏에 등장하도록 허용되지 않은 수하의 어깨와 몸통 일부를 왼쪽 전경에 표시했다."
        ],
        "physics": "박철진의 머리는 목과 몸통에 자연스럽게 연결되고, 모자는 머리에 얹혀 있으며 턱끈은 아래로 늘어진다. 하체와 좌판은 프레임 밖이라 착석 접촉은 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다. 전경 인물 역시 몸 일부만 잘렸을 뿐 떠 있는 것으로 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "박철진의 얼굴과 양쪽 눈이 화면 밖 왼쪽 위의 수하를 향한다. 렌즈를 응시하지 않으며 왼쪽에 시선 여백이 확보되어 있다. 커진 눈과 벌어진 입이 보고를 듣고 놀라 멈춘 순간으로 읽힌다.",
        "built_space": "왼쪽에 세로 유리창과 손잡이가 있는 회색 문 하나, 그 옆 벽 부착 장치 하나, 오른쪽에 금속 프레임 창과 책상 하나, 그 앞에 검은 의자 등받이 일부가 보인다. 화면 아래 왼쪽에는 박철진 좌석의 금속 테두리 일부가 보인다. 벽과 창, 책상의 재질 및 야간 배경은 참조와 연속성이 있다. 다만 문은 닫혀 있어 열린 문 조건을 어긴다. 오른쪽 중심의 얼굴과 사선 구도는 맞고, 시점은 눈높이에 가깝게 읽혀 약간 높은 시점이라는 조건은 뚜렷하지 않다.",
        "entities": "허용된 박철진 한 명만 보인다. 중년 한국인 남성의 얼굴, 짧은 검은 머리, 남색 모자와 턱끈, 낡은 남색 전투복 및 오른쪽 아래에 걸친 붉은 완장이 참조와 대체로 일치한다. 눈은 해부학적으로 정상이며 표정 연기로 놀람을 나타낸다. 추가 인물이나 글자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "상체 아래 왼쪽에 좌석의 금속 테두리가 보이며 몸은 앉은 상태로 자연스럽게 이어진다. 엉덩이와 발의 접촉은 클로즈업 밖이라 확인할 수 없다. 머리와 모자는 정상적으로 지지되고 턱끈은 중력 방향으로 늘어지며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.775
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.525
   },
   "violations": {
    "B": [
     "[gemini-pro] 프레임 밖에 있어야 할 수하의 어깨 실루엣이 화면 좌측에 포함되어, '허용된 인물 외 다른 인물 또는 신체 일부 추가 불가' 규칙 위반",
     "[gpt-high] 화면 밖에 있어야 하고 이번 숏에 등장하도록 허용되지 않은 수하의 어깨와 몸통 일부를 왼쪽 전경에 표시했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 525
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 박철진의 놀란 표정과 클로즈업 구도를 완벽하게 구현했으며, 프레임 내에 다른 인물을 배제하라는 엄격한 지시를 정확히 따랐습니다."
   },
   {
    "label": "B",
    "score": 525,
    "verdict_ko": "박철진의 표정과 배경은 훌륭하게 묘사되었으나, 프레임 밖에 있어야 할 수하의 어깨가 화면 좌측을 침범하여 인물 등장 제한 규칙을 어겼습니다.  ★위반: [gemini-pro] 프레임 밖에 있어야 할 수하의 어깨 실루엣이 화면 좌측에 포함되어, '허용된 인물 외 다른 인물 또는 신체 일부 추가 불가' 규칙 위반 / [gpt-high] 화면 밖에 있어야 하고 이번 숏에 등장하도록 허용되지 않은 수하의 어깨와 몸통 일부를 왼쪽 전경에 표시했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh10_sel.png",
    "asset_id": "d083bd45-de05-491a-ab29-7879b5167f47",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09f7-5a30-7b28-bcae-ac745ab3c1ad",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh10"
  }
 },
 "S32sh1::signage": {
  "fp": "26ebd74d2b0b3df7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh1": {
  "input_fingerprint": "493076bb92cdd91a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지하실 원탁 앞에 구도환, 찰리, 앰버, 신부가 앉아 있고 이현우가 팔짱을 낀 채 벽에 기대선 넓은 구도.\n\nLOCATION (lock): Around the table in the church's sparsely furnished basement prayer room, with an adjacent wall available for standing. The previously activated lanterns illuminate the space. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high, diagonal entry position outside the seated group, tilting down across the round table while retaining 이현우 against the wall at the right rear. Arrange 구도환 at the far-left side, 찰리 near-left, 앰버 near-center in rear three-quarter view, and 신부 beyond her, with varied torso inclinations rather than a symmetrical lineup; the table occupies less than a third of the image. 앰버 turns her question toward 구도환, 찰리 lowers his attention toward the speaking 앰버, and 신부 listens toward her, while 이현우 watches the discussion from his separate wall position with arms crossed.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: round table with seated group in the middle-left of the frame, midground; wall supporting 이현우 apart from the table in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Round table (Surrounded by the seated group) — Its top is visible from the elevated diagonal camera position; used as Organize the shared discussion without concealing the separation from 이현우; Basement wall (Supporting 이현우 as he leans apart from the group) — Visible behind 이현우 on the right side of the composition; used as Provide a distinct spatial anchor for his refusal to join the table.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the basement, with subdued brightness and enough facial separation to read the group's differing attitudes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, small organ, cross, desk, chairs, and basement surfaces from the reference. Exclude the transient burst of light that first activated the lanterns.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, table, chairs, cross and small organ, and the old radio is now functional. Charlie is seated, retaining his old coat and hat disguise, blue-lit eyes and worn chest logo. 구도환: He is seated at the table. 앰버: She is seated, retaining her mask and waist tool pouch. 신부: He is seated and wears his clerical collar. 이현우: He stands against the wall with folded arms, facial bruises and the persistent leg injury. His outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지하실 원탁 앞에 구도환, 찰리, 앰버, 신부가 앉아 있고 이현우가 팔짱을 낀 채 벽에 기대선 넓은 구도.\n\nLOCATION (lock): Around the table in the church's sparsely furnished basement prayer room, with an adjacent wall available for standing. The previously activated lanterns illuminate the space. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high, diagonal entry position outside the seated group, tilting down across the round table while retaining 이현우 against the wall at the right rear. Arrange 구도환 at the far-left side, 찰리 near-left, 앰버 near-center in rear three-quarter view, and 신부 beyond her, with varied torso inclinations rather than a symmetrical lineup; the table occupies less than a third of the image. 앰버 turns her question toward 구도환, 찰리 lowers his attention toward the speaking 앰버, and 신부 listens toward her, while 이현우 watches the discussion from his separate wall position with arms crossed.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: round table with seated group in the middle-left of the frame, midground; wall supporting 이현우 apart from the table in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Round table (Surrounded by the seated group) — Its top is visible from the elevated diagonal camera position; used as Organize the shared discussion without concealing the separation from 이현우; Basement wall (Supporting 이현우 as he leans apart from the group) — Visible behind 이현우 on the right side of the composition; used as Provide a distinct spatial anchor for his refusal to join the table.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the basement, with subdued brightness and enough facial separation to read the group's differing attitudes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, small organ, cross, desk, chairs, and basement surfaces from the reference. Exclude the transient burst of light that first activated the lanterns.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, table, chairs, cross and small organ, and the old radio is now functional. Charlie is seated, retaining his old coat and hat disguise, blue-lit eyes and worn chest logo. 구도환: He is seated at the table. 앰버: She is seated, retaining her mask and waist tool pouch. 신부: He is seated and wears his clerical collar. 이현우: He stands against the wall with folded arms, facial bruises and the persistent leg injury. His outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지하실 원탁 앞에 구도환, 찰리, 앰버, 신부가 앉아 있고 이현우가 팔짱을 낀 채 벽에 기대선 넓은 구도.\n\nLOCATION (lock): Around the table in the church's sparsely furnished basement prayer room, with an adjacent wall available for standing. The previously activated lanterns illuminate the space. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high, diagonal entry position outside the seated group, tilting down across the round table while retaining 이현우 against the wall at the right rear. Arrange 구도환 at the far-left side, 찰리 near-left, 앰버 near-center in rear three-quarter view, and 신부 beyond her, with varied torso inclinations rather than a symmetrical lineup; the table occupies less than a third of the image. 앰버 turns her question toward 구도환, 찰리 lowers his attention toward the speaking 앰버, and 신부 listens toward her, while 이현우 watches the discussion from his separate wall position with arms crossed.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: round table with seated group in the middle-left of the frame, midground; wall supporting 이현우 apart from the table in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Round table (Surrounded by the seated group) — Its top is visible from the elevated diagonal camera position; used as Organize the shared discussion without concealing the separation from 이현우; Basement wall (Supporting 이현우 as he leans apart from the group) — Visible behind 이현우 on the right side of the composition; used as Provide a distinct spatial anchor for his refusal to join the table.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the basement, with subdued brightness and enough facial separation to read the group's differing attitudes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the lit lanterns, small organ, cross, desk, chairs, and basement surfaces from the reference. Exclude the transient burst of light that first activated the lanterns.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The basement retains its lit lanterns, table, chairs, cross and small organ, and the old radio is now functional. Charlie is seated, retaining his old coat and hat disguise, blue-lit eyes and worn chest logo. 구도환: He is seated at the table. 앰버: She is seated, retaining her mask and waist tool pouch. 신부: He is seated and wears his clerical collar. 이현우: He stands against the wall with folded arms, facial bruises and the persistent leg injury. His outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 구도환을, 찰리와 신부는 앰버를, 이현우는 원탁 그룹을 주시하여 모든 시선이 완벽히 일치함.",
    "built_space": "원탁 등 지시된 배치를 따랐으며, 이전 샷의 물주전자 탁자와 계단의 푸른 조명까지 정확히 유지됨.",
    "entities": "5명 모두 레퍼런스의 인상 및 복장(마스크, 무전기, 사제복 등)과 정확히 일치함.",
    "hard_violations": [
     "[gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다.",
     "[gpt-high] 지정되지 않은 지도 형태의 인쇄물을 원탁에 추가해 새로운 도해 소품을 만들었다."
    ],
    "physics": "모든 인물의 무게 중심과 접촉면이 바닥 및 가구와 자연스럽고 안정적으로 연결됨."
   },
   {
    "label": "B",
    "direction": "모든 인물의 시선이 지시문대로 교환되거나 그룹을 향하고 있음.",
    "built_space": "주요 기물은 유지되었으나 좌측 물주전자 탁자가 랜턴 스툴로 대체되고 계단 조명색이 변경됨.",
    "entities": "캐릭터 외형은 전반적으로 일치하나 이현우의 디테일과 유사도가 다소 부족함.",
    "hard_violations": [
     "[gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다."
    ],
    "physics": "이현우의 교차한 오른발이 바닥에서 살짝 떠 있어 체중 지지가 다소 불안정해 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 시선과 구도를 완벽히 구현했으며, 이전 샷의 미세한 공간 디테일까지 정확하게 유지하여 완성도가 높습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 캐릭터는 양호하나 배경의 연속성이 일부 훼손되었고 인물의 자세가 상대적으로 다소 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 구도환을, 찰리와 신부는 앰버를, 이현우는 원탁 그룹을 주시하여 모든 시선이 완벽히 일치함.",
        "built_space": "원탁 등 지시된 배치를 따랐으며, 이전 샷의 물주전자 탁자와 계단의 푸른 조명까지 정확히 유지됨.",
        "entities": "5명 모두 레퍼런스의 인상 및 복장(마스크, 무전기, 사제복 등)과 정확히 일치함.",
        "hard_violations": [],
        "physics": "모든 인물의 무게 중심과 접촉면이 바닥 및 가구와 자연스럽고 안정적으로 연결됨."
       },
       {
        "label": "B",
        "direction": "모든 인물의 시선이 지시문대로 교환되거나 그룹을 향하고 있음.",
        "built_space": "주요 기물은 유지되었으나 좌측 물주전자 탁자가 랜턴 스툴로 대체되고 계단 조명색이 변경됨.",
        "entities": "캐릭터 외형은 전반적으로 일치하나 이현우의 디테일과 유사도가 다소 부족함.",
        "hard_violations": [],
        "physics": "이현우의 교차한 오른발이 바닥에서 살짝 떠 있어 체중 지지가 다소 불안정해 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 시선과 구도를 완벽히 구현했으며, 이전 샷의 미세한 공간 디테일까지 정확하게 유지하여 완성도가 높습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 캐릭터는 양호하나 배경의 연속성이 일부 훼손되었고 인물의 자세가 상대적으로 다소 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 구도환을, 찰리와 신부는 앰버를, 이현우는 원탁 그룹을 주시하여 모든 시선이 완벽히 일치함.",
        "built_space": "원탁 등 지시된 배치를 따랐으며, 이전 샷의 물주전자 탁자와 계단의 푸른 조명까지 정확히 유지됨.",
        "entities": "5명 모두 레퍼런스의 인상 및 복장(마스크, 무전기, 사제복 등)과 정확히 일치함.",
        "hard_violations": [],
        "physics": "모든 인물의 무게 중심과 접촉면이 바닥 및 가구와 자연스럽고 안정적으로 연결됨."
       },
       {
        "label": "B",
        "direction": "모든 인물의 시선이 지시문대로 교환되거나 그룹을 향하고 있음.",
        "built_space": "주요 기물은 유지되었으나 좌측 물주전자 탁자가 랜턴 스툴로 대체되고 계단 조명색이 변경됨.",
        "entities": "캐릭터 외형은 전반적으로 일치하나 이현우의 디테일과 유사도가 다소 부족함.",
        "hard_violations": [],
        "physics": "이현우의 교차한 오른발이 바닥에서 살짝 떠 있어 체중 지지가 다소 불안정해 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "높은 대각선 와이드 구도와 이현우의 분리된 벽 자세, 찰리의 흰 얼굴은 더 충실하지만, 앰버에게 모여야 할 시선과 파란 눈이 어긋나고 지정되지 않은 컵이 추가됐다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "기본 구도와 벽에 기댄 자세는 맞지만, 찰리가 앰버보다 신부 쪽을 향하며 흰 마스크형 얼굴도 약해졌고, 컵과 지도 형태의 인쇄물이 추가됐다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 뒤쪽 사선 모습으로 앉아 고개를 왼쪽 구도환에게 돌리고 손으로 말하는 동작을 한다. 구도환도 앰버 쪽을 본다. 찰리는 얼굴을 아래로 숙였지만 방향은 앰버보다 구도환과 탁자 왼쪽에 가깝다. 신부 역시 얼굴이 앰버보다 왼쪽 구도환 쪽으로 돌아가 있어 앰버를 경청하는 시선이 명확하지 않다. 오른쪽 이현우는 고개를 왼쪽 아래 토론 무리로 향한다. 무기나 이동 중인 물체는 없다.",
        "built_space": "높은 대각선 시점에서 원탁 상판이 보이며 원탁은 화면의 3분의 1보다 작다. 네 인물은 왼쪽 중경에, 이현우는 오른쪽 뒤 벽에 분리되어 있다. 구도환은 화면 가장 왼쪽, 찰리는 그보다 뒤쪽, 앰버는 앞쪽 중앙, 신부는 앰버 너머에 있다. 착석자의 목재 의자 등받이는 세 개가 뚜렷하고 찰리의 좌석은 몸에 가려진다. 왼쪽 벽의 빈 의자 세 개와 뒤쪽 책상 주변 의자 두 개가 보인다. 십자가 하나, 뒤쪽 사각 책상 하나, 오른쪽 앞 소형 오르간 하나, 오른쪽 라디오 하나, 책장 두 개가 있으며 기존 벽면과 계단, 배관, 깔개의 연속성이 강하다. 켜진 랜턴은 가장자리의 부분 노출을 포함해 열한 개가 보인다. 불가능한 반사는 없다.",
        "entities": "지정된 다섯 인물만 보인다. 구도환은 중년 동아시아계 남성과 갈색 점퍼로 표현되지만 참고보다 머리가 헝클어지고 얼굴이 더 거칠다. 찰리는 모자와 낡은 코트, 육중한 기계 팔, 흰 마스크형 얼굴과 푸른 가슴 원자로를 갖췄으나 눈은 요구된 파랑이 아니라 주황색이다. 앰버는 어린 금발 소녀이며 방진 마스크, 카키 작업복, 허리 공구 주머니가 보인다. 가려진 얼굴로 세부 정체성은 확인하기 어렵다. 신부는 회색 짧은 머리의 노년 동아시아계 남성으로 검은 사제복과 흰 로만칼라가 맞는다. 이현우는 젊은 동아시아계 남성으로 어두운 셔츠와 바지를 입고 겉옷 없이 팔짱을 꼈으며 얼굴의 상처가 보인다. 인이어와 지속적인 다리 부상은 명확히 식별되지 않는다. 라디오에는 불빛이 들어온다. 원탁 위 금속 컵은 지정되거나 참고에 확립된 소품이 아니다.",
        "hard_violations": [
         "프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다."
        ],
        "physics": "구도환, 앰버, 신부는 각자의 의자에 엉덩이를 두고 등받이를 뒤로 둔 자연스러운 착석 상태다. 찰리의 좌석은 가려져 있지만 몸통과 굽힌 다리, 바닥에 닿는 기계 발은 착석으로 읽히며 공중에 뜬 증거는 없다. 이현우는 등과 어깨를 벽에 기대고 두 발을 바닥에 둔다. 팔짱과 체중 지지는 가능하다. 앰버의 들린 손은 팔에 연결된 발화 제스처이며 소품을 공중에 띄우지 않는다. 책, 컵, 랜턴과 라디오는 각각 탁자나 가구 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "앰버는 고개를 왼쪽 구도환 쪽으로 돌리고 한 손을 들어 질문하는 모습이며, 구도환은 앰버를 본다. 찰리의 얼굴은 오른쪽 신부 쪽으로 돌아가 있어 앰버에게 주의를 낮추라는 지시와 다르다. 신부의 시선도 앰버보다 왼쪽의 찰리 또는 구도환 쪽으로 읽힌다. 이현우는 오른쪽 벽에서 왼쪽 토론 무리를 바라본다. 무기나 이동하는 물체는 없다.",
        "built_space": "원탁 상판을 내려다보는 높은 대각선 와이드 구도이며 원탁 면적은 화면의 3분의 1 미만이다. 구도환은 가장 왼쪽 앞, 찰리는 왼쪽 뒤, 앰버는 앞쪽 중앙의 후면 사선, 신부는 그 너머에 앉아 있다. 이현우는 오른쪽 뒤 벽에 분리되어 있다. 착석자의 목재 의자 등받이 세 개가 명확하고 찰리 좌석은 가려진다. 왼쪽 벽의 빈 의자 세 개와 뒤쪽 책상 주변 의자 두 개가 보인다. 십자가 하나, 뒤쪽 사각 책상 하나, 소형 오르간 하나, 라디오 하나, 책장 두 개와 기존 계단·배관·깔개가 유지된다. 켜진 랜턴은 가장자리 부분 노출을 포함해 열한 개가 보이지만 참고의 오른쪽 별도 받침대 랜턴 배치는 그대로 확인되지 않는다. 불가능한 반사는 없다.",
        "entities": "지정된 다섯 인물만 있다. 구도환의 중년 동아시아계 외형과 갈색 점퍼, 신부의 노년 외형과 검은 사제복·흰 칼라는 대체로 맞는다. 앰버는 어린 금발 소녀로 마스크, 카키 작업복과 공구 벨트를 유지한다. 찰리는 모자와 코트, 굵은 기계 팔과 푸른 가슴 원자로를 갖췄지만 얼굴이 참고의 귀여운 흰 마스크보다 각진 베이지색 금속판에 가깝고 눈도 주황색이다. 이현우는 젊은 동아시아계 남성으로 겉옷 없이 어두운 셔츠 차림이며 팔짱을 꼈다. 얼굴 멍과 인이어는 뚜렷하지 않고 다리 자세만으로 부상의 지속 여부를 확정할 수 없다. 라디오에는 불빛이 보인다. 원탁에는 지정되지 않은 금속 컵과 지도처럼 선이 인쇄된 종이가 추가되어 있다.",
        "hard_violations": [
         "프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다.",
         "지정되지 않은 지도 형태의 인쇄물을 원탁에 추가해 새로운 도해 소품을 만들었다."
        ],
        "physics": "세 인간 착석자는 각자 의자 좌판에 앉아 있고, 찰리도 굽힌 다리와 바닥에 놓인 기계 발을 가진 착석 상태다. 찰리의 좌판 자체는 가려져 있으나 부유하는 모습은 아니다. 이현우는 벽에 상체를 기대고 한쪽 무릎을 굽혀 다른 다리에 체중을 싣는다. 발과 벽의 지지가 있어 가능한 자세지만 이것만으로 다리 부상을 표현했다고 단정할 수는 없다. 앰버의 든 손은 자연스러운 제스처다. 종이와 책, 컵, 랜턴은 상판이나 받침 가구에 놓여 있으며 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "높은 대각선 와이드 구도와 이현우의 분리된 벽 자세, 찰리의 흰 얼굴은 더 충실하지만, 앰버에게 모여야 할 시선과 파란 눈이 어긋나고 지정되지 않은 컵이 추가됐다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "기본 구도와 벽에 기댄 자세는 맞지만, 찰리가 앰버보다 신부 쪽을 향하며 흰 마스크형 얼굴도 약해졌고, 컵과 지도 형태의 인쇄물이 추가됐다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 뒤쪽 사선 모습으로 앉아 고개를 왼쪽 구도환에게 돌리고 손으로 말하는 동작을 한다. 구도환도 앰버 쪽을 본다. 찰리는 얼굴을 아래로 숙였지만 방향은 앰버보다 구도환과 탁자 왼쪽에 가깝다. 신부 역시 얼굴이 앰버보다 왼쪽 구도환 쪽으로 돌아가 있어 앰버를 경청하는 시선이 명확하지 않다. 오른쪽 이현우는 고개를 왼쪽 아래 토론 무리로 향한다. 무기나 이동 중인 물체는 없다.",
        "built_space": "높은 대각선 시점에서 원탁 상판이 보이며 원탁은 화면의 3분의 1보다 작다. 네 인물은 왼쪽 중경에, 이현우는 오른쪽 뒤 벽에 분리되어 있다. 구도환은 화면 가장 왼쪽, 찰리는 그보다 뒤쪽, 앰버는 앞쪽 중앙, 신부는 앰버 너머에 있다. 착석자의 목재 의자 등받이는 세 개가 뚜렷하고 찰리의 좌석은 몸에 가려진다. 왼쪽 벽의 빈 의자 세 개와 뒤쪽 책상 주변 의자 두 개가 보인다. 십자가 하나, 뒤쪽 사각 책상 하나, 오른쪽 앞 소형 오르간 하나, 오른쪽 라디오 하나, 책장 두 개가 있으며 기존 벽면과 계단, 배관, 깔개의 연속성이 강하다. 켜진 랜턴은 가장자리의 부분 노출을 포함해 열한 개가 보인다. 불가능한 반사는 없다.",
        "entities": "지정된 다섯 인물만 보인다. 구도환은 중년 동아시아계 남성과 갈색 점퍼로 표현되지만 참고보다 머리가 헝클어지고 얼굴이 더 거칠다. 찰리는 모자와 낡은 코트, 육중한 기계 팔, 흰 마스크형 얼굴과 푸른 가슴 원자로를 갖췄으나 눈은 요구된 파랑이 아니라 주황색이다. 앰버는 어린 금발 소녀이며 방진 마스크, 카키 작업복, 허리 공구 주머니가 보인다. 가려진 얼굴로 세부 정체성은 확인하기 어렵다. 신부는 회색 짧은 머리의 노년 동아시아계 남성으로 검은 사제복과 흰 로만칼라가 맞는다. 이현우는 젊은 동아시아계 남성으로 어두운 셔츠와 바지를 입고 겉옷 없이 팔짱을 꼈으며 얼굴의 상처가 보인다. 인이어와 지속적인 다리 부상은 명확히 식별되지 않는다. 라디오에는 불빛이 들어온다. 원탁 위 금속 컵은 지정되거나 참고에 확립된 소품이 아니다.",
        "hard_violations": [
         "프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다."
        ],
        "physics": "구도환, 앰버, 신부는 각자의 의자에 엉덩이를 두고 등받이를 뒤로 둔 자연스러운 착석 상태다. 찰리의 좌석은 가려져 있지만 몸통과 굽힌 다리, 바닥에 닿는 기계 발은 착석으로 읽히며 공중에 뜬 증거는 없다. 이현우는 등과 어깨를 벽에 기대고 두 발을 바닥에 둔다. 팔짱과 체중 지지는 가능하다. 앰버의 들린 손은 팔에 연결된 발화 제스처이며 소품을 공중에 띄우지 않는다. 책, 컵, 랜턴과 라디오는 각각 탁자나 가구 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "앰버는 고개를 왼쪽 구도환 쪽으로 돌리고 한 손을 들어 질문하는 모습이며, 구도환은 앰버를 본다. 찰리의 얼굴은 오른쪽 신부 쪽으로 돌아가 있어 앰버에게 주의를 낮추라는 지시와 다르다. 신부의 시선도 앰버보다 왼쪽의 찰리 또는 구도환 쪽으로 읽힌다. 이현우는 오른쪽 벽에서 왼쪽 토론 무리를 바라본다. 무기나 이동하는 물체는 없다.",
        "built_space": "원탁 상판을 내려다보는 높은 대각선 와이드 구도이며 원탁 면적은 화면의 3분의 1 미만이다. 구도환은 가장 왼쪽 앞, 찰리는 왼쪽 뒤, 앰버는 앞쪽 중앙의 후면 사선, 신부는 그 너머에 앉아 있다. 이현우는 오른쪽 뒤 벽에 분리되어 있다. 착석자의 목재 의자 등받이 세 개가 명확하고 찰리 좌석은 가려진다. 왼쪽 벽의 빈 의자 세 개와 뒤쪽 책상 주변 의자 두 개가 보인다. 십자가 하나, 뒤쪽 사각 책상 하나, 소형 오르간 하나, 라디오 하나, 책장 두 개와 기존 계단·배관·깔개가 유지된다. 켜진 랜턴은 가장자리 부분 노출을 포함해 열한 개가 보이지만 참고의 오른쪽 별도 받침대 랜턴 배치는 그대로 확인되지 않는다. 불가능한 반사는 없다.",
        "entities": "지정된 다섯 인물만 있다. 구도환의 중년 동아시아계 외형과 갈색 점퍼, 신부의 노년 외형과 검은 사제복·흰 칼라는 대체로 맞는다. 앰버는 어린 금발 소녀로 마스크, 카키 작업복과 공구 벨트를 유지한다. 찰리는 모자와 코트, 굵은 기계 팔과 푸른 가슴 원자로를 갖췄지만 얼굴이 참고의 귀여운 흰 마스크보다 각진 베이지색 금속판에 가깝고 눈도 주황색이다. 이현우는 젊은 동아시아계 남성으로 겉옷 없이 어두운 셔츠 차림이며 팔짱을 꼈다. 얼굴 멍과 인이어는 뚜렷하지 않고 다리 자세만으로 부상의 지속 여부를 확정할 수 없다. 라디오에는 불빛이 보인다. 원탁에는 지정되지 않은 금속 컵과 지도처럼 선이 인쇄된 종이가 추가되어 있다.",
        "hard_violations": [
         "프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다.",
         "지정되지 않은 지도 형태의 인쇄물을 원탁에 추가해 새로운 도해 소품을 만들었다."
        ],
        "physics": "세 인간 착석자는 각자 의자 좌판에 앉아 있고, 찰리도 굽힌 다리와 바닥에 놓인 기계 발을 가진 착석 상태다. 찰리의 좌판 자체는 가려져 있으나 부유하는 모습은 아니다. 이현우는 벽에 상체를 기대고 한쪽 무릎을 굽혀 다른 다리에 체중을 싣는다. 발과 벽의 지지가 있어 가능한 자세지만 이것만으로 다리 부상을 표현했다고 단정할 수는 없다. 앰버의 든 손은 자연스러운 제스처다. 종이와 책, 컵, 랜턴은 상판이나 받침 가구에 놓여 있으며 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.55,
    "B": 1.607
   },
   "violations": {
    "B": [
     "[gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다."
    ],
    "A": [
     "[gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다.",
     "[gpt-high] 지정되지 않은 지도 형태의 인쇄물을 원탁에 추가해 새로운 도해 소품을 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1550,
   "B": 1607
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1550,
    "verdict_ko": "지시된 시선과 구도를 완벽히 구현했으며, 이전 샷의 미세한 공간 디테일까지 정확하게 유지하여 완성도가 높습니다.  ★위반: [gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다. / [gpt-high] 지정되지 않은 지도 형태의 인쇄물을 원탁에 추가해 새로운 도해 소품을 만들었다."
   },
   {
    "label": "B",
    "score": 1607,
    "verdict_ko": "구도와 캐릭터는 양호하나 배경의 연속성이 일부 훼손되었고 인물의 자세가 상대적으로 다소 어색합니다.  ★위반: [gpt-high] 프롬프트와 장소 참고에 없는 금속 컵을 원탁 위에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S30sh6_sel.png",
    "asset_id": "39664a01-ac18-43ee-9d36-27bbfaa65114",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab09fb-af18-7a97-9a35-d168d049d84a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh6"
  }
 },
 "S32sh5::signage": {
  "fp": "e63cfe2cb13b0601",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh5": {
  "input_fingerprint": "f3426b4cc572624d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 낡은 라디오를 내려다본 채 충격으로 두 눈을 크게 뜬 찰리의 굳은 얼굴.\n\nLOCATION (lock): Beside the table inside the lantern-lit church basement prayer room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the table's outer edge on 찰리's three-quarter side, placing the lens just below his seated eye line and angling gently toward his lowered face. Hold his rigid face and widened eyes near the center, with his hands and the old radio confined to the lower edge and a narrow table margin preserving spatial context. 찰리 looks down at the radio without lifting his head, his grip arrested by the news; the nearer distance, rather than a new lighting or frontal angle, carries the emotional emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old radio (Held in 찰리's hands) — Seen obliquely beneath his lowered face, without requiring legible markings; used as Anchor the downward eyeline while occupying only a small portion of the lower frame; Table edge (Beside the seated 찰리) — A narrow oblique segment remains at the bottom of the composition; used as Maintain continuity with the preceding table arrangement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement's restrained ambient treatment, preserving expressive eye detail and precise hard-surface contours without adding emitted light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie holds the activated old radio, retaining his old coat and hat disguise, worn chest logo and blue-lit eyes. The basement lanterns remain lit around the table, chairs, cross and small organ.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 낡은 라디오를 내려다본 채 충격으로 두 눈을 크게 뜬 찰리의 굳은 얼굴.\n\nLOCATION (lock): Beside the table inside the lantern-lit church basement prayer room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the table's outer edge on 찰리's three-quarter side, placing the lens just below his seated eye line and angling gently toward his lowered face. Hold his rigid face and widened eyes near the center, with his hands and the old radio confined to the lower edge and a narrow table margin preserving spatial context. 찰리 looks down at the radio without lifting his head, his grip arrested by the news; the nearer distance, rather than a new lighting or frontal angle, carries the emotional emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old radio (Held in 찰리's hands) — Seen obliquely beneath his lowered face, without requiring legible markings; used as Anchor the downward eyeline while occupying only a small portion of the lower frame; Table edge (Beside the seated 찰리) — A narrow oblique segment remains at the bottom of the composition; used as Maintain continuity with the preceding table arrangement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement's restrained ambient treatment, preserving expressive eye detail and precise hard-surface contours without adding emitted light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie holds the activated old radio, retaining his old coat and hat disguise, worn chest logo and blue-lit eyes. The basement lanterns remain lit around the table, chairs, cross and small organ.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 낡은 라디오를 내려다본 채 충격으로 두 눈을 크게 뜬 찰리의 굳은 얼굴.\n\nLOCATION (lock): Beside the table inside the lantern-lit church basement prayer room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the table's outer edge on 찰리's three-quarter side, placing the lens just below his seated eye line and angling gently toward his lowered face. Hold his rigid face and widened eyes near the center, with his hands and the old radio confined to the lower edge and a narrow table margin preserving spatial context. 찰리 looks down at the radio without lifting his head, his grip arrested by the news; the nearer distance, rather than a new lighting or frontal angle, carries the emotional emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Old radio (Held in 찰리's hands) — Seen obliquely beneath his lowered face, without requiring legible markings; used as Anchor the downward eyeline while occupying only a small portion of the lower frame; Table edge (Beside the seated 찰리) — A narrow oblique segment remains at the bottom of the composition; used as Maintain continuity with the preceding table arrangement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the basement's restrained ambient treatment, preserving expressive eye detail and precise hard-surface contours without adding emitted light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie holds the activated old radio, retaining his old coat and hat disguise, worn chest logo and blue-lit eyes. The basement lanterns remain lit around the table, chairs, cross and small organ.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선이 손에 쥐어진 낡은 라디오를 정확히 향하고 있음.",
    "built_space": "배경에 나무 십자가, 촛불이 놓인 작은 테이블, 의자 등 이전 샷의 지하실 요소들이 적절한 원근감으로 배치됨.",
    "entities": "찰리의 고릴라형 기계 몸체, 흰색 마스크(눈 2개, 입 1개), 가슴의 원자로, 낡은 코트와 모자가 기준과 일치하며 낡은 라디오와 테이블 모서리가 확인됨.",
    "hard_violations": [],
    "physics": "기계 손이 라디오를 물리적으로 자연스럽게 파지하고 있으며, 팔의 무게중심이 안정적임."
   },
   {
    "label": "B",
    "direction": "찰리의 시선이 프레임 하단의 라디오로 향하고 있음.",
    "built_space": "십자가, 책장과 성모상 등 지하실 요소가 보이나, 기준 이미지의 넓은 공간감에 비해 배경 사물들이 찰리 바로 뒤로 압축되어 묘사됨.",
    "entities": "찰리의 복장과 기계 외형 요소가 일치하며, 하단에 라디오와 이전 샷에 있던 금속 컵이 보임. 단, 우측 기계 손가락 관절 형태가 다소 불분명함.",
    "hard_violations": [],
    "physics": "두 손으로 라디오를 감싸 쥐고 있는 자세가 지지점 없이 자연스럽게 유지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로우 앵글 클로즈업 구도를 완벽히 따랐으며, 기계 손의 구조와 배경 요소의 스케일을 기준 이미지에 맞게 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷의 테이블 위 컵을 묘사하여 연속성을 살렸으나, 기계 손가락의 디테일이 다소 뭉개지고 배경 공간이 좁게 압축된 아쉬움이 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 손에 쥐어진 낡은 라디오를 정확히 향하고 있음.",
        "built_space": "배경에 나무 십자가, 촛불이 놓인 작은 테이블, 의자 등 이전 샷의 지하실 요소들이 적절한 원근감으로 배치됨.",
        "entities": "찰리의 고릴라형 기계 몸체, 흰색 마스크(눈 2개, 입 1개), 가슴의 원자로, 낡은 코트와 모자가 기준과 일치하며 낡은 라디오와 테이블 모서리가 확인됨.",
        "hard_violations": [],
        "physics": "기계 손이 라디오를 물리적으로 자연스럽게 파지하고 있으며, 팔의 무게중심이 안정적임."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 프레임 하단의 라디오로 향하고 있음.",
        "built_space": "십자가, 책장과 성모상 등 지하실 요소가 보이나, 기준 이미지의 넓은 공간감에 비해 배경 사물들이 찰리 바로 뒤로 압축되어 묘사됨.",
        "entities": "찰리의 복장과 기계 외형 요소가 일치하며, 하단에 라디오와 이전 샷에 있던 금속 컵이 보임. 단, 우측 기계 손가락 관절 형태가 다소 불분명함.",
        "hard_violations": [],
        "physics": "두 손으로 라디오를 감싸 쥐고 있는 자세가 지지점 없이 자연스럽게 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로우 앵글 클로즈업 구도를 완벽히 따랐으며, 기계 손의 구조와 배경 요소의 스케일을 기준 이미지에 맞게 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷의 테이블 위 컵을 묘사하여 연속성을 살렸으나, 기계 손가락의 디테일이 다소 뭉개지고 배경 공간이 좁게 압축된 아쉬움이 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 손에 쥐어진 낡은 라디오를 정확히 향하고 있음.",
        "built_space": "배경에 나무 십자가, 촛불이 놓인 작은 테이블, 의자 등 이전 샷의 지하실 요소들이 적절한 원근감으로 배치됨.",
        "entities": "찰리의 고릴라형 기계 몸체, 흰색 마스크(눈 2개, 입 1개), 가슴의 원자로, 낡은 코트와 모자가 기준과 일치하며 낡은 라디오와 테이블 모서리가 확인됨.",
        "hard_violations": [],
        "physics": "기계 손이 라디오를 물리적으로 자연스럽게 파지하고 있으며, 팔의 무게중심이 안정적임."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 프레임 하단의 라디오로 향하고 있음.",
        "built_space": "십자가, 책장과 성모상 등 지하실 요소가 보이나, 기준 이미지의 넓은 공간감에 비해 배경 사물들이 찰리 바로 뒤로 압축되어 묘사됨.",
        "entities": "찰리의 복장과 기계 외형 요소가 일치하며, 하단에 라디오와 이전 샷에 있던 금속 컵이 보임. 단, 우측 기계 손가락 관절 형태가 다소 불분명함.",
        "hard_violations": [],
        "physics": "두 손으로 라디오를 감싸 쥐고 있는 자세가 지지점 없이 자연스럽게 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "내려다보는 얼굴과 기계 정체성은 맞지만 라디오 조작면이 관객을 향하며, 손·가슴·라디오가 지나치게 커서 요구한 얼굴 중심 클로즈업에 미달한다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "라디오의 뒷면을 관객에게 보이고 조작부를 찰리 쪽에 두는 방향과 지하실 배치가 더 충실하지만, 얼굴보다 넓은 상체 구도와 주황색 눈은 요구와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 머리를 숙여 화면 오른쪽 아래의 라디오를 바라본다. 다만 동공 없는 발광 눈이라 정확한 시선은 머리 방향으로 읽힌다. 라디오의 스피커와 작은 표시창이 있는 전면이 카메라를 향해, 기능면을 자신의 눈 쪽으로 두라는 지시와 반대다. 안테나는 라디오에서 오른쪽 위로 뻗는다.",
        "built_space": "전경 탁자 한 개의 가장자리와 컵 하나가 보인다. 배경에는 십자가 하나, 성화 액자 하나, 책장 하나, 성상 하나, 의자 등받이 하나, 작은 탁자 하나가 보이며 좌우에 랜턴이 하나씩 있다. 십자가 위에도 별도의 밝은 점이 보인다. 낡은 벽과 목재 가구는 참고 장소에 부합하지만, 참고 장면의 제한된 랜턴 조명보다 추가 광원이 있는 듯하다. 좌석과 하체는 프레임 밖이므로 착석 접촉은 확인할 수 없다. 얼굴 외에 가슴과 양팔 전체에 가까운 면적을 담아 요구한 좁은 클로즈업보다 넓다.",
        "entities": "찰리 한 명만 있으며 인간 피부나 인간 손은 없다. 흰 마스크형 얼굴, 두 원형 눈과 선형 입, 샌드 베이지 장갑판, 낡은 검은 코트와 모자, 푸른 원형 가슴 장치가 참고와 부합한다. 눈은 참고 이미지처럼 주황색이지만 이번 장면의 명시적인 푸른 눈 지시에는 어긋난다. 눈을 크게 뜬 충격은 둥근 눈과 굳은 자세로만 약하게 읽힌다. 낡은 휴대용 라디오는 있으나 켜져 있는지는 명확하지 않다. 손과 라디오가 하단의 작은 부분을 넘어 크게 드러난다.",
        "hard_violations": [],
        "physics": "양쪽 기계 손가락이 라디오의 양옆을 감싸고, 라디오 하단과 팔은 탁자에 닿거나 바로 위에 놓인다. 손과 탁자라는 지지가 확인되며 공중에 떠 있는 물체는 없다. 안테나는 본체에 연결되어 있다. 고개를 숙이고 물건을 쥔 채 멈춘 자세는 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "숙인 얼굴이 두 손 사이의 라디오를 향하며, 카메라를 정면 응시하지 않는다. 카메라에는 표시창 없는 통풍구 형태의 후면이 주로 보이고 상단 조작부가 찰리의 내려다보는 방향에 놓여 A보다 사용 방향이 자연스럽다. 안테나는 본체에서 오른쪽 위로 뻗는다.",
        "built_space": "전경 탁자 한 개, 배경 작은 탁자 한 개와 촛불 하나, 십자가 하나, 성화 액자 하나, 책장과 낮은 수납 가구, 의자 등받이 하나가 보인다. 랜턴은 왼쪽 벽, 중앙의 낮은 가구, 그 뒤 높은 책장, 오른쪽 의자 위에 각각 하나씩 네 개가 보인다. 왼쪽의 푸른 계단 입구 일부와 벽·책장·십자가의 관계가 이전 장면과 잘 이어진다. 불가능한 반사나 명백한 중복 고정물은 없다. 다만 얼굴이 중앙보다 왼쪽에 있고 가슴·팔·라디오까지 크게 담아 얼굴 중심 클로즈업보다 넓다.",
        "entities": "등장 인물은 찰리 한 명이며, 흰 기계 마스크와 선형 입, 베이지 금속 손과 장갑판, 낡은 코트와 모자, 푸른 가슴 장치가 참고 정체성을 유지한다. 인간 해부학을 덧붙이지 않았다. 눈은 주황색으로 참고에는 맞지만 명시된 푸른 눈 조건은 충족하지 않는다. 굳은 표정은 보이나 눈이 특별히 확대된 충격 표현은 제한적이다. 낡은 라디오가 있으며 작동 여부는 확실하지 않다. 작은 오르간과 하체는 이 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [],
        "physics": "두 기계 손이 라디오 양쪽을 실제로 감싸 지지하고 있으며, 팔은 전경 탁자에 기대어 있다. 라디오의 기울기는 손으로 유지할 수 있는 범위다. 머리와 목의 연결, 손목과 팔의 연결도 자연스럽다. 지지 없이 떠 있는 몸이나 물체는 없고, 소식을 듣다가 움직임을 멈춘 자세로 가능하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "내려다보는 얼굴과 기계 정체성은 맞지만 라디오 조작면이 관객을 향하며, 손·가슴·라디오가 지나치게 커서 요구한 얼굴 중심 클로즈업에 미달한다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "라디오의 뒷면을 관객에게 보이고 조작부를 찰리 쪽에 두는 방향과 지하실 배치가 더 충실하지만, 얼굴보다 넓은 상체 구도와 주황색 눈은 요구와 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 머리를 숙여 화면 오른쪽 아래의 라디오를 바라본다. 다만 동공 없는 발광 눈이라 정확한 시선은 머리 방향으로 읽힌다. 라디오의 스피커와 작은 표시창이 있는 전면이 카메라를 향해, 기능면을 자신의 눈 쪽으로 두라는 지시와 반대다. 안테나는 라디오에서 오른쪽 위로 뻗는다.",
        "built_space": "전경 탁자 한 개의 가장자리와 컵 하나가 보인다. 배경에는 십자가 하나, 성화 액자 하나, 책장 하나, 성상 하나, 의자 등받이 하나, 작은 탁자 하나가 보이며 좌우에 랜턴이 하나씩 있다. 십자가 위에도 별도의 밝은 점이 보인다. 낡은 벽과 목재 가구는 참고 장소에 부합하지만, 참고 장면의 제한된 랜턴 조명보다 추가 광원이 있는 듯하다. 좌석과 하체는 프레임 밖이므로 착석 접촉은 확인할 수 없다. 얼굴 외에 가슴과 양팔 전체에 가까운 면적을 담아 요구한 좁은 클로즈업보다 넓다.",
        "entities": "찰리 한 명만 있으며 인간 피부나 인간 손은 없다. 흰 마스크형 얼굴, 두 원형 눈과 선형 입, 샌드 베이지 장갑판, 낡은 검은 코트와 모자, 푸른 원형 가슴 장치가 참고와 부합한다. 눈은 참고 이미지처럼 주황색이지만 이번 장면의 명시적인 푸른 눈 지시에는 어긋난다. 눈을 크게 뜬 충격은 둥근 눈과 굳은 자세로만 약하게 읽힌다. 낡은 휴대용 라디오는 있으나 켜져 있는지는 명확하지 않다. 손과 라디오가 하단의 작은 부분을 넘어 크게 드러난다.",
        "hard_violations": [],
        "physics": "양쪽 기계 손가락이 라디오의 양옆을 감싸고, 라디오 하단과 팔은 탁자에 닿거나 바로 위에 놓인다. 손과 탁자라는 지지가 확인되며 공중에 떠 있는 물체는 없다. 안테나는 본체에 연결되어 있다. 고개를 숙이고 물건을 쥔 채 멈춘 자세는 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "숙인 얼굴이 두 손 사이의 라디오를 향하며, 카메라를 정면 응시하지 않는다. 카메라에는 표시창 없는 통풍구 형태의 후면이 주로 보이고 상단 조작부가 찰리의 내려다보는 방향에 놓여 A보다 사용 방향이 자연스럽다. 안테나는 본체에서 오른쪽 위로 뻗는다.",
        "built_space": "전경 탁자 한 개, 배경 작은 탁자 한 개와 촛불 하나, 십자가 하나, 성화 액자 하나, 책장과 낮은 수납 가구, 의자 등받이 하나가 보인다. 랜턴은 왼쪽 벽, 중앙의 낮은 가구, 그 뒤 높은 책장, 오른쪽 의자 위에 각각 하나씩 네 개가 보인다. 왼쪽의 푸른 계단 입구 일부와 벽·책장·십자가의 관계가 이전 장면과 잘 이어진다. 불가능한 반사나 명백한 중복 고정물은 없다. 다만 얼굴이 중앙보다 왼쪽에 있고 가슴·팔·라디오까지 크게 담아 얼굴 중심 클로즈업보다 넓다.",
        "entities": "등장 인물은 찰리 한 명이며, 흰 기계 마스크와 선형 입, 베이지 금속 손과 장갑판, 낡은 코트와 모자, 푸른 가슴 장치가 참고 정체성을 유지한다. 인간 해부학을 덧붙이지 않았다. 눈은 주황색으로 참고에는 맞지만 명시된 푸른 눈 조건은 충족하지 않는다. 굳은 표정은 보이나 눈이 특별히 확대된 충격 표현은 제한적이다. 낡은 라디오가 있으며 작동 여부는 확실하지 않다. 작은 오르간과 하체는 이 구도 밖이므로 누락으로 판단하지 않는다.",
        "hard_violations": [],
        "physics": "두 기계 손이 라디오 양쪽을 실제로 감싸 지지하고 있으며, 팔은 전경 탁자에 기대어 있다. 라디오의 기울기는 손으로 유지할 수 있는 범위다. 머리와 목의 연결, 손목과 팔의 연결도 자연스럽다. 지지 없이 떠 있는 몸이나 물체는 없고, 소식을 듣다가 움직임을 멈춘 자세로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.589
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.589
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1589
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 로우 앵글 클로즈업 구도를 완벽히 따랐으며, 기계 손의 구조와 배경 요소의 스케일을 기준 이미지에 맞게 사실적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 1589,
    "verdict_ko": "이전 샷의 테이블 위 컵을 묘사하여 연속성을 살렸으나, 기계 손가락의 디테일이 다소 뭉개지고 배경 공간이 좁게 압축된 아쉬움이 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S32sh1_sel.png",
    "asset_id": "3bfabda1-7855-4c66-996d-4d029eea52bc",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a03-7768-75b9-b281-84fd0b6918ec",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  }
 },
 "S32sh11::signage": {
  "fp": "6851293c42577b34",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh11": {
  "input_fingerprint": "368675d00f2062bf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 뻗은 팔이 어두운 비밀 통로 입구를 똑바로 가리키는 구도.\n\nLOCATION (lock): At the entrance to the secret escape passage adjoining the church basement. A flashlight has been taken for the dark route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track behind and to one side of 신부 at upper-torso height, looking obliquely past his shoulder toward the secret passage. Keep his rear three-quarter upper body on the lower-left side and his extended arm running diagonally toward the entrance at upper right, with the fingertip stopping short of its visible opening. His head and attention follow the indicated route as he leads the escape, and the arm's placement—not a change in lighting—makes the destination unmistakable.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: secret passage entrance indicated by 신부 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Secret passage entrance (Visible as the escape route indicated by 신부) — The opening is viewed obliquely beyond his extended arm; used as Receive the pointing diagonal without being obscured by the hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the passage entrance dark as described while retaining enough ambient separation to read the pointing arm and the opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신부 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lit basement connects to a secret escape passage. Charlie retains the old radio in hand, his old coat and hat disguise, worn chest logo and blue-lit eyes. 신부: He is up and preparing to lead through the passage, carrying a flashlight and wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 뻗은 팔이 어두운 비밀 통로 입구를 똑바로 가리키는 구도.\n\nLOCATION (lock): At the entrance to the secret escape passage adjoining the church basement. A flashlight has been taken for the dark route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track behind and to one side of 신부 at upper-torso height, looking obliquely past his shoulder toward the secret passage. Keep his rear three-quarter upper body on the lower-left side and his extended arm running diagonally toward the entrance at upper right, with the fingertip stopping short of its visible opening. His head and attention follow the indicated route as he leads the escape, and the arm's placement—not a change in lighting—makes the destination unmistakable.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: secret passage entrance indicated by 신부 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Secret passage entrance (Visible as the escape route indicated by 신부) — The opening is viewed obliquely beyond his extended arm; used as Receive the pointing diagonal without being obscured by the hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the passage entrance dark as described while retaining enough ambient separation to read the pointing arm and the opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신부 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lit basement connects to a secret escape passage. Charlie retains the old radio in hand, his old coat and hat disguise, worn chest logo and blue-lit eyes. 신부: He is up and preparing to lead through the passage, carrying a flashlight and wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 뻗은 팔이 어두운 비밀 통로 입구를 똑바로 가리키는 구도.\n\nLOCATION (lock): At the entrance to the secret escape passage adjoining the church basement. A flashlight has been taken for the dark route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the track behind and to one side of 신부 at upper-torso height, looking obliquely past his shoulder toward the secret passage. Keep his rear three-quarter upper body on the lower-left side and his extended arm running diagonally toward the entrance at upper right, with the fingertip stopping short of its visible opening. His head and attention follow the indicated route as he leads the escape, and the arm's placement—not a change in lighting—makes the destination unmistakable.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: secret passage entrance indicated by 신부 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Secret passage entrance (Visible as the escape route indicated by 신부) — The opening is viewed obliquely beyond his extended arm; used as Receive the pointing diagonal without being obscured by the hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the passage entrance dark as described while retaining enough ambient separation to read the pointing arm and the opening.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신부 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lit basement connects to a secret escape passage. Charlie retains the old radio in hand, his old coat and hat disguise, worn chest logo and blue-lit eyes. 신부: He is up and preparing to lead through the passage, carrying a flashlight and wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "신부의 오른팔이 화면 우측에 위치한 어두운 통로 입구를 향해 대각선으로 정확히 뻗어 있으며, 고개와 시선 역시 그곳을 향하고 있습니다. 손끝은 입구의 어두운 영역 직전에 멈춰 있습니다.",
    "built_space": "화면 우측에 위로 향하는 계단 입구가 있고 좌측에 마리아상이 있는 책장이 보입니다. 하지만 레퍼런스 이미지에서 책장은 계단 반대편 벽에 위치하므로, 두 요소가 나란히 있는 것은 원본 공간 구조와 맞지 않습니다.",
    "entities": "60대 한국인 남성 신부가 짧은 백발에 하얀 로만칼라와 낡은 검은색 사제복을 입고 있으며, 요구된 대로 왼손에 손전등을 들고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 계단 반대편에 위치했던 마리아상 책장이 계단 입구 바로 옆으로 이동하여 물리적으로 불가능한 무대 구조가 형성됨."
    ],
    "physics": "신부가 두 발로 바닥을 딛고 서서 왼손으로 손전등을 단단히 쥐고 있으며, 오른팔을 자연스럽게 뻗어 지탱하는 등 물리적 오류 없이 안정적인 자세를 보여줍니다."
   },
   {
    "label": "B",
    "direction": "신부의 오른팔이 수평에 가깝게 우측을 향해 뻗어 있으나, 통로 입구가 화면 중앙에 있어 손끝이 엉뚱한 우측 벽면을 가리키고 있습니다.",
    "built_space": "화면 중앙에 위로 올라가는 계단이 있고 좌측에 마리아상 책장, 우측에 성화 액자가 배치되어 있습니다. 이는 레퍼런스의 방 구조를 완전히 해체하여 임의로 재조합한 형태입니다.",
    "entities": "60대 한국인 남성 신부가 지시된 사제복을 입고 있으며 손전등을 들고 있으나, 손전등을 쥔 왼손 엄지손가락의 형태와 관절 묘사가 어색합니다.",
    "hard_violations": [
     "[gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 서로 멀리 떨어져 있던 마리아상 책장과 성화 액자가 계단 입구를 둘러싸고 나란히 배치되는 등 공간 기하학이 완전히 파괴됨."
    ],
    "physics": "서 있는 자세와 팔을 뻗은 동작 자체는 지지대가 확보되어 있으나, 왼손 손가락이 손전등을 감싸 쥐는 형태가 해부학적으로 부자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간 배치를 완벽히 유지하지는 못했으나, 프롬프트가 엄격하게 요구한 피사체의 좌측 하단 배치와 우측 상단의 입구를 향해 대각선으로 뻗은 팔의 구도를 가장 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 구도를 무시하고 통로를 화면 중앙에 배치했으며, 뻗은 팔이 통로가 아닌 우측 벽을 향하고 있어 방향성과 프레이밍 모두 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부의 오른팔이 화면 우측에 위치한 어두운 통로 입구를 향해 대각선으로 정확히 뻗어 있으며, 고개와 시선 역시 그곳을 향하고 있습니다. 손끝은 입구의 어두운 영역 직전에 멈춰 있습니다.",
        "built_space": "화면 우측에 위로 향하는 계단 입구가 있고 좌측에 마리아상이 있는 책장이 보입니다. 하지만 레퍼런스 이미지에서 책장은 계단 반대편 벽에 위치하므로, 두 요소가 나란히 있는 것은 원본 공간 구조와 맞지 않습니다.",
        "entities": "60대 한국인 남성 신부가 짧은 백발에 하얀 로만칼라와 낡은 검은색 사제복을 입고 있으며, 요구된 대로 왼손에 손전등을 들고 있습니다.",
        "hard_violations": [
         "공간 배치 오류: 레퍼런스 샷에서 계단 반대편에 위치했던 마리아상 책장이 계단 입구 바로 옆으로 이동하여 물리적으로 불가능한 무대 구조가 형성됨."
        ],
        "physics": "신부가 두 발로 바닥을 딛고 서서 왼손으로 손전등을 단단히 쥐고 있으며, 오른팔을 자연스럽게 뻗어 지탱하는 등 물리적 오류 없이 안정적인 자세를 보여줍니다."
       },
       {
        "label": "B",
        "direction": "신부의 오른팔이 수평에 가깝게 우측을 향해 뻗어 있으나, 통로 입구가 화면 중앙에 있어 손끝이 엉뚱한 우측 벽면을 가리키고 있습니다.",
        "built_space": "화면 중앙에 위로 올라가는 계단이 있고 좌측에 마리아상 책장, 우측에 성화 액자가 배치되어 있습니다. 이는 레퍼런스의 방 구조를 완전히 해체하여 임의로 재조합한 형태입니다.",
        "entities": "60대 한국인 남성 신부가 지시된 사제복을 입고 있으며 손전등을 들고 있으나, 손전등을 쥔 왼손 엄지손가락의 형태와 관절 묘사가 어색합니다.",
        "hard_violations": [
         "공간 배치 오류: 레퍼런스 샷에서 서로 멀리 떨어져 있던 마리아상 책장과 성화 액자가 계단 입구를 둘러싸고 나란히 배치되는 등 공간 기하학이 완전히 파괴됨."
        ],
        "physics": "서 있는 자세와 팔을 뻗은 동작 자체는 지지대가 확보되어 있으나, 왼손 손가락이 손전등을 감싸 쥐는 형태가 해부학적으로 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간 배치를 완벽히 유지하지는 못했으나, 프롬프트가 엄격하게 요구한 피사체의 좌측 하단 배치와 우측 상단의 입구를 향해 대각선으로 뻗은 팔의 구도를 가장 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시된 구도를 무시하고 통로를 화면 중앙에 배치했으며, 뻗은 팔이 통로가 아닌 우측 벽을 향하고 있어 방향성과 프레이밍 모두 실패했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "신부의 오른팔이 화면 우측에 위치한 어두운 통로 입구를 향해 대각선으로 정확히 뻗어 있으며, 고개와 시선 역시 그곳을 향하고 있습니다. 손끝은 입구의 어두운 영역 직전에 멈춰 있습니다.",
        "built_space": "화면 우측에 위로 향하는 계단 입구가 있고 좌측에 마리아상이 있는 책장이 보입니다. 하지만 레퍼런스 이미지에서 책장은 계단 반대편 벽에 위치하므로, 두 요소가 나란히 있는 것은 원본 공간 구조와 맞지 않습니다.",
        "entities": "60대 한국인 남성 신부가 짧은 백발에 하얀 로만칼라와 낡은 검은색 사제복을 입고 있으며, 요구된 대로 왼손에 손전등을 들고 있습니다.",
        "hard_violations": [
         "공간 배치 오류: 레퍼런스 샷에서 계단 반대편에 위치했던 마리아상 책장이 계단 입구 바로 옆으로 이동하여 물리적으로 불가능한 무대 구조가 형성됨."
        ],
        "physics": "신부가 두 발로 바닥을 딛고 서서 왼손으로 손전등을 단단히 쥐고 있으며, 오른팔을 자연스럽게 뻗어 지탱하는 등 물리적 오류 없이 안정적인 자세를 보여줍니다."
       },
       {
        "label": "B",
        "direction": "신부의 오른팔이 수평에 가깝게 우측을 향해 뻗어 있으나, 통로 입구가 화면 중앙에 있어 손끝이 엉뚱한 우측 벽면을 가리키고 있습니다.",
        "built_space": "화면 중앙에 위로 올라가는 계단이 있고 좌측에 마리아상 책장, 우측에 성화 액자가 배치되어 있습니다. 이는 레퍼런스의 방 구조를 완전히 해체하여 임의로 재조합한 형태입니다.",
        "entities": "60대 한국인 남성 신부가 지시된 사제복을 입고 있으며 손전등을 들고 있으나, 손전등을 쥔 왼손 엄지손가락의 형태와 관절 묘사가 어색합니다.",
        "hard_violations": [
         "공간 배치 오류: 레퍼런스 샷에서 서로 멀리 떨어져 있던 마리아상 책장과 성화 액자가 계단 입구를 둘러싸고 나란히 배치되는 등 공간 기하학이 완전히 파괴됨."
        ],
        "physics": "서 있는 자세와 팔을 뻗은 동작 자체는 지지대가 확보되어 있으나, 왼손 손가락이 손전등을 감싸 쥐는 형태가 해부학적으로 부자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "신부의 시선과 손끝은 실제 계단 입구를 향하지만, 입구가 오른쪽 중앙까지 크게 내려오고 팔의 상승 대각선도 B보다 약해 지정 구도에 덜 정확하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "후측면 상반신에서 오른쪽 위의 어두운 입구로 이어지는 팔의 대각선과 입구를 가리지 않는 손끝이 지정된 탈출 지시 구도를 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부는 고개를 오른쪽 계단 입구로 돌리고, 뻗은 검지 역시 문 안쪽의 어두운 계단을 향한다. 손끝은 개구부 왼쪽 가장자리 직전에 멈춰 입구를 가리지 않는다. 팔은 화면 오른쪽 위로 향하지만 상승 각도가 비교적 완만하다. 다른 손의 손전등 전면은 오른쪽 위를 향하며 입구를 밝히는 뚜렷한 광선은 없다.",
        "built_space": "오른쪽에 위로 올라가는 계단 입구 하나와 벽면 난간 하나가 보인다. 왼쪽에는 책장 하나, 그 위 성모상 하나와 꽃병 하나, 등불 하나가 있고 입구 오른쪽 벽에 등불 하나와 성화 액자 하나가 있다. 의자는 왼쪽 하나와 오른쪽 두 개가 부분적으로 보이며 탁자와 러그도 일부 보인다. 낡은 벽, 배관, 목재 가구와 따뜻한 등불은 참조와 유사하지만, 성모상이 놓인 책장과 계단 입구의 인접 배치는 이전 장면의 공간 배치를 정확히 재현했다고 보기 어렵다. 신부는 입구 앞 전경에 서 있고 통행 공간을 막는 고정물은 없다.",
        "entities": "인물은 신부 한 명뿐이다. 보이는 귀와 옆얼굴, 짧은 회색 머리, 체격은 참조의 고령 한국인 남성에 대체로 부합한다. 검은 사제복과 흰 로만칼라가 보이고 한 손에는 검은 손전등이 있다. 뒷모습 위주라 얼굴의 세부 동일성은 확인하기 어렵다. 다른 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "지시하는 팔은 어깨에서 자연스럽게 뻗어 있고 손전등은 반대 손의 손가락으로 잡혀 있다. 하체와 발은 화면 밖이므로 접지는 직접 보이지 않지만, 상체는 정상적인 선 자세이며 공중에 뜬 징후가 없다. 책장 위 물건과 의자 위 책은 받침면에 놓여 있고 벽 등불은 걸쇠에 매달려 있다."
       },
       {
        "label": "B",
        "direction": "신부의 고개와 주의는 오른쪽 위의 어두운 계단 입구를 향한다. 어깨에서 검지까지 이어지는 팔의 대각선을 연장하면 개구부 안쪽에 도달하며, 손끝 자체는 왼쪽 문설주 앞에서 멈춰 통로를 가리지 않는다. 손전등은 다른 손에 들려 오른쪽 위를 향하지만 통로 내부를 강하게 비추지 않는다.",
        "built_space": "오른쪽에 계단 입구 하나와 안쪽 벽의 난간 하나가 있고 계단은 뒤쪽 위로 올라간다. 왼쪽에는 책장 하나, 성모상 하나, 꽃병 하나, 책장 위 등불 하나가 보인다. 작은 받침대 위 등불 하나와 입구 오른쪽 벽 등불 하나까지 등불은 세 개다. 왼쪽 아래에는 의자 일부와 러그가 보인다. 벽과 바닥의 마모, 목재와 배관, 따뜻한 실내광은 참조의 재질감을 이어 가지만 성모상 책장이 계단 바로 왼쪽에 놓인 관계는 이전 장면과 공간적 연속성이 불완전하다. 카메라는 신부의 상반신 높이에서 어깨 너머로 입구를 비스듬히 본다.",
        "entities": "신부 한 명만 등장하며 짧은 회색 머리와 고령 남성의 옆얼굴, 검은 사제복, 흰 로만칼라가 참조와 대체로 맞는다. 후측면 촬영이므로 정면 얼굴의 정확한 일치 여부는 제한적으로만 판단할 수 있다. 검은 손전등을 손에 들고 있고 추가 인물, 문자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "뻗은 팔과 검지는 가능한 관절 각도로 연결되어 있고, 반대 손은 손전등 몸통을 확실히 감싸 쥔다. 발은 프레임 밖이지만 상체의 자세에서 부유나 불가능한 지지 관계는 보이지 않는다. 성모상과 꽃병은 책장에, 작은 등불은 받침대에 놓여 있으며 오른쪽 등불은 벽 걸이에 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "신부의 시선과 손끝은 실제 계단 입구를 향하지만, 입구가 오른쪽 중앙까지 크게 내려오고 팔의 상승 대각선도 B보다 약해 지정 구도에 덜 정확하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "후측면 상반신에서 오른쪽 위의 어두운 입구로 이어지는 팔의 대각선과 입구를 가리지 않는 손끝이 지정된 탈출 지시 구도를 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 고개를 오른쪽 계단 입구로 돌리고, 뻗은 검지 역시 문 안쪽의 어두운 계단을 향한다. 손끝은 개구부 왼쪽 가장자리 직전에 멈춰 입구를 가리지 않는다. 팔은 화면 오른쪽 위로 향하지만 상승 각도가 비교적 완만하다. 다른 손의 손전등 전면은 오른쪽 위를 향하며 입구를 밝히는 뚜렷한 광선은 없다.",
        "built_space": "오른쪽에 위로 올라가는 계단 입구 하나와 벽면 난간 하나가 보인다. 왼쪽에는 책장 하나, 그 위 성모상 하나와 꽃병 하나, 등불 하나가 있고 입구 오른쪽 벽에 등불 하나와 성화 액자 하나가 있다. 의자는 왼쪽 하나와 오른쪽 두 개가 부분적으로 보이며 탁자와 러그도 일부 보인다. 낡은 벽, 배관, 목재 가구와 따뜻한 등불은 참조와 유사하지만, 성모상이 놓인 책장과 계단 입구의 인접 배치는 이전 장면의 공간 배치를 정확히 재현했다고 보기 어렵다. 신부는 입구 앞 전경에 서 있고 통행 공간을 막는 고정물은 없다.",
        "entities": "인물은 신부 한 명뿐이다. 보이는 귀와 옆얼굴, 짧은 회색 머리, 체격은 참조의 고령 한국인 남성에 대체로 부합한다. 검은 사제복과 흰 로만칼라가 보이고 한 손에는 검은 손전등이 있다. 뒷모습 위주라 얼굴의 세부 동일성은 확인하기 어렵다. 다른 인물이나 자막은 없다.",
        "hard_violations": [],
        "physics": "지시하는 팔은 어깨에서 자연스럽게 뻗어 있고 손전등은 반대 손의 손가락으로 잡혀 있다. 하체와 발은 화면 밖이므로 접지는 직접 보이지 않지만, 상체는 정상적인 선 자세이며 공중에 뜬 징후가 없다. 책장 위 물건과 의자 위 책은 받침면에 놓여 있고 벽 등불은 걸쇠에 매달려 있다."
       },
       {
        "label": "A",
        "direction": "신부의 고개와 주의는 오른쪽 위의 어두운 계단 입구를 향한다. 어깨에서 검지까지 이어지는 팔의 대각선을 연장하면 개구부 안쪽에 도달하며, 손끝 자체는 왼쪽 문설주 앞에서 멈춰 통로를 가리지 않는다. 손전등은 다른 손에 들려 오른쪽 위를 향하지만 통로 내부를 강하게 비추지 않는다.",
        "built_space": "오른쪽에 계단 입구 하나와 안쪽 벽의 난간 하나가 있고 계단은 뒤쪽 위로 올라간다. 왼쪽에는 책장 하나, 성모상 하나, 꽃병 하나, 책장 위 등불 하나가 보인다. 작은 받침대 위 등불 하나와 입구 오른쪽 벽 등불 하나까지 등불은 세 개다. 왼쪽 아래에는 의자 일부와 러그가 보인다. 벽과 바닥의 마모, 목재와 배관, 따뜻한 실내광은 참조의 재질감을 이어 가지만 성모상 책장이 계단 바로 왼쪽에 놓인 관계는 이전 장면과 공간적 연속성이 불완전하다. 카메라는 신부의 상반신 높이에서 어깨 너머로 입구를 비스듬히 본다.",
        "entities": "신부 한 명만 등장하며 짧은 회색 머리와 고령 남성의 옆얼굴, 검은 사제복, 흰 로만칼라가 참조와 대체로 맞는다. 후측면 촬영이므로 정면 얼굴의 정확한 일치 여부는 제한적으로만 판단할 수 있다. 검은 손전등을 손에 들고 있고 추가 인물, 문자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "뻗은 팔과 검지는 가능한 관절 각도로 연결되어 있고, 반대 손은 손전등 몸통을 확실히 감싸 쥔다. 발은 프레임 밖이지만 상체의 자세에서 부유나 불가능한 지지 관계는 보이지 않는다. 성모상과 꽃병은 책장에, 작은 등불은 받침대에 놓여 있으며 오른쪽 등불은 벽 걸이에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.304
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.054
   },
   "violations": {
    "A": [
     "[gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 계단 반대편에 위치했던 마리아상 책장이 계단 입구 바로 옆으로 이동하여 물리적으로 불가능한 무대 구조가 형성됨."
    ],
    "B": [
     "[gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 서로 멀리 떨어져 있던 마리아상 책장과 성화 액자가 계단 입구를 둘러싸고 나란히 배치되는 등 공간 기하학이 완전히 파괴됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1054
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 공간 배치를 완벽히 유지하지는 못했으나, 프롬프트가 엄격하게 요구한 피사체의 좌측 하단 배치와 우측 상단의 입구를 향해 대각선으로 뻗은 팔의 구도를 가장 정확하게 구현했습니다.  ★위반: [gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 계단 반대편에 위치했던 마리아상 책장이 계단 입구 바로 옆으로 이동하여 물리적으로 불가능한 무대 구조가 형성됨."
   },
   {
    "label": "B",
    "score": 1054,
    "verdict_ko": "지시된 구도를 무시하고 통로를 화면 중앙에 배치했으며, 뻗은 팔이 통로가 아닌 우측 벽을 향하고 있어 방향성과 프레이밍 모두 실패했습니다.  ★위반: [gemini-pro] 공간 배치 오류: 레퍼런스 샷에서 서로 멀리 떨어져 있던 마리아상 책장과 성화 액자가 계단 입구를 둘러싸고 나란히 배치되는 등 공간 기하학이 완전히 파괴됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 신부 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S32sh1_sel.png",
    "asset_id": "3bfabda1-7855-4c66-996d-4d029eea52bc",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a08-85a7-7f2c-93f1-83d2952f8b33",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  }
 },
 "S33sh7::confined_fp_apt": {
  "applies": true,
  "reason_ko": "헬리콥터 조종석이라는 통제 장치가 집중된 밀폐된 공간 내부이며, 특정 좌석에 앉은 인물이 아래쪽을 향해 방향성을 가지고 삿대질하는 장면이므로 인물의 위치 및 방향 설정이 정확해야 합니다.",
  "input_fingerprint": "87e67cd2721bd252"
 },
 "S33sh7::signage": {
  "fp": "373adcb65fef627f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::44f9d884a39b": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_44f9d884a39b.png",
  "place_text": "At the front passenger position inside a combat helicopter above the refugee settlement. The dark exterior is visible through the forward glazing, with only minimal cockpit illumination.",
  "input_fingerprint": "258a0f00c1faa094"
 },
 "S33sh7::confined_fp": {
  "reads": {
   "controls": "A Control Stick is attached to the Pilot seat (front left).",
   "mirrors": "No mirrors or reflective surfaces are indicated on the diagram.",
   "camera": "The camera is positioned inboard, between the front and rear seats, pointing forward and right towards the Front Passenger seat.",
   "occupants": "The Front Passenger seat is occupied by 박철진. The Pilot, Left Crew, and Right Crew seats are unoccupied."
  },
  "mismatches": [],
  "scene_description_en": "The camera is positioned inboard within the helicopter cockpit, directed towards the front passenger seat on the right. Park Chul-jin occupies this seat, appearing in the upper-left foreground in a near-profile view of his left face, looking rightward and down. His arm extends forcefully across the frame, pointing downwards into the lower-right foreground. The inboard side of the passenger seat is visible closely behind his torso on the left. In the background on the right, the dark exterior is visible through a section of the forward glazing. No other seats, controls, or mirrors are visible in this framing.",
  "fixed": false,
  "input_fingerprint": "f6bb382a492f8c57"
 },
 "S33sh7": {
  "input_fingerprint": "77381ec18f6e021a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 향해 거칠게 삿대질한 채 매섭게 공격을 지시하는 박철진의 옆얼굴.\n\nLOCATION (lock): At the front passenger position inside a combat helicopter above the refugee settlement. The dark exterior is visible through the forward glazing, with only minimal cockpit illumination. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static position inside the helicopter, inboard of 박철진's passenger seat at seated chest height, with a slight upward angle onto his near-profile face. Place his head in the upper-left portion and retain his forcefully extended forearm and downward-pointing finger across the lower-right portion without enlarging the hand through foreground perspective. He leans into the command with his gaze directed down toward the activity below and outside the frame, preserving the established helicopter side view.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Helicopter passenger seat (Occupied by 박철진) — A partial inboard side is visible behind his torso; used as Establish the confined airborne command position without competing with his gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained illumination appropriate to the nighttime helicopter interior, maintaining readable facial tension without inventing colored instrument light or flashing effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The combat helicopter is airborne over the camp, with a convoy traveling along the seawall road below; the seawall retains its cracks and seepage. The church's basement access has been discovered, and the management office remains illuminated. 박철진: He occupies the helicopter's passenger seat and is directing the operation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned inboard within the helicopter cockpit, directed towards the front passenger seat on the right. Park Chul-jin occupies this seat, appearing in the upper-left foreground in a near-profile view of his left face, looking rightward and down. His arm extends forcefully across the frame, pointing downwards into the lower-right foreground. The inboard side of the passenger seat is visible closely behind his torso on the left. In the background on the right, the dark exterior is visible through a section of the forward glazing. No other seats, controls, or mirrors are visible in this framing.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 향해 거칠게 삿대질한 채 매섭게 공격을 지시하는 박철진의 옆얼굴.\n\nLOCATION (lock): At the front passenger position inside a combat helicopter above the refugee settlement. The dark exterior is visible through the forward glazing, with only minimal cockpit illumination. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained illumination appropriate to the nighttime helicopter interior, maintaining readable facial tension without inventing colored instrument light or flashing effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The combat helicopter is airborne over the camp, with a convoy traveling along the seawall road below; the seawall retains its cracks and seepage. The church's basement access has been discovered, and the management office remains illuminated. 박철진: He occupies the helicopter's passenger seat and is directing the operation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned inboard within the helicopter cockpit, directed towards the front passenger seat on the right. Park Chul-jin occupies this seat, appearing in the upper-left foreground in a near-profile view of his left face, looking rightward and down. His arm extends forcefully across the frame, pointing downwards into the lower-right foreground. The inboard side of the passenger seat is visible closely behind his torso on the left. In the background on the right, the dark exterior is visible through a section of the forward glazing. No other seats, controls, or mirrors are visible in this framing.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 향해 거칠게 삿대질한 채 매섭게 공격을 지시하는 박철진의 옆얼굴.\n\nLOCATION (lock): At the front passenger position inside a combat helicopter above the refugee settlement. The dark exterior is visible through the forward glazing, with only minimal cockpit illumination. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained illumination appropriate to the nighttime helicopter interior, maintaining readable facial tension without inventing colored instrument light or flashing effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The combat helicopter is airborne over the camp, with a convoy traveling along the seawall road below; the seawall retains its cracks and seepage. The church's basement access has been discovered, and the management office remains illuminated. 박철진: He occupies the helicopter's passenger seat and is directing the operation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "박철진",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "박철진",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 날카로운 시선과 거칠게 뻗은 손가락이 창밖 아래쪽의 방파제 도로와 정착촌을 정확하게 향하고 있습니다.",
    "built_space": "프롬프트의 도면 및 텍스트 지시대로 헬기 내부 조수석 안쪽에서 창밖을 내다보는 위치에 카메라가 안착되어 있으며, 인물 뒤로 조수석 시트가 올바르게 배치되어 있습니다.",
    "entities": "박철진의 얼굴과 복장(모자, 셔츠)이 레퍼런스와 일치합니다(완장의 마크 형태만 다소 다름). 창밖 배경에는 지시된 대로 방파제 도로 위를 달리는 차량 행렬과 십자가가 있는 조명 켜진 교회가 명확하게 존재합니다.",
    "hard_violations": [],
    "physics": "조수석에 올바르게 앉아 몸을 지탱하고 있으며, 아래로 뻗은 팔과 손의 동작이 물리적, 해부학적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선과 뻗은 손가락이 창밖 아래쪽을 향하고 있습니다.",
    "built_space": "헬기 내부 조수석 공간을 잘 구성하였고, 요구된 카메라 앵글과 헬기 측면 창문의 프레이밍이 지시사항에 부합합니다.",
    "entities": "인물의 얼굴과 완장의 십자 문양은 레퍼런스와 잘 맞으나, 레퍼런스에 없는 두꺼운 검은색 전술 하네스(스트랩)를 가슴에 추가로 착용하고 있습니다. 창밖 배경에는 방파제가 보이나 차량 행렬은 찾아볼 수 없습니다.",
    "hard_violations": [],
    "physics": "자세는 지탱되고 있으나, 아래를 가리키는 검지손가락의 마디가 비정상적으로 길고 고무처럼 휘어져 있어 해부학적인 오류가 관찰됩니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "카메라 구도, 인물의 자세 및 복장이 매우 정확하며, 특히 창밖의 차량 행렬(convoy)과 조명이 켜진 교회 등 디테일한 배경 지시사항을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 앵글은 좋으나 레퍼런스에 존재하지 않는 전술 하네스를 착용시켰고, 요구된 배경 요소인 차량 행렬이 누락되었으며 손가락의 묘사가 다소 기괴합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 날카로운 시선과 거칠게 뻗은 손가락이 창밖 아래쪽의 방파제 도로와 정착촌을 정확하게 향하고 있습니다.",
        "built_space": "프롬프트의 도면 및 텍스트 지시대로 헬기 내부 조수석 안쪽에서 창밖을 내다보는 위치에 카메라가 안착되어 있으며, 인물 뒤로 조수석 시트가 올바르게 배치되어 있습니다.",
        "entities": "박철진의 얼굴과 복장(모자, 셔츠)이 레퍼런스와 일치합니다(완장의 마크 형태만 다소 다름). 창밖 배경에는 지시된 대로 방파제 도로 위를 달리는 차량 행렬과 십자가가 있는 조명 켜진 교회가 명확하게 존재합니다.",
        "hard_violations": [],
        "physics": "조수석에 올바르게 앉아 몸을 지탱하고 있으며, 아래로 뻗은 팔과 손의 동작이 물리적, 해부학적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선과 뻗은 손가락이 창밖 아래쪽을 향하고 있습니다.",
        "built_space": "헬기 내부 조수석 공간을 잘 구성하였고, 요구된 카메라 앵글과 헬기 측면 창문의 프레이밍이 지시사항에 부합합니다.",
        "entities": "인물의 얼굴과 완장의 십자 문양은 레퍼런스와 잘 맞으나, 레퍼런스에 없는 두꺼운 검은색 전술 하네스(스트랩)를 가슴에 추가로 착용하고 있습니다. 창밖 배경에는 방파제가 보이나 차량 행렬은 찾아볼 수 없습니다.",
        "hard_violations": [],
        "physics": "자세는 지탱되고 있으나, 아래를 가리키는 검지손가락의 마디가 비정상적으로 길고 고무처럼 휘어져 있어 해부학적인 오류가 관찰됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "카메라 구도, 인물의 자세 및 복장이 매우 정확하며, 특히 창밖의 차량 행렬(convoy)과 조명이 켜진 교회 등 디테일한 배경 지시사항을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 앵글은 좋으나 레퍼런스에 존재하지 않는 전술 하네스를 착용시켰고, 요구된 배경 요소인 차량 행렬이 누락되었으며 손가락의 묘사가 다소 기괴합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 날카로운 시선과 거칠게 뻗은 손가락이 창밖 아래쪽의 방파제 도로와 정착촌을 정확하게 향하고 있습니다.",
        "built_space": "프롬프트의 도면 및 텍스트 지시대로 헬기 내부 조수석 안쪽에서 창밖을 내다보는 위치에 카메라가 안착되어 있으며, 인물 뒤로 조수석 시트가 올바르게 배치되어 있습니다.",
        "entities": "박철진의 얼굴과 복장(모자, 셔츠)이 레퍼런스와 일치합니다(완장의 마크 형태만 다소 다름). 창밖 배경에는 지시된 대로 방파제 도로 위를 달리는 차량 행렬과 십자가가 있는 조명 켜진 교회가 명확하게 존재합니다.",
        "hard_violations": [],
        "physics": "조수석에 올바르게 앉아 몸을 지탱하고 있으며, 아래로 뻗은 팔과 손의 동작이 물리적, 해부학적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선과 뻗은 손가락이 창밖 아래쪽을 향하고 있습니다.",
        "built_space": "헬기 내부 조수석 공간을 잘 구성하였고, 요구된 카메라 앵글과 헬기 측면 창문의 프레이밍이 지시사항에 부합합니다.",
        "entities": "인물의 얼굴과 완장의 십자 문양은 레퍼런스와 잘 맞으나, 레퍼런스에 없는 두꺼운 검은색 전술 하네스(스트랩)를 가슴에 추가로 착용하고 있습니다. 창밖 배경에는 방파제가 보이나 차량 행렬은 찾아볼 수 없습니다.",
        "hard_violations": [],
        "physics": "자세는 지탱되고 있으나, 아래를 가리키는 검지손가락의 마디가 비정상적으로 길고 고무처럼 휘어져 있어 해부학적인 오류가 관찰됩니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "좌상단 옆얼굴과 우하단으로 강하게 뻗은 팔, 명확한 하향 삿대질이 핵심 지시에 더 충실하나, 전방 유리와 약한 올려다보기 시점은 불분명하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "좌석에서 몸을 기울여 명령하는 연기는 맞지만, 시선과 검지가 전방에 가까워 아래쪽을 거칠게 지시하는 동작이 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 오른쪽을 향한 거의 옆모습이고 눈은 오른쪽 아래 창밖으로 향한다. 팔은 우하단으로 뻗고 검지는 손목에서 더 가파르게 아래로 꺾여 창밖 수면 방향을 가리킨다. 특정 공격 대상은 식별되지 않지만 아래쪽을 지시한다는 방향은 명확하다. 총기나 다른 지시 물체는 없다.",
        "built_space": "인물 뒤 왼쪽 가장자리에 좌석 등받이 한 개의 일부가 보이고, 오른쪽에는 금속 테두리로 둘러싸인 큰 유리창 한 면이 보인다. 인물은 등받이 앞에서 몸을 앞으로 기울이고 있으며 중복 좌석이나 다른 탑승자는 없다. 내부에서 옆얼굴을 보는 배치는 맞지만, 창이 전방 유리인지 측면 유리인지는 확정하기 어렵다. 카메라는 뚜렷한 낮은 시점보다는 얼굴 높이에 가까워 보인다. 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리가 모자 아래 드러나는 중년 동아시아계 남성 한 명이며, 얼굴과 체격은 박철진 참고와 대체로 일치한다. 남색 전투복, 짙은 챙모자, 흰 표식이 있는 붉은 상완 완장이 참고에 부합한다. 몸통에는 좌석 안전띠가 보인다. 창밖에는 밤의 해안 건물과 도로 불빛, 수면이 보이지만 교회·관리사무소·차량 행렬은 명확히 구별되지 않는다. 이들의 세부를 보여주기 위한 별도 확대나 삽입 화면은 없다.",
        "hard_violations": [],
        "physics": "몸통 뒤 좌석과 몸에 걸린 안전띠가 착석 상태를 뒷받침한다. 엉덩이는 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 뻗은 팔은 어깨와 팔꿈치에 정상적으로 연결되고 손목과 검지의 하향 자세도 실제로 가능한 동작이다. 손은 과도하게 확대되지 않았으며 창틀을 관통하는 모습도 없다."
       },
       {
        "label": "B",
        "direction": "얼굴은 오른쪽을 향하고 시선은 창밖 전방에서 약간 아래쪽으로 향한다. 팔은 오른쪽 아래로 뻗지만 검지는 거의 오른쪽을 향하면서 조금 내려가며, 손끝 앞에는 방조제 아래 수면이 보인다. 아래를 노려보며 가파르게 삿대질하기보다는 전방을 지시하는 인상이 더 강하다. 특정 공격 대상은 식별되지 않는다.",
        "built_space": "왼쪽 뒤에 좌석 등받이 한 개가 부분적으로 보이고 오른쪽에는 금속 창틀 안의 큰 유리창 한 면이 있다. 남성은 그 좌석 앞에서 몸을 기울이며, 추가 탑승자나 중복된 좌석은 보이지 않는다. 내부 측면 관찰 구도는 성립하지만 전방 유리와 앞 승객석의 관계는 확실하지 않다. 약한 올려다보기보다는 얼굴과 비슷한 높이에서 본 인상이 있다. 창밖 방조제와 정착지가 A보다 선명하게 드러나며 불가능한 반사는 없다.",
        "entities": "중년 동아시아계 남성 한 명의 얼굴, 모자 아래 짧은 검은 머리, 남색 전투복과 붉은 완장은 참고의 박철진과 대체로 맞는다. 모자와 완장의 흰 표식도 유지된다. 창밖에는 정착지, 십자가가 있는 교회 건물, 불 켜진 건물들과 방조제 도로 위 여러 차량이 보인다. 특정 건물을 관리사무소라고 확인할 근거와 지하실 입구는 보이지 않는다. 자막이나 도면 표시는 없다.",
        "hard_violations": [],
        "physics": "뒤쪽 등받이와 낮게 잘린 몸통 배치가 좌석에 앉아 앞으로 기울인 자세로 읽힌다. 하체가 잘려 지지면 자체는 보이지 않지만 무지지 부유로 보이지는 않는다. 팔과 손은 어깨에서 정상적으로 이어지며 지시 동작으로 가능한 자세다. 창밖 차량들은 방조제 도로 위에 놓여 있고 건물도 지면에 연결되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "좌상단 옆얼굴과 우하단으로 강하게 뻗은 팔, 명확한 하향 삿대질이 핵심 지시에 더 충실하나, 전방 유리와 약한 올려다보기 시점은 불분명하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "좌석에서 몸을 기울여 명령하는 연기는 맞지만, 시선과 검지가 전방에 가까워 아래쪽을 거칠게 지시하는 동작이 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 오른쪽을 향한 거의 옆모습이고 눈은 오른쪽 아래 창밖으로 향한다. 팔은 우하단으로 뻗고 검지는 손목에서 더 가파르게 아래로 꺾여 창밖 수면 방향을 가리킨다. 특정 공격 대상은 식별되지 않지만 아래쪽을 지시한다는 방향은 명확하다. 총기나 다른 지시 물체는 없다.",
        "built_space": "인물 뒤 왼쪽 가장자리에 좌석 등받이 한 개의 일부가 보이고, 오른쪽에는 금속 테두리로 둘러싸인 큰 유리창 한 면이 보인다. 인물은 등받이 앞에서 몸을 앞으로 기울이고 있으며 중복 좌석이나 다른 탑승자는 없다. 내부에서 옆얼굴을 보는 배치는 맞지만, 창이 전방 유리인지 측면 유리인지는 확정하기 어렵다. 카메라는 뚜렷한 낮은 시점보다는 얼굴 높이에 가까워 보인다. 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리가 모자 아래 드러나는 중년 동아시아계 남성 한 명이며, 얼굴과 체격은 박철진 참고와 대체로 일치한다. 남색 전투복, 짙은 챙모자, 흰 표식이 있는 붉은 상완 완장이 참고에 부합한다. 몸통에는 좌석 안전띠가 보인다. 창밖에는 밤의 해안 건물과 도로 불빛, 수면이 보이지만 교회·관리사무소·차량 행렬은 명확히 구별되지 않는다. 이들의 세부를 보여주기 위한 별도 확대나 삽입 화면은 없다.",
        "hard_violations": [],
        "physics": "몸통 뒤 좌석과 몸에 걸린 안전띠가 착석 상태를 뒷받침한다. 엉덩이는 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 뻗은 팔은 어깨와 팔꿈치에 정상적으로 연결되고 손목과 검지의 하향 자세도 실제로 가능한 동작이다. 손은 과도하게 확대되지 않았으며 창틀을 관통하는 모습도 없다."
       },
       {
        "label": "A",
        "direction": "얼굴은 오른쪽을 향하고 시선은 창밖 전방에서 약간 아래쪽으로 향한다. 팔은 오른쪽 아래로 뻗지만 검지는 거의 오른쪽을 향하면서 조금 내려가며, 손끝 앞에는 방조제 아래 수면이 보인다. 아래를 노려보며 가파르게 삿대질하기보다는 전방을 지시하는 인상이 더 강하다. 특정 공격 대상은 식별되지 않는다.",
        "built_space": "왼쪽 뒤에 좌석 등받이 한 개가 부분적으로 보이고 오른쪽에는 금속 창틀 안의 큰 유리창 한 면이 있다. 남성은 그 좌석 앞에서 몸을 기울이며, 추가 탑승자나 중복된 좌석은 보이지 않는다. 내부 측면 관찰 구도는 성립하지만 전방 유리와 앞 승객석의 관계는 확실하지 않다. 약한 올려다보기보다는 얼굴과 비슷한 높이에서 본 인상이 있다. 창밖 방조제와 정착지가 A보다 선명하게 드러나며 불가능한 반사는 없다.",
        "entities": "중년 동아시아계 남성 한 명의 얼굴, 모자 아래 짧은 검은 머리, 남색 전투복과 붉은 완장은 참고의 박철진과 대체로 맞는다. 모자와 완장의 흰 표식도 유지된다. 창밖에는 정착지, 십자가가 있는 교회 건물, 불 켜진 건물들과 방조제 도로 위 여러 차량이 보인다. 특정 건물을 관리사무소라고 확인할 근거와 지하실 입구는 보이지 않는다. 자막이나 도면 표시는 없다.",
        "hard_violations": [],
        "physics": "뒤쪽 등받이와 낮게 잘린 몸통 배치가 좌석에 앉아 앞으로 기울인 자세로 읽힌다. 하체가 잘려 지지면 자체는 보이지 않지만 무지지 부유로 보이지는 않는다. 팔과 손은 어깨에서 정상적으로 이어지며 지시 동작으로 가능한 자세다. 창밖 차량들은 방조제 도로 위에 놓여 있고 건물도 지면에 연결되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1667
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "카메라 구도, 인물의 자세 및 복장이 매우 정확하며, 특히 창밖의 차량 행렬(convoy)과 조명이 켜진 교회 등 디테일한 배경 지시사항을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "구도와 앵글은 좋으나 레퍼런스에 존재하지 않는 전술 하네스를 착용시켰고, 요구된 배경 요소인 차량 행렬이 누락되었으며 손가락의 묘사가 다소 기괴합니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh7_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "박철진",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a0e-75bf-7d60-810e-b880ed888d05",
  "confined_fp": {
   "base_key": "confinedfp::44f9d884a39b",
   "apt_reason": "헬리콥터 조종석이라는 통제 장치가 집중된 밀폐된 공간 내부이며, 특정 좌석에 앉은 인물이 아래쪽을 향해 방향성을 가지고 삿대질하는 장면이므로 인물의 위치 및 방향 설정이 정확해야 합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S33sh12::signage": {
  "fp": "0bd9b075aeef726c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::admin_yard_steps": {
  "input_fingerprint": "0a6f8ec0dfe67b2d",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "admin_yard_steps",
    "tags": [
     "S33sh12"
    ]
   },
   "context_sig": "e08dc8abb219dfa7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 인공제방·보수 공사장, 제방 도로: 바닷물을 막기 위해 세워진 거대한 콘크리트 벽과 지지대로, 금이 가고 물이 새는 노후된 구조물이다. (특징: 벽면 곳곳에 심하게 금이 가고 물줄기가 새어 나오는 거대한 콘크리트 인공제방; 자재를 옮기는 소형 지게차를 운전하는 미연(현우와 앰버의 엄마); 로만칼라 복장을 입은 60대 남자 신부; 군용 트럭과 깔끔한 정장·구두를 착용한 관리소장 및 무장 정규 경비병; 제방 벽에 다닥다닥 달라붙어 연쇄 폭발을 일으키는 자폭 드론들; 상단부가 터져나가며 쏟아져 들어오는 흙탕물과 휩쓸리는 중장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /관리사무소 마당\n- 그 틈을 타 잽싸게 사무소 계단을 빠르게 올라가는 현우.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 인공제방·보수 공사장, 제방 도로: 바닷물을 막기 위해 세워진 거대한 콘크리트 벽과 지지대로, 금이 가고 물이 새는 노후된 구조물이다. (특징: 벽면 곳곳에 심하게 금이 가고 물줄기가 새어 나오는 거대한 콘크리트 인공제방; 자재를 옮기는 소형 지게차를 운전하는 미연(현우와 앰버의 엄마); 로만칼라 복장을 입은 60대 남자 신부; 군용 트럭과 깔끔한 정장·구두를 착용한 관리소장 및 무장 정규 경비병; 제방 벽에 다닥다닥 달라붙어 연쇄 폭발을 일으키는 자폭 드론들; 상단부가 터져나가며 쏟아져 들어오는 흙탕물과 휩쓸리는 중장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /관리사무소 마당\n- 그 틈을 타 잽싸게 사무소 계단을 빠르게 올라가는 현우.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_yard_steps_e00afd.png",
  "asset_id": "783fc97e-558e-4b67-9322-6cfaf7c82440",
  "input_asset_ids": [
   "3e589345-49dc-482d-9456-b9f786948626"
  ],
  "origin_tag": "S33sh12",
  "place_text": "On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.",
  "origin_inputs": {
   "place_text": "On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.",
   "time_of_day_en": "night",
   "conti_asset_id": "3e589345-49dc-482d-9456-b9f786948626"
  }
 },
 "S33sh12::bgfirst_bg": {
  "input_fingerprint": "ba917b84f5344feb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 출동을 서두르는 경비병들의 등 뒤로 이현우가 한 발로 관리사무소 계단을 딛고 몸이 붕 떠오른 mid-action 자세의 역동적인 측면.\n\nLOCATION (lock): On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral track parallel to 이현우's route, looking slightly upward from behind the departing guards while retaining his full side-view action at the stair approach. Keep the guards' backs across the left foreground and 이현우 unobstructed at right-center, one foot driving against the first stair while his trailing foot lifts and his torso rises toward the upper-right continuation. The guards attend to boarding the truck in staggered phases—one transferring weight upward, another approaching with a different stride—while 이현우 watches his next stair, making his movement behind their backs the single emphasized positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: first management-office stair beneath 이현우's planted foot in the lower-right of the frame, midground; upward continuation of management-office stairs in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Management-office stairs (Being ascended by 이현우) — Viewed from the side, rising toward the upper right; used as Expose the first planted foot and provide a clear continuation for his escape movement; Truck (Being boarded by guards preparing to depart) — Only the boarding-side portion is retained at the left edge; used as Anchor the guards' activity and explain why their backs are turned to 이현우.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient illumination with sufficient separation to read 이현우's planted foot, rising body, and the foreground guards without introducing a new visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 출동을 서두르는 경비병들의 등 뒤로 이현우가 한 발로 관리사무소 계단을 딛고 몸이 붕 떠오른 mid-action 자세의 역동적인 측면.\n\nLOCATION (lock): On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral track parallel to 이현우's route, looking slightly upward from behind the departing guards while retaining his full side-view action at the stair approach. Keep the guards' backs across the left foreground and 이현우 unobstructed at right-center, one foot driving against the first stair while his trailing foot lifts and his torso rises toward the upper-right continuation. The guards attend to boarding the truck in staggered phases—one transferring weight upward, another approaching with a different stride—while 이현우 watches his next stair, making his movement behind their backs the single emphasized positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: first management-office stair beneath 이현우's planted foot in the lower-right of the frame, midground; upward continuation of management-office stairs in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Management-office stairs (Being ascended by 이현우) — Viewed from the side, rising toward the upper right; used as Expose the first planted foot and provide a clear continuation for his escape movement; Truck (Being boarded by guards preparing to depart) — Only the boarding-side portion is retained at the left edge; used as Anchor the guards' activity and explain why their backs are turned to 이현우.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient illumination with sufficient separation to read 이현우's planted foot, rising body, and the foreground guards without introducing a new visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh12__bgfirst_bg.png",
  "asset_id": "1dae5d47-29d7-4487-b106-e9cb5e14e5b4",
  "input_asset_ids": [
   "3e589345-49dc-482d-9456-b9f786948626",
   "783fc97e-558e-4b67-9322-6cfaf7c82440"
  ]
 },
 "S33sh12": {
  "input_fingerprint": "ed26da89e79d9b18",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 출동을 서두르는 경비병들의 등 뒤로 이현우가 한 발로 관리사무소 계단을 딛고 몸이 붕 떠오른 mid-action 자세의 역동적인 측면.\n\nLOCATION (lock): On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral track parallel to 이현우's route, looking slightly upward from behind the departing guards while retaining his full side-view action at the stair approach. Keep the guards' backs across the left foreground and 이현우 unobstructed at right-center, one foot driving against the first stair while his trailing foot lifts and his torso rises toward the upper-right continuation. The guards attend to boarding the truck in staggered phases—one transferring weight upward, another approaching with a different stride—while 이현우 watches his next stair, making his movement behind their backs the single emphasized positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: first management-office stair beneath 이현우's planted foot in the lower-right of the frame, midground; upward continuation of management-office stairs in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Management-office stairs (Being ascended by 이현우) — Viewed from the side, rising toward the upper right; used as Expose the first planted foot and provide a clear continuation for his escape movement; Truck (Being boarded by guards preparing to depart) — Only the boarding-side portion is retained at the left edge; used as Anchor the guards' activity and explain why their backs are turned to 이현우.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient illumination with sufficient separation to read 이현우's planted foot, rising body, and the foreground guards without introducing a new visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the illuminated management office is open, and deployment trucks are leaving the grounds. The third-floor detention-room door remains locked and its handle is still intact. 이현우: He is ascending the management-office stairs with facial bruises and the persistent leg injury; his outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 출동을 서두르는 경비병들의 등 뒤로 이현우가 한 발로 관리사무소 계단을 딛고 몸이 붕 떠오른 mid-action 자세의 역동적인 측면.\n\nLOCATION (lock): On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral track parallel to 이현우's route, looking slightly upward from behind the departing guards while retaining his full side-view action at the stair approach. Keep the guards' backs across the left foreground and 이현우 unobstructed at right-center, one foot driving against the first stair while his trailing foot lifts and his torso rises toward the upper-right continuation. The guards attend to boarding the truck in staggered phases—one transferring weight upward, another approaching with a different stride—while 이현우 watches his next stair, making his movement behind their backs the single emphasized positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: first management-office stair beneath 이현우's planted foot in the lower-right of the frame, midground; upward continuation of management-office stairs in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Management-office stairs (Being ascended by 이현우) — Viewed from the side, rising toward the upper right; used as Expose the first planted foot and provide a clear continuation for his escape movement; Truck (Being boarded by guards preparing to depart) — Only the boarding-side portion is retained at the left edge; used as Anchor the guards' activity and explain why their backs are turned to 이현우.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient illumination with sufficient separation to read 이현우's planted foot, rising body, and the foreground guards without introducing a new visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the illuminated management office is open, and deployment trucks are leaving the grounds. The third-floor detention-room door remains locked and its handle is still intact. 이현우: He is ascending the management-office stairs with facial bruises and the persistent leg injury; his outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 출동을 서두르는 경비병들의 등 뒤로 이현우가 한 발로 관리사무소 계단을 딛고 몸이 붕 떠오른 mid-action 자세의 역동적인 측면.\n\nLOCATION (lock): On the entrance steps leading up from the camp administration building's yard, beside the departing guards and trucks. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the low lateral track parallel to 이현우's route, looking slightly upward from behind the departing guards while retaining his full side-view action at the stair approach. Keep the guards' backs across the left foreground and 이현우 unobstructed at right-center, one foot driving against the first stair while his trailing foot lifts and his torso rises toward the upper-right continuation. The guards attend to boarding the truck in staggered phases—one transferring weight upward, another approaching with a different stride—while 이현우 watches his next stair, making his movement behind their backs the single emphasized positional change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: first management-office stair beneath 이현우's planted foot in the lower-right of the frame, midground; upward continuation of management-office stairs in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Management-office stairs (Being ascended by 이현우) — Viewed from the side, rising toward the upper right; used as Expose the first planted foot and provide a clear continuation for his escape movement; Truck (Being boarded by guards preparing to depart) — Only the boarding-side portion is retained at the left edge; used as Anchor the guards' activity and explain why their backs are turned to 이현우.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient illumination with sufficient separation to read 이현우's planted foot, rising body, and the foreground guards without introducing a new visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The manhole beside the illuminated management office is open, and deployment trucks are leaving the grounds. The third-floor detention-room door remains locked and its handle is still intact. 이현우: He is ascending the management-office stairs with facial bruises and the persistent leg injury; his outer garment remains removed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh12__bgfirst_bg.png",
     "asset_id": "1dae5d47-29d7-4487-b106-e9cb5e14e5b4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S33sh12.png",
     "asset_id": "3e589345-49dc-482d-9456-b9f786948626",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_yard_steps_e00afd.png",
     "asset_id": "783fc97e-558e-4b67-9322-6cfaf7c82440",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 몸통은 계단 위쪽 도주로를 향하고 있으며, 세 명의 경비병들은 모두 화면 좌측의 트럭 쪽으로 몸을 향하고 있음.",
    "built_space": "우측에 관리사무소 건물과 위로 이어지는 계단, 좌측에 군용 트럭, 배경에 댐과 조명 탑, 앞마당 바닥에 열린 맨홀이 레퍼런스와 일치하는 위치와 스케일로 올바르게 배치됨.",
    "entities": "이현우는 핏자국과 흙먼지가 묻은 낡은 셔츠와 바지를 입고 상처 난 얼굴을 한 10대 후반의 외모로 레퍼런스와 일치하나, 귀의 소형 인이어 무전기는 보이지 않음. 경비병들은 군복과 헬멧, 무기를 갖추고 있음.",
    "hard_violations": [],
    "physics": "이현우의 오른발이 계단을 단단히 딛고 체중을 지탱하며, 왼발은 공중에 떠 있어 도약하는 힘이 자연스럽게 표현됨. 경비병들의 보행 및 탑승 자세도 바닥과 트럭에 의해 안정적으로 지지됨. 공중에 떠 있는 물체나 신체 부위는 없음."
   },
   {
    "label": "B",
    "direction": "이현우의 시선과 뻗은 몸의 방향은 우측 계단 위를 향하고 있으며, 경비병들은 화면 좌측 트럭의 탑승구를 향해 등진 상태로 움직임.",
    "built_space": "우측의 관리사무소 건물 및 계단, 좌측에 배치된 트럭의 뒷부분, 뒷배경의 거대한 댐 구조물, 마당의 열린 맨홀 뚜껑 등 공간적 요소들이 모두 적절히 구성됨.",
    "entities": "이현우는 레퍼런스의 인상착의와 복장을 잘 따르고 있으며 귀에 소형 인이어 무전기가 명확하게 묘사됨. 경비병들 역시 지문이 요구하는 복장과 무장을 갖춤.",
    "hard_violations": [],
    "physics": "이현우의 오른발은 계단에, 왼발은 마당 바닥(경계석)에 닿아 있어 두 발 모두 지면의 지지를 받고 있음. 물리적으로 불가능한 자세는 없으나, 공중에 떠오르며 뒤쪽 발이 들린다는 지문 내용과 달리 양발이 바닥에 닿은 채 보폭만 넓힌 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "이현우의 뒤쪽 발이 허공에 들려 몸이 붕 떠오른 역동적인 도약 액션을 정확히 구현하여 높은 우선순위의 동작 지침을 훌륭하게 충족했습니다 (인이어 무전기가 생략되었으나 동작 묘사가 더 중요하게 평가됨)."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인이어 무전기 등의 디테일은 훌륭하게 살렸으나, 이현우의 뒤쪽 발이 바닥에 그대로 닿아 있어 '몸이 붕 떠오른' 공중 동작이라는 가장 핵심적인 액션 지침을 전혀 구현하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸통은 계단 위쪽 도주로를 향하고 있으며, 세 명의 경비병들은 모두 화면 좌측의 트럭 쪽으로 몸을 향하고 있음.",
        "built_space": "우측에 관리사무소 건물과 위로 이어지는 계단, 좌측에 군용 트럭, 배경에 댐과 조명 탑, 앞마당 바닥에 열린 맨홀이 레퍼런스와 일치하는 위치와 스케일로 올바르게 배치됨.",
        "entities": "이현우는 핏자국과 흙먼지가 묻은 낡은 셔츠와 바지를 입고 상처 난 얼굴을 한 10대 후반의 외모로 레퍼런스와 일치하나, 귀의 소형 인이어 무전기는 보이지 않음. 경비병들은 군복과 헬멧, 무기를 갖추고 있음.",
        "hard_violations": [],
        "physics": "이현우의 오른발이 계단을 단단히 딛고 체중을 지탱하며, 왼발은 공중에 떠 있어 도약하는 힘이 자연스럽게 표현됨. 경비병들의 보행 및 탑승 자세도 바닥과 트럭에 의해 안정적으로 지지됨. 공중에 떠 있는 물체나 신체 부위는 없음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 뻗은 몸의 방향은 우측 계단 위를 향하고 있으며, 경비병들은 화면 좌측 트럭의 탑승구를 향해 등진 상태로 움직임.",
        "built_space": "우측의 관리사무소 건물 및 계단, 좌측에 배치된 트럭의 뒷부분, 뒷배경의 거대한 댐 구조물, 마당의 열린 맨홀 뚜껑 등 공간적 요소들이 모두 적절히 구성됨.",
        "entities": "이현우는 레퍼런스의 인상착의와 복장을 잘 따르고 있으며 귀에 소형 인이어 무전기가 명확하게 묘사됨. 경비병들 역시 지문이 요구하는 복장과 무장을 갖춤.",
        "hard_violations": [],
        "physics": "이현우의 오른발은 계단에, 왼발은 마당 바닥(경계석)에 닿아 있어 두 발 모두 지면의 지지를 받고 있음. 물리적으로 불가능한 자세는 없으나, 공중에 떠오르며 뒤쪽 발이 들린다는 지문 내용과 달리 양발이 바닥에 닿은 채 보폭만 넓힌 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "이현우의 뒤쪽 발이 허공에 들려 몸이 붕 떠오른 역동적인 도약 액션을 정확히 구현하여 높은 우선순위의 동작 지침을 훌륭하게 충족했습니다 (인이어 무전기가 생략되었으나 동작 묘사가 더 중요하게 평가됨)."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인이어 무전기 등의 디테일은 훌륭하게 살렸으나, 이현우의 뒤쪽 발이 바닥에 그대로 닿아 있어 '몸이 붕 떠오른' 공중 동작이라는 가장 핵심적인 액션 지침을 전혀 구현하지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸통은 계단 위쪽 도주로를 향하고 있으며, 세 명의 경비병들은 모두 화면 좌측의 트럭 쪽으로 몸을 향하고 있음.",
        "built_space": "우측에 관리사무소 건물과 위로 이어지는 계단, 좌측에 군용 트럭, 배경에 댐과 조명 탑, 앞마당 바닥에 열린 맨홀이 레퍼런스와 일치하는 위치와 스케일로 올바르게 배치됨.",
        "entities": "이현우는 핏자국과 흙먼지가 묻은 낡은 셔츠와 바지를 입고 상처 난 얼굴을 한 10대 후반의 외모로 레퍼런스와 일치하나, 귀의 소형 인이어 무전기는 보이지 않음. 경비병들은 군복과 헬멧, 무기를 갖추고 있음.",
        "hard_violations": [],
        "physics": "이현우의 오른발이 계단을 단단히 딛고 체중을 지탱하며, 왼발은 공중에 떠 있어 도약하는 힘이 자연스럽게 표현됨. 경비병들의 보행 및 탑승 자세도 바닥과 트럭에 의해 안정적으로 지지됨. 공중에 떠 있는 물체나 신체 부위는 없음."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 뻗은 몸의 방향은 우측 계단 위를 향하고 있으며, 경비병들은 화면 좌측 트럭의 탑승구를 향해 등진 상태로 움직임.",
        "built_space": "우측의 관리사무소 건물 및 계단, 좌측에 배치된 트럭의 뒷부분, 뒷배경의 거대한 댐 구조물, 마당의 열린 맨홀 뚜껑 등 공간적 요소들이 모두 적절히 구성됨.",
        "entities": "이현우는 레퍼런스의 인상착의와 복장을 잘 따르고 있으며 귀에 소형 인이어 무전기가 명확하게 묘사됨. 경비병들 역시 지문이 요구하는 복장과 무장을 갖춤.",
        "hard_violations": [],
        "physics": "이현우의 오른발은 계단에, 왼발은 마당 바닥(경계석)에 닿아 있어 두 발 모두 지면의 지지를 받고 있음. 물리적으로 불가능한 자세는 없으나, 공중에 떠오르며 뒤쪽 발이 들린다는 지문 내용과 달리 양발이 바닥에 닿은 채 보폭만 넓힌 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 측면 와이드 구도와 경비병 뒤에서 계단을 밀고 상승하는 동작이 더 명확하지만, 디딘 곳이 첫 단이 아니고 시선도 다음 발판보다 높다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "장소와 좌우 배치는 충실하지만, 앞발이 계단 모서리에 걸쳐 지지가 불명확하고 첫 단을 힘껏 밀어 올라가는 지정 동작이 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 몸과 얼굴은 오른쪽 위로 이어지는 계단을 향한다. 다만 눈길은 바로 다음 발판보다 상부 출입구 쪽에 가깝다. 왼쪽 경비병들은 등을 보이며 트럭이나 마당 안쪽으로 이동하고 이현우를 보지 않는다. 가까운 경비병의 휴대 소총은 아래를 향하며 누구를 겨누지 않는다.",
        "built_space": "오른쪽에 콘크리트 계단 한 줄과 양쪽 금속 난간, 상부 출입문 하나, 처마 조명 하나가 보인다. 왼쪽 가장자리의 승차용 트럭 한 대와 중앙 배경의 트럭 한 대, 열린 맨홀 하나와 옆으로 기울어진 뚜껑 하나가 있다. 콘크리트 외벽·창·화단·외곽 장벽은 장소 참조와 잘 연결된다. 경비병 세 명 중 한 명은 승차 중이고 나머지는 서로 다른 보폭으로 이동한다. 이현우는 오른쪽 중앙에서 가리지 않고 보이지만, 발을 댄 단 아래에 낮은 단들이 남아 있어 지정된 첫 단 접촉은 아니다.",
        "entities": "이현우 한 명과 장면에 명시된 경비병 세 명이 보인다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 얼굴 윤곽이 참조에 대체로 맞는다. 어두운 오염된 셔츠와 바지, 귀의 소형 인이어 장치, 얼굴의 상처가 보이고 겉옷은 없다. 지속적인 다리 부상은 동작만으로 확정하기 어렵다. 경비병들은 헬멧과 군복을 착용한다. 화면 밖인 3층 구금실의 잠금 상태는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽 부츠 밑창이 계단 수평면에 닿아 체중을 지지하고, 반대쪽 발은 지면에서 들려 있다. 굽힌 앞무릎과 전방으로 기울어진 몸통은 계단을 밀어 상승하는 동작으로 성립하며 무근거한 공중 부양은 아니다. 승차 중인 경비병은 트럭 발판에 발을 두고 있고, 접근 중인 경비병들은 각각 한 발로 지면을 지지한다. 소총은 몸에 멘 끈으로 지지되며 맨홀 뚜껑은 지면과 개구부 가장자리에 걸쳐 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 위 계단 방향으로 이동하지만 얼굴과 시선은 다음 발판보다 높은 출입구 방향을 향한다. 경비병 세 명은 왼쪽 트럭과 앞쪽 마당을 향해 등을 돌리고 있다. 멘 소총들의 총구는 대체로 위를 향하고, 특정 인물을 겨누지 않는다.",
        "built_space": "오른쪽의 단일 콘크리트 계단, 양쪽 금속 난간, 출입문 하나와 처마 조명 하나가 보인다. 트럭은 왼쪽 가장자리와 중앙 배경에 한 대씩 있으며 열린 맨홀과 뚜껑도 각각 하나다. 건물 외벽과 창문, 화단, 장벽의 재료와 배치는 참조와 가깝다. 낮은 측면 와이드 구도에서 경비병들의 등이 왼쪽, 이현우의 전신이 오른쪽 중앙에 놓인다. 다만 이현우의 앞발 아래로 여러 단이 보여 첫 단을 딛는 배치는 충족하지 않는다.",
        "entities": "젊은 동아시아계 남성 이현우 한 명과 군복·헬멧 차림의 경비병 세 명이 보인다. 이현우의 검은 헝클어진 머리, 마른 체격, 어두운 낡은 셔츠와 바지는 참조에 대체로 부합한다. 얼굴과 바지에 상처 또는 오염 흔적은 있으나 인이어 장치와 지속적인 다리 부상은 A보다 판독하기 어렵다. 겉옷은 입지 않았다. 3층 구금실 문과 손잡이는 화면에 없다.",
        "hard_violations": [],
        "physics": "이현우의 뒤쪽 발은 공중에 있고, 앞쪽 부츠는 발끝 부근이 계단 모서리에 걸친 모습이다. 뒤꿈치가 발판 높이보다 아래로 내려와 밑창 전체로 밀어내는 지지는 읽히지 않지만, 앞꿈치 접촉 가능성이 있어 완전한 무지지 부양으로 단정할 수는 없다. A보다 계단에 힘을 전달하는 순간이 불명확하다. 승차 경비병은 트럭 발판으로, 나머지 경비병들은 지면에 닿은 발로 지지된다. 소총은 어깨끈으로 지지되고 맨홀 뚜껑은 바닥에 걸쳐 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 측면 와이드 구도와 경비병 뒤에서 계단을 밀고 상승하는 동작이 더 명확하지만, 디딘 곳이 첫 단이 아니고 시선도 다음 발판보다 높다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "장소와 좌우 배치는 충실하지만, 앞발이 계단 모서리에 걸쳐 지지가 불명확하고 첫 단을 힘껏 밀어 올라가는 지정 동작이 A보다 약하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 몸과 얼굴은 오른쪽 위로 이어지는 계단을 향한다. 다만 눈길은 바로 다음 발판보다 상부 출입구 쪽에 가깝다. 왼쪽 경비병들은 등을 보이며 트럭이나 마당 안쪽으로 이동하고 이현우를 보지 않는다. 가까운 경비병의 휴대 소총은 아래를 향하며 누구를 겨누지 않는다.",
        "built_space": "오른쪽에 콘크리트 계단 한 줄과 양쪽 금속 난간, 상부 출입문 하나, 처마 조명 하나가 보인다. 왼쪽 가장자리의 승차용 트럭 한 대와 중앙 배경의 트럭 한 대, 열린 맨홀 하나와 옆으로 기울어진 뚜껑 하나가 있다. 콘크리트 외벽·창·화단·외곽 장벽은 장소 참조와 잘 연결된다. 경비병 세 명 중 한 명은 승차 중이고 나머지는 서로 다른 보폭으로 이동한다. 이현우는 오른쪽 중앙에서 가리지 않고 보이지만, 발을 댄 단 아래에 낮은 단들이 남아 있어 지정된 첫 단 접촉은 아니다.",
        "entities": "이현우 한 명과 장면에 명시된 경비병 세 명이 보인다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 얼굴 윤곽이 참조에 대체로 맞는다. 어두운 오염된 셔츠와 바지, 귀의 소형 인이어 장치, 얼굴의 상처가 보이고 겉옷은 없다. 지속적인 다리 부상은 동작만으로 확정하기 어렵다. 경비병들은 헬멧과 군복을 착용한다. 화면 밖인 3층 구금실의 잠금 상태는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽 부츠 밑창이 계단 수평면에 닿아 체중을 지지하고, 반대쪽 발은 지면에서 들려 있다. 굽힌 앞무릎과 전방으로 기울어진 몸통은 계단을 밀어 상승하는 동작으로 성립하며 무근거한 공중 부양은 아니다. 승차 중인 경비병은 트럭 발판에 발을 두고 있고, 접근 중인 경비병들은 각각 한 발로 지면을 지지한다. 소총은 몸에 멘 끈으로 지지되며 맨홀 뚜껑은 지면과 개구부 가장자리에 걸쳐 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 위 계단 방향으로 이동하지만 얼굴과 시선은 다음 발판보다 높은 출입구 방향을 향한다. 경비병 세 명은 왼쪽 트럭과 앞쪽 마당을 향해 등을 돌리고 있다. 멘 소총들의 총구는 대체로 위를 향하고, 특정 인물을 겨누지 않는다.",
        "built_space": "오른쪽의 단일 콘크리트 계단, 양쪽 금속 난간, 출입문 하나와 처마 조명 하나가 보인다. 트럭은 왼쪽 가장자리와 중앙 배경에 한 대씩 있으며 열린 맨홀과 뚜껑도 각각 하나다. 건물 외벽과 창문, 화단, 장벽의 재료와 배치는 참조와 가깝다. 낮은 측면 와이드 구도에서 경비병들의 등이 왼쪽, 이현우의 전신이 오른쪽 중앙에 놓인다. 다만 이현우의 앞발 아래로 여러 단이 보여 첫 단을 딛는 배치는 충족하지 않는다.",
        "entities": "젊은 동아시아계 남성 이현우 한 명과 군복·헬멧 차림의 경비병 세 명이 보인다. 이현우의 검은 헝클어진 머리, 마른 체격, 어두운 낡은 셔츠와 바지는 참조에 대체로 부합한다. 얼굴과 바지에 상처 또는 오염 흔적은 있으나 인이어 장치와 지속적인 다리 부상은 A보다 판독하기 어렵다. 겉옷은 입지 않았다. 3층 구금실 문과 손잡이는 화면에 없다.",
        "hard_violations": [],
        "physics": "이현우의 뒤쪽 발은 공중에 있고, 앞쪽 부츠는 발끝 부근이 계단 모서리에 걸친 모습이다. 뒤꿈치가 발판 높이보다 아래로 내려와 밑창 전체로 밀어내는 지지는 읽히지 않지만, 앞꿈치 접촉 가능성이 있어 완전한 무지지 부양으로 단정할 수는 없다. A보다 계단에 힘을 전달하는 순간이 불명확하다. 승차 경비병은 트럭 발판으로, 나머지 경비병들은 지면에 닿은 발로 지지된다. 소총은 어깨끈으로 지지되고 맨홀 뚜껑은 바닥에 걸쳐 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1667
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "이현우의 뒤쪽 발이 허공에 들려 몸이 붕 떠오른 역동적인 도약 액션을 정확히 구현하여 높은 우선순위의 동작 지침을 훌륭하게 충족했습니다 (인이어 무전기가 생략되었으나 동작 묘사가 더 중요하게 평가됨)."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "인이어 무전기 등의 디테일은 훌륭하게 살렸으나, 이현우의 뒤쪽 발이 바닥에 그대로 닿아 있어 '몸이 붕 떠오른' 공중 동작이라는 가장 핵심적인 액션 지침을 전혀 구현하지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_yard_steps_e00afd.png",
    "asset_id": "783fc97e-558e-4b67-9322-6cfaf7c82440",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a17-4b1f-702c-93e5-791ae7f2938e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S33sh12__bgfirst_bg.png",
   "bg_asset_id": "1dae5d47-29d7-4487-b106-e9cb5e14e5b4",
   "bg_record_key": "S33sh12::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "admin_yard_steps",
   "groupbg_asset_id": "783fc97e-558e-4b67-9322-6cfaf7c82440"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S33sh15::signage": {
  "fp": "6b3e7e42cca6c522",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh15": {
  "input_fingerprint": "6fefcdf940d860b2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발로 산산이 부서진 제방 상단부 틈새로 거대한 바닷물이 폭포수처럼 쏟아져 내리는 웅장한 전경.\n\nLOCATION (lock): At the breached upper section of the artificial seawall, with seawater cascading down toward the construction area below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated inland position, hold the midpoint of the crane rise and look obliquely downward across the entire breach, keeping both broken shoulders of the embankment inside the frame. The descending seawater occupies the central third, with the broken crest above and repair machinery small along the lower edge; the widening camera distance is the principal change, while the machinery remains secondary. This is a direct external observation, with no visible people.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (The upper section has been shattered by the explosions) — The inland face and broken upper edges are visible obliquely from above; used as The surviving edges bracket the full width of the opening and establish the cascade's scale; Seawater cascade (A massive volume of seawater is pouring through the breach) — The water descends from the upper opening toward the lower foreground; used as Central moving subject connecting the broken crest to the ground below; Repair and construction machinery (Positioned near the embankment as the incoming seawater reaches the work area) — Seen from above at the foot of the embankment; used as Small lower-frame scale references, not yet the principal subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime exposure and controlled contrast preserve the separation of the broken embankment, descending water, and ground without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The already cracked and leaking seawall now has a blasted-open upper section, with seawater pouring through the breach. Repair machinery and construction equipment at its base are being swept away.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발로 산산이 부서진 제방 상단부 틈새로 거대한 바닷물이 폭포수처럼 쏟아져 내리는 웅장한 전경.\n\nLOCATION (lock): At the breached upper section of the artificial seawall, with seawater cascading down toward the construction area below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated inland position, hold the midpoint of the crane rise and look obliquely downward across the entire breach, keeping both broken shoulders of the embankment inside the frame. The descending seawater occupies the central third, with the broken crest above and repair machinery small along the lower edge; the widening camera distance is the principal change, while the machinery remains secondary. This is a direct external observation, with no visible people.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (The upper section has been shattered by the explosions) — The inland face and broken upper edges are visible obliquely from above; used as The surviving edges bracket the full width of the opening and establish the cascade's scale; Seawater cascade (A massive volume of seawater is pouring through the breach) — The water descends from the upper opening toward the lower foreground; used as Central moving subject connecting the broken crest to the ground below; Repair and construction machinery (Positioned near the embankment as the incoming seawater reaches the work area) — Seen from above at the foot of the embankment; used as Small lower-frame scale references, not yet the principal subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime exposure and controlled contrast preserve the separation of the broken embankment, descending water, and ground without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The already cracked and leaking seawall now has a blasted-open upper section, with seawater pouring through the breach. Repair machinery and construction equipment at its base are being swept away.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 폭발로 산산이 부서진 제방 상단부 틈새로 거대한 바닷물이 폭포수처럼 쏟아져 내리는 웅장한 전경.\n\nLOCATION (lock): At the breached upper section of the artificial seawall, with seawater cascading down toward the construction area below. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated inland position, hold the midpoint of the crane rise and look obliquely downward across the entire breach, keeping both broken shoulders of the embankment inside the frame. The descending seawater occupies the central third, with the broken crest above and repair machinery small along the lower edge; the widening camera distance is the principal change, while the machinery remains secondary. This is a direct external observation, with no visible people.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Breached artificial embankment (The upper section has been shattered by the explosions) — The inland face and broken upper edges are visible obliquely from above; used as The surviving edges bracket the full width of the opening and establish the cascade's scale; Seawater cascade (A massive volume of seawater is pouring through the breach) — The water descends from the upper opening toward the lower foreground; used as Central moving subject connecting the broken crest to the ground below; Repair and construction machinery (Positioned near the embankment as the incoming seawater reaches the work area) — Seen from above at the foot of the embankment; used as Small lower-frame scale references, not yet the principal subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime exposure and controlled contrast preserve the separation of the broken embankment, descending water, and ground without introducing a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The already cracked and leaking seawall now has a blasted-open upper section, with seawater pouring through the breach. Repair machinery and construction equipment at its base are being swept away.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 내륙의 높은 곳에서 부서진 제방과 쏟아지는 물줄기를 비스듬히 내려다보고 있음.",
    "built_space": "양쪽으로 부서진 거대한 콘크리트 제방이 프레임을 채우고 있으며, 그 틈으로 바닷물이 쏟아짐. 하단에는 크레인과 굴삭기 등 수리 장비들이 배치됨.",
    "entities": "사람은 보이지 않음. 부서진 제방, 거대한 바닷물 폭포, 하단의 소형 중장비들이 프롬프트와 일치함.",
    "hard_violations": [],
    "physics": "쏟아지는 물줄기의 중력과 하단에서 물에 휩쓸리거나 버티는 중장비들의 물리적 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "카메라는 내륙의 높은 곳에서 제방과 바닷물을 비스듬히 내려다봄.",
    "built_space": "부서진 제방 양쪽과 중앙의 틈새로 쏟아지는 바닷물이 보임. 하단에 굴삭기들이 배치되어 있음. 좌측 하단에 정체불명의 어두운 지형이 프레임을 일부 가림.",
    "entities": "사람은 보이지 않음. 부서진 제방과 바닷물, 중장비들이 존재하나, 좌측 하단 전경에 큰 바위 같은 질감이 시야를 방해함.",
    "hard_violations": [],
    "physics": "물줄기와 굴삭기들의 배치는 무리가 없으나, 쏟아지는 물의 흐름이 다소 거칠고 덜 수직적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 야간 시간대, 부서진 제방과 폭포수처럼 쏟아지는 바닷물, 하단부의 중장비들이 정확한 스케일과 구도로 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구사항을 전반적으로 충족하나, 좌측 하단의 검은 덩어리가 시야를 가리며 물줄기의 표현이 다소 덜 폭포수 같습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 내륙의 높은 곳에서 부서진 제방과 쏟아지는 물줄기를 비스듬히 내려다보고 있음.",
        "built_space": "양쪽으로 부서진 거대한 콘크리트 제방이 프레임을 채우고 있으며, 그 틈으로 바닷물이 쏟아짐. 하단에는 크레인과 굴삭기 등 수리 장비들이 배치됨.",
        "entities": "사람은 보이지 않음. 부서진 제방, 거대한 바닷물 폭포, 하단의 소형 중장비들이 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "쏟아지는 물줄기의 중력과 하단에서 물에 휩쓸리거나 버티는 중장비들의 물리적 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "카메라는 내륙의 높은 곳에서 제방과 바닷물을 비스듬히 내려다봄.",
        "built_space": "부서진 제방 양쪽과 중앙의 틈새로 쏟아지는 바닷물이 보임. 하단에 굴삭기들이 배치되어 있음. 좌측 하단에 정체불명의 어두운 지형이 프레임을 일부 가림.",
        "entities": "사람은 보이지 않음. 부서진 제방과 바닷물, 중장비들이 존재하나, 좌측 하단 전경에 큰 바위 같은 질감이 시야를 방해함.",
        "hard_violations": [],
        "physics": "물줄기와 굴삭기들의 배치는 무리가 없으나, 쏟아지는 물의 흐름이 다소 거칠고 덜 수직적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 야간 시간대, 부서진 제방과 폭포수처럼 쏟아지는 바닷물, 하단부의 중장비들이 정확한 스케일과 구도로 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구사항을 전반적으로 충족하나, 좌측 하단의 검은 덩어리가 시야를 가리며 물줄기의 표현이 다소 덜 폭포수 같습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 내륙의 높은 곳에서 부서진 제방과 쏟아지는 물줄기를 비스듬히 내려다보고 있음.",
        "built_space": "양쪽으로 부서진 거대한 콘크리트 제방이 프레임을 채우고 있으며, 그 틈으로 바닷물이 쏟아짐. 하단에는 크레인과 굴삭기 등 수리 장비들이 배치됨.",
        "entities": "사람은 보이지 않음. 부서진 제방, 거대한 바닷물 폭포, 하단의 소형 중장비들이 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "쏟아지는 물줄기의 중력과 하단에서 물에 휩쓸리거나 버티는 중장비들의 물리적 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "카메라는 내륙의 높은 곳에서 제방과 바닷물을 비스듬히 내려다봄.",
        "built_space": "부서진 제방 양쪽과 중앙의 틈새로 쏟아지는 바닷물이 보임. 하단에 굴삭기들이 배치되어 있음. 좌측 하단에 정체불명의 어두운 지형이 프레임을 일부 가림.",
        "entities": "사람은 보이지 않음. 부서진 제방과 바닷물, 중장비들이 존재하나, 좌측 하단 전경에 큰 바위 같은 질감이 시야를 방해함.",
        "hard_violations": [],
        "physics": "물줄기와 굴삭기들의 배치는 무리가 없으나, 쏟아지는 물의 흐름이 다소 거칠고 덜 수직적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "내륙의 높은 사선 시점에서 양쪽 파손부와 낙하수를 함께 담고 장비를 하단의 작은 척도로 유지해 우세하지만, 장비가 휩쓸리는 상태는 약하고 작업등을 추가했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "중앙 폭포와 침수·전도된 장비는 잘 구현했지만, 전경의 대형 크레인이 화면을 크게 차지해 장비는 하단의 작은 보조 요소여야 한다는 구도를 약화하며 작업등도 추가했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "상단 뒤쪽 바다에서 중앙의 파손된 개구부로 물이 밀려와 아래쪽 내륙 작업장으로 떨어진다. 낙하와 하단으로 퍼지는 흐름의 목적지가 요청한 작업 구역과 일치한다. 사람, 시선, 무기는 없으며 굴착기 붐은 주변 지면 쪽으로 향한다.",
        "built_space": "하나의 콘크리트 제방에 큰 개구부 하나가 있고, 좌우의 부서진 어깨 두 곳과 내륙 벽면이 모두 보인다. 높은 내륙 사선 시점에서 상단 단면과 아래 작업장을 함께 내려다본다. 양쪽 상단에 끊긴 난간이 있고 하단에는 적어도 세 대의 굴착기와 여러 작업 차량이 분산되어 있다. 장비는 대체로 작지만 물줄기는 중앙 3분의 1보다 넓게 퍼진다. 여러 작업등이 보여 새 광원을 도입하지 말라는 조건에는 어긋난다.",
        "entities": "파쇄된 콘크리트와 노출 철근을 가진 인공 제방, 거대한 해수 낙하, 제방 아래 건설·보수 장비가 확인된다. 사람이 없어 이 프레임의 인물 배제 지시를 지킨다. 야간이며 표면은 젖은 콘크리트·금속·포말로 읽힌다. 비교할 장소 참조 사진은 없고, 읽을 수 있는 추가 문구나 도식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "해수는 높은 바다 수위에서 파손부를 넘어 중력으로 떨어지고 하부 지면에 충돌해 포말과 확산류를 만든다. 제방 잔존부는 양쪽 구조에 연결되어 있으며 장비들은 궤도나 바퀴로 작업장 바닥에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다. 다만 대부분의 장비가 바로 서 있어 이미 휩쓸리고 있다는 동작은 분명하지 않다."
       },
       {
        "label": "B",
        "direction": "바닷물이 화면 위쪽 개구부에서 중앙을 따라 아래 작업장으로 낙하하고 전경으로 퍼진다. 요구한 바다에서 내륙으로의 흐름이 맞다. 사람이나 무기는 없고 크레인 붐들은 위쪽으로, 매달린 케이블은 아래로 향한다.",
        "built_space": "하나의 제방과 중앙 개구부 하나, 좌우의 파쇄된 어깨 두 곳이 보인다. 위에서 내려다보는 시점이지만 A보다 벽면을 정면에 가깝게 바라본다. 중앙 왼쪽의 큰 크레인 한 대에 더해 양쪽 가장자리에 크레인 붐이 하나씩 보인다. 하단에는 여러 굴착기와 차량, 작업 시설이 배치되어 있다. 특히 왼쪽 전경 크레인은 화면 높이 상당 부분을 차지해 하단의 작은 장비라는 배치에서 벗어난다. 양쪽 작업등과 조명 기둥은 별도 광원으로 보인다.",
        "entities": "상부가 크게 부서진 인공 콘크리트 제방, 노출 철근, 중앙의 대규모 해수 폭포, 침수된 건설 장비가 모두 있다. 보이는 사람은 없으며 야간 조건도 맞는다. 전경에는 기울어진 청색 장비와 잔해가 보여 장비 피해가 A보다 명확하다. 읽을 수 있는 추가 문자나 그래픽 표시는 보이지 않는다.",
        "hard_violations": [],
        "physics": "물은 바다에서 개구부를 넘어 하강한 뒤 하부 작업장에 부딪혀 퍼진다. 크레인 붐은 지상의 차체와 케이블로 지지되고 굴착기 버킷은 붐에 연결되어 있다. 기울어진 하단 장비는 침수된 바닥과 물에 걸쳐 있어 근거 없이 공중에 뜬 상태는 아니다. 강한 흐름 속의 전도·침수는 장비가 휩쓸린다는 상황과 양립하지만, 일부 굴착기는 여전히 정상 자세를 유지한다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "내륙의 높은 사선 시점에서 양쪽 파손부와 낙하수를 함께 담고 장비를 하단의 작은 척도로 유지해 우세하지만, 장비가 휩쓸리는 상태는 약하고 작업등을 추가했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "중앙 폭포와 침수·전도된 장비는 잘 구현했지만, 전경의 대형 크레인이 화면을 크게 차지해 장비는 하단의 작은 보조 요소여야 한다는 구도를 약화하며 작업등도 추가했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "상단 뒤쪽 바다에서 중앙의 파손된 개구부로 물이 밀려와 아래쪽 내륙 작업장으로 떨어진다. 낙하와 하단으로 퍼지는 흐름의 목적지가 요청한 작업 구역과 일치한다. 사람, 시선, 무기는 없으며 굴착기 붐은 주변 지면 쪽으로 향한다.",
        "built_space": "하나의 콘크리트 제방에 큰 개구부 하나가 있고, 좌우의 부서진 어깨 두 곳과 내륙 벽면이 모두 보인다. 높은 내륙 사선 시점에서 상단 단면과 아래 작업장을 함께 내려다본다. 양쪽 상단에 끊긴 난간이 있고 하단에는 적어도 세 대의 굴착기와 여러 작업 차량이 분산되어 있다. 장비는 대체로 작지만 물줄기는 중앙 3분의 1보다 넓게 퍼진다. 여러 작업등이 보여 새 광원을 도입하지 말라는 조건에는 어긋난다.",
        "entities": "파쇄된 콘크리트와 노출 철근을 가진 인공 제방, 거대한 해수 낙하, 제방 아래 건설·보수 장비가 확인된다. 사람이 없어 이 프레임의 인물 배제 지시를 지킨다. 야간이며 표면은 젖은 콘크리트·금속·포말로 읽힌다. 비교할 장소 참조 사진은 없고, 읽을 수 있는 추가 문구나 도식은 보이지 않는다.",
        "hard_violations": [],
        "physics": "해수는 높은 바다 수위에서 파손부를 넘어 중력으로 떨어지고 하부 지면에 충돌해 포말과 확산류를 만든다. 제방 잔존부는 양쪽 구조에 연결되어 있으며 장비들은 궤도나 바퀴로 작업장 바닥에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다. 다만 대부분의 장비가 바로 서 있어 이미 휩쓸리고 있다는 동작은 분명하지 않다."
       },
       {
        "label": "A",
        "direction": "바닷물이 화면 위쪽 개구부에서 중앙을 따라 아래 작업장으로 낙하하고 전경으로 퍼진다. 요구한 바다에서 내륙으로의 흐름이 맞다. 사람이나 무기는 없고 크레인 붐들은 위쪽으로, 매달린 케이블은 아래로 향한다.",
        "built_space": "하나의 제방과 중앙 개구부 하나, 좌우의 파쇄된 어깨 두 곳이 보인다. 위에서 내려다보는 시점이지만 A보다 벽면을 정면에 가깝게 바라본다. 중앙 왼쪽의 큰 크레인 한 대에 더해 양쪽 가장자리에 크레인 붐이 하나씩 보인다. 하단에는 여러 굴착기와 차량, 작업 시설이 배치되어 있다. 특히 왼쪽 전경 크레인은 화면 높이 상당 부분을 차지해 하단의 작은 장비라는 배치에서 벗어난다. 양쪽 작업등과 조명 기둥은 별도 광원으로 보인다.",
        "entities": "상부가 크게 부서진 인공 콘크리트 제방, 노출 철근, 중앙의 대규모 해수 폭포, 침수된 건설 장비가 모두 있다. 보이는 사람은 없으며 야간 조건도 맞는다. 전경에는 기울어진 청색 장비와 잔해가 보여 장비 피해가 A보다 명확하다. 읽을 수 있는 추가 문자나 그래픽 표시는 보이지 않는다.",
        "hard_violations": [],
        "physics": "물은 바다에서 개구부를 넘어 하강한 뒤 하부 작업장에 부딪혀 퍼진다. 크레인 붐은 지상의 차체와 케이블로 지지되고 굴착기 버킷은 붐에 연결되어 있다. 기울어진 하단 장비는 침수된 바닥과 물에 걸쳐 있어 근거 없이 공중에 뜬 상태는 아니다. 강한 흐름 속의 전도·침수는 장비가 휩쓸린다는 상황과 양립하지만, 일부 굴착기는 여전히 정상 자세를 유지한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 야간 시간대, 부서진 제방과 폭포수처럼 쏟아지는 바닷물, 하단부의 중장비들이 정확한 스케일과 구도로 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "요구사항을 전반적으로 충족하나, 좌측 하단의 검은 덩어리가 시야를 가리며 물줄기의 표현이 다소 덜 폭포수 같습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a20-4a80-7dc3-8345-21c7e71ea7ae",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S34sh4::signage": {
  "fp": "9989ab5a842e3926",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh4": {
  "input_fingerprint": "26984dcbe7d297c7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 힘껏 내리친 붉은 소화기 바닥이 문의 철제 문고리에 정통으로 부딪히는 찰나.\n\nLOCATION (lock): In the third-floor corridor of the camp administration building, directly outside the locked detention-room door. The building remains brightly lit at night. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Keep the camera static on the corridor side at handle height, looking diagonally across the door and safely outside the descending extinguisher's path. Capture the red extinguisher base contacting the metal handle near the lower center, with 이현우's hands and bent forearms crossing from the left and a cropped portion of his turned-away torso providing scale; the extinguisher occupies less than a third of the frame. His head remains outside the crop, his attention directed downward toward the impact, and the closer camera distance—not a new lighting treatment—carries the emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office door (Still closed at the instant of the blow) — Its corridor-facing side is seen at an oblique angle; used as Provides a stable plane behind the impact and anchors the action spatially; Metal door handle (Being struck directly by the extinguisher base) — Seen side-on enough to distinguish the handle from the door surface; used as Small impact target near the lower center; Red fire extinguisher (Held in both hands and descending into contact with the handle) — The base leads the downward strike, with the body crossing the frame diagonally; used as Moving foreground object, kept in realistic proportion to the hands and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior ambient illumination and controlled highlights keep the hand action and impact legible without adding sparks or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is still shut and resisting entry at this impact; the handle must not yet be shown as detached. A fire extinguisher is being used against the handle.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 힘껏 내리친 붉은 소화기 바닥이 문의 철제 문고리에 정통으로 부딪히는 찰나.\n\nLOCATION (lock): In the third-floor corridor of the camp administration building, directly outside the locked detention-room door. The building remains brightly lit at night. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Keep the camera static on the corridor side at handle height, looking diagonally across the door and safely outside the descending extinguisher's path. Capture the red extinguisher base contacting the metal handle near the lower center, with 이현우's hands and bent forearms crossing from the left and a cropped portion of his turned-away torso providing scale; the extinguisher occupies less than a third of the frame. His head remains outside the crop, his attention directed downward toward the impact, and the closer camera distance—not a new lighting treatment—carries the emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office door (Still closed at the instant of the blow) — Its corridor-facing side is seen at an oblique angle; used as Provides a stable plane behind the impact and anchors the action spatially; Metal door handle (Being struck directly by the extinguisher base) — Seen side-on enough to distinguish the handle from the door surface; used as Small impact target near the lower center; Red fire extinguisher (Held in both hands and descending into contact with the handle) — The base leads the downward strike, with the body crossing the frame diagonally; used as Moving foreground object, kept in realistic proportion to the hands and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior ambient illumination and controlled highlights keep the hand action and impact legible without adding sparks or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is still shut and resisting entry at this impact; the handle must not yet be shown as detached. A fire extinguisher is being used against the handle.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 힘껏 내리친 붉은 소화기 바닥이 문의 철제 문고리에 정통으로 부딪히는 찰나.\n\nLOCATION (lock): In the third-floor corridor of the camp administration building, directly outside the locked detention-room door. The building remains brightly lit at night. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Keep the camera static on the corridor side at handle height, looking diagonally across the door and safely outside the descending extinguisher's path. Capture the red extinguisher base contacting the metal handle near the lower center, with 이현우's hands and bent forearms crossing from the left and a cropped portion of his turned-away torso providing scale; the extinguisher occupies less than a third of the frame. His head remains outside the crop, his attention directed downward toward the impact, and the closer camera distance—not a new lighting treatment—carries the emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office door (Still closed at the instant of the blow) — Its corridor-facing side is seen at an oblique angle; used as Provides a stable plane behind the impact and anchors the action spatially; Metal door handle (Being struck directly by the extinguisher base) — Seen side-on enough to distinguish the handle from the door surface; used as Small impact target near the lower center; Red fire extinguisher (Held in both hands and descending into contact with the handle) — The base leads the downward strike, with the body crossing the frame diagonally; used as Moving foreground object, kept in realistic proportion to the hands and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior ambient illumination and controlled highlights keep the hand action and impact legible without adding sparks or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is still shut and resisting entry at this impact; the handle must not yet be shown as detached. A fire extinguisher is being used against the handle.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 정확히 조준되어 맞닿아 있음.",
    "built_space": "밤에도 밝게 조명이 켜진 복도 환경. 우측에 사선 각도로 닫혀 있는 문과 금속 손잡이가 위치하며, 배경으로 복도의 깊이감이 나타남.",
    "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 젊은 남성(레퍼런스와 일치하는 피부톤과 의상)의 양팔과 잘린 몸통 일부.",
    "hard_violations": [
     "[gpt-high] 같은 문짝의 좌우 가장자리에 경첩을 배치해 정상적으로 여닫을 수 없는 문 구조를 만들었다.",
     "[gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기에 지정되지 않은 설명 문구 라벨을 생성했다."
    ],
    "physics": "양손이 소화기의 손잡이와 몸통을 단단히 쥐어 하중을 지탱하고 있으며, 문고리를 내리치는 물리적 충돌 직전의 힘이 자연스럽게 연출됨."
   },
   {
    "label": "B",
    "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 조준되어 맞닿아 있음.",
    "built_space": "조명이 켜진 복도 공간. 우측에 닫힌 문과 금속 문고리가 위치함.",
    "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 남성의 양팔과 잘린 몸통.",
    "hard_violations": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조 (소화기 몸통을 쥐고 있는 아래쪽 손의 손가락과 엄지가 기형적으로 융합되고 뭉개짐)",
     "[gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기 표면에 지정되지 않은 설명 문구와 도식 라벨을 생성했다."
    ],
    "physics": "양팔이 소화기를 받치고 있으나, 아래쪽 손의 해부학적 붕괴로 인해 정상적으로 물체를 쥐고 지탱하는 힘이 성립하지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 클로즈업 앵글과 샷 스케일을 정확히 준수했으며, 소화기로 문고리를 내리치는 찰나의 동작과 질감이 매우 사실적으로 표현되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "제시된 프레이밍과 구도는 전반적으로 잘 따랐으나, 소화기 몸통을 쥐고 있는 아래쪽 손의 구조가 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 정확히 조준되어 맞닿아 있음.",
        "built_space": "밤에도 밝게 조명이 켜진 복도 환경. 우측에 사선 각도로 닫혀 있는 문과 금속 손잡이가 위치하며, 배경으로 복도의 깊이감이 나타남.",
        "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 젊은 남성(레퍼런스와 일치하는 피부톤과 의상)의 양팔과 잘린 몸통 일부.",
        "hard_violations": [],
        "physics": "양손이 소화기의 손잡이와 몸통을 단단히 쥐어 하중을 지탱하고 있으며, 문고리를 내리치는 물리적 충돌 직전의 힘이 자연스럽게 연출됨."
       },
       {
        "label": "B",
        "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 조준되어 맞닿아 있음.",
        "built_space": "조명이 켜진 복도 공간. 우측에 닫힌 문과 금속 문고리가 위치함.",
        "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 남성의 양팔과 잘린 몸통.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조 (소화기 몸통을 쥐고 있는 아래쪽 손의 손가락과 엄지가 기형적으로 융합되고 뭉개짐)"
        ],
        "physics": "양팔이 소화기를 받치고 있으나, 아래쪽 손의 해부학적 붕괴로 인해 정상적으로 물체를 쥐고 지탱하는 힘이 성립하지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 클로즈업 앵글과 샷 스케일을 정확히 준수했으며, 소화기로 문고리를 내리치는 찰나의 동작과 질감이 매우 사실적으로 표현되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "제시된 프레이밍과 구도는 전반적으로 잘 따랐으나, 소화기 몸통을 쥐고 있는 아래쪽 손의 구조가 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 정확히 조준되어 맞닿아 있음.",
        "built_space": "밤에도 밝게 조명이 켜진 복도 환경. 우측에 사선 각도로 닫혀 있는 문과 금속 손잡이가 위치하며, 배경으로 복도의 깊이감이 나타남.",
        "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 젊은 남성(레퍼런스와 일치하는 피부톤과 의상)의 양팔과 잘린 몸통 일부.",
        "hard_violations": [],
        "physics": "양손이 소화기의 손잡이와 몸통을 단단히 쥐어 하중을 지탱하고 있으며, 문고리를 내리치는 물리적 충돌 직전의 힘이 자연스럽게 연출됨."
       },
       {
        "label": "B",
        "direction": "소화기의 밑바닥이 화면 우측 하단의 금속 문고리를 향해 조준되어 맞닿아 있음.",
        "built_space": "조명이 켜진 복도 공간. 우측에 닫힌 문과 금속 문고리가 위치함.",
        "entities": "붉은색 소화기, 금속 문고리, 네이비색 반팔 셔츠를 입은 남성의 양팔과 잘린 몸통.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조 (소화기 몸통을 쥐고 있는 아래쪽 손의 손가락과 엄지가 기형적으로 융합되고 뭉개짐)"
        ],
        "physics": "양팔이 소화기를 받치고 있으나, 아래쪽 손의 해부학적 붕괴로 인해 정상적으로 물체를 쥐고 지탱하는 힘이 성립하지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "양손으로 잡은 소화기 바닥이 닫힌 문의 손잡이에 닿는 순간은 잘 구현했지만, 금지된 임의 라벨 문구를 추가했고 충돌점도 하단 중앙보다 오른쪽에 치우쳤다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "바닥과 손잡이의 접촉은 보이지만, 같은 문 양쪽의 경첩 배치가 개폐 구조를 모순되게 만들고 임의 라벨까지 추가해 A보다 충실도가 낮다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "소화기는 왼쪽 위에서 오른쪽 아래로 기울어 있으며, 바닥이 오른쪽 문에 달린 금속 손잡이에 직접 닿아 있다. 타격 대상은 맞지만 충돌점은 하단 중앙보다 오른쪽이다. 머리는 화면 밖이므로 아래를 향한 시선은 확인할 수 없다.",
        "built_space": "오른쪽에 닫힌 문 하나가 비스듬히 보이고, 그 문에는 금속 레버 손잡이 하나, 오른쪽 경첩 하나, 상단 망입 유리창 하나가 보인다. 뒤쪽 복도에는 별도의 문과 손잡이가 하나 더 보인다. 인물은 문 바깥 왼쪽에 있어 타격 위치가 성립한다. 밝은 실내 복도는 맞지만, 창과 구체적인 문 설비는 제공된 장소 정보로 확인할 수 없는 추가 설정이다. 3층과 야간 여부는 화면만으로 판별되지 않는다.",
        "entities": "한 사람의 손 두 개와 팔, 남색 반소매 상의 및 몸통·바지 일부가 보이며 얼굴과 다른 인물은 없다. 피부색과 남색 소매는 참조와 대체로 어울리지만, 참조에는 손이 없어 손 자체의 동일성은 확정할 수 없다. 붉은 소화기 하나와 아직 붙어 있는 금속 손잡이는 요구와 맞는다. 소화기에 참조나 지시가 지정하지 않은 문구와 도식이 든 라벨이 추가되어 있다.",
        "hard_violations": [
         "문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기 표면에 지정되지 않은 설명 문구와 도식 라벨을 생성했다."
        ],
        "physics": "한 손은 소화기 상부를, 다른 손은 몸통 쪽을 잡아 무게를 지탱한다. 바닥이 손잡이에 접촉하고 손잡이는 문에 부착되어 있어 지지 없는 부유물은 없다. 팔과 용기의 기울기는 아래로 내리치는 동작으로 가능하지만, 정지 화면에서는 강한 타격보다 눌러 대는 순간처럼도 보인다. 불꽃이나 별도의 충격 광원은 없다."
       },
       {
        "label": "B",
        "direction": "소화기 바닥이 오른쪽 아래로 향하고 금속 레버 위에 직접 접촉한다. 목표인 손잡이는 맞혔으나 충돌점은 하단 중앙에서 오른쪽으로 벗어나 있다. 머리가 잘려 있어 주시 방향은 보이지 않는다.",
        "built_space": "오른쪽 닫힌 문에는 손잡이 한 조와 상단 망입 유리창 하나가 보인다. 그런데 손잡이 왼쪽 문 가장자리와 문 오른쪽 가장자리에 각각 경첩이 보여 같은 문짝을 양쪽에서 고정한 구조로 읽힌다. 뒤로 길게 드러난 복도에는 여러 문과 천장 조명이 있으며, A보다 배경 공간을 더 많이 보여 준다. 인물은 복도 쪽 왼쪽에 있다. 층수와 야간은 직접 확인되지 않는다.",
        "entities": "한 사람의 팔과 손, 남색 상의와 바지 일부가 보이며 얼굴이나 미연은 등장하지 않는다. 가까운 손은 소화기 몸통 쪽을 잡고 있고, 다른 손은 상부 장치 뒤에 일부만 드러나 양손 파지가 A보다 불명확하다. 피부색과 소매는 참조와 대체로 부합하지만 연령·민족적 정체성은 이 부분 크롭만으로 확정할 수 없다. 붉은 소화기와 부착 상태의 금속 손잡이는 맞으며, 소화기에는 지정되지 않은 글자 라벨이 있다.",
        "hard_violations": [
         "같은 문짝의 좌우 가장자리에 경첩을 배치해 정상적으로 여닫을 수 없는 문 구조를 만들었다.",
         "문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기에 지정되지 않은 설명 문구 라벨을 생성했다."
        ],
        "physics": "가까운 손의 파지와 상부에 일부 보이는 다른 손이 소화기를 지지하며, 바닥은 문손잡이에 걸쳐 접촉한다. 소화기가 근거 없이 떠 있지는 않고 팔의 자세도 가능한 범위다. 손잡이는 아직 문에 붙어 있고 불꽃은 없다. 다만 문 자체는 양쪽 경첩 때문에 정상적인 개폐 기구로 성립하지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "양손으로 잡은 소화기 바닥이 닫힌 문의 손잡이에 닿는 순간은 잘 구현했지만, 금지된 임의 라벨 문구를 추가했고 충돌점도 하단 중앙보다 오른쪽에 치우쳤다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "바닥과 손잡이의 접촉은 보이지만, 같은 문 양쪽의 경첩 배치가 개폐 구조를 모순되게 만들고 임의 라벨까지 추가해 A보다 충실도가 낮다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "소화기는 왼쪽 위에서 오른쪽 아래로 기울어 있으며, 바닥이 오른쪽 문에 달린 금속 손잡이에 직접 닿아 있다. 타격 대상은 맞지만 충돌점은 하단 중앙보다 오른쪽이다. 머리는 화면 밖이므로 아래를 향한 시선은 확인할 수 없다.",
        "built_space": "오른쪽에 닫힌 문 하나가 비스듬히 보이고, 그 문에는 금속 레버 손잡이 하나, 오른쪽 경첩 하나, 상단 망입 유리창 하나가 보인다. 뒤쪽 복도에는 별도의 문과 손잡이가 하나 더 보인다. 인물은 문 바깥 왼쪽에 있어 타격 위치가 성립한다. 밝은 실내 복도는 맞지만, 창과 구체적인 문 설비는 제공된 장소 정보로 확인할 수 없는 추가 설정이다. 3층과 야간 여부는 화면만으로 판별되지 않는다.",
        "entities": "한 사람의 손 두 개와 팔, 남색 반소매 상의 및 몸통·바지 일부가 보이며 얼굴과 다른 인물은 없다. 피부색과 남색 소매는 참조와 대체로 어울리지만, 참조에는 손이 없어 손 자체의 동일성은 확정할 수 없다. 붉은 소화기 하나와 아직 붙어 있는 금속 손잡이는 요구와 맞는다. 소화기에 참조나 지시가 지정하지 않은 문구와 도식이 든 라벨이 추가되어 있다.",
        "hard_violations": [
         "문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기 표면에 지정되지 않은 설명 문구와 도식 라벨을 생성했다."
        ],
        "physics": "한 손은 소화기 상부를, 다른 손은 몸통 쪽을 잡아 무게를 지탱한다. 바닥이 손잡이에 접촉하고 손잡이는 문에 부착되어 있어 지지 없는 부유물은 없다. 팔과 용기의 기울기는 아래로 내리치는 동작으로 가능하지만, 정지 화면에서는 강한 타격보다 눌러 대는 순간처럼도 보인다. 불꽃이나 별도의 충격 광원은 없다."
       },
       {
        "label": "A",
        "direction": "소화기 바닥이 오른쪽 아래로 향하고 금속 레버 위에 직접 접촉한다. 목표인 손잡이는 맞혔으나 충돌점은 하단 중앙에서 오른쪽으로 벗어나 있다. 머리가 잘려 있어 주시 방향은 보이지 않는다.",
        "built_space": "오른쪽 닫힌 문에는 손잡이 한 조와 상단 망입 유리창 하나가 보인다. 그런데 손잡이 왼쪽 문 가장자리와 문 오른쪽 가장자리에 각각 경첩이 보여 같은 문짝을 양쪽에서 고정한 구조로 읽힌다. 뒤로 길게 드러난 복도에는 여러 문과 천장 조명이 있으며, A보다 배경 공간을 더 많이 보여 준다. 인물은 복도 쪽 왼쪽에 있다. 층수와 야간은 직접 확인되지 않는다.",
        "entities": "한 사람의 팔과 손, 남색 상의와 바지 일부가 보이며 얼굴이나 미연은 등장하지 않는다. 가까운 손은 소화기 몸통 쪽을 잡고 있고, 다른 손은 상부 장치 뒤에 일부만 드러나 양손 파지가 A보다 불명확하다. 피부색과 소매는 참조와 대체로 부합하지만 연령·민족적 정체성은 이 부분 크롭만으로 확정할 수 없다. 붉은 소화기와 부착 상태의 금속 손잡이는 맞으며, 소화기에는 지정되지 않은 글자 라벨이 있다.",
        "hard_violations": [
         "같은 문짝의 좌우 가장자리에 경첩을 배치해 정상적으로 여닫을 수 없는 문 구조를 만들었다.",
         "문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기에 지정되지 않은 설명 문구 라벨을 생성했다."
        ],
        "physics": "가까운 손의 파지와 상부에 일부 보이는 다른 손이 소화기를 지지하며, 바닥은 문손잡이에 걸쳐 접촉한다. 소화기가 근거 없이 떠 있지는 않고 팔의 자세도 가능한 범위다. 손잡이는 아직 문에 붙어 있고 불꽃은 없다. 다만 문 자체는 양쪽 경첩 때문에 정상적인 개폐 기구로 성립하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조 (소화기 몸통을 쥐고 있는 아래쪽 손의 손가락과 엄지가 기형적으로 융합되고 뭉개짐)",
     "[gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기 표면에 지정되지 않은 설명 문구와 도식 라벨을 생성했다."
    ],
    "A": [
     "[gpt-high] 같은 문짝의 좌우 가장자리에 경첩을 배치해 정상적으로 여닫을 수 없는 문 구조를 만들었다.",
     "[gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기에 지정되지 않은 설명 문구 라벨을 생성했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1350,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "지정된 클로즈업 앵글과 샷 스케일을 정확히 준수했으며, 소화기로 문고리를 내리치는 찰나의 동작과 질감이 매우 사실적으로 표현되었습니다.  ★위반: [gpt-high] 같은 문짝의 좌우 가장자리에 경첩을 배치해 정상적으로 여닫을 수 없는 문 구조를 만들었다. / [gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기에 지정되지 않은 설명 문구 라벨을 생성했다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "제시된 프레이밍과 구도는 전반적으로 잘 따랐으나, 소화기 몸통을 쥐고 있는 아래쪽 손의 구조가 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체 구조 (소화기 몸통을 쥐고 있는 아래쪽 손의 손가락과 엄지가 기형적으로 융합되고 뭉개짐) / [gpt-high] 문구를 새로 만들지 말라는 명시적 금지에도 불구하고 소화기 표면에 지정되지 않은 설명 문구와 도식 라벨을 생성했다."
   }
  ],
  "refs": [
   {
    "label": "HAND OWNER REFERENCE — 이현우: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a25-3b02-7214-9690-8013b173112a",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S34sh6::signage": {
  "fp": "e8c3c6a6d092914d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh6": {
  "input_fingerprint": "5b6f358121d81d70",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 안쪽에서 이현우의 두 팔이 멍든 미연의 상체를 와락 끌어안은 구도.\n\nLOCATION (lock): Just inside the newly opened detention-room doorway on the administration building's third floor, under the building's nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just inside the office, begin the embrace-stage rise at upper-chest height, looking diagonally across 이현우's arms from one side of the pair's interaction axis. Place his near shoulder at the left edge and 미연's bruised upper body center-right, capturing his arms closing around her as she leans into him; both lower their attention toward the contact between them rather than the lens. Their reunion is the sole positional emphasis, with the open doorway remaining a narrow background reference and the camera not yet at face height.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (The door has been opened after its handle was broken) — Seen from inside the office, diagonally back toward the corridor; used as A narrow background boundary preserving the route of 이현우's entrance.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior ambient illumination and gentle tonal separation so the embrace feels tender without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the detention room's interior surfaces, chairs, and nighttime lighting from the reference. Exclude the interrogator and his subordinate, who are no longer in the room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is now open, and its handle has been broken off. The broken handle must not reappear intact. 이현우: He stands inside the opened doorway with both arms extended in an embrace; his facial bruises, untreated leg wound, and removed outer garment remain unchanged. The stiff contact card stays concealed inside his shoe. 미연: She is at the opened doorway, still bearing the pronounced facial swelling and injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 안쪽에서 이현우의 두 팔이 멍든 미연의 상체를 와락 끌어안은 구도.\n\nLOCATION (lock): Just inside the newly opened detention-room doorway on the administration building's third floor, under the building's nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just inside the office, begin the embrace-stage rise at upper-chest height, looking diagonally across 이현우's arms from one side of the pair's interaction axis. Place his near shoulder at the left edge and 미연's bruised upper body center-right, capturing his arms closing around her as she leans into him; both lower their attention toward the contact between them rather than the lens. Their reunion is the sole positional emphasis, with the open doorway remaining a narrow background reference and the camera not yet at face height.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (The door has been opened after its handle was broken) — Seen from inside the office, diagonally back toward the corridor; used as A narrow background boundary preserving the route of 이현우's entrance.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior ambient illumination and gentle tonal separation so the embrace feels tender without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the detention room's interior surfaces, chairs, and nighttime lighting from the reference. Exclude the interrogator and his subordinate, who are no longer in the room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is now open, and its handle has been broken off. The broken handle must not reappear intact. 이현우: He stands inside the opened doorway with both arms extended in an embrace; his facial bruises, untreated leg wound, and removed outer garment remain unchanged. The stiff contact card stays concealed inside his shoe. 미연: She is at the opened doorway, still bearing the pronounced facial swelling and injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 안쪽에서 이현우의 두 팔이 멍든 미연의 상체를 와락 끌어안은 구도.\n\nLOCATION (lock): Just inside the newly opened detention-room doorway on the administration building's third floor, under the building's nighttime lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just inside the office, begin the embrace-stage rise at upper-chest height, looking diagonally across 이현우's arms from one side of the pair's interaction axis. Place his near shoulder at the left edge and 미연's bruised upper body center-right, capturing his arms closing around her as she leans into him; both lower their attention toward the contact between them rather than the lens. Their reunion is the sole positional emphasis, with the open doorway remaining a narrow background reference and the camera not yet at face height.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (The door has been opened after its handle was broken) — Seen from inside the office, diagonally back toward the corridor; used as A narrow background boundary preserving the route of 이현우's entrance.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained interior ambient illumination and gentle tonal separation so the embrace feels tender without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the detention room's interior surfaces, chairs, and nighttime lighting from the reference. Exclude the interrogator and his subordinate, who are no longer in the room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The third-floor door is now open, and its handle has been broken off. The broken handle must not reappear intact. 이현우: He stands inside the opened doorway with both arms extended in an embrace; his facial bruises, untreated leg wound, and removed outer garment remain unchanged. The stiff contact card stays concealed inside his shoe. 미연: She is at the opened doorway, still bearing the pronounced facial swelling and injuries from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 미연 모두 시선을 아래로 향하며 서로 맞닿은 부분에 집중하고 있음.",
    "built_space": "사무실 내부. 왼쪽 가장자리에 손잡이가 부서진 열린 문이 배경의 경계로 배치됨. 뒤편에는 참고 이미지의 책상, 의자, 창밖의 야간 불빛이 보임 (카메라가 복도가 아닌 사무실 안쪽을 향함).",
    "entities": "이현우는 인이어 무전기를 착용하고 상처 입은 얼굴과 낡은 어두운 셔츠를 입고 있음. 미연 역시 멍든 얼굴과 해진 회색 셔츠로 참고 이미지의 외형과 완벽히 일치함.",
    "hard_violations": [
     "[gpt-high] 카메라가 방 안에서 복도를 향해야 하는데, 열린 문을 전경에 두고 구금실 안쪽을 바라보는 반대쪽 시점을 사용했습니다."
    ],
    "physics": "두 인물이 지면에 서서 서로를 안고 있으며, 이현우의 갈색 소매 팔과 미연의 회색 소매 팔이 해부학적으로 올바르게 얽혀 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 시선을 아래로 향하며 포옹하고 있음.",
    "built_space": "사무실 내부. 오른쪽 가장자리에 손잡이 구멍이 난 열린 문이 보이며, 배경에 책상과 창문이 위치함 (역시 사무실 안쪽을 향함).",
    "entities": "미연은 부은 얼굴과 회색 셔츠로 일치하나, 이현우의 귀에 요구된 소형 인이어 무전기가 없음.",
    "hard_violations": [
     "[gemini-pro] 미연의 등을 감싸는 이현우의 팔에 이현우의 갈색 셔츠가 아닌 미연의 회색 셔츠 소매가 렌더링된 물리적/해부학적 오류 (physically impossible anatomy / leaked wardrobe)",
     "[gpt-high] 실내에서 열린 문을 지나 복도 쪽으로 바라보라는 고정 시점을 뒤집어, 문 너머 구금실의 책상과 창문을 바라보는 구도로 촬영했습니다."
    ],
    "physics": "미연의 등과 어깨를 감싸는 팔의 소매 색상이 껴안는 주체(이현우)의 의상과 불일치하여 물리적 오류가 발생함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 숄더 너머의 구도와 인이어 무전기 등의 디테일을 정확히 구현했으며, 두 인물이 포옹하는 팔의 해부학적 구조와 의상이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 인이어 무전기가 누락되었으며, 미연을 안고 있는 팔에 이현우의 셔츠가 아닌 미연과 같은 회색 소매가 렌더링되는 심각한 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 미연 모두 시선을 아래로 향하며 서로 맞닿은 부분에 집중하고 있음.",
        "built_space": "사무실 내부. 왼쪽 가장자리에 손잡이가 부서진 열린 문이 배경의 경계로 배치됨. 뒤편에는 참고 이미지의 책상, 의자, 창밖의 야간 불빛이 보임 (카메라가 복도가 아닌 사무실 안쪽을 향함).",
        "entities": "이현우는 인이어 무전기를 착용하고 상처 입은 얼굴과 낡은 어두운 셔츠를 입고 있음. 미연 역시 멍든 얼굴과 해진 회색 셔츠로 참고 이미지의 외형과 완벽히 일치함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 서로를 안고 있으며, 이현우의 갈색 소매 팔과 미연의 회색 소매 팔이 해부학적으로 올바르게 얽혀 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선을 아래로 향하며 포옹하고 있음.",
        "built_space": "사무실 내부. 오른쪽 가장자리에 손잡이 구멍이 난 열린 문이 보이며, 배경에 책상과 창문이 위치함 (역시 사무실 안쪽을 향함).",
        "entities": "미연은 부은 얼굴과 회색 셔츠로 일치하나, 이현우의 귀에 요구된 소형 인이어 무전기가 없음.",
        "hard_violations": [
         "미연의 등을 감싸는 이현우의 팔에 이현우의 갈색 셔츠가 아닌 미연의 회색 셔츠 소매가 렌더링된 물리적/해부학적 오류 (physically impossible anatomy / leaked wardrobe)"
        ],
        "physics": "미연의 등과 어깨를 감싸는 팔의 소매 색상이 껴안는 주체(이현우)의 의상과 불일치하여 물리적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 숄더 너머의 구도와 인이어 무전기 등의 디테일을 정확히 구현했으며, 두 인물이 포옹하는 팔의 해부학적 구조와 의상이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우의 인이어 무전기가 누락되었으며, 미연을 안고 있는 팔에 이현우의 셔츠가 아닌 미연과 같은 회색 소매가 렌더링되는 심각한 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 미연 모두 시선을 아래로 향하며 서로 맞닿은 부분에 집중하고 있음.",
        "built_space": "사무실 내부. 왼쪽 가장자리에 손잡이가 부서진 열린 문이 배경의 경계로 배치됨. 뒤편에는 참고 이미지의 책상, 의자, 창밖의 야간 불빛이 보임 (카메라가 복도가 아닌 사무실 안쪽을 향함).",
        "entities": "이현우는 인이어 무전기를 착용하고 상처 입은 얼굴과 낡은 어두운 셔츠를 입고 있음. 미연 역시 멍든 얼굴과 해진 회색 셔츠로 참고 이미지의 외형과 완벽히 일치함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 서로를 안고 있으며, 이현우의 갈색 소매 팔과 미연의 회색 소매 팔이 해부학적으로 올바르게 얽혀 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선을 아래로 향하며 포옹하고 있음.",
        "built_space": "사무실 내부. 오른쪽 가장자리에 손잡이 구멍이 난 열린 문이 보이며, 배경에 책상과 창문이 위치함 (역시 사무실 안쪽을 향함).",
        "entities": "미연은 부은 얼굴과 회색 셔츠로 일치하나, 이현우의 귀에 요구된 소형 인이어 무전기가 없음.",
        "hard_violations": [
         "미연의 등을 감싸는 이현우의 팔에 이현우의 갈색 셔츠가 아닌 미연의 회색 셔츠 소매가 렌더링된 물리적/해부학적 오류 (physically impossible anatomy / leaked wardrobe)"
        ],
        "physics": "미연의 등과 어깨를 감싸는 팔의 소매 색상이 껴안는 주체(이현우)의 의상과 불일치하여 물리적 오류가 발생함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "왼쪽 어깨와 중앙 오른쪽의 미연 배치는 가깝지만, 미디엄 숏보다 타이트하고 복도가 아닌 구금실 내부를 배경으로 삼아 지정된 카메라 방향을 어겼습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 팔의 포옹과 허리 부근까지 담은 미디엄 숏, 인이어는 더 충실하지만, 문밖에서 실내를 보는 역방향 시점이라 두 후보 모두 재촬영이 필요합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 고개와 눈을 미연의 머리 및 두 사람의 접촉 부위 쪽으로 내립니다. 미연은 눈을 감고 얼굴을 그의 가슴에 기댑니다. 두 팔과 손은 미연의 어깨와 위팔을 감싸며, 렌즈를 보는 사람은 없습니다. 다만 카메라의 관찰 방향은 복도 쪽이 아니라 책상과 야간 창문이 있는 방 안쪽입니다.",
        "built_space": "양쪽 가장자리에 문과 문틀이 크게 걸리고, 뒤에는 창문 한 세트, 책상 하나, 검은 의자 두 개, 천장 조명 두 개와 천장형 냉방기 하나가 보입니다. 회색 벽과 사무용 가구, 야간 창밖 풍경은 이전 장면과 유사합니다. 오른쪽 문에는 손잡이가 뜯긴 구멍이 있습니다. 두 인물은 문턱 부근에 있으나, 문이 좁은 후경 경계가 아니라 전경을 둘러싸고 방 내부가 그 뒤에 펼쳐집니다.",
        "entities": "인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐입니다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 얼굴 상처와 피·먼지가 묻은 어두운 셔츠는 부합하지만, 드러난 귀에서 인이어는 확인되지 않습니다. 미연의 검은 단발, 회색 낡은 셔츠, 부은 눈가와 짙은 볼 멍은 참조와 대체로 맞습니다. 심문관이나 부하, 추가 문구는 없습니다. 다리 상처와 신발 속 카드는 프레임 밖이라 판단하지 않습니다.",
        "hard_violations": [
         "실내에서 열린 문을 지나 복도 쪽으로 바라보라는 고정 시점을 뒤집어, 문 너머 구금실의 책상과 창문을 바라보는 구도로 촬영했습니다."
        ],
        "physics": "이현우의 한 손은 미연의 어깨 뒤를 감싸고 다른 손은 위팔에 닿아 상체를 끌어안습니다. 미연의 머리는 그의 가슴에 실제로 접촉하며 기대고 있습니다. 하체는 잘려 있지만 보이는 상체는 자연스러운 직립 포옹이고, 공중에 떠 있거나 지지 없이 매달린 부분은 없습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 미연의 머리와 품 안쪽으로 내려가고, 미연은 눈을 감은 채 그의 가슴 쪽으로 얼굴을 숙입니다. 그의 두 팔은 미연의 등과 위팔을 향해 닫혀 있으며 미연도 그의 몸통을 감쌉니다. 시선과 포옹의 대상은 정확하지만, 카메라는 복도가 아니라 실내 창문과 책상을 향합니다.",
        "built_space": "왼쪽에는 망입유리 세로창이 있는 열린 금속문 하나, 오른쪽에는 문틀이 보입니다. 후경에는 창문 한 세트, 책상 하나, 검은 의자 두 개, 천장 조명 하나와 냉방기 하나가 보입니다. 문 손잡이 자리에는 파손된 금속 부품이 남아 있고 원래의 수평 레버는 보이지 않습니다. 재료와 야간 조명은 참조를 잘 따르지만, 열린 문이 전경의 넓은 테두리가 되고 방 내부가 후경이 되어 요구된 공간 방향과 반대입니다.",
        "entities": "젊은 동아시아계 남성과 중년 동아시아계 여성, 두 사람만 있습니다. 이현우의 얼굴과 짧은 검은 머리, 마른 체격, 상처 난 뺨, 오염된 어두운 셔츠가 참조에 가깝고 귀에는 작은 검은 인이어가 보입니다. 미연의 검은 단발과 해진 회색 셔츠, 눈가 부종과 볼의 멍도 유지됩니다. 추가 인물이나 문구는 없습니다. 하체와 신발 속 소품은 보이지 않아 평가 대상에서 제외합니다.",
        "hard_violations": [
         "카메라가 방 안에서 복도를 향해야 하는데, 열린 문을 전경에 두고 구금실 안쪽을 바라보는 반대쪽 시점을 사용했습니다."
        ],
        "physics": "이현우의 양손은 미연의 어깨 뒤와 위팔에 접촉하고, 미연의 팔도 그의 허리와 옆구리를 감쌉니다. 미연의 머리와 상체는 그의 가슴과 팔에 기대어 지지됩니다. 두 몸통의 연결과 팔의 굽힘은 가능한 포옹 자세이며, 발이 프레임 밖이라는 이유만으로 부유를 의심할 근거는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "왼쪽 어깨와 중앙 오른쪽의 미연 배치는 가깝지만, 미디엄 숏보다 타이트하고 복도가 아닌 구금실 내부를 배경으로 삼아 지정된 카메라 방향을 어겼습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 팔의 포옹과 허리 부근까지 담은 미디엄 숏, 인이어는 더 충실하지만, 문밖에서 실내를 보는 역방향 시점이라 두 후보 모두 재촬영이 필요합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 고개와 눈을 미연의 머리 및 두 사람의 접촉 부위 쪽으로 내립니다. 미연은 눈을 감고 얼굴을 그의 가슴에 기댑니다. 두 팔과 손은 미연의 어깨와 위팔을 감싸며, 렌즈를 보는 사람은 없습니다. 다만 카메라의 관찰 방향은 복도 쪽이 아니라 책상과 야간 창문이 있는 방 안쪽입니다.",
        "built_space": "양쪽 가장자리에 문과 문틀이 크게 걸리고, 뒤에는 창문 한 세트, 책상 하나, 검은 의자 두 개, 천장 조명 두 개와 천장형 냉방기 하나가 보입니다. 회색 벽과 사무용 가구, 야간 창밖 풍경은 이전 장면과 유사합니다. 오른쪽 문에는 손잡이가 뜯긴 구멍이 있습니다. 두 인물은 문턱 부근에 있으나, 문이 좁은 후경 경계가 아니라 전경을 둘러싸고 방 내부가 그 뒤에 펼쳐집니다.",
        "entities": "인물은 젊은 동아시아계 남성 한 명과 중년 동아시아계 여성 한 명뿐입니다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 얼굴 상처와 피·먼지가 묻은 어두운 셔츠는 부합하지만, 드러난 귀에서 인이어는 확인되지 않습니다. 미연의 검은 단발, 회색 낡은 셔츠, 부은 눈가와 짙은 볼 멍은 참조와 대체로 맞습니다. 심문관이나 부하, 추가 문구는 없습니다. 다리 상처와 신발 속 카드는 프레임 밖이라 판단하지 않습니다.",
        "hard_violations": [
         "실내에서 열린 문을 지나 복도 쪽으로 바라보라는 고정 시점을 뒤집어, 문 너머 구금실의 책상과 창문을 바라보는 구도로 촬영했습니다."
        ],
        "physics": "이현우의 한 손은 미연의 어깨 뒤를 감싸고 다른 손은 위팔에 닿아 상체를 끌어안습니다. 미연의 머리는 그의 가슴에 실제로 접촉하며 기대고 있습니다. 하체는 잘려 있지만 보이는 상체는 자연스러운 직립 포옹이고, 공중에 떠 있거나 지지 없이 매달린 부분은 없습니다."
       },
       {
        "label": "A",
        "direction": "이현우의 시선은 미연의 머리와 품 안쪽으로 내려가고, 미연은 눈을 감은 채 그의 가슴 쪽으로 얼굴을 숙입니다. 그의 두 팔은 미연의 등과 위팔을 향해 닫혀 있으며 미연도 그의 몸통을 감쌉니다. 시선과 포옹의 대상은 정확하지만, 카메라는 복도가 아니라 실내 창문과 책상을 향합니다.",
        "built_space": "왼쪽에는 망입유리 세로창이 있는 열린 금속문 하나, 오른쪽에는 문틀이 보입니다. 후경에는 창문 한 세트, 책상 하나, 검은 의자 두 개, 천장 조명 하나와 냉방기 하나가 보입니다. 문 손잡이 자리에는 파손된 금속 부품이 남아 있고 원래의 수평 레버는 보이지 않습니다. 재료와 야간 조명은 참조를 잘 따르지만, 열린 문이 전경의 넓은 테두리가 되고 방 내부가 후경이 되어 요구된 공간 방향과 반대입니다.",
        "entities": "젊은 동아시아계 남성과 중년 동아시아계 여성, 두 사람만 있습니다. 이현우의 얼굴과 짧은 검은 머리, 마른 체격, 상처 난 뺨, 오염된 어두운 셔츠가 참조에 가깝고 귀에는 작은 검은 인이어가 보입니다. 미연의 검은 단발과 해진 회색 셔츠, 눈가 부종과 볼의 멍도 유지됩니다. 추가 인물이나 문구는 없습니다. 하체와 신발 속 소품은 보이지 않아 평가 대상에서 제외합니다.",
        "hard_violations": [
         "카메라가 방 안에서 복도를 향해야 하는데, 열린 문을 전경에 두고 구금실 안쪽을 바라보는 반대쪽 시점을 사용했습니다."
        ],
        "physics": "이현우의 양손은 미연의 어깨 뒤와 위팔에 접촉하고, 미연의 팔도 그의 허리와 옆구리를 감쌉니다. 미연의 머리와 상체는 그의 가슴과 팔에 기대어 지지됩니다. 두 몸통의 연결과 팔의 굽힘은 가능한 포옹 자세이며, 발이 프레임 밖이라는 이유만으로 부유를 의심할 근거는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 미연의 등을 감싸는 이현우의 팔에 이현우의 갈색 셔츠가 아닌 미연의 회색 셔츠 소매가 렌더링된 물리적/해부학적 오류 (physically impossible anatomy / leaked wardrobe)",
     "[gpt-high] 실내에서 열린 문을 지나 복도 쪽으로 바라보라는 고정 시점을 뒤집어, 문 너머 구금실의 책상과 창문을 바라보는 구도로 촬영했습니다."
    ],
    "A": [
     "[gpt-high] 카메라가 방 안에서 복도를 향해야 하는데, 열린 문을 전경에 두고 구금실 안쪽을 바라보는 반대쪽 시점을 사용했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 요구한 숄더 너머의 구도와 인이어 무전기 등의 디테일을 정확히 구현했으며, 두 인물이 포옹하는 팔의 해부학적 구조와 의상이 자연스럽습니다.  ★위반: [gpt-high] 카메라가 방 안에서 복도를 향해야 하는데, 열린 문을 전경에 두고 구금실 안쪽을 바라보는 반대쪽 시점을 사용했습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "이현우의 인이어 무전기가 누락되었으며, 미연을 안고 있는 팔에 이현우의 셔츠가 아닌 미연과 같은 회색 소매가 렌더링되는 심각한 오류가 발생했습니다.  ★위반: [gemini-pro] 미연의 등을 감싸는 이현우의 팔에 이현우의 갈색 셔츠가 아닌 미연의 회색 셔츠 소매가 렌더링된 물리적/해부학적 오류 (physically impossible anatomy / leaked wardrobe) / [gpt-high] 실내에서 열린 문을 지나 복도 쪽으로 바라보라는 고정 시점을 뒤집어, 문 너머 구금실의 책상과 창문을 바라보는 구도로 촬영했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S31sh2_sel.png",
    "asset_id": "be7c227b-686c-438e-afbe-a2905c8bbcbd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a2a-1417-73d9-84a8-fe74fed72bac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh2"
  }
 },
 "S34sh8::signage": {
  "fp": "81488cb00fa7ce80",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh8": {
  "input_fingerprint": "90c3ff79f4d3341f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 귀를 찌르는 강렬한 사이렌 소리에 놀라 복도 창문 쪽으로 고개를 확 돌린 이현우와 미연의 굳은 표정.\n\nLOCATION (lock): At the detention-room doorway adjoining the third-floor corridor of the lit administration building, facing the corridor windows. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle at face height beside the doorway inside the office, holding an oblique view from the same side of the pair rather than moving onto their eyeline. Keep 이현우 left-center and 미연 center, with open look-room on the right as both snap their attention toward the corridor window outside the frame, their shoulders still caught in the interrupted embrace. The changed gaze is the only emphasized axis; hold the established subject scale and lighting while their expressions harden.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (Open to the corridor) — Its interior edge is visible beside the pair, with the corridor beyond; used as Maintains spatial continuity and indicates the direction of their off-screen attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding interior exposure and controlled contrast unchanged, allowing the alarm reaction to register without invented flashing light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open doorway, damaged door hardware, and the room's nighttime appearance from the reference. Exclude the earlier embrace as a frozen event and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forced-open third-floor door retains its broken-off handle. No alarm-light effect is established by the siren alone. 이현우: He remains on the third floor with facial bruises and the untreated leg wound, without his outer garment. The stiff contact card remains concealed inside his shoe. 미연: She is no longer confined in the room, but her face remains badly swollen and injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 귀를 찌르는 강렬한 사이렌 소리에 놀라 복도 창문 쪽으로 고개를 확 돌린 이현우와 미연의 굳은 표정.\n\nLOCATION (lock): At the detention-room doorway adjoining the third-floor corridor of the lit administration building, facing the corridor windows. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle at face height beside the doorway inside the office, holding an oblique view from the same side of the pair rather than moving onto their eyeline. Keep 이현우 left-center and 미연 center, with open look-room on the right as both snap their attention toward the corridor window outside the frame, their shoulders still caught in the interrupted embrace. The changed gaze is the only emphasized axis; hold the established subject scale and lighting while their expressions harden.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (Open to the corridor) — Its interior edge is visible beside the pair, with the corridor beyond; used as Maintains spatial continuity and indicates the direction of their off-screen attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding interior exposure and controlled contrast unchanged, allowing the alarm reaction to register without invented flashing light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open doorway, damaged door hardware, and the room's nighttime appearance from the reference. Exclude the earlier embrace as a frozen event and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forced-open third-floor door retains its broken-off handle. No alarm-light effect is established by the siren alone. 이현우: He remains on the third floor with facial bruises and the untreated leg wound, without his outer garment. The stiff contact card remains concealed inside his shoe. 미연: She is no longer confined in the room, but her face remains badly swollen and injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 귀를 찌르는 강렬한 사이렌 소리에 놀라 복도 창문 쪽으로 고개를 확 돌린 이현우와 미연의 굳은 표정.\n\nLOCATION (lock): At the detention-room doorway adjoining the third-floor corridor of the lit administration building, facing the corridor windows. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle at face height beside the doorway inside the office, holding an oblique view from the same side of the pair rather than moving onto their eyeline. Keep 이현우 left-center and 미연 center, with open look-room on the right as both snap their attention toward the corridor window outside the frame, their shoulders still caught in the interrupted embrace. The changed gaze is the only emphasized axis; hold the established subject scale and lighting while their expressions harden.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Open office doorway (Open to the corridor) — Its interior edge is visible beside the pair, with the corridor beyond; used as Maintains spatial continuity and indicates the direction of their off-screen attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding interior exposure and controlled contrast unchanged, allowing the alarm reaction to register without invented flashing light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open doorway, damaged door hardware, and the room's nighttime appearance from the reference. Exclude the earlier embrace as a frozen event and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forced-open third-floor door retains its broken-off handle. No alarm-light effect is established by the siren alone. 이현우: He remains on the third floor with facial bruises and the untreated leg wound, without his outer garment. The stiff contact card remains concealed inside his shoe. 미연: She is no longer confined in the room, but her face remains badly swollen and injured.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 향함.",
    "built_space": "사무실 내부 시점. 우측에 열린 문과 복도가 보이나 복도 창문은 명확하지 않음.",
    "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 지지하여 서 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 둠.",
    "built_space": "사무실 내부 시점. 우측에 열린 문과 밤 시간대의 복도 창문이 명확히 보임.",
    "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "서 있는 자세이며, 미연이 이현우의 팔을 잡은 상태로 바닥의 지지를 받음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 인물 외양을 잘 따랐으며, 중단된 포옹의 신체 접촉과 복도의 야간 창문 배경이 더 정확히 묘사되었습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프레임과 인물의 기본 외양은 지시를 따랐으나, 중단된 포옹의 동작 디테일과 복도의 창문 배경 묘사가 다소 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 향함.",
        "built_space": "사무실 내부 시점. 우측에 열린 문과 복도가 보이나 복도 창문은 명확하지 않음.",
        "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 지지하여 서 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 둠.",
        "built_space": "사무실 내부 시점. 우측에 열린 문과 밤 시간대의 복도 창문이 명확히 보임.",
        "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "서 있는 자세이며, 미연이 이현우의 팔을 잡은 상태로 바닥의 지지를 받음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 인물 외양을 잘 따랐으며, 중단된 포옹의 신체 접촉과 복도의 야간 창문 배경이 더 정확히 묘사되었습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프레임과 인물의 기본 외양은 지시를 따랐으나, 중단된 포옹의 동작 디테일과 복도의 창문 배경 묘사가 다소 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 향함.",
        "built_space": "사무실 내부 시점. 우측에 열린 문과 복도가 보이나 복도 창문은 명확하지 않음.",
        "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 지지하여 서 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 화면 우측 밖(복도 창문 방향)으로 시선을 둠.",
        "built_space": "사무실 내부 시점. 우측에 열린 문과 밤 시간대의 복도 창문이 명확히 보임.",
        "entities": "이현우(인이어, 얼굴 상처, 어두운 셔츠), 미연(얼굴 상처, 회색 셔츠) 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "서 있는 자세이며, 미연이 이현우의 팔을 잡은 상태로 바닥의 지지를 받음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 중앙의 현우와 중앙의 미연을 비스듬한 미디엄 숏으로 담아, 오른쪽으로 돌린 시선과 중단된 포옹, 열린 문 너머 복도의 연결을 더 충실하게 구현했다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽을 향한 두 사람의 긴장과 부상은 잘 맞지만, 더 정면에 가까운 타이트한 구도와 오른쪽을 크게 차지하는 문짝 때문에 지정된 사선 시점과 시선 여백이 약해졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우와 미연 모두 얼굴과 눈을 화면 오른쪽으로 돌려 카메라 밖의 대상을 본다. 보이는 복도 창문들은 배경에 있고 실제 주시점은 오른쪽 프레임 밖으로 읽혀, 화면 밖 복도 창문을 향한 반응과 양립한다. 몸통과 팔은 서로 붙어 있어 시선만 갑자기 바뀐 순간이 드러난다.",
        "built_space": "오른쪽에 열린 회색 금속문 한 짝, 그 안의 망입유리 한 장, 경첩 두 개가 보인다. 두 사람 바로 옆 문틀 너머로 천장등과 창문이 반복되는 복도가 이어지며, 왼쪽 실내에는 서류가 놓인 수납장 일부가 있다. 실내 문 옆에서 복도를 비스듬히 보는 위치가 설득력 있다. 문 하단의 철물은 일부만 보여 파손 상태를 확정하기 어렵지만 온전한 레버 손잡이가 복구된 모습은 보이지 않는다.",
        "entities": "등장인물은 젊은 동아시아계 남성과 중년 동아시아계 여성 두 명뿐이다. 현우의 짧고 헝클어진 검은 머리, 마른 체격, 얼굴 상처, 인이어 장치, 피와 먼지가 묻은 어두운 셔츠가 맞는다. 미연의 검은 단발, 중년 얼굴, 멍과 부기, 해진 회색 셔츠도 참조와 대체로 맞는다. 외투는 없고, 다리 상처와 신발 속 카드는 구도 밖이다. 야간 복도 조명이 유지되며 경보 점멸광은 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 세운 채 서로 밀착해 있고, 미연의 팔이 현우 앞을 감싸며 현우의 손이 그 팔에 닿아 있다. 발은 프레임 밖이지만 공중에 뜬 자세나 지지 없는 물체는 보이지 않는다. 어깨와 팔의 접촉을 유지한 채 목만 오른쪽으로 돌리는 동작은 가능하다. 문은 보이는 경첩으로 지지된다."
       },
       {
        "label": "B",
        "direction": "두 사람의 눈과 얼굴이 모두 화면 오른쪽 바깥을 향한다. 렌즈를 바라보지 않으며 같은 방향의 소리에 반응하는 것으로 읽힌다. 다만 얼굴이 A보다 카메라 쪽에 더 열려 있어, 같은 편에서 비스듬히 관찰하는 시점은 상대적으로 약하다.",
        "built_space": "오른쪽 전경에 회색 금속문 한 짝과 망입유리 한 장, 경첩 두 개가 보이고, 문틀 뒤 좁은 틈으로 조명 켜진 복도가 이어진다. 두 사람은 열린 출입구의 실내 쪽에 붙어 있다. 참조의 금속문과 복도 재질은 이어지지만 문짝이 화면 오른쪽을 크게 차지해 열린 시선 여백이 줄었다. 손잡이 위치는 화면 아래로 잘려 파손 여부를 확인할 수 없다.",
        "entities": "현우와 미연으로 읽히는 두 사람만 있다. 현우의 젊은 얼굴, 헝클어진 검은 머리, 인이어 장치, 얼굴 멍, 피 묻은 어두운 셔츠가 맞는다. 미연은 검은 단발의 중년 여성으로, 눈 주위의 심한 멍과 부기 및 찢어진 회색 셔츠가 잘 드러난다. 두 사람의 얼굴과 의상은 참조에 대체로 부합한다. 하체와 숨긴 카드는 프레임 밖이며, 추가 경보광은 없다.",
        "hard_violations": [],
        "physics": "현우의 팔은 미연 뒤로 이어지고 미연의 팔은 현우 앞쪽으로 뻗어 있어, 포옹 중 접촉을 유지하는 자세로 읽힌다. 손의 일부는 하단에서 잘렸으나 보이는 팔과 몸통 연결에 명백한 불가능성은 없다. 발은 보이지 않지만 상체는 자연스러운 직립 자세이며 부유하는 몸이나 물체는 없다. 열린 문은 경첩에 매달려 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 중앙의 현우와 중앙의 미연을 비스듬한 미디엄 숏으로 담아, 오른쪽으로 돌린 시선과 중단된 포옹, 열린 문 너머 복도의 연결을 더 충실하게 구현했다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽을 향한 두 사람의 긴장과 부상은 잘 맞지만, 더 정면에 가까운 타이트한 구도와 오른쪽을 크게 차지하는 문짝 때문에 지정된 사선 시점과 시선 여백이 약해졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우와 미연 모두 얼굴과 눈을 화면 오른쪽으로 돌려 카메라 밖의 대상을 본다. 보이는 복도 창문들은 배경에 있고 실제 주시점은 오른쪽 프레임 밖으로 읽혀, 화면 밖 복도 창문을 향한 반응과 양립한다. 몸통과 팔은 서로 붙어 있어 시선만 갑자기 바뀐 순간이 드러난다.",
        "built_space": "오른쪽에 열린 회색 금속문 한 짝, 그 안의 망입유리 한 장, 경첩 두 개가 보인다. 두 사람 바로 옆 문틀 너머로 천장등과 창문이 반복되는 복도가 이어지며, 왼쪽 실내에는 서류가 놓인 수납장 일부가 있다. 실내 문 옆에서 복도를 비스듬히 보는 위치가 설득력 있다. 문 하단의 철물은 일부만 보여 파손 상태를 확정하기 어렵지만 온전한 레버 손잡이가 복구된 모습은 보이지 않는다.",
        "entities": "등장인물은 젊은 동아시아계 남성과 중년 동아시아계 여성 두 명뿐이다. 현우의 짧고 헝클어진 검은 머리, 마른 체격, 얼굴 상처, 인이어 장치, 피와 먼지가 묻은 어두운 셔츠가 맞는다. 미연의 검은 단발, 중년 얼굴, 멍과 부기, 해진 회색 셔츠도 참조와 대체로 맞는다. 외투는 없고, 다리 상처와 신발 속 카드는 구도 밖이다. 야간 복도 조명이 유지되며 경보 점멸광은 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 세운 채 서로 밀착해 있고, 미연의 팔이 현우 앞을 감싸며 현우의 손이 그 팔에 닿아 있다. 발은 프레임 밖이지만 공중에 뜬 자세나 지지 없는 물체는 보이지 않는다. 어깨와 팔의 접촉을 유지한 채 목만 오른쪽으로 돌리는 동작은 가능하다. 문은 보이는 경첩으로 지지된다."
       },
       {
        "label": "A",
        "direction": "두 사람의 눈과 얼굴이 모두 화면 오른쪽 바깥을 향한다. 렌즈를 바라보지 않으며 같은 방향의 소리에 반응하는 것으로 읽힌다. 다만 얼굴이 A보다 카메라 쪽에 더 열려 있어, 같은 편에서 비스듬히 관찰하는 시점은 상대적으로 약하다.",
        "built_space": "오른쪽 전경에 회색 금속문 한 짝과 망입유리 한 장, 경첩 두 개가 보이고, 문틀 뒤 좁은 틈으로 조명 켜진 복도가 이어진다. 두 사람은 열린 출입구의 실내 쪽에 붙어 있다. 참조의 금속문과 복도 재질은 이어지지만 문짝이 화면 오른쪽을 크게 차지해 열린 시선 여백이 줄었다. 손잡이 위치는 화면 아래로 잘려 파손 여부를 확인할 수 없다.",
        "entities": "현우와 미연으로 읽히는 두 사람만 있다. 현우의 젊은 얼굴, 헝클어진 검은 머리, 인이어 장치, 얼굴 멍, 피 묻은 어두운 셔츠가 맞는다. 미연은 검은 단발의 중년 여성으로, 눈 주위의 심한 멍과 부기 및 찢어진 회색 셔츠가 잘 드러난다. 두 사람의 얼굴과 의상은 참조에 대체로 부합한다. 하체와 숨긴 카드는 프레임 밖이며, 추가 경보광은 없다.",
        "hard_violations": [],
        "physics": "현우의 팔은 미연 뒤로 이어지고 미연의 팔은 현우 앞쪽으로 뻗어 있어, 포옹 중 접촉을 유지하는 자세로 읽힌다. 손의 일부는 하단에서 잘렸으나 보이는 팔과 몸통 연결에 명백한 불가능성은 없다. 발은 보이지 않지만 상체는 자연스러운 직립 자세이며 부유하는 몸이나 물체는 없다. 열린 문은 경첩에 매달려 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.746,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.746,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1746
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 구도와 인물 외양을 잘 따랐으며, 중단된 포옹의 신체 접촉과 복도의 야간 창문 배경이 더 정확히 묘사되었습니다."
   },
   {
    "label": "A",
    "score": 1746,
    "verdict_ko": "프레임과 인물의 기본 외양은 지시를 따랐으나, 중단된 포옹의 동작 디테일과 복도의 창문 배경 묘사가 다소 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S34sh4_sel.png",
    "asset_id": "a9dd47ee-e712-477a-8816-835c1decc6d2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a30-56c9-7528-9b66-db86ae4919ac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S34sh4"
  }
 },
 "S35sh2::signage": {
  "fp": "fab0d3269a22268d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S35sh2": {
  "input_fingerprint": "e8e6dfc91c325380",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 아래 난민촌을 향해 솟아오른 거대한 쓰나미 벽의 압도적인 전경.\n\nLOCATION (lock): Over the exposed streets and container roofs of the refugee settlement, facing the enormous incoming seawater surge beneath the night sky. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the upper endpoint of the crane rise over the camp's outer edge, look obliquely across the settlement with a slight downward pitch, remaining laterally offset from the incoming wave's axis. Place the camp across the lower band and the immense wave across the middle third beneath the night sky, using their relative scale rather than an exaggerated foreground object to convey its height; no individuals are visible in this selected composition. The camera-distance expansion completes the reveal, leaving the geometry ready for the subsequent backward retreat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp (In the path of the rapidly approaching seawater) — Its layout is viewed obliquely from above, with the incoming wave beyond; used as Lower-frame environmental scale reference; Tsunami wall (Rising and advancing toward the camp) — Its advancing face is visible across the middle of the image; used as Primary threat, held in relation to the camp and sky; Night sky (Dark above the approaching wave); used as Upper-frame negative space defining the wave's crest.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime exposure and restrained tonal separation keep the wave readable against the dark sky without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At night, seawater from the breached seawall is flooding the refugee settlement, with a rapidly advancing tsunami front. The seawall breach remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 아래 난민촌을 향해 솟아오른 거대한 쓰나미 벽의 압도적인 전경.\n\nLOCATION (lock): Over the exposed streets and container roofs of the refugee settlement, facing the enormous incoming seawater surge beneath the night sky. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the upper endpoint of the crane rise over the camp's outer edge, look obliquely across the settlement with a slight downward pitch, remaining laterally offset from the incoming wave's axis. Place the camp across the lower band and the immense wave across the middle third beneath the night sky, using their relative scale rather than an exaggerated foreground object to convey its height; no individuals are visible in this selected composition. The camera-distance expansion completes the reveal, leaving the geometry ready for the subsequent backward retreat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp (In the path of the rapidly approaching seawater) — Its layout is viewed obliquely from above, with the incoming wave beyond; used as Lower-frame environmental scale reference; Tsunami wall (Rising and advancing toward the camp) — Its advancing face is visible across the middle of the image; used as Primary threat, held in relation to the camp and sky; Night sky (Dark above the approaching wave); used as Upper-frame negative space defining the wave's crest.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime exposure and restrained tonal separation keep the wave readable against the dark sky without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At night, seawater from the breached seawall is flooding the refugee settlement, with a rapidly advancing tsunami front. The seawall breach remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤하늘 아래 난민촌을 향해 솟아오른 거대한 쓰나미 벽의 압도적인 전경.\n\nLOCATION (lock): Over the exposed streets and container roofs of the refugee settlement, facing the enormous incoming seawater surge beneath the night sky. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the upper endpoint of the crane rise over the camp's outer edge, look obliquely across the settlement with a slight downward pitch, remaining laterally offset from the incoming wave's axis. Place the camp across the lower band and the immense wave across the middle third beneath the night sky, using their relative scale rather than an exaggerated foreground object to convey its height; no individuals are visible in this selected composition. The camera-distance expansion completes the reveal, leaving the geometry ready for the subsequent backward retreat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee camp (In the path of the rapidly approaching seawater) — Its layout is viewed obliquely from above, with the incoming wave beyond; used as Lower-frame environmental scale reference; Tsunami wall (Rising and advancing toward the camp) — Its advancing face is visible across the middle of the image; used as Primary threat, held in relation to the camp and sky; Night sky (Dark above the approaching wave); used as Upper-frame negative space defining the wave's crest.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime exposure and restrained tonal separation keep the wave readable against the dark sky without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At night, seawater from the breached seawall is flooding the refugee settlement, with a rapidly advancing tsunami front. The seawall breach remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 컨테이너 난민촌을 약간 내려다보는 각도로, 다가오는 거대한 해일을 정면으로 향해 있음.",
    "built_space": "화면 하단에 조명이 켜진 컨테이너 캠프가 넓게 퍼져 있고, 중단에 거대한 파도가, 상단에 어두운 밤하늘이 배치됨.",
    "entities": "컨테이너, 가로등, 파도, 밤하늘. 화면 중앙 하단과 우측 하단에 사람의 실루엣이 분명하게 존재함.",
    "hard_violations": [
     "[gemini-pro] 사람이 화면에 나타나서는 안 된다는 지시(NO PEOPLE IN THIS SHOT)를 위반하여 두 명 이상의 사람 실루엣이 포함됨.",
     "[gpt-high] 중앙 하단과 오른쪽 아래에 최소 두 명의 사람을 추가해, 어떠한 살아 있는 인물도 등장하지 말라는 명시적 금지 조건을 위반했다."
    ],
    "physics": "거대한 해일이 물리적인 무게감을 가지며 캠프를 향해 솟아오름."
   },
   {
    "label": "B",
    "direction": "캠프 외곽에서 비스듬히 내려다보는 구도로, 펜스 너머로 밀려오는 거대한 해일을 향하고 있음.",
    "built_space": "화면 하단 및 우측에 컨테이너와 펜스로 이루어진 캠프가 배치되고, 중앙 전체를 거대한 파도가 가로지르며 압도적인 크기를 보여줌.",
    "entities": "컨테이너, 펜스, 조명등, 밀려오는 파도, 밤하늘. 요구사항대로 화면 내에 사람이 전혀 없음.",
    "hard_violations": [
     "[gpt-high] 장소 설명에 없는 높은 경계 감시탑 한 개를 추가했다. 컨테이너 지붕과 노출된 거리만으로 장소를 구성하고 그 밖의 시설을 발명하지 말라는 제한에 어긋난다."
    ],
    "physics": "파도의 형태와 캠프 바닥에 고인 물의 반사광 등이 자연스럽게 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글과 피사체 비율을 잘 구현했으며, 프롬프트의 '사람이 없어야 한다'는 핵심 규칙을 완벽하게 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 어떤 인물도 나타나서는 안 된다는 명시적인 지시에도 불구하고 사람의 실루엣이 묘사되어 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 컨테이너 난민촌을 약간 내려다보는 각도로, 다가오는 거대한 해일을 정면으로 향해 있음.",
        "built_space": "화면 하단에 조명이 켜진 컨테이너 캠프가 넓게 퍼져 있고, 중단에 거대한 파도가, 상단에 어두운 밤하늘이 배치됨.",
        "entities": "컨테이너, 가로등, 파도, 밤하늘. 화면 중앙 하단과 우측 하단에 사람의 실루엣이 분명하게 존재함.",
        "hard_violations": [
         "사람이 화면에 나타나서는 안 된다는 지시(NO PEOPLE IN THIS SHOT)를 위반하여 두 명 이상의 사람 실루엣이 포함됨."
        ],
        "physics": "거대한 해일이 물리적인 무게감을 가지며 캠프를 향해 솟아오름."
       },
       {
        "label": "B",
        "direction": "캠프 외곽에서 비스듬히 내려다보는 구도로, 펜스 너머로 밀려오는 거대한 해일을 향하고 있음.",
        "built_space": "화면 하단 및 우측에 컨테이너와 펜스로 이루어진 캠프가 배치되고, 중앙 전체를 거대한 파도가 가로지르며 압도적인 크기를 보여줌.",
        "entities": "컨테이너, 펜스, 조명등, 밀려오는 파도, 밤하늘. 요구사항대로 화면 내에 사람이 전혀 없음.",
        "hard_violations": [],
        "physics": "파도의 형태와 캠프 바닥에 고인 물의 반사광 등이 자연스럽게 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 앵글과 피사체 비율을 잘 구현했으며, 프롬프트의 '사람이 없어야 한다'는 핵심 규칙을 완벽하게 준수했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면에 어떤 인물도 나타나서는 안 된다는 명시적인 지시에도 불구하고 사람의 실루엣이 묘사되어 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 컨테이너 난민촌을 약간 내려다보는 각도로, 다가오는 거대한 해일을 정면으로 향해 있음.",
        "built_space": "화면 하단에 조명이 켜진 컨테이너 캠프가 넓게 퍼져 있고, 중단에 거대한 파도가, 상단에 어두운 밤하늘이 배치됨.",
        "entities": "컨테이너, 가로등, 파도, 밤하늘. 화면 중앙 하단과 우측 하단에 사람의 실루엣이 분명하게 존재함.",
        "hard_violations": [
         "사람이 화면에 나타나서는 안 된다는 지시(NO PEOPLE IN THIS SHOT)를 위반하여 두 명 이상의 사람 실루엣이 포함됨."
        ],
        "physics": "거대한 해일이 물리적인 무게감을 가지며 캠프를 향해 솟아오름."
       },
       {
        "label": "B",
        "direction": "캠프 외곽에서 비스듬히 내려다보는 구도로, 펜스 너머로 밀려오는 거대한 해일을 향하고 있음.",
        "built_space": "화면 하단 및 우측에 컨테이너와 펜스로 이루어진 캠프가 배치되고, 중앙 전체를 거대한 파도가 가로지르며 압도적인 크기를 보여줌.",
        "entities": "컨테이너, 펜스, 조명등, 밀려오는 파도, 밤하늘. 요구사항대로 화면 내에 사람이 전혀 없음.",
        "hard_violations": [],
        "physics": "파도의 형태와 캠프 바닥에 고인 물의 반사광 등이 자연스럽게 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "하단 난민촌·중앙 파도·상단 밤하늘과 비스듬한 상승 시점은 더 정확하지만, 지정되지 않은 감시탑을 추가해 장소 제한을 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "난민촌으로 밀려드는 해수는 잘 보이지만, 금지된 인물이 최소 두 명 등장하고 난민촌이 화면 하단 절반까지 확대되어 지정 구도에서도 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "거대한 파도의 전면과 무너지는 포말이 화면 아래 난민촌을 향한다. 난민촌 경계가 왼쪽 아래에서 오른쪽 위로 이어져, 파도의 진행축에서 옆으로 비켜선 사선 관찰이 읽힌다. 확인 가능한 시선이나 조준 대상은 없다.",
        "built_space": "여러 줄의 컨테이너 지붕과 그 사이 노출된 도로가 하단에 배치되어 있다. 난민촌 바깥쪽에는 연속된 경계 울타리, 오른쪽 중경에는 높은 감시탑 한 개가 보이며 조명 기둥도 여러 개 있다. 높은 외곽 시점에서 지붕을 내려다보는 구도는 맞지만, 감시탑은 지정된 장소 설명에 없는 별도 시설이다. 파도는 중앙부, 어두운 하늘은 상단에 자리한다.",
        "entities": "컨테이너 난민촌, 거대한 바닷물 벽, 어두운 밤하늘이 모두 보인다. 확실하게 식별되는 사람이나 얼굴, 읽을 수 있는 문구와 화면 자막은 없다. 도로에는 젖은 반사가 있지만, 해수가 이미 난민촌 안으로 유입되는 상태는 B보다 덜 명확하다. 방파제의 미복구 파손부는 식별되지 않는다.",
        "hard_violations": [
         "장소 설명에 없는 높은 경계 감시탑 한 개를 추가했다. 컨테이너 지붕과 노출된 거리만으로 장소를 구성하고 그 밖의 시설을 발명하지 말라는 제한에 어긋난다."
        ],
        "physics": "컨테이너는 지면에 놓여 있고 감시탑과 조명은 기둥으로 지지된다. 파도는 아래의 연속된 해수면에 연결되어 있으며, 마루에서 튀는 물보라와 아래로 떨어지는 포말에 운동 원인이 보인다. 도로의 빛 반사는 젖은 표면과 주변 조명으로 설명된다. 지지 없이 공중에 매달린 물체는 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "파도의 전면과 선행 침수대가 화면 아래 컨테이너 구역으로 밀려온다. 중앙 하단 인물은 파도 쪽을 향한 뒷모습으로 보이고, 오른쪽 아래 인물의 정확한 시선은 식별하기 어렵다. 카메라는 높은 위치에서 내려다보지만 A보다 파도에 정면으로 마주한 인상이 강하다.",
        "built_space": "컨테이너와 통로가 화면 하단 절반가량을 차지하며, 오른쪽에는 이층으로 쌓인 컨테이너가 있다. 여러 조명 기둥과 컨테이너 출입구 등이 보인다. 중앙 하단 통로에 한 명, 오른쪽 아래 통로에 한 명이 서 있다. 지붕을 내려다보는 시점은 맞지만 난민촌의 화면 점유가 크고 파도 마루가 위로 올라가 밤하늘의 여백이 좁다.",
        "entities": "컨테이너 난민촌, 어두운 밤하늘, 거대한 파도와 침수된 통로가 보인다. 사람은 최소 두 명이 명확하게 등장해 무인 장면 조건을 위반한다. 인물은 작고 어두워 나이·성별·민족성을 판단할 수 없다. 읽을 수 있는 문구나 자막은 없으며, 방파제 파손부는 식별되지 않는다.",
        "hard_violations": [
         "중앙 하단과 오른쪽 아래에 최소 두 명의 사람을 추가해, 어떠한 살아 있는 인물도 등장하지 말라는 명시적 금지 조건을 위반했다."
        ],
        "physics": "두 인물은 발을 통로 지면에 딛고 있으며 공중에 떠 있지 않다. 컨테이너는 지면 또는 아래층 컨테이너로 지지된다. 파도는 연속된 해수면에서 솟고, 포말과 침수 흐름이 난민촌까지 이어진다. 젖은 통로의 반사는 주변 조명과 양립하며, 지지 없는 부유 물체는 확인되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "하단 난민촌·중앙 파도·상단 밤하늘과 비스듬한 상승 시점은 더 정확하지만, 지정되지 않은 감시탑을 추가해 장소 제한을 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "난민촌으로 밀려드는 해수는 잘 보이지만, 금지된 인물이 최소 두 명 등장하고 난민촌이 화면 하단 절반까지 확대되어 지정 구도에서도 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "거대한 파도의 전면과 무너지는 포말이 화면 아래 난민촌을 향한다. 난민촌 경계가 왼쪽 아래에서 오른쪽 위로 이어져, 파도의 진행축에서 옆으로 비켜선 사선 관찰이 읽힌다. 확인 가능한 시선이나 조준 대상은 없다.",
        "built_space": "여러 줄의 컨테이너 지붕과 그 사이 노출된 도로가 하단에 배치되어 있다. 난민촌 바깥쪽에는 연속된 경계 울타리, 오른쪽 중경에는 높은 감시탑 한 개가 보이며 조명 기둥도 여러 개 있다. 높은 외곽 시점에서 지붕을 내려다보는 구도는 맞지만, 감시탑은 지정된 장소 설명에 없는 별도 시설이다. 파도는 중앙부, 어두운 하늘은 상단에 자리한다.",
        "entities": "컨테이너 난민촌, 거대한 바닷물 벽, 어두운 밤하늘이 모두 보인다. 확실하게 식별되는 사람이나 얼굴, 읽을 수 있는 문구와 화면 자막은 없다. 도로에는 젖은 반사가 있지만, 해수가 이미 난민촌 안으로 유입되는 상태는 B보다 덜 명확하다. 방파제의 미복구 파손부는 식별되지 않는다.",
        "hard_violations": [
         "장소 설명에 없는 높은 경계 감시탑 한 개를 추가했다. 컨테이너 지붕과 노출된 거리만으로 장소를 구성하고 그 밖의 시설을 발명하지 말라는 제한에 어긋난다."
        ],
        "physics": "컨테이너는 지면에 놓여 있고 감시탑과 조명은 기둥으로 지지된다. 파도는 아래의 연속된 해수면에 연결되어 있으며, 마루에서 튀는 물보라와 아래로 떨어지는 포말에 운동 원인이 보인다. 도로의 빛 반사는 젖은 표면과 주변 조명으로 설명된다. 지지 없이 공중에 매달린 물체는 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "파도의 전면과 선행 침수대가 화면 아래 컨테이너 구역으로 밀려온다. 중앙 하단 인물은 파도 쪽을 향한 뒷모습으로 보이고, 오른쪽 아래 인물의 정확한 시선은 식별하기 어렵다. 카메라는 높은 위치에서 내려다보지만 A보다 파도에 정면으로 마주한 인상이 강하다.",
        "built_space": "컨테이너와 통로가 화면 하단 절반가량을 차지하며, 오른쪽에는 이층으로 쌓인 컨테이너가 있다. 여러 조명 기둥과 컨테이너 출입구 등이 보인다. 중앙 하단 통로에 한 명, 오른쪽 아래 통로에 한 명이 서 있다. 지붕을 내려다보는 시점은 맞지만 난민촌의 화면 점유가 크고 파도 마루가 위로 올라가 밤하늘의 여백이 좁다.",
        "entities": "컨테이너 난민촌, 어두운 밤하늘, 거대한 파도와 침수된 통로가 보인다. 사람은 최소 두 명이 명확하게 등장해 무인 장면 조건을 위반한다. 인물은 작고 어두워 나이·성별·민족성을 판단할 수 없다. 읽을 수 있는 문구나 자막은 없으며, 방파제 파손부는 식별되지 않는다.",
        "hard_violations": [
         "중앙 하단과 오른쪽 아래에 최소 두 명의 사람을 추가해, 어떠한 살아 있는 인물도 등장하지 말라는 명시적 금지 조건을 위반했다."
        ],
        "physics": "두 인물은 발을 통로 지면에 딛고 있으며 공중에 떠 있지 않다. 컨테이너는 지면 또는 아래층 컨테이너로 지지된다. 파도는 연속된 해수면에서 솟고, 포말과 침수 흐름이 난민촌까지 이어진다. 젖은 통로의 반사는 주변 조명과 양립하며, 지지 없는 부유 물체는 확인되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 사람이 화면에 나타나서는 안 된다는 지시(NO PEOPLE IN THIS SHOT)를 위반하여 두 명 이상의 사람 실루엣이 포함됨.",
     "[gpt-high] 중앙 하단과 오른쪽 아래에 최소 두 명의 사람을 추가해, 어떠한 살아 있는 인물도 등장하지 말라는 명시적 금지 조건을 위반했다."
    ],
    "B": [
     "[gpt-high] 장소 설명에 없는 높은 경계 감시탑 한 개를 추가했다. 컨테이너 지붕과 노출된 거리만으로 장소를 구성하고 그 밖의 시설을 발명하지 말라는 제한에 어긋난다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 579
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 앵글과 피사체 비율을 잘 구현했으며, 프롬프트의 '사람이 없어야 한다'는 핵심 규칙을 완벽하게 준수했습니다.  ★위반: [gpt-high] 장소 설명에 없는 높은 경계 감시탑 한 개를 추가했다. 컨테이너 지붕과 노출된 거리만으로 장소를 구성하고 그 밖의 시설을 발명하지 말라는 제한에 어긋난다."
   },
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "화면에 어떤 인물도 나타나서는 안 된다는 명시적인 지시에도 불구하고 사람의 실루엣이 묘사되어 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 사람이 화면에 나타나서는 안 된다는 지시(NO PEOPLE IN THIS SHOT)를 위반하여 두 명 이상의 사람 실루엣이 포함됨. / [gpt-high] 중앙 하단과 오른쪽 아래에 최소 두 명의 사람을 추가해, 어떠한 살아 있는 인물도 등장하지 말라는 명시적 금지 조건을 위반했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a35-38e7-7243-98d6-d2ba95b6a048",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S36sh4::signage": {
  "fp": "a179971e6bea01d6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::secret_sewer_route": {
  "input_fingerprint": "0f088a037051a31f",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "secret_sewer_route",
    "tags": [
     "S36sh4"
    ]
   },
   "context_sig": "c0bde0b04c313d21"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 밖으로 나가는 비밀 통로가 있어. 이 길은 아무도 몰라.\n- 맨홀 하수도를 달리던 구도환, 앰버, 찰리, 신부님.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 밖으로 나가는 비밀 통로가 있어. 이 길은 아무도 몰라.\n- 맨홀 하수도를 달리던 구도환, 앰버, 찰리, 신부님.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_secret_sewer_route_03a6aa.png",
  "asset_id": "f261e43b-2b4c-4e43-81c2-4fd7e438de90",
  "input_asset_ids": [
   "6677edce-34aa-4d18-b33a-1106606c8648"
  ],
  "origin_tag": "S36sh4",
  "place_text": "Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.",
  "origin_inputs": {
   "place_text": "Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.",
   "time_of_day_en": "night",
   "conti_asset_id": "6677edce-34aa-4d18-b33a-1106606c8648"
  }
 },
 "S36sh4::bgfirst_bg": {
  "input_fingerprint": "a7ebab582d8d0dde",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구도환의 등 뒤 하수도 터널 양쪽에서 거대한 바닷물이 폭포수처럼 쏟아져 들어오는 찰나.\n\nLOCATION (lock): Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated front-quarter position inside the sewer, continue the restrained crane rise and angle downward past 구도환's near shoulder into the tunnel behind him. Catch him mid-stride along the lower-left edge, looking toward his escape route beyond the camera's right side, while separate torrents enter behind him from the rear-left and rear-right and converge toward the central passage. The widened viewing distance is the reveal's emphasis: preserve clear depth between the runner and the incoming water, with the other companions outside this crop.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left torrent behind the runner in the middle-left of the frame, background; Right torrent behind the runner in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Sewer tunnel (Beginning to fill as seawater enters from multiple directions) — The passage recedes behind 구도환, with both lateral sides visible; used as Establishes the depth and enclosure necessary to read the approaching flood; Rear-left seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the left rear toward the central passage; used as One side of the converging threat, kept distinct from the opposite torrent; Rear-right seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the right rear toward the central passage; used as Completes the bilateral flood reveal without obscuring the runner.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the sewer maintains readable separation between 구도환, the tunnel, and the water without introducing visible fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 구도환의 등 뒤 하수도 터널 양쪽에서 거대한 바닷물이 폭포수처럼 쏟아져 들어오는 찰나.\n\nLOCATION (lock): Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated front-quarter position inside the sewer, continue the restrained crane rise and angle downward past 구도환's near shoulder into the tunnel behind him. Catch him mid-stride along the lower-left edge, looking toward his escape route beyond the camera's right side, while separate torrents enter behind him from the rear-left and rear-right and converge toward the central passage. The widened viewing distance is the reveal's emphasis: preserve clear depth between the runner and the incoming water, with the other companions outside this crop.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left torrent behind the runner in the middle-left of the frame, background; Right torrent behind the runner in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Sewer tunnel (Beginning to fill as seawater enters from multiple directions) — The passage recedes behind 구도환, with both lateral sides visible; used as Establishes the depth and enclosure necessary to read the approaching flood; Rear-left seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the left rear toward the central passage; used as One side of the converging threat, kept distinct from the opposite torrent; Rear-right seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the right rear toward the central passage; used as Completes the bilateral flood reveal without obscuring the runner.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the sewer maintains readable separation between 구도환, the tunnel, and the water without introducing visible fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh4__bgfirst_bg.png",
  "asset_id": "1247f538-7085-407c-85da-0a6f36f7a2d1",
  "input_asset_ids": [
   "6677edce-34aa-4d18-b33a-1106606c8648",
   "f261e43b-2b4c-4e43-81c2-4fd7e438de90"
  ]
 },
 "S36sh4": {
  "input_fingerprint": "77a0eca1703d6843",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환의 등 뒤 하수도 터널 양쪽에서 거대한 바닷물이 폭포수처럼 쏟아져 들어오는 찰나.\n\nLOCATION (lock): Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated front-quarter position inside the sewer, continue the restrained crane rise and angle downward past 구도환's near shoulder into the tunnel behind him. Catch him mid-stride along the lower-left edge, looking toward his escape route beyond the camera's right side, while separate torrents enter behind him from the rear-left and rear-right and converge toward the central passage. The widened viewing distance is the reveal's emphasis: preserve clear depth between the runner and the incoming water, with the other companions outside this crop.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left torrent behind the runner in the middle-left of the frame, background; Right torrent behind the runner in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Sewer tunnel (Beginning to fill as seawater enters from multiple directions) — The passage recedes behind 구도환, with both lateral sides visible; used as Establishes the depth and enclosure necessary to read the approaching flood; Rear-left seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the left rear toward the central passage; used as One side of the converging threat, kept distinct from the opposite torrent; Rear-right seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the right rear toward the central passage; used as Completes the bilateral flood reveal without obscuring the runner.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the sewer maintains readable separation between 구도환, the tunnel, and the water without introducing visible fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground sewer is beginning to fill with seawater; this is the same flood released by the destroyed seawall. 구도환: He is on foot inside the sewer tunnel during the escape. Huge streams of water pour into the tunnel from multiple directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환의 등 뒤 하수도 터널 양쪽에서 거대한 바닷물이 폭포수처럼 쏟아져 들어오는 찰나.\n\nLOCATION (lock): Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated front-quarter position inside the sewer, continue the restrained crane rise and angle downward past 구도환's near shoulder into the tunnel behind him. Catch him mid-stride along the lower-left edge, looking toward his escape route beyond the camera's right side, while separate torrents enter behind him from the rear-left and rear-right and converge toward the central passage. The widened viewing distance is the reveal's emphasis: preserve clear depth between the runner and the incoming water, with the other companions outside this crop.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left torrent behind the runner in the middle-left of the frame, background; Right torrent behind the runner in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Sewer tunnel (Beginning to fill as seawater enters from multiple directions) — The passage recedes behind 구도환, with both lateral sides visible; used as Establishes the depth and enclosure necessary to read the approaching flood; Rear-left seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the left rear toward the central passage; used as One side of the converging threat, kept distinct from the opposite torrent; Rear-right seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the right rear toward the central passage; used as Completes the bilateral flood reveal without obscuring the runner.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the sewer maintains readable separation between 구도환, the tunnel, and the water without introducing visible fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground sewer is beginning to fill with seawater; this is the same flood released by the destroyed seawall. 구도환: He is on foot inside the sewer tunnel during the escape. Huge streams of water pour into the tunnel from multiple directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 구도환의 등 뒤 하수도 터널 양쪽에서 거대한 바닷물이 폭포수처럼 쏟아져 들어오는 찰나.\n\nLOCATION (lock): Inside the underground sewer tunnel beneath the settlement, where seawater bursts in from both sides. The passage has no established fixed light source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From an elevated front-quarter position inside the sewer, continue the restrained crane rise and angle downward past 구도환's near shoulder into the tunnel behind him. Catch him mid-stride along the lower-left edge, looking toward his escape route beyond the camera's right side, while separate torrents enter behind him from the rear-left and rear-right and converge toward the central passage. The widened viewing distance is the reveal's emphasis: preserve clear depth between the runner and the incoming water, with the other companions outside this crop.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left torrent behind the runner in the middle-left of the frame, background; Right torrent behind the runner in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Sewer tunnel (Beginning to fill as seawater enters from multiple directions) — The passage recedes behind 구도환, with both lateral sides visible; used as Establishes the depth and enclosure necessary to read the approaching flood; Rear-left seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the left rear toward the central passage; used as One side of the converging threat, kept distinct from the opposite torrent; Rear-right seawater torrent (Pouring into the tunnel behind 구도환) — Enters from the right rear toward the central passage; used as Completes the bilateral flood reveal without obscuring the runner.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination appropriate to the sewer maintains readable separation between 구도환, the tunnel, and the water without introducing visible fixtures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground sewer is beginning to fill with seawater; this is the same flood released by the destroyed seawall. 구도환: He is on foot inside the sewer tunnel during the escape. Huge streams of water pour into the tunnel from multiple directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh4__bgfirst_bg.png",
     "asset_id": "1247f538-7085-407c-85da-0a6f36f7a2d1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S36sh4.png",
     "asset_id": "6677edce-34aa-4d18-b33a-1106606c8648",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_secret_sewer_route_03a6aa.png",
     "asset_id": "f261e43b-2b4c-4e43-81c2-4fd7e438de90",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물은 우측 도주로를 바라보고, 양쪽의 물줄기는 터널 중앙을 향해 쏟아짐.",
    "built_space": "레퍼런스와 동일한 아치형 하수도 터널, 파이프, 사다리. 인물은 화면 좌측 하단에 위치.",
    "entities": "구도환의 얼굴, 헤어스타일, 의상(낡은 갈색 점퍼, 올리브색 티셔츠와 바지)이 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "물 아래 바닥을 딛고 달리는 동작이 자연스러우며, 물보라가 물리적으로 타당함."
   },
   {
    "label": "B",
    "direction": "인물은 우측을 응시하고, 양쪽 폭포수는 중앙으로 유입됨.",
    "built_space": "터널 내부 구조(아치, 배관, 사다리)가 일치하며, 인물은 좌측 하단에 배치됨.",
    "entities": "구도환의 외모와 복장이 레퍼런스와 잘 일치함.",
    "hard_violations": [],
    "physics": "물살을 헤치며 달리는 자세와 발의 지지 상태가 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글, 화면 배치, 인물의 외모와 의상 레퍼런스를 정확히 반영했으며, 사실적인 물의 질감과 조명이 돋보임."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물 설정 등 요구사항을 잘 충족했으나, A에 비해 동작의 자연스러움과 시네마틱한 디테일이 다소 떨어짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 도주로를 바라보고, 양쪽의 물줄기는 터널 중앙을 향해 쏟아짐.",
        "built_space": "레퍼런스와 동일한 아치형 하수도 터널, 파이프, 사다리. 인물은 화면 좌측 하단에 위치.",
        "entities": "구도환의 얼굴, 헤어스타일, 의상(낡은 갈색 점퍼, 올리브색 티셔츠와 바지)이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "물 아래 바닥을 딛고 달리는 동작이 자연스러우며, 물보라가 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "인물은 우측을 응시하고, 양쪽 폭포수는 중앙으로 유입됨.",
        "built_space": "터널 내부 구조(아치, 배관, 사다리)가 일치하며, 인물은 좌측 하단에 배치됨.",
        "entities": "구도환의 외모와 복장이 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "물살을 헤치며 달리는 자세와 발의 지지 상태가 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글, 화면 배치, 인물의 외모와 의상 레퍼런스를 정확히 반영했으며, 사실적인 물의 질감과 조명이 돋보임."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물 설정 등 요구사항을 잘 충족했으나, A에 비해 동작의 자연스러움과 시네마틱한 디테일이 다소 떨어짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 도주로를 바라보고, 양쪽의 물줄기는 터널 중앙을 향해 쏟아짐.",
        "built_space": "레퍼런스와 동일한 아치형 하수도 터널, 파이프, 사다리. 인물은 화면 좌측 하단에 위치.",
        "entities": "구도환의 얼굴, 헤어스타일, 의상(낡은 갈색 점퍼, 올리브색 티셔츠와 바지)이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "물 아래 바닥을 딛고 달리는 동작이 자연스러우며, 물보라가 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "인물은 우측을 응시하고, 양쪽 폭포수는 중앙으로 유입됨.",
        "built_space": "터널 내부 구조(아치, 배관, 사다리)가 일치하며, 인물은 좌측 하단에 배치됨.",
        "entities": "구도환의 외모와 복장이 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "물살을 헤치며 달리는 자세와 발의 지지 상태가 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "더 높은 하향 시점과 작은 인물 비중으로 좌하단의 구도환, 뒤쪽 양 갈래 급류, 그 사이의 깊이를 보여 주어 지정된 와이드 공개 구도에 더 충실하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽 탈출로를 보는 시선과 양측 유입은 맞지만, 인물이 더 크게 잡히고 하향 시점이 약해 넓어진 관찰 거리와 터널 깊이를 강조하라는 지시에서 뒤처진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구도환은 얼굴과 시선을 화면 오른쪽 밖으로 돌리고 있으며, 몸은 카메라 쪽 전경으로 달려 나온다. 왼쪽 뒤 개구부의 물은 오른쪽 아래 중앙 수로로, 오른쪽 뒤 개구부의 물은 왼쪽 아래 중앙 수로로 쏟아진다. 두 급류의 출발점이 분리되어 있고 인물 뒤에서 합류한다.",
        "built_space": "젖은 콘크리트 아치형 주 통로 하나, 좌우 측면의 큰 아치형 유입구 각 하나, 양 벽 상단의 주 배관 각 한 줄, 양쪽 낮은 턱, 천장 개구부 하나와 그 아래 사다리 하나가 보인다. 장소 참조의 주요 구조와 수량이 맞는다. 인물은 왼쪽 턱 위가 아니라 물이 찬 중앙 통로의 왼쪽 전경에 있다. 카메라는 터널 내부에서 내려다보며, 인물과 뒤쪽 유입구 사이에 수면이 충분히 드러난다. 별도 조명기구나 불가능한 반사는 보이지 않는다.",
        "entities": "사람은 구도환 한 명뿐이다. 짧은 검은 머리의 중년 동아시아계 남성으로, 참조의 얼굴과 체격에 대체로 부합한다. 한국 국적 자체는 영상으로 확인할 수 없다. 낡은 갈색 점퍼, 짙은 올리브색 속옷, 주름진 올리브색 면바지가 보이며 참조 복장과 잘 맞는다. 거대한 두 물줄기와 차오르는 물은 확인되지만 해수 여부는 외관만으로 판별할 수 없다. 추가 인물, 휴대 소품, 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "구도환의 한쪽 다리는 아래로 뻗고 반대쪽 다리는 뒤로 굽혀지며, 양팔은 달리기에 맞게 교차한다. 발과 바닥의 접촉은 물보라에 가려져 직접 보이지 않지만, 다리가 수면 아래로 이어지고 주변에 물보라가 생겨 침수된 바닥을 딛는 달리기로 읽힌다. 공중에 매달린 몸은 아니다. 물은 양측 통로에서 연속적으로 유입되어 중력에 따라 떨어지고 중앙 수면에 부딪혀 퍼진다."
       },
       {
        "label": "B",
        "direction": "구도환의 얼굴과 눈은 화면 오른쪽 밖의 탈출 방향을 향한다. 몸은 전경으로 달려오며 양팔을 굽혀 움직인다. 뒤쪽 왼쪽 물줄기는 중앙 오른쪽으로, 뒤쪽 오른쪽 물줄기는 중앙 왼쪽으로 낙하하여 주 통로에서 합쳐진다. 급류가 인물 뒤에 있다는 관계는 명확하다.",
        "built_space": "아치형 주 통로 하나와 측면 유입구 좌우 각 하나, 상단 주 배관 좌우 각 한 줄, 양쪽 낮은 턱, 천장 개구부 하나와 사다리 하나가 보인다. 콘크리트 표면과 배관 배치는 장소 참조에 부합한다. 인물은 물이 찬 통로의 왼쪽 전경에 있지만 A보다 화면을 크게 차지한다. 시점도 더 낮고 정면에 가까워, 어깨 너머 아래로 내려다보는 와이드 공개 구도의 효과가 약하다. 추가 조명기구나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명만 보이며 참조의 구도환과 얼굴·체격이 대체로 맞는다. 국적은 외관만으로 확정할 수 없다. 갈색의 해진 점퍼와 올리브색 속옷·면바지는 맞지만 점퍼 소매가 걷혀 있어 참조와 착용 상태가 다르다. 양측의 거대한 물줄기와 침수 중인 하수도는 분명하다. 해수 성분은 시각적으로 확인할 수 없다. 동료나 별도 소품, 글자는 없다.",
        "hard_violations": [],
        "physics": "앞으로 나오는 허벅지와 뒤쪽으로 이어지는 반대 다리, 교차하는 팔이 달리기 동작을 이룬다. 발은 화면 하단과 물보라에 가려져 접지 자체는 확인되지 않지만, 하체가 수면으로 이어져 침수된 바닥을 딛고 움직이는 모습으로 읽힌다. 부유나 불가능한 관절은 보이지 않는다. 두 급류는 측면 유입구에 연결되어 있으며 낙하 후 중앙 수면에 충돌하는 흐름도 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "더 높은 하향 시점과 작은 인물 비중으로 좌하단의 구도환, 뒤쪽 양 갈래 급류, 그 사이의 깊이를 보여 주어 지정된 와이드 공개 구도에 더 충실하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽 탈출로를 보는 시선과 양측 유입은 맞지만, 인물이 더 크게 잡히고 하향 시점이 약해 넓어진 관찰 거리와 터널 깊이를 강조하라는 지시에서 뒤처진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "구도환은 얼굴과 시선을 화면 오른쪽 밖으로 돌리고 있으며, 몸은 카메라 쪽 전경으로 달려 나온다. 왼쪽 뒤 개구부의 물은 오른쪽 아래 중앙 수로로, 오른쪽 뒤 개구부의 물은 왼쪽 아래 중앙 수로로 쏟아진다. 두 급류의 출발점이 분리되어 있고 인물 뒤에서 합류한다.",
        "built_space": "젖은 콘크리트 아치형 주 통로 하나, 좌우 측면의 큰 아치형 유입구 각 하나, 양 벽 상단의 주 배관 각 한 줄, 양쪽 낮은 턱, 천장 개구부 하나와 그 아래 사다리 하나가 보인다. 장소 참조의 주요 구조와 수량이 맞는다. 인물은 왼쪽 턱 위가 아니라 물이 찬 중앙 통로의 왼쪽 전경에 있다. 카메라는 터널 내부에서 내려다보며, 인물과 뒤쪽 유입구 사이에 수면이 충분히 드러난다. 별도 조명기구나 불가능한 반사는 보이지 않는다.",
        "entities": "사람은 구도환 한 명뿐이다. 짧은 검은 머리의 중년 동아시아계 남성으로, 참조의 얼굴과 체격에 대체로 부합한다. 한국 국적 자체는 영상으로 확인할 수 없다. 낡은 갈색 점퍼, 짙은 올리브색 속옷, 주름진 올리브색 면바지가 보이며 참조 복장과 잘 맞는다. 거대한 두 물줄기와 차오르는 물은 확인되지만 해수 여부는 외관만으로 판별할 수 없다. 추가 인물, 휴대 소품, 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "구도환의 한쪽 다리는 아래로 뻗고 반대쪽 다리는 뒤로 굽혀지며, 양팔은 달리기에 맞게 교차한다. 발과 바닥의 접촉은 물보라에 가려져 직접 보이지 않지만, 다리가 수면 아래로 이어지고 주변에 물보라가 생겨 침수된 바닥을 딛는 달리기로 읽힌다. 공중에 매달린 몸은 아니다. 물은 양측 통로에서 연속적으로 유입되어 중력에 따라 떨어지고 중앙 수면에 부딪혀 퍼진다."
       },
       {
        "label": "A",
        "direction": "구도환의 얼굴과 눈은 화면 오른쪽 밖의 탈출 방향을 향한다. 몸은 전경으로 달려오며 양팔을 굽혀 움직인다. 뒤쪽 왼쪽 물줄기는 중앙 오른쪽으로, 뒤쪽 오른쪽 물줄기는 중앙 왼쪽으로 낙하하여 주 통로에서 합쳐진다. 급류가 인물 뒤에 있다는 관계는 명확하다.",
        "built_space": "아치형 주 통로 하나와 측면 유입구 좌우 각 하나, 상단 주 배관 좌우 각 한 줄, 양쪽 낮은 턱, 천장 개구부 하나와 사다리 하나가 보인다. 콘크리트 표면과 배관 배치는 장소 참조에 부합한다. 인물은 물이 찬 통로의 왼쪽 전경에 있지만 A보다 화면을 크게 차지한다. 시점도 더 낮고 정면에 가까워, 어깨 너머 아래로 내려다보는 와이드 공개 구도의 효과가 약하다. 추가 조명기구나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명만 보이며 참조의 구도환과 얼굴·체격이 대체로 맞는다. 국적은 외관만으로 확정할 수 없다. 갈색의 해진 점퍼와 올리브색 속옷·면바지는 맞지만 점퍼 소매가 걷혀 있어 참조와 착용 상태가 다르다. 양측의 거대한 물줄기와 침수 중인 하수도는 분명하다. 해수 성분은 시각적으로 확인할 수 없다. 동료나 별도 소품, 글자는 없다.",
        "hard_violations": [],
        "physics": "앞으로 나오는 허벅지와 뒤쪽으로 이어지는 반대 다리, 교차하는 팔이 달리기 동작을 이룬다. 발은 화면 하단과 물보라에 가려져 접지 자체는 확인되지 않지만, 하체가 수면으로 이어져 침수된 바닥을 딛고 움직이는 모습으로 읽힌다. 부유나 불가능한 관절은 보이지 않는다. 두 급류는 측면 유입구에 연결되어 있으며 낙하 후 중앙 수면에 충돌하는 흐름도 자연스럽다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.778,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.778,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1778,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1778,
    "verdict_ko": "지정된 앵글, 화면 배치, 인물의 외모와 의상 레퍼런스를 정확히 반영했으며, 사실적인 물의 질감과 조명이 돋보임."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "구도와 인물 설정 등 요구사항을 잘 충족했으나, A에 비해 동작의 자연스러움과 시네마틱한 디테일이 다소 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_secret_sewer_route_03a6aa.png",
    "asset_id": "f261e43b-2b4c-4e43-81c2-4fd7e438de90",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a3a-4244-7263-b730-19387048fb92",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh4__bgfirst_bg.png",
   "bg_asset_id": "1247f538-7085-407c-85da-0a6f36f7a2d1",
   "bg_record_key": "S36sh4::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "secret_sewer_route",
   "groupbg_asset_id": "f261e43b-2b4c-4e43-81c2-4fd7e438de90"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S36sh7::signage": {
  "fp": "654d21d7ad60bd20",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::flood_outlet_water": {
  "input_fingerprint": "f9249fe7727776b6",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "flood_outlet_water",
    "tags": [
     "S36sh7"
    ]
   },
   "context_sig": "517ae2077e017e0a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 하수도 밖으로 밀려나와 바닷물에 잠기며 허우적거리는 앰버!!\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 지하 하수도 터널·갈림길, 침수된 수중: 지하에 위치한 어둡고 축축한 하수관로로 두 갈래 길과 사다리가 있다. (특징: 천장에서 빛이 떨어지는 사다리 입구; 더러운 물이 흐르는 콘크리트 벽면과 바닥; 하수관을 가득 메우며 쏟아져 내리는 거대한 바닷물 해일; 물에 완전히 잠긴 하수도 내부에서 허우적대는 앰버와 기계 관절로 수영하는 찰리)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 하수도 밖으로 밀려나와 바닷물에 잠기며 허우적거리는 앰버!!\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_outlet_water_33110b.png",
  "asset_id": "822689d4-88c5-4fa3-b7e3-f31bee70a02f",
  "input_asset_ids": [
   "cc262e8c-bc85-49aa-b36c-8afa0f294417"
  ],
  "origin_tag": "S36sh7",
  "place_text": "In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.",
  "origin_inputs": {
   "place_text": "In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.",
   "time_of_day_en": "night",
   "conti_asset_id": "cc262e8c-bc85-49aa-b36c-8afa0f294417"
  }
 },
 "S36sh7::bgfirst_bg": {
  "input_fingerprint": "46bf1aef699ed3df",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 물속에서 찰리의 거대한 기계 손이 앰버의 작은 손을 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Underwater outside the sewer, remain on the established side of 찰리 and 앰버, slightly below their joining hands and looking obliquely upward along their forearms as the tracking move begins. Place 찰리's mechanical hand left of center, occupying roughly a quarter of the image, as its fingers close firmly around 앰버's smaller hand from the right; cropped arm and torso edges retain bodily scale without including either face. Both are engaged in securing the rescue grip, their eyes outside the crop, and the tightened camera distance makes this physical connection the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Both characters are submerged after being swept out of the sewer); used as Continuous surrounding space for the joined hands and cropped bodies, with no additional background structures introduced.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained underwater visibility and controlled highlights distinguish the mechanical fingers from 앰버's hand without adding a colored source or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 물속에서 찰리의 거대한 기계 손이 앰버의 작은 손을 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Underwater outside the sewer, remain on the established side of 찰리 and 앰버, slightly below their joining hands and looking obliquely upward along their forearms as the tracking move begins. Place 찰리's mechanical hand left of center, occupying roughly a quarter of the image, as its fingers close firmly around 앰버's smaller hand from the right; cropped arm and torso edges retain bodily scale without including either face. Both are engaged in securing the rescue grip, their eyes outside the crop, and the tightened camera distance makes this physical connection the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Both characters are submerged after being swept out of the sewer); used as Continuous surrounding space for the joined hands and cropped bodies, with no additional background structures introduced.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained underwater visibility and controlled highlights distinguish the mechanical fingers from 앰버's hand without adding a colored source or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh7__bgfirst_bg.png",
  "asset_id": "020902ca-091e-4c8c-b820-5a60dc0cea20",
  "input_asset_ids": [
   "cc262e8c-bc85-49aa-b36c-8afa0f294417",
   "822689d4-88c5-4fa3-b7e3-f31bee70a02f"
  ]
 },
 "S36sh7": {
  "input_fingerprint": "2c7e44e50b66545f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 물속에서 찰리의 거대한 기계 손이 앰버의 작은 손을 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Underwater outside the sewer, remain on the established side of 찰리 and 앰버, slightly below their joining hands and looking obliquely upward along their forearms as the tracking move begins. Place 찰리's mechanical hand left of center, occupying roughly a quarter of the image, as its fingers close firmly around 앰버's smaller hand from the right; cropped arm and torso edges retain bodily scale without including either face. Both are engaged in securing the rescue grip, their eyes outside the crop, and the tightened camera distance makes this physical connection the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Both characters are submerged after being swept out of the sewer); used as Continuous surrounding space for the joined hands and cropped bodies, with no additional background structures introduced.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained underwater visibility and controlled highlights distinguish the mechanical fingers from 앰버's hand without adding a colored source or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has been overwhelmed by seawater, and the flood continues outside its outlet. 앰버: She is soaked and submerged outside the sewer, struggling to stay afloat with one hand held out in the water. 찰리: He is swimming in the floodwater with one hand closed in a firm grip. His old coat and hat are soaked; no removal has been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 물속에서 찰리의 거대한 기계 손이 앰버의 작은 손을 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Underwater outside the sewer, remain on the established side of 찰리 and 앰버, slightly below their joining hands and looking obliquely upward along their forearms as the tracking move begins. Place 찰리's mechanical hand left of center, occupying roughly a quarter of the image, as its fingers close firmly around 앰버's smaller hand from the right; cropped arm and torso edges retain bodily scale without including either face. Both are engaged in securing the rescue grip, their eyes outside the crop, and the tightened camera distance makes this physical connection the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Both characters are submerged after being swept out of the sewer); used as Continuous surrounding space for the joined hands and cropped bodies, with no additional background structures introduced.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained underwater visibility and controlled highlights distinguish the mechanical fingers from 앰버's hand without adding a colored source or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has been overwhelmed by seawater, and the flood continues outside its outlet. 앰버: She is soaked and submerged outside the sewer, struggling to stay afloat with one hand held out in the water. 찰리: He is swimming in the floodwater with one hand closed in a firm grip. His old coat and hat are soaked; no removal has been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 물속에서 찰리의 거대한 기계 손이 앰버의 작은 손을 꽉 움켜쥔 근접 찰나.\n\nLOCATION (lock): In open floodwater outside the sewer outlet, beneath the surface of the inundated settlement. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Underwater outside the sewer, remain on the established side of 찰리 and 앰버, slightly below their joining hands and looking obliquely upward along their forearms as the tracking move begins. Place 찰리's mechanical hand left of center, occupying roughly a quarter of the image, as its fingers close firmly around 앰버's smaller hand from the right; cropped arm and torso edges retain bodily scale without including either face. Both are engaged in securing the rescue grip, their eyes outside the crop, and the tightened camera distance makes this physical connection the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater outside the sewer (Both characters are submerged after being swept out of the sewer); used as Continuous surrounding space for the joined hands and cropped bodies, with no additional background structures introduced.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained underwater visibility and controlled highlights distinguish the mechanical fingers from 앰버's hand without adding a colored source or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sewer has been overwhelmed by seawater, and the flood continues outside its outlet. 앰버: She is soaked and submerged outside the sewer, struggling to stay afloat with one hand held out in the water. 찰리: He is swimming in the floodwater with one hand closed in a firm grip. His old coat and hat are soaked; no removal has been established.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh7__bgfirst_bg.png",
     "asset_id": "020902ca-091e-4c8c-b820-5a60dc0cea20",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S36sh7.png",
     "asset_id": "cc262e8c-bc85-49aa-b36c-8afa0f294417",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_outlet_water_33110b.png",
     "asset_id": "822689d4-88c5-4fa3-b7e3-f31bee70a02f",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 팔이 수면의 빛을 향해 대각선 위로 뻗어 있으며, 카메라는 비스듬히 올려다보는 시점입니다.",
    "built_space": "추가 구조물이 없는 탁 트인 수중 공간이며 위쪽으로 수면이 보입니다.",
    "entities": "좌측에 찰리의 샌드 베이지 기계 팔과 젖은 어두운 옷소매, 우측에 앰버의 작은 맨손과 카키색 소매, 금발 머리가 확인됩니다.",
    "hard_violations": [],
    "physics": "완전히 물에 잠겨 부력을 받고 있으며, 기포가 수면을 향해 위로 자연스럽게 상승합니다. 두 손이 단단히 지탱하며 맞잡고 있습니다."
   },
   {
    "label": "B",
    "direction": "두 팔이 거의 수평으로 마주 뻗어 있으며, 카메라 시선도 수평에 가깝습니다.",
    "built_space": "수중 환경이며 상단에 수면이 보입니다.",
    "entities": "좌측에 찰리의 기계 팔과 흉부 원자로, 우측에 앰버의 카키색 작업복과 마스크, 금발 머리가 보입니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 연출: 완전한 수중 환경임에도 불구하고 맞잡은 손 주위로 공기 중에서처럼 물방울이 아래로 떨어지고 있습니다."
    ],
    "physics": "배경은 수중이나, 손 주위의 물보라와 떨어지는 물방울은 공기 중의 물리 법칙을 따르고 있어 모순됩니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 비스듬히 올려다보는 카메라 앵글과 클로즈업 프레이밍을 정확히 따랐으며, 수중 물리 법칙을 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "카메라 앵글이 수평에 가깝고, 물속임에도 물방울이 아래로 떨어지는 물리적 오류가 있어 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 팔이 수면의 빛을 향해 대각선 위로 뻗어 있으며, 카메라는 비스듬히 올려다보는 시점입니다.",
        "built_space": "추가 구조물이 없는 탁 트인 수중 공간이며 위쪽으로 수면이 보입니다.",
        "entities": "좌측에 찰리의 샌드 베이지 기계 팔과 젖은 어두운 옷소매, 우측에 앰버의 작은 맨손과 카키색 소매, 금발 머리가 확인됩니다.",
        "hard_violations": [],
        "physics": "완전히 물에 잠겨 부력을 받고 있으며, 기포가 수면을 향해 위로 자연스럽게 상승합니다. 두 손이 단단히 지탱하며 맞잡고 있습니다."
       },
       {
        "label": "B",
        "direction": "두 팔이 거의 수평으로 마주 뻗어 있으며, 카메라 시선도 수평에 가깝습니다.",
        "built_space": "수중 환경이며 상단에 수면이 보입니다.",
        "entities": "좌측에 찰리의 기계 팔과 흉부 원자로, 우측에 앰버의 카키색 작업복과 마스크, 금발 머리가 보입니다.",
        "hard_violations": [
         "물리적으로 불가능한 연출: 완전한 수중 환경임에도 불구하고 맞잡은 손 주위로 공기 중에서처럼 물방울이 아래로 떨어지고 있습니다."
        ],
        "physics": "배경은 수중이나, 손 주위의 물보라와 떨어지는 물방울은 공기 중의 물리 법칙을 따르고 있어 모순됩니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 비스듬히 올려다보는 카메라 앵글과 클로즈업 프레이밍을 정확히 따랐으며, 수중 물리 법칙을 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "카메라 앵글이 수평에 가깝고, 물속임에도 물방울이 아래로 떨어지는 물리적 오류가 있어 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 팔이 수면의 빛을 향해 대각선 위로 뻗어 있으며, 카메라는 비스듬히 올려다보는 시점입니다.",
        "built_space": "추가 구조물이 없는 탁 트인 수중 공간이며 위쪽으로 수면이 보입니다.",
        "entities": "좌측에 찰리의 샌드 베이지 기계 팔과 젖은 어두운 옷소매, 우측에 앰버의 작은 맨손과 카키색 소매, 금발 머리가 확인됩니다.",
        "hard_violations": [],
        "physics": "완전히 물에 잠겨 부력을 받고 있으며, 기포가 수면을 향해 위로 자연스럽게 상승합니다. 두 손이 단단히 지탱하며 맞잡고 있습니다."
       },
       {
        "label": "B",
        "direction": "두 팔이 거의 수평으로 마주 뻗어 있으며, 카메라 시선도 수평에 가깝습니다.",
        "built_space": "수중 환경이며 상단에 수면이 보입니다.",
        "entities": "좌측에 찰리의 기계 팔과 흉부 원자로, 우측에 앰버의 카키색 작업복과 마스크, 금발 머리가 보입니다.",
        "hard_violations": [
         "물리적으로 불가능한 연출: 완전한 수중 환경임에도 불구하고 맞잡은 손 주위로 공기 중에서처럼 물방울이 아래로 떨어지고 있습니다."
        ],
        "physics": "배경은 수중이나, 손 주위의 물보라와 떨어지는 물방울은 공기 중의 물리 법칙을 따르고 있어 모순됩니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "수중 구조 악력과 얼굴 없는 근접 구도는 맞지만, 기계 손이 중앙에 놓이고 상대적으로 작아 지정된 중앙 왼쪽의 큰 손 강조는 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙 왼쪽의 거대한 기계 손이 오른쪽에서 뻗은 작은 손을 감싸며, 낮은 수중 시점과 잘린 몸통으로 구조의 연결만 강조하는 구도를 더 정확히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 팔은 왼쪽 위에서 중앙 아래로, 앰버의 팔은 오른쪽에서 중앙으로 뻗어 서로의 손에 도달한다. 기계 손가락은 앰버의 손을 감싸 닫혀 있다. 두 얼굴과 눈은 화면 밖이다. 카메라는 손 아래에서 팔과 수면을 비스듬히 올려다본다.",
        "built_space": "배경은 기포가 섞인 물과 위쪽 수면뿐이며 터널, 사다리, 배관 등 고정 구조물은 보이지 않는다. 구조물을 추가하지 말라는 외부 수중 장면 지시와 맞는다. 찰리의 몸통 일부는 왼쪽, 앰버의 몸통 일부는 오른쪽에 잘려 들어온다. 다만 잡은 손은 중앙 왼쪽보다 화면 중앙에 가깝다.",
        "entities": "샌드 베이지 장갑판과 검은 관절의 기계 팔, 왼쪽 가장자리의 푸른 원자로 일부가 찰리의 설정과 맞는다. 어깨에는 젖은 낡은 천이 보인다. 앰버는 작은 창백한 손, 금발 일부, 때 묻은 카키 작업복과 마스크 일부로 나타난다. 손의 크기는 아동 설정과 부합하지만 얼굴이 없어 정확한 나이와 혼혈 정체성은 확인할 수 없다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 팔에 연결되어 있고 기계 손가락이 작은 손 주위를 실제로 감싼다. 몸통은 화면 양쪽으로 이어져 있어 분리되거나 떠 있는 손이 아니다. 잠긴 신체와 천, 머리카락은 주변 물의 부력과 저항을 받는 모습이며 기포는 수면 쪽에 분포한다. 수영의 추진 동작은 크롭 밖이므로 확인할 수 없지만 불가능한 지지 상태는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 팔은 왼쪽 위에서 중앙 왼쪽으로 내려오며, 앰버의 팔은 오른쪽 아래에서 그 손을 향해 뻗는다. 거대한 기계 손의 엄지와 굽힌 손가락이 앰버의 작은 손을 위아래로 감싸 구조 악력을 형성한다. 눈과 얼굴은 모두 제외되어 있다. 위쪽 수면과 팔의 단축감이 손 아래에서 비스듬히 올려다보는 시점을 뒷받침한다.",
        "built_space": "주변은 연속된 수중 공간이고 상단에 수면이 보인다. 건축물과 고정 설비는 하나도 도입되지 않았다. 찰리의 잘린 몸통과 외투는 왼쪽, 앰버의 어깨와 머리카락 일부는 오른쪽 아래에 있다. 기계 손이 중앙 왼쪽에서 크게 차지하여 지정된 화면 배치와 손 중심 근접 구도에 더 가깝다.",
        "entities": "찰리의 손과 전완은 마모된 샌드 베이지 금속 장갑판, 굵은 관절, 육중한 비율로 표현되어 참조의 기계 정체성과 맞는다. 낡은 외투의 젖은 천도 왼쪽에 보인다. 앰버에게는 작은 창백한 손과 가는 손목, 금발, 오염된 카키색 소매가 보인다. 얼굴이 제외되어 성별·정확한 연령·혼혈 외모를 독립적으로 확인할 수는 없다. 모자와 공구 벨트 등은 크롭 밖이므로 누락으로 판단하지 않는다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "작은 손이 기계 손바닥 안으로 들어가고 여러 기계 손가락이 그 주위를 닫아 접촉과 지지가 분명하다. 양쪽 손목과 팔의 연결도 자연스럽다. 신체는 물에 잠겨 있고 머리카락과 옷자락은 물의 저항을 받으며 퍼져 있다. 발이나 추진 동작은 보이지 않지만 이 근접 장면에서 무지지 공중 부유로 볼 요소는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "수중 구조 악력과 얼굴 없는 근접 구도는 맞지만, 기계 손이 중앙에 놓이고 상대적으로 작아 지정된 중앙 왼쪽의 큰 손 강조는 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙 왼쪽의 거대한 기계 손이 오른쪽에서 뻗은 작은 손을 감싸며, 낮은 수중 시점과 잘린 몸통으로 구조의 연결만 강조하는 구도를 더 정확히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 팔은 왼쪽 위에서 중앙 아래로, 앰버의 팔은 오른쪽에서 중앙으로 뻗어 서로의 손에 도달한다. 기계 손가락은 앰버의 손을 감싸 닫혀 있다. 두 얼굴과 눈은 화면 밖이다. 카메라는 손 아래에서 팔과 수면을 비스듬히 올려다본다.",
        "built_space": "배경은 기포가 섞인 물과 위쪽 수면뿐이며 터널, 사다리, 배관 등 고정 구조물은 보이지 않는다. 구조물을 추가하지 말라는 외부 수중 장면 지시와 맞는다. 찰리의 몸통 일부는 왼쪽, 앰버의 몸통 일부는 오른쪽에 잘려 들어온다. 다만 잡은 손은 중앙 왼쪽보다 화면 중앙에 가깝다.",
        "entities": "샌드 베이지 장갑판과 검은 관절의 기계 팔, 왼쪽 가장자리의 푸른 원자로 일부가 찰리의 설정과 맞는다. 어깨에는 젖은 낡은 천이 보인다. 앰버는 작은 창백한 손, 금발 일부, 때 묻은 카키 작업복과 마스크 일부로 나타난다. 손의 크기는 아동 설정과 부합하지만 얼굴이 없어 정확한 나이와 혼혈 정체성은 확인할 수 없다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 팔에 연결되어 있고 기계 손가락이 작은 손 주위를 실제로 감싼다. 몸통은 화면 양쪽으로 이어져 있어 분리되거나 떠 있는 손이 아니다. 잠긴 신체와 천, 머리카락은 주변 물의 부력과 저항을 받는 모습이며 기포는 수면 쪽에 분포한다. 수영의 추진 동작은 크롭 밖이므로 확인할 수 없지만 불가능한 지지 상태는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 팔은 왼쪽 위에서 중앙 왼쪽으로 내려오며, 앰버의 팔은 오른쪽 아래에서 그 손을 향해 뻗는다. 거대한 기계 손의 엄지와 굽힌 손가락이 앰버의 작은 손을 위아래로 감싸 구조 악력을 형성한다. 눈과 얼굴은 모두 제외되어 있다. 위쪽 수면과 팔의 단축감이 손 아래에서 비스듬히 올려다보는 시점을 뒷받침한다.",
        "built_space": "주변은 연속된 수중 공간이고 상단에 수면이 보인다. 건축물과 고정 설비는 하나도 도입되지 않았다. 찰리의 잘린 몸통과 외투는 왼쪽, 앰버의 어깨와 머리카락 일부는 오른쪽 아래에 있다. 기계 손이 중앙 왼쪽에서 크게 차지하여 지정된 화면 배치와 손 중심 근접 구도에 더 가깝다.",
        "entities": "찰리의 손과 전완은 마모된 샌드 베이지 금속 장갑판, 굵은 관절, 육중한 비율로 표현되어 참조의 기계 정체성과 맞는다. 낡은 외투의 젖은 천도 왼쪽에 보인다. 앰버에게는 작은 창백한 손과 가는 손목, 금발, 오염된 카키색 소매가 보인다. 얼굴이 제외되어 성별·정확한 연령·혼혈 외모를 독립적으로 확인할 수는 없다. 모자와 공구 벨트 등은 크롭 밖이므로 누락으로 판단하지 않는다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "작은 손이 기계 손바닥 안으로 들어가고 여러 기계 손가락이 그 주위를 닫아 접촉과 지지가 분명하다. 양쪽 손목과 팔의 연결도 자연스럽다. 신체는 물에 잠겨 있고 머리카락과 옷자락은 물의 저항을 받으며 퍼져 있다. 발이나 추진 동작은 보이지 않지만 이 근접 장면에서 무지지 공중 부유로 볼 요소는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.317
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.067
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 연출: 완전한 수중 환경임에도 불구하고 맞잡은 손 주위로 공기 중에서처럼 물방울이 아래로 떨어지고 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1067
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 비스듬히 올려다보는 카메라 앵글과 클로즈업 프레이밍을 정확히 따랐으며, 수중 물리 법칙을 자연스럽게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1067,
    "verdict_ko": "카메라 앵글이 수평에 가깝고, 물속임에도 물방울이 아래로 떨어지는 물리적 오류가 있어 감점되었습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 연출: 완전한 수중 환경임에도 불구하고 맞잡은 손 주위로 공기 중에서처럼 물방울이 아래로 떨어지고 있습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_outlet_water_33110b.png",
    "asset_id": "822689d4-88c5-4fa3-b7e3-f31bee70a02f",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a43-0baf-7501-ac31-0c6c9ebcf8c9",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S36sh7__bgfirst_bg.png",
   "bg_asset_id": "020902ca-091e-4c8c-b820-5a60dc0cea20",
   "bg_record_key": "S36sh7::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "flood_outlet_water",
   "groupbg_asset_id": "822689d4-88c5-4fa3-b7e3-f31bee70a02f"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S37sh3::signage": {
  "fp": "a28a038fd367a12e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S37sh3": {
  "input_fingerprint": "03b4b41127e9814e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 창문 너머로 거대한 산더미 같은 해일이 건물을 향해 솟구친 압도적인 시점 구도, 거센 물보라가 허공에 맺혀 있는 찰나.\n\nLOCATION (lock): At a corridor window inside the camp administration building, looking out at the approaching wave. The building's nighttime lights remain on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just inside the corridor window at chest height, complete the upward tilt while looking obliquely outward past one edge of the window, without assigning the view to either character. Retain a narrow window edge on the left and place the rising wave across the central third of the exterior view, with suspended spray ahead of its advancing face and open exterior space around it. No people appear; the upward redirection of the view, rather than a lighting change, makes the approaching height register.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Not yet shattered) — Viewed obliquely from inside, with the approaching wave visible through it; used as A narrow edge anchors the interior vantage without enclosing the image in a centered window frame; Approaching wave (Rising toward the building) — Its approaching face is seen from below through the window; used as Primary exterior threat, held against surrounding space to convey scale; Water spray (Suspended ahead of the advancing wave at this instant); used as Separates the leading edge of the threat from the larger water mass behind it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination and controlled contrast retain the wave and spray as physically present exterior details rather than a dreamlike apparition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave is approaching the management-office building at night. The corridor windows are still intact before the later impact.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 창문 너머로 거대한 산더미 같은 해일이 건물을 향해 솟구친 압도적인 시점 구도, 거센 물보라가 허공에 맺혀 있는 찰나.\n\nLOCATION (lock): At a corridor window inside the camp administration building, looking out at the approaching wave. The building's nighttime lights remain on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just inside the corridor window at chest height, complete the upward tilt while looking obliquely outward past one edge of the window, without assigning the view to either character. Retain a narrow window edge on the left and place the rising wave across the central third of the exterior view, with suspended spray ahead of its advancing face and open exterior space around it. No people appear; the upward redirection of the view, rather than a lighting change, makes the approaching height register.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Not yet shattered) — Viewed obliquely from inside, with the approaching wave visible through it; used as A narrow edge anchors the interior vantage without enclosing the image in a centered window frame; Approaching wave (Rising toward the building) — Its approaching face is seen from below through the window; used as Primary exterior threat, held against surrounding space to convey scale; Water spray (Suspended ahead of the advancing wave at this instant); used as Separates the leading edge of the threat from the larger water mass behind it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination and controlled contrast retain the wave and spray as physically present exterior details rather than a dreamlike apparition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave is approaching the management-office building at night. The corridor windows are still intact before the later impact.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 창문 너머로 거대한 산더미 같은 해일이 건물을 향해 솟구친 압도적인 시점 구도, 거센 물보라가 허공에 맺혀 있는 찰나.\n\nLOCATION (lock): At a corridor window inside the camp administration building, looking out at the approaching wave. The building's nighttime lights remain on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just inside the corridor window at chest height, complete the upward tilt while looking obliquely outward past one edge of the window, without assigning the view to either character. Retain a narrow window edge on the left and place the rising wave across the central third of the exterior view, with suspended spray ahead of its advancing face and open exterior space around it. No people appear; the upward redirection of the view, rather than a lighting change, makes the approaching height register.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Corridor window (Not yet shattered) — Viewed obliquely from inside, with the approaching wave visible through it; used as A narrow edge anchors the interior vantage without enclosing the image in a centered window frame; Approaching wave (Rising toward the building) — Its approaching face is seen from below through the window; used as Primary exterior threat, held against surrounding space to convey scale; Water spray (Suspended ahead of the advancing wave at this instant); used as Separates the leading edge of the threat from the larger water mass behind it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination and controlled contrast retain the wave and spray as physically present exterior details rather than a dreamlike apparition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A massive wave is approaching the management-office building at night. The corridor windows are still intact before the later impact.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 창문 너머로 다가오는 거대한 파도를 향해 올려다보는 시점을 취하고 있습니다.",
    "built_space": "화면 좌측에 창틀과 유리창이 배치되어 실내 조명이 반사되고 있으나, 창밖의 풍경이 캠프라기보다는 산업 시설처럼 보입니다.",
    "entities": "거대한 해일과 허공에 맺힌 물보라가 확인되며 인물은 없습니다.",
    "hard_violations": [],
    "physics": "건물을 덮치기 직전 솟구친 파도와 공중에 흩날리는 물보라의 움직임이 자연스럽게 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 창문 밖으로 솟구치는 거대한 해일을 향해 위로 올려다보고 있습니다.",
    "built_space": "좌측에 좁은 창틀이 위치하고 창문에 실내 형광등이 반사되며, 외부에는 철책과 임시 구조물 등 캠프 시설이 배치되어 있습니다.",
    "entities": "산더미 같은 파도와 물보라가 명확히 보이며, 지시대로 인물은 등장하지 않습니다.",
    "hard_violations": [],
    "physics": "거대한 파도가 건물을 향해 밀려오고 물보라가 허공에 맺혀 있는 찰나가 물리적으로 설득력 있게 묘사되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "창틀을 좌측에 배치한 구도와 실내 조명 반사, 난민 캠프의 디테일 및 압도적인 해일의 묘사가 지시문과 훌륭하게 일치합니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 시점과 파도의 형태는 잘 구현되었으나, 외부 풍경이 난민 캠프보다는 일반 산업 시설에 가깝게 표현되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 창문 밖으로 솟구치는 거대한 해일을 향해 위로 올려다보고 있습니다.",
        "built_space": "좌측에 좁은 창틀이 위치하고 창문에 실내 형광등이 반사되며, 외부에는 철책과 임시 구조물 등 캠프 시설이 배치되어 있습니다.",
        "entities": "산더미 같은 파도와 물보라가 명확히 보이며, 지시대로 인물은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "거대한 파도가 건물을 향해 밀려오고 물보라가 허공에 맺혀 있는 찰나가 물리적으로 설득력 있게 묘사되었습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 창문 너머로 다가오는 거대한 파도를 향해 올려다보는 시점을 취하고 있습니다.",
        "built_space": "화면 좌측에 창틀과 유리창이 배치되어 실내 조명이 반사되고 있으나, 창밖의 풍경이 캠프라기보다는 산업 시설처럼 보입니다.",
        "entities": "거대한 해일과 허공에 맺힌 물보라가 확인되며 인물은 없습니다.",
        "hard_violations": [],
        "physics": "건물을 덮치기 직전 솟구친 파도와 공중에 흩날리는 물보라의 움직임이 자연스럽게 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "창틀을 좌측에 배치한 구도와 실내 조명 반사, 난민 캠프의 디테일 및 압도적인 해일의 묘사가 지시문과 훌륭하게 일치합니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 시점과 파도의 형태는 잘 구현되었으나, 외부 풍경이 난민 캠프보다는 일반 산업 시설에 가깝게 표현되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 창문 밖으로 솟구치는 거대한 해일을 향해 위로 올려다보고 있습니다.",
        "built_space": "좌측에 좁은 창틀이 위치하고 창문에 실내 형광등이 반사되며, 외부에는 철책과 임시 구조물 등 캠프 시설이 배치되어 있습니다.",
        "entities": "산더미 같은 파도와 물보라가 명확히 보이며, 지시대로 인물은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "거대한 파도가 건물을 향해 밀려오고 물보라가 허공에 맺혀 있는 찰나가 물리적으로 설득력 있게 묘사되었습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 창문 너머로 다가오는 거대한 파도를 향해 올려다보는 시점을 취하고 있습니다.",
        "built_space": "화면 좌측에 창틀과 유리창이 배치되어 실내 조명이 반사되고 있으나, 창밖의 풍경이 캠프라기보다는 산업 시설처럼 보입니다.",
        "entities": "거대한 해일과 허공에 맺힌 물보라가 확인되며 인물은 없습니다.",
        "hard_violations": [],
        "physics": "건물을 덮치기 직전 솟구친 파도와 공중에 흩날리는 물보라의 움직임이 자연스럽게 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "왼쪽 창 가장자리만으로 실내 시점을 고정하는 구도가 더 가깝지만, 해일이 외부 시야 대부분을 차지하고 지상 시설도 많이 보여 중앙 3분의 1 배치와 상향 구도를 완전히 충족하지는 못한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "온전한 창과 켜진 복도 조명은 이어지지만, 왼쪽 복도와 아래 창턱을 넓게 포함하고 해일도 중앙 영역을 크게 벗어나 지정된 창가 구도에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "해일의 어두운 전면과 말리는 능선이 전경의 수용소 및 카메라가 있는 건물 쪽을 향한다. 물보라는 능선 위와 전면으로 퍼져 있다. 사람이나 시선, 무기는 없다. 높은 파면을 아래에서 보지만 지상 수용소도 넓게 내려다보여 상향 전환만으로 높이를 강조한 구도는 다소 약하다.",
        "built_space": "왼쪽에 밝은 벽 일부와 어두운 세로 창틀 하나, 아래쪽에 비스듬한 창턱이 보인다. 창 전체를 정면으로 둘러싼 구도는 아니며 유리 파손도 없다. 유리 왼쪽에는 형광등 반사로 읽히는 밝은 무리가 세 군데 있어 실내 조명이 켜진 상태와 양립한다. 외부에는 낮은 숙소, 울타리, 조명 기둥과 오른쪽 감시탑 하나가 보이지만 이 외부 배치는 참고 사진만으로 동일성을 확인할 수 없다.",
        "entities": "요구된 거대한 해일, 공중에 흩어진 물보라, 온전한 복도 창, 야간 환경이 모두 보인다. 사람·얼굴·문구·화면 위 표식은 없다. 창틀과 벽은 참고의 어두운 금속 및 밝은 도장면 계열을 따른다. 다만 해일은 중앙 3분의 1에 제한되지 않고 외부 시야 대부분을 채운다.",
        "hard_violations": [],
        "physics": "해일의 하부는 지상의 연속된 수괴와 이어지고, 능선의 쇄파가 물방울을 공중으로 내보내는 모습이다. 물보라는 그 운동의 순간으로 설명되며 지지나 발사 원인이 없는 부유 물체는 없다. 외부 건물과 감시탑은 지면에 놓이고 창틀과 창턱은 벽에 연결되어 있다."
       },
       {
        "label": "B",
        "direction": "해일 전면이 외부 마당과 창이 있는 건물 쪽으로 밀려오는 모습이며, 물보라는 능선 위와 앞쪽으로 분출한다. 사람의 시선이나 조준 대상은 없다. 파면을 올려다보는 요소는 있으나 넓은 마당과 건물 하부까지 포함해 위로 기울인 시선의 강조가 약하다.",
        "built_space": "왼쪽에 복도 창의 여러 유리 구획, 넓은 밝은 벽기둥 하나와 주 창의 세로틀이 보이고, 아래에는 두꺼운 창턱이 길게 놓인다. 좁은 창 가장자리만 남기라는 지시보다 실내 구조의 점유가 크다. 왼쪽 유리에 두 군데의 뚜렷한 형광등 반사가 있으며 참고의 켜진 복도 조명과 양립한다. 밖에는 울타리로 나뉜 마당, 낮은 건물들과 오른쪽의 불 켜진 건물 한 동이 보인다. 유리는 깨지지 않았다.",
        "entities": "해일, 비산하는 물보라, 야간의 온전한 창과 점등된 건물들이 보이며 사람이나 얼굴은 없다. 참고의 밝은 복도 벽과 금속 창호 재질을 대체로 유지한다. 해일은 중앙 영역을 넘어 오른쪽 끝까지 크게 펼쳐져 주변의 열린 외부 공간이 제한된다. 읽을 수 있는 추가 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "물벽 하단은 뒤쪽의 거품과 수괴에 연결되고, 상단에서 분리된 물방울은 쇄파로 튀어 오른 것으로 읽힌다. 근거 없이 공중에 떠 있는 물체는 없다. 창은 벽과 창턱에 고정되고 외부 건물, 울타리와 조명 기둥은 지면에 지지된다. 젖은 마당의 빛 반사도 표면과 조명 위치에 맞게 읽힌다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 창 가장자리만으로 실내 시점을 고정하는 구도가 더 가깝지만, 해일이 외부 시야 대부분을 차지하고 지상 시설도 많이 보여 중앙 3분의 1 배치와 상향 구도를 완전히 충족하지는 못한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "온전한 창과 켜진 복도 조명은 이어지지만, 왼쪽 복도와 아래 창턱을 넓게 포함하고 해일도 중앙 영역을 크게 벗어나 지정된 창가 구도에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "해일의 어두운 전면과 말리는 능선이 전경의 수용소 및 카메라가 있는 건물 쪽을 향한다. 물보라는 능선 위와 전면으로 퍼져 있다. 사람이나 시선, 무기는 없다. 높은 파면을 아래에서 보지만 지상 수용소도 넓게 내려다보여 상향 전환만으로 높이를 강조한 구도는 다소 약하다.",
        "built_space": "왼쪽에 밝은 벽 일부와 어두운 세로 창틀 하나, 아래쪽에 비스듬한 창턱이 보인다. 창 전체를 정면으로 둘러싼 구도는 아니며 유리 파손도 없다. 유리 왼쪽에는 형광등 반사로 읽히는 밝은 무리가 세 군데 있어 실내 조명이 켜진 상태와 양립한다. 외부에는 낮은 숙소, 울타리, 조명 기둥과 오른쪽 감시탑 하나가 보이지만 이 외부 배치는 참고 사진만으로 동일성을 확인할 수 없다.",
        "entities": "요구된 거대한 해일, 공중에 흩어진 물보라, 온전한 복도 창, 야간 환경이 모두 보인다. 사람·얼굴·문구·화면 위 표식은 없다. 창틀과 벽은 참고의 어두운 금속 및 밝은 도장면 계열을 따른다. 다만 해일은 중앙 3분의 1에 제한되지 않고 외부 시야 대부분을 채운다.",
        "hard_violations": [],
        "physics": "해일의 하부는 지상의 연속된 수괴와 이어지고, 능선의 쇄파가 물방울을 공중으로 내보내는 모습이다. 물보라는 그 운동의 순간으로 설명되며 지지나 발사 원인이 없는 부유 물체는 없다. 외부 건물과 감시탑은 지면에 놓이고 창틀과 창턱은 벽에 연결되어 있다."
       },
       {
        "label": "A",
        "direction": "해일 전면이 외부 마당과 창이 있는 건물 쪽으로 밀려오는 모습이며, 물보라는 능선 위와 앞쪽으로 분출한다. 사람의 시선이나 조준 대상은 없다. 파면을 올려다보는 요소는 있으나 넓은 마당과 건물 하부까지 포함해 위로 기울인 시선의 강조가 약하다.",
        "built_space": "왼쪽에 복도 창의 여러 유리 구획, 넓은 밝은 벽기둥 하나와 주 창의 세로틀이 보이고, 아래에는 두꺼운 창턱이 길게 놓인다. 좁은 창 가장자리만 남기라는 지시보다 실내 구조의 점유가 크다. 왼쪽 유리에 두 군데의 뚜렷한 형광등 반사가 있으며 참고의 켜진 복도 조명과 양립한다. 밖에는 울타리로 나뉜 마당, 낮은 건물들과 오른쪽의 불 켜진 건물 한 동이 보인다. 유리는 깨지지 않았다.",
        "entities": "해일, 비산하는 물보라, 야간의 온전한 창과 점등된 건물들이 보이며 사람이나 얼굴은 없다. 참고의 밝은 복도 벽과 금속 창호 재질을 대체로 유지한다. 해일은 중앙 영역을 넘어 오른쪽 끝까지 크게 펼쳐져 주변의 열린 외부 공간이 제한된다. 읽을 수 있는 추가 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "물벽 하단은 뒤쪽의 거품과 수괴에 연결되고, 상단에서 분리된 물방울은 쇄파로 튀어 오른 것으로 읽힌다. 근거 없이 공중에 떠 있는 물체는 없다. 창은 벽과 창턱에 고정되고 외부 건물, 울타리와 조명 기둥은 지면에 지지된다. 젖은 마당의 빛 반사도 표면과 조명 위치에 맞게 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.714,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "창틀을 좌측에 배치한 구도와 실내 조명 반사, 난민 캠프의 디테일 및 압도적인 해일의 묘사가 지시문과 훌륭하게 일치합니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "지시된 시점과 파도의 형태는 잘 구현되었으나, 외부 풍경이 난민 캠프보다는 일반 산업 시설에 가깝게 표현되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S34sh8_sel.png",
    "asset_id": "4d0e9db4-7ce0-43c5-b928-7168ecf7a6c1",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a4c-1729-7e6e-b2fd-73d2f25cf8af",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S34sh8"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S37sh6::signage": {
  "fp": "dc2326bb9e283b3f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::admin_upper_landing": {
  "input_fingerprint": "f1ee47ae2631f9a8",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "admin_upper_landing",
    "tags": [
     "S37sh6",
     "S37sh7"
    ]
   },
   "context_sig": "ca82e492adcd95cb"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 관리사무소 내 미연 감금실, 3층 복도·창가, 내부 계단: 구타가 자행되는 취조실과 비상사이렌이 울리는 복도 공간이다. (특징: 얼굴이 퉁퉁 부은 미연과 심문하는 박철진; 복도 구석에 비치된 붉은 소화기; 소화기로 부서진 금속 문고리; 창문 유리를 와장창 깨부수며 복도 안으로 밀어닥치는 흙빛 해일)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 아뿔싸! 더 높은 층으로 계속 올라가는 두 사람.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n인천 난민촌 관리사무소 내 미연 감금실, 3층 복도·창가, 내부 계단: 구타가 자행되는 취조실과 비상사이렌이 울리는 복도 공간이다. (특징: 얼굴이 퉁퉁 부은 미연과 심문하는 박철진; 복도 구석에 비치된 붉은 소화기; 소화기로 부서진 금속 문고리; 창문 유리를 와장창 깨부수며 복도 안으로 밀어닥치는 흙빛 해일)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 아뿔싸! 더 높은 층으로 계속 올라가는 두 사람.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_upper_landing_0054fc.png",
  "asset_id": "13d19c76-e99d-4455-a38e-95fb0ecda129",
  "input_asset_ids": [
   "fbd824bf-9ec3-4690-afb9-11453c2b1b32"
  ],
  "origin_tag": "S37sh6",
  "place_text": "In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.",
  "origin_inputs": {
   "place_text": "In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.",
   "time_of_day_en": "night",
   "conti_asset_id": "fbd824bf-9ec3-4690-afb9-11453c2b1b32"
  }
 },
 "S37sh6::bgfirst_bg": {
  "input_fingerprint": "4f226090a99c7d93",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미연이 이현우의 머리를 자신의 가슴에 깊이 파묻은 채 두 팔로 꽉 껴안은 구도.\n\nLOCATION (lock): In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the pair's interior flank, holding a three-quarter side view from just above 미연's head with a gentle downward tilt. Place 미연 center-left as she folds both arms tightly around 이현우, whose bent neck and partially hidden profile lead into her chest at lower center; her attention is lowered to his tucked head, while his face turns down into the embrace. Keep the window beyond them at the right edge, preserving the axis for the coming retreat and emphasizing only the reduced camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Upper-floor corridor window (Still intact immediately before the wave breaks through) — Seen obliquely beyond the embracing pair from their interior flank; used as A restrained right-background reference establishing the direction of the coming impact; Upper-floor corridor (The pair have reached it after climbing higher) — Seen along the diagonal from the interior side toward the window; used as Maintains the spatial axis for the following backward camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued interior ambient illumination and gentle facial contrast, letting the protective contact provide warmth without changing the established lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미연이 이현우의 머리를 자신의 가슴에 깊이 파묻은 채 두 팔로 꽉 껴안은 구도.\n\nLOCATION (lock): In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the pair's interior flank, holding a three-quarter side view from just above 미연's head with a gentle downward tilt. Place 미연 center-left as she folds both arms tightly around 이현우, whose bent neck and partially hidden profile lead into her chest at lower center; her attention is lowered to his tucked head, while his face turns down into the embrace. Keep the window beyond them at the right edge, preserving the axis for the coming retreat and emphasizing only the reduced camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Upper-floor corridor window (Still intact immediately before the wave breaks through) — Seen obliquely beyond the embracing pair from their interior flank; used as A restrained right-background reference establishing the direction of the coming impact; Upper-floor corridor (The pair have reached it after climbing higher) — Seen along the diagonal from the interior side toward the window; used as Maintains the spatial axis for the following backward camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued interior ambient illumination and gentle facial contrast, letting the protective contact provide warmth without changing the established lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S37sh6__bgfirst_bg.png",
  "asset_id": "7bb25d14-8695-42cc-bd94-0523aeca4b7c",
  "input_asset_ids": [
   "fbd824bf-9ec3-4690-afb9-11453c2b1b32",
   "13d19c76-e99d-4455-a38e-95fb0ecda129"
  ]
 },
 "S37sh6": {
  "input_fingerprint": "04ff7f348b0a8444",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 이현우의 머리를 자신의 가슴에 깊이 파묻은 채 두 팔로 꽉 껴안은 구도.\n\nLOCATION (lock): In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the pair's interior flank, holding a three-quarter side view from just above 미연's head with a gentle downward tilt. Place 미연 center-left as she folds both arms tightly around 이현우, whose bent neck and partially hidden profile lead into her chest at lower center; her attention is lowered to his tucked head, while his face turns down into the embrace. Keep the window beyond them at the right edge, preserving the axis for the coming retreat and emphasizing only the reduced camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Upper-floor corridor window (Still intact immediately before the wave breaks through) — Seen obliquely beyond the embracing pair from their interior flank; used as A restrained right-background reference establishing the direction of the coming impact; Upper-floor corridor (The pair have reached it after climbing higher) — Seen along the diagonal from the interior side toward the window; used as Maintains the spatial axis for the following backward camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued interior ambient illumination and gentle facial contrast, letting the protective contact provide warmth without changing the established lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The enormous wave has reached the building's upper-floor vicinity. The windows have not yet shattered. 이현우: He is on a higher floor with his head tucked inward, still bearing facial bruises and the untreated leg wound. His outer garment remains removed, and the stiff contact card remains inside his shoe. 미연: She holds both arms tightly inward in a protective embrace, with her face still swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 이현우의 머리를 자신의 가슴에 깊이 파묻은 채 두 팔로 꽉 껴안은 구도.\n\nLOCATION (lock): In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the pair's interior flank, holding a three-quarter side view from just above 미연's head with a gentle downward tilt. Place 미연 center-left as she folds both arms tightly around 이현우, whose bent neck and partially hidden profile lead into her chest at lower center; her attention is lowered to his tucked head, while his face turns down into the embrace. Keep the window beyond them at the right edge, preserving the axis for the coming retreat and emphasizing only the reduced camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Upper-floor corridor window (Still intact immediately before the wave breaks through) — Seen obliquely beyond the embracing pair from their interior flank; used as A restrained right-background reference establishing the direction of the coming impact; Upper-floor corridor (The pair have reached it after climbing higher) — Seen along the diagonal from the interior side toward the window; used as Maintains the spatial axis for the following backward camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued interior ambient illumination and gentle facial contrast, letting the protective contact provide warmth without changing the established lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The enormous wave has reached the building's upper-floor vicinity. The windows have not yet shattered. 이현우: He is on a higher floor with his head tucked inward, still bearing facial bruises and the untreated leg wound. His outer garment remains removed, and the stiff contact card remains inside his shoe. 미연: She holds both arms tightly inward in a protective embrace, with her face still swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미연이 이현우의 머리를 자신의 가슴에 깊이 파묻은 채 두 팔로 꽉 껴안은 구도.\n\nLOCATION (lock): In the administration building's upper-floor circulation area beside the windows, after the pair have climbed higher. The interior is still lit at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach along the pair's interior flank, holding a three-quarter side view from just above 미연's head with a gentle downward tilt. Place 미연 center-left as she folds both arms tightly around 이현우, whose bent neck and partially hidden profile lead into her chest at lower center; her attention is lowered to his tucked head, while his face turns down into the embrace. Keep the window beyond them at the right edge, preserving the axis for the coming retreat and emphasizing only the reduced camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Upper-floor corridor window (Still intact immediately before the wave breaks through) — Seen obliquely beyond the embracing pair from their interior flank; used as A restrained right-background reference establishing the direction of the coming impact; Upper-floor corridor (The pair have reached it after climbing higher) — Seen along the diagonal from the interior side toward the window; used as Maintains the spatial axis for the following backward camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued interior ambient illumination and gentle facial contrast, letting the protective contact provide warmth without changing the established lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The enormous wave has reached the building's upper-floor vicinity. The windows have not yet shattered. 이현우: He is on a higher floor with his head tucked inward, still bearing facial bruises and the untreated leg wound. His outer garment remains removed, and the stiff contact card remains inside his shoe. 미연: She holds both arms tightly inward in a protective embrace, with her face still swollen from the beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S37sh6__bgfirst_bg.png",
     "asset_id": "7bb25d14-8695-42cc-bd94-0523aeca4b7c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S37sh6.png",
     "asset_id": "fbd824bf-9ec3-4690-afb9-11453c2b1b32",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_upper_landing_0054fc.png",
     "asset_id": "13d19c76-e99d-4455-a38e-95fb0ecda129",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:814924>",
     "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "미연은 이현우의 머리를 내려다보고 있으며, 이현우는 미연의 가슴에 얼굴을 묻은 채 시선을 아래로 향함.",
    "built_space": "우측에 창문과 야경이 보이고 좌측으로 복도 문들이 이어지는 구조가 참조 이미지와 정확히 일치함.",
    "entities": "미연의 부은 얼굴과 낡은 회색 셔츠가 일치하며, 이현우의 귀에 요구된 검은색 인이어 무전기가 뚜렷하게 묘사됨.",
    "hard_violations": [],
    "physics": "두 사람이 서로의 무게를 지탱하며 두 팔로 단단히 껴안고 서 있는 자세가 자연스럽게 표현됨."
   },
   {
    "label": "B",
    "direction": "미연은 이현우를 내려다보고, 이현우는 얼굴을 파묻고 안겨 있음.",
    "built_space": "참조 이미지에 제시된 복도와 우측 창문 밖의 야경이 올바르게 배치됨.",
    "entities": "인물들의 의상과 외모는 대체로 맞으나, 이현우의 귀에 있어야 할 인이어 무전기가 없음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (이현우의 머리를 감싼 미연의 손 방향이 비정상적임)"
    ],
    "physics": "서로 안고 서 있는 자세이나, 미연의 왼팔에서 이어지는 손의 구조가 해부학적으로 어색함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인이어 무전기를 정확히 묘사했으며, 자연스러운 포옹 구도와 해부학적 형태를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "필수 요소인 이현우의 인이어 무전기가 누락되었으며, 인물을 감싸는 손의 해부학적 오류가 발생하여 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 이현우의 머리를 내려다보고 있으며, 이현우는 미연의 가슴에 얼굴을 묻은 채 시선을 아래로 향함.",
        "built_space": "우측에 창문과 야경이 보이고 좌측으로 복도 문들이 이어지는 구조가 참조 이미지와 정확히 일치함.",
        "entities": "미연의 부은 얼굴과 낡은 회색 셔츠가 일치하며, 이현우의 귀에 요구된 검은색 인이어 무전기가 뚜렷하게 묘사됨.",
        "hard_violations": [],
        "physics": "두 사람이 서로의 무게를 지탱하며 두 팔로 단단히 껴안고 서 있는 자세가 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "미연은 이현우를 내려다보고, 이현우는 얼굴을 파묻고 안겨 있음.",
        "built_space": "참조 이미지에 제시된 복도와 우측 창문 밖의 야경이 올바르게 배치됨.",
        "entities": "인물들의 의상과 외모는 대체로 맞으나, 이현우의 귀에 있어야 할 인이어 무전기가 없음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (이현우의 머리를 감싼 미연의 손 방향이 비정상적임)"
        ],
        "physics": "서로 안고 서 있는 자세이나, 미연의 왼팔에서 이어지는 손의 구조가 해부학적으로 어색함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 인이어 무전기를 정확히 묘사했으며, 자연스러운 포옹 구도와 해부학적 형태를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "필수 요소인 이현우의 인이어 무전기가 누락되었으며, 인물을 감싸는 손의 해부학적 오류가 발생하여 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "미연은 이현우의 머리를 내려다보고 있으며, 이현우는 미연의 가슴에 얼굴을 묻은 채 시선을 아래로 향함.",
        "built_space": "우측에 창문과 야경이 보이고 좌측으로 복도 문들이 이어지는 구조가 참조 이미지와 정확히 일치함.",
        "entities": "미연의 부은 얼굴과 낡은 회색 셔츠가 일치하며, 이현우의 귀에 요구된 검은색 인이어 무전기가 뚜렷하게 묘사됨.",
        "hard_violations": [],
        "physics": "두 사람이 서로의 무게를 지탱하며 두 팔로 단단히 껴안고 서 있는 자세가 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "미연은 이현우를 내려다보고, 이현우는 얼굴을 파묻고 안겨 있음.",
        "built_space": "참조 이미지에 제시된 복도와 우측 창문 밖의 야경이 올바르게 배치됨.",
        "entities": "인물들의 의상과 외모는 대체로 맞으나, 이현우의 귀에 있어야 할 인이어 무전기가 없음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (이현우의 머리를 감싼 미연의 손 방향이 비정상적임)"
        ],
        "physics": "서로 안고 서 있는 자세이나, 미연의 왼팔에서 이어지는 손의 구조가 해부학적으로 어색함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "내려다보는 측면 시점과 가슴에 머리를 묻은 포옹은 정확하지만, 미연의 정수리까지 잘리는 타이트한 구도가 지정된 미디엄 숏에서 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "미연을 중앙 왼쪽에 둔 미디엄 숏과 두 팔의 보호적 포옹을 더 충실히 구현하고 인이어도 보이지만, 카메라의 하향각은 약하고 창밖에 도달한 거대한 파도는 없다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 눈과 고개를 아래로 내려 현우의 머리를 보고 있다. 현우는 목을 굽혀 얼굴을 미연의 가슴 안쪽으로 파묻었으며 귀와 뺨 일부만 보인다. 미연의 한 손은 뒤통수를, 다른 손은 등을 감싸므로 포옹의 방향과 대상이 명확하다.",
        "built_space": "왼쪽 벽에는 소화전함 한 개와 바닥의 소화기 한 개, 뒤로 이어지는 여러 출입문이 보인다. 오른쪽에는 가까운 큰 창 구획 두 개와 뒤쪽 창 일부가 있으며 유리는 온전하다. 두 사람은 창 안쪽 복도에 서 있고 창턱이나 벽과 충돌하지 않는다. 회색 투톤 벽과 광택 바닥, 야간 항구 풍경은 장소 참조와 맞는다. 창 반사에 불가능한 인물이나 중복 설비는 없다. 카메라는 미연보다 높아 보이나 정수리가 잘리고 상체가 크게 차지해 미디엄 숏보다 타이트하다.",
        "entities": "인물은 미연과 현우 두 명뿐이다. 미연은 검은 단발의 중년 동아시아계 여성으로 참조의 외모와 해진 회색 셔츠를 대체로 유지하며 얼굴에 타박상도 보인다. 현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 가려진 얼굴의 일부에 상처가 있고 어두운 오염된 셔츠를 입었다. 얼굴이 숨겨져 정확한 얼굴 일치는 제한적으로만 판단할 수 있다. 노출된 귀에서 인이어는 식별되지 않는다. 다리 상처와 신발 속 카드는 프레임 밖이라 평가하지 않는다. 창밖은 평온한 항구로 보이며 상층부에 도달한 거대한 파도는 보이지 않는다.",
        "hard_violations": [],
        "physics": "미연의 양팔은 현우의 머리와 등을 실제로 감싸고 손바닥이 접촉한다. 현우는 상체를 앞으로 숙여 미연에게 기대는 자세이며 목과 몸통의 연결이 자연스럽다. 발은 프레임 밖이지만 몸통은 아래로 이어져 서 있는 포옹으로 읽히며 공중에 뜬 신체는 없다."
       },
       {
        "label": "B",
        "direction": "미연은 고개를 숙여 현우의 정수리에 얼굴을 가까이 대고 주의를 그의 머리에 두고 있다. 현우의 얼굴은 아래쪽으로 돌아 미연의 가슴에 묻혀 있다. 미연은 한 손으로 뒤통수를 받치고 다른 팔로 어깨와 등을 당겨 안으며, 현우도 미연의 허리를 감싼다.",
        "built_space": "왼쪽에 소화전함 한 개와 소화기 한 개, 뒤로 이어지는 출입문들이 있고 천장에는 네 개의 밝은 직사각 조명판이 뚜렷하다. 오른쪽에는 온전한 큰 창 구획 두 개와 뒤쪽 창 일부가 보인다. 투톤 벽, 금속 창틀과 창턱, 광택 바닥 및 야간 항구가 참조 장소와 일치한다. 두 사람은 창 옆 복도 내부에 자연스럽게 배치되어 있으며 설비 중복이나 불가능한 반사는 없다. 미연의 머리 전체와 허리 부근까지 포함하는 미디엄 숏이고 중앙 왼쪽 배치도 맞지만, 지정된 머리 위 시점의 하향각은 약하다.",
        "entities": "미연과 현우만 등장한다. 미연의 중년 여성 얼굴, 검은 단발, 해지고 먼지 묻은 회색 셔츠가 참조와 부합하며 뺨의 타박상과 부기가 보인다. 현우는 젊은 동아시아계 남성의 짧은 검은 머리와 마른 체격으로 읽히고, 부분적으로 드러난 뺨에 상처가 있다. 어두운 셔츠에는 먼지와 핏자국이 있으며 귀에는 작은 검은 인이어가 보인다. 숨겨진 얼굴 전체와 프레임 밖 다리 상처·신발 속 카드는 판단 대상에서 제외한다. 창밖에는 항구 불빛과 수면만 보이고 상층부까지 온 파도는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "미연의 손은 현우의 뒤통수와 어깨에 닿고 두 팔이 몸을 밀착시킨다. 현우의 아래쪽 팔과 손도 미연의 허리를 감싸 상호 접촉이 분명하다. 앞으로 굽힌 현우의 목과 몸통은 자연스럽게 연결되며 두 사람의 하체는 프레임 아래로 이어진다. 발을 보여주지 않아도 서서 기대는 자세로 성립하고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "내려다보는 측면 시점과 가슴에 머리를 묻은 포옹은 정확하지만, 미연의 정수리까지 잘리는 타이트한 구도가 지정된 미디엄 숏에서 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "미연을 중앙 왼쪽에 둔 미디엄 숏과 두 팔의 보호적 포옹을 더 충실히 구현하고 인이어도 보이지만, 카메라의 하향각은 약하고 창밖에 도달한 거대한 파도는 없다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 눈과 고개를 아래로 내려 현우의 머리를 보고 있다. 현우는 목을 굽혀 얼굴을 미연의 가슴 안쪽으로 파묻었으며 귀와 뺨 일부만 보인다. 미연의 한 손은 뒤통수를, 다른 손은 등을 감싸므로 포옹의 방향과 대상이 명확하다.",
        "built_space": "왼쪽 벽에는 소화전함 한 개와 바닥의 소화기 한 개, 뒤로 이어지는 여러 출입문이 보인다. 오른쪽에는 가까운 큰 창 구획 두 개와 뒤쪽 창 일부가 있으며 유리는 온전하다. 두 사람은 창 안쪽 복도에 서 있고 창턱이나 벽과 충돌하지 않는다. 회색 투톤 벽과 광택 바닥, 야간 항구 풍경은 장소 참조와 맞는다. 창 반사에 불가능한 인물이나 중복 설비는 없다. 카메라는 미연보다 높아 보이나 정수리가 잘리고 상체가 크게 차지해 미디엄 숏보다 타이트하다.",
        "entities": "인물은 미연과 현우 두 명뿐이다. 미연은 검은 단발의 중년 동아시아계 여성으로 참조의 외모와 해진 회색 셔츠를 대체로 유지하며 얼굴에 타박상도 보인다. 현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 가려진 얼굴의 일부에 상처가 있고 어두운 오염된 셔츠를 입었다. 얼굴이 숨겨져 정확한 얼굴 일치는 제한적으로만 판단할 수 있다. 노출된 귀에서 인이어는 식별되지 않는다. 다리 상처와 신발 속 카드는 프레임 밖이라 평가하지 않는다. 창밖은 평온한 항구로 보이며 상층부에 도달한 거대한 파도는 보이지 않는다.",
        "hard_violations": [],
        "physics": "미연의 양팔은 현우의 머리와 등을 실제로 감싸고 손바닥이 접촉한다. 현우는 상체를 앞으로 숙여 미연에게 기대는 자세이며 목과 몸통의 연결이 자연스럽다. 발은 프레임 밖이지만 몸통은 아래로 이어져 서 있는 포옹으로 읽히며 공중에 뜬 신체는 없다."
       },
       {
        "label": "A",
        "direction": "미연은 고개를 숙여 현우의 정수리에 얼굴을 가까이 대고 주의를 그의 머리에 두고 있다. 현우의 얼굴은 아래쪽으로 돌아 미연의 가슴에 묻혀 있다. 미연은 한 손으로 뒤통수를 받치고 다른 팔로 어깨와 등을 당겨 안으며, 현우도 미연의 허리를 감싼다.",
        "built_space": "왼쪽에 소화전함 한 개와 소화기 한 개, 뒤로 이어지는 출입문들이 있고 천장에는 네 개의 밝은 직사각 조명판이 뚜렷하다. 오른쪽에는 온전한 큰 창 구획 두 개와 뒤쪽 창 일부가 보인다. 투톤 벽, 금속 창틀과 창턱, 광택 바닥 및 야간 항구가 참조 장소와 일치한다. 두 사람은 창 옆 복도 내부에 자연스럽게 배치되어 있으며 설비 중복이나 불가능한 반사는 없다. 미연의 머리 전체와 허리 부근까지 포함하는 미디엄 숏이고 중앙 왼쪽 배치도 맞지만, 지정된 머리 위 시점의 하향각은 약하다.",
        "entities": "미연과 현우만 등장한다. 미연의 중년 여성 얼굴, 검은 단발, 해지고 먼지 묻은 회색 셔츠가 참조와 부합하며 뺨의 타박상과 부기가 보인다. 현우는 젊은 동아시아계 남성의 짧은 검은 머리와 마른 체격으로 읽히고, 부분적으로 드러난 뺨에 상처가 있다. 어두운 셔츠에는 먼지와 핏자국이 있으며 귀에는 작은 검은 인이어가 보인다. 숨겨진 얼굴 전체와 프레임 밖 다리 상처·신발 속 카드는 판단 대상에서 제외한다. 창밖에는 항구 불빛과 수면만 보이고 상층부까지 온 파도는 확인되지 않는다.",
        "hard_violations": [],
        "physics": "미연의 손은 현우의 뒤통수와 어깨에 닿고 두 팔이 몸을 밀착시킨다. 현우의 아래쪽 팔과 손도 미연의 허리를 감싸 상호 접촉이 분명하다. 앞으로 굽힌 현우의 목과 몸통은 자연스럽게 연결되며 두 사람의 하체는 프레임 아래로 이어진다. 발을 보여주지 않아도 서서 기대는 자세로 성립하고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.304
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.054
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (이현우의 머리를 감싼 미연의 손 방향이 비정상적임)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1054
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 인이어 무전기를 정확히 묘사했으며, 자연스러운 포옹 구도와 해부학적 형태를 성공적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 1054,
    "verdict_ko": "필수 요소인 이현우의 인이어 무전기가 누락되었으며, 인물을 감싸는 손의 해부학적 오류가 발생하여 크게 감점되었습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (이현우의 머리를 감싼 미연의 손 방향이 비정상적임)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_admin_upper_landing_0054fc.png",
    "asset_id": "13d19c76-e99d-4455-a38e-95fb0ecda129",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a51-3ff4-7741-867e-698314e94641",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S37sh6__bgfirst_bg.png",
   "bg_asset_id": "7bb25d14-8695-42cc-bd94-0523aeca4b7c",
   "bg_record_key": "S37sh6::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "admin_upper_landing",
   "groupbg_asset_id": "13d19c76-e99d-4455-a38e-95fb0ecda129"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S37sh7::signage": {
  "fp": "8c811030654d9263",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S37sh7": {
  "input_fingerprint": "79935cc816c59e2d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산산조각 난 창문의 유리 파편들과 거대한 바닷물이 이현우와 미연을 덮치기 직전 허공에 쏟아지는 폭발적인 찰나.\n\nLOCATION (lock): Inside the administration building's upper-floor window-side passage, as the glazing gives way to the wave. Interior lighting is overwhelmed by spray and broken glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Dolly backward along the same interior diagonal, retaining the camera height just above 미연's head and the slight downward tilt, so increased camera distance is the only emphasized change. Keep the embracing pair left of center as 미연 lowers her cheek toward 이현우's crown and he tucks his face fully inward, both still absorbed in the protective hold rather than turning toward the lens. The shattered window remains in the right background, with glass fragments and incoming seawater crossing the visible gap toward them but not yet making contact; keep this advancing mass within roughly the right third of the frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Shattered window behind and right of the embracing pair in the upper-right of the frame, background; Incoming water and fragments in the gap, not yet touching the pair in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Shattered upper-floor window (Breaking inward under the impact of the wave) — The opening is seen diagonally from inside the corridor, behind and to the right of the pair; used as Identifies the origin of the incoming water and fragments; Airborne glass fragments (Scattered inward between the window and the embracing pair); used as Marks the advancing impact front without obscuring the embrace; Incoming seawater (Bursting through the window immediately before reaching the pair) — Advancing diagonally from the right background toward the left midground; used as Makes the remaining separation before contact visible; Upper-floor corridor (Being breached by the incoming wave) — The interior diagonal remains continuous with the preceding embrace shot; used as Provides the visible space between the pair and the broken window.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding interior exposure and restrained contrast, keeping the airborne glass and water legible without an invented flash or lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The upper-floor windows are now shattering, leaving broken glazing and airborne fragments as seawater breaks into the building. 이현우: He remains curled inward at the moment of impact, with his earlier facial bruises and leg wound unchanged. The stiff contact card is still concealed inside his shoe. 미연: She maintains the tight protective embrace through the impact, with the earlier facial swelling still present. A mass of seawater and shattered window glass bursts into the interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산산조각 난 창문의 유리 파편들과 거대한 바닷물이 이현우와 미연을 덮치기 직전 허공에 쏟아지는 폭발적인 찰나.\n\nLOCATION (lock): Inside the administration building's upper-floor window-side passage, as the glazing gives way to the wave. Interior lighting is overwhelmed by spray and broken glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Dolly backward along the same interior diagonal, retaining the camera height just above 미연's head and the slight downward tilt, so increased camera distance is the only emphasized change. Keep the embracing pair left of center as 미연 lowers her cheek toward 이현우's crown and he tucks his face fully inward, both still absorbed in the protective hold rather than turning toward the lens. The shattered window remains in the right background, with glass fragments and incoming seawater crossing the visible gap toward them but not yet making contact; keep this advancing mass within roughly the right third of the frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Shattered window behind and right of the embracing pair in the upper-right of the frame, background; Incoming water and fragments in the gap, not yet touching the pair in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Shattered upper-floor window (Breaking inward under the impact of the wave) — The opening is seen diagonally from inside the corridor, behind and to the right of the pair; used as Identifies the origin of the incoming water and fragments; Airborne glass fragments (Scattered inward between the window and the embracing pair); used as Marks the advancing impact front without obscuring the embrace; Incoming seawater (Bursting through the window immediately before reaching the pair) — Advancing diagonally from the right background toward the left midground; used as Makes the remaining separation before contact visible; Upper-floor corridor (Being breached by the incoming wave) — The interior diagonal remains continuous with the preceding embrace shot; used as Provides the visible space between the pair and the broken window.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding interior exposure and restrained contrast, keeping the airborne glass and water legible without an invented flash or lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The upper-floor windows are now shattering, leaving broken glazing and airborne fragments as seawater breaks into the building. 이현우: He remains curled inward at the moment of impact, with his earlier facial bruises and leg wound unchanged. The stiff contact card is still concealed inside his shoe. 미연: She maintains the tight protective embrace through the impact, with the earlier facial swelling still present. A mass of seawater and shattered window glass bursts into the interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산산조각 난 창문의 유리 파편들과 거대한 바닷물이 이현우와 미연을 덮치기 직전 허공에 쏟아지는 폭발적인 찰나.\n\nLOCATION (lock): Inside the administration building's upper-floor window-side passage, as the glazing gives way to the wave. Interior lighting is overwhelmed by spray and broken glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Dolly backward along the same interior diagonal, retaining the camera height just above 미연's head and the slight downward tilt, so increased camera distance is the only emphasized change. Keep the embracing pair left of center as 미연 lowers her cheek toward 이현우's crown and he tucks his face fully inward, both still absorbed in the protective hold rather than turning toward the lens. The shattered window remains in the right background, with glass fragments and incoming seawater crossing the visible gap toward them but not yet making contact; keep this advancing mass within roughly the right third of the frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Shattered window behind and right of the embracing pair in the upper-right of the frame, background; Incoming water and fragments in the gap, not yet touching the pair in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Shattered upper-floor window (Breaking inward under the impact of the wave) — The opening is seen diagonally from inside the corridor, behind and to the right of the pair; used as Identifies the origin of the incoming water and fragments; Airborne glass fragments (Scattered inward between the window and the embracing pair); used as Marks the advancing impact front without obscuring the embrace; Incoming seawater (Bursting through the window immediately before reaching the pair) — Advancing diagonally from the right background toward the left midground; used as Makes the remaining separation before contact visible; Upper-floor corridor (Being breached by the incoming wave) — The interior diagonal remains continuous with the preceding embrace shot; used as Provides the visible space between the pair and the broken window.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding interior exposure and restrained contrast, keeping the airborne glass and water legible without an invented flash or lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The upper-floor windows are now shattering, leaving broken glazing and airborne fragments as seawater breaks into the building. 이현우: He remains curled inward at the moment of impact, with his earlier facial bruises and leg wound unchanged. The stiff contact card is still concealed inside his shoe. 미연: She maintains the tight protective embrace through the impact, with the earlier facial swelling still present. A mass of seawater and shattered window glass bursts into the interior.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — wearing: 거친 작업과 생활고에 찌들어 먼지가 가득 묻고 해진 회색 셔츠와 어두운 작업 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "우측 창문에서 쏟아지는 바닷물과 유리 파편들이 좌측의 두 사람을 향해 덮치듯 이동함.",
    "built_space": "복도 좌측의 문들과 소화기, 우측의 창문, 천장 조명들이 기준 이미지와 동일한 구조이며 카메라는 뒤로 후진함.",
    "entities": "미연과 이현우의 복장, 상처, 인이어 무전기 등 기준 이미지의 특징이 정확히 일치함.",
    "hard_violations": [],
    "physics": "두 사람은 바닥을 딛고 서로를 안고 있으며, 바닷물과 유리 파편들은 충격으로 허공에 흩뿌려져 있음."
   },
   {
    "label": "B",
    "direction": "우측 창문에서 시작된 물살과 파편이 좌측의 인물들을 향해 쏟아짐.",
    "built_space": "기준 이미지의 복도 구조(소화기, 문, 조명 등)를 유지하며 카메라가 후진함.",
    "entities": "미연과 이현우의 외형이 일치하나, 손가락 등의 해부학적 형태가 다소 뭉개짐.",
    "hard_violations": [],
    "physics": "인물들은 바닥에 서서 포옹 중이고, 파도와 유리는 공중에 떠 있는 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기준 이미지의 인물과 조명을 완벽히 유지하면서, 뒤로 후진한 카메라 구도와 우측에서 밀려오는 물살의 공간감을 지시문대로 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 구도와 파도의 움직임을 따랐으나, 인물의 손 묘사가 어색하고 유리 파편의 질감이 다소 평면적으로 표현되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "우측 창문에서 쏟아지는 바닷물과 유리 파편들이 좌측의 두 사람을 향해 덮치듯 이동함.",
        "built_space": "복도 좌측의 문들과 소화기, 우측의 창문, 천장 조명들이 기준 이미지와 동일한 구조이며 카메라는 뒤로 후진함.",
        "entities": "미연과 이현우의 복장, 상처, 인이어 무전기 등 기준 이미지의 특징이 정확히 일치함.",
        "hard_violations": [],
        "physics": "두 사람은 바닥을 딛고 서로를 안고 있으며, 바닷물과 유리 파편들은 충격으로 허공에 흩뿌려져 있음."
       },
       {
        "label": "B",
        "direction": "우측 창문에서 시작된 물살과 파편이 좌측의 인물들을 향해 쏟아짐.",
        "built_space": "기준 이미지의 복도 구조(소화기, 문, 조명 등)를 유지하며 카메라가 후진함.",
        "entities": "미연과 이현우의 외형이 일치하나, 손가락 등의 해부학적 형태가 다소 뭉개짐.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 서서 포옹 중이고, 파도와 유리는 공중에 떠 있는 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기준 이미지의 인물과 조명을 완벽히 유지하면서, 뒤로 후진한 카메라 구도와 우측에서 밀려오는 물살의 공간감을 지시문대로 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 구도와 파도의 움직임을 따랐으나, 인물의 손 묘사가 어색하고 유리 파편의 질감이 다소 평면적으로 표현되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "우측 창문에서 쏟아지는 바닷물과 유리 파편들이 좌측의 두 사람을 향해 덮치듯 이동함.",
        "built_space": "복도 좌측의 문들과 소화기, 우측의 창문, 천장 조명들이 기준 이미지와 동일한 구조이며 카메라는 뒤로 후진함.",
        "entities": "미연과 이현우의 복장, 상처, 인이어 무전기 등 기준 이미지의 특징이 정확히 일치함.",
        "hard_violations": [],
        "physics": "두 사람은 바닥을 딛고 서로를 안고 있으며, 바닷물과 유리 파편들은 충격으로 허공에 흩뿌려져 있음."
       },
       {
        "label": "B",
        "direction": "우측 창문에서 시작된 물살과 파편이 좌측의 인물들을 향해 쏟아짐.",
        "built_space": "기준 이미지의 복도 구조(소화기, 문, 조명 등)를 유지하며 카메라가 후진함.",
        "entities": "미연과 이현우의 외형이 일치하나, 손가락 등의 해부학적 형태가 다소 뭉개짐.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 서서 포옹 중이고, 파도와 유리는 공중에 떠 있는 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물이 조금 더 작고 왼쪽에 놓여 후퇴한 구도와 충돌 전 간격을 더 잘 살리지만, 요구한 와이드숏에는 부족하고 물과 파편이 오른쪽 3분의 1을 크게 벗어난다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "보호하는 포옹과 장소의 연속성은 맞지만, 인물이 더 크고 중앙에 가까우며 물보라가 머리 바로 뒤까지 차올라 요구한 와이드숏과 분리된 충돌 전 공간이 더 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연은 눈을 내리감고 이현우의 정수리 쪽으로 얼굴을 낮춘다. 이현우는 얼굴을 미연의 가슴 안쪽에 묻으며 둘 다 렌즈를 보지 않는다. 물과 유리 조각은 오른쪽 창에서 왼쪽 실내와 두 사람 쪽으로 분출한다. 주된 물덩어리와 포옹 사이에는 간격이 있으나, 잔물방울과 파편은 화면 중앙과 왼쪽까지 퍼져 있다.",
        "built_space": "왼쪽 벽의 금속 설비함 1개, 그 위 작은 붉은 장치 1개, 바닥 소화기 1개가 유지된다. 왼쪽 문은 최소 4개, 천장의 직사각 조명은 최소 5개가 깊이 방향으로 이어진다. 오른쪽에는 가까운 큰 창 개구부 2개와 뒤쪽의 반복 창틀이 보인다. 회색 하부 벽, 밝은 상부 벽, 반사되는 복도 바닥과 야간 항구가 참조 장소에 부합한다. 두 사람은 창가 통로의 중앙 왼쪽에 있고 실내 대각선도 유지되지만, 하체가 잘려 충분히 멀어진 와이드숏으로 보이지는 않는다.",
        "entities": "등장인물은 두 명뿐이다. 미연의 중년 동아시아계 여성 외형, 짧은 검은 머리, 뺨의 상처와 해진 회색 셔츠가 참조에 부합한다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 더럽고 어두운 셔츠와 귀의 작은 검은 장치도 맞는다. 숨겨진 얼굴로 정확한 연령과 멍은 확인하기 어렵고, 다리 상처와 신발 속 카드는 화면 밖이어서 판정할 수 없다. 깨진 창유리, 공중 파편, 바닷물이 모두 보이며 추가 인물이나 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "미연의 손은 이현우의 뒤통수와 어깨를 잡고, 이현우의 팔은 미연의 허리를 감싼다. 서로 접촉한 상체는 자연스러운 포옹으로 읽힌다. 발은 화면 밖이므로 바닥 접점은 확인할 수 없지만 공중에 뜬 자세는 아니다. 공중의 물과 유리는 창을 깨고 들어오는 파도의 충격으로 설명되며, 바닥의 물과 파편은 바닥이 받친다. 주된 파도는 아직 몸에 닿지 않았지만 이미 넓게 젖은 바닥과 흩어진 잔물보라는 최초 접촉 직전이라는 순간을 다소 흐린다."
       },
       {
        "label": "B",
        "direction": "미연은 이현우의 정수리를 향해 고개와 시선을 낮추고, 이현우는 얼굴을 완전히 안쪽으로 숨긴다. 두 사람 모두 카메라나 파도를 향해 돌아보지 않는다. 바닷물과 파편은 오른쪽 뒤 창에서 왼쪽의 포옹 쪽으로 향하지만, 물보라의 앞부분이 머리 바로 뒤까지 넓게 번져 남은 간격이 덜 명료하다.",
        "built_space": "왼쪽 금속 설비함 1개, 작은 붉은 벽 장치 1개, 소화기 1개가 보이고, 왼쪽 문은 최소 4개, 천장 조명은 최소 5개가 이어진다. 오른쪽의 가까운 창 개구부 2개와 뒤로 이어지는 창틀, 이색 벽면과 반사 바닥은 참조 복도의 구조를 따른다. 두 사람은 중앙 왼쪽이지만 A보다 중앙에 가깝고 크게 잡혔다. 실내의 약한 내려다보기와 대각선은 유지되나, 하체를 자른 구도로 요구한 와이드숏과 거리를 둔 관찰점이 충분히 구현되지 않았다.",
        "entities": "미연과 이현우에 해당하는 두 사람만 보인다. 미연의 중년 동아시아계 외형, 검은 단발, 뺨의 상처, 먼지 묻고 찢어진 회색 셔츠가 유지된다. 이현우의 검은 머리와 어두운 오염된 셔츠, 귀의 작은 검은 인이어 장치도 참조와 맞는다. 얼굴 대부분이 숨겨져 세부 얼굴 정체성과 멍은 확인이 제한되며, 다리 상처와 숨긴 카드는 확인 대상 밖이다. 깨진 유리와 바닷물, 야간 항구가 보이고 추가 인물이나 그래픽 문구는 없다.",
        "hard_violations": [],
        "physics": "미연의 두 손이 뒤통수와 어깨에 접촉하고 이현우의 팔이 허리를 감싸므로 포옹의 지지 관계가 성립한다. 하체와 발은 잘려 있지만 몸이 떠 있다고 볼 근거는 없다. 유리 조각과 물의 비행은 창밖 파도의 충격에서 시작되며, 아래쪽 조각들은 낙하 경로에 있다. 바닥에 고인 물은 바닥이 지지한다. 불가능한 부유는 없으나 물보라가 인물 윤곽 가까이 겹쳐, 아직 접촉하지 않은 순간을 명확하게 분리하지 못한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물이 조금 더 작고 왼쪽에 놓여 후퇴한 구도와 충돌 전 간격을 더 잘 살리지만, 요구한 와이드숏에는 부족하고 물과 파편이 오른쪽 3분의 1을 크게 벗어난다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "보호하는 포옹과 장소의 연속성은 맞지만, 인물이 더 크고 중앙에 가까우며 물보라가 머리 바로 뒤까지 차올라 요구한 와이드숏과 분리된 충돌 전 공간이 더 약하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "미연은 눈을 내리감고 이현우의 정수리 쪽으로 얼굴을 낮춘다. 이현우는 얼굴을 미연의 가슴 안쪽에 묻으며 둘 다 렌즈를 보지 않는다. 물과 유리 조각은 오른쪽 창에서 왼쪽 실내와 두 사람 쪽으로 분출한다. 주된 물덩어리와 포옹 사이에는 간격이 있으나, 잔물방울과 파편은 화면 중앙과 왼쪽까지 퍼져 있다.",
        "built_space": "왼쪽 벽의 금속 설비함 1개, 그 위 작은 붉은 장치 1개, 바닥 소화기 1개가 유지된다. 왼쪽 문은 최소 4개, 천장의 직사각 조명은 최소 5개가 깊이 방향으로 이어진다. 오른쪽에는 가까운 큰 창 개구부 2개와 뒤쪽의 반복 창틀이 보인다. 회색 하부 벽, 밝은 상부 벽, 반사되는 복도 바닥과 야간 항구가 참조 장소에 부합한다. 두 사람은 창가 통로의 중앙 왼쪽에 있고 실내 대각선도 유지되지만, 하체가 잘려 충분히 멀어진 와이드숏으로 보이지는 않는다.",
        "entities": "등장인물은 두 명뿐이다. 미연의 중년 동아시아계 여성 외형, 짧은 검은 머리, 뺨의 상처와 해진 회색 셔츠가 참조에 부합한다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 더럽고 어두운 셔츠와 귀의 작은 검은 장치도 맞는다. 숨겨진 얼굴로 정확한 연령과 멍은 확인하기 어렵고, 다리 상처와 신발 속 카드는 화면 밖이어서 판정할 수 없다. 깨진 창유리, 공중 파편, 바닷물이 모두 보이며 추가 인물이나 삽입 문구는 없다.",
        "hard_violations": [],
        "physics": "미연의 손은 이현우의 뒤통수와 어깨를 잡고, 이현우의 팔은 미연의 허리를 감싼다. 서로 접촉한 상체는 자연스러운 포옹으로 읽힌다. 발은 화면 밖이므로 바닥 접점은 확인할 수 없지만 공중에 뜬 자세는 아니다. 공중의 물과 유리는 창을 깨고 들어오는 파도의 충격으로 설명되며, 바닥의 물과 파편은 바닥이 받친다. 주된 파도는 아직 몸에 닿지 않았지만 이미 넓게 젖은 바닥과 흩어진 잔물보라는 최초 접촉 직전이라는 순간을 다소 흐린다."
       },
       {
        "label": "A",
        "direction": "미연은 이현우의 정수리를 향해 고개와 시선을 낮추고, 이현우는 얼굴을 완전히 안쪽으로 숨긴다. 두 사람 모두 카메라나 파도를 향해 돌아보지 않는다. 바닷물과 파편은 오른쪽 뒤 창에서 왼쪽의 포옹 쪽으로 향하지만, 물보라의 앞부분이 머리 바로 뒤까지 넓게 번져 남은 간격이 덜 명료하다.",
        "built_space": "왼쪽 금속 설비함 1개, 작은 붉은 벽 장치 1개, 소화기 1개가 보이고, 왼쪽 문은 최소 4개, 천장 조명은 최소 5개가 이어진다. 오른쪽의 가까운 창 개구부 2개와 뒤로 이어지는 창틀, 이색 벽면과 반사 바닥은 참조 복도의 구조를 따른다. 두 사람은 중앙 왼쪽이지만 A보다 중앙에 가깝고 크게 잡혔다. 실내의 약한 내려다보기와 대각선은 유지되나, 하체를 자른 구도로 요구한 와이드숏과 거리를 둔 관찰점이 충분히 구현되지 않았다.",
        "entities": "미연과 이현우에 해당하는 두 사람만 보인다. 미연의 중년 동아시아계 외형, 검은 단발, 뺨의 상처, 먼지 묻고 찢어진 회색 셔츠가 유지된다. 이현우의 검은 머리와 어두운 오염된 셔츠, 귀의 작은 검은 인이어 장치도 참조와 맞는다. 얼굴 대부분이 숨겨져 세부 얼굴 정체성과 멍은 확인이 제한되며, 다리 상처와 숨긴 카드는 확인 대상 밖이다. 깨진 유리와 바닷물, 야간 항구가 보이고 추가 인물이나 그래픽 문구는 없다.",
        "hard_violations": [],
        "physics": "미연의 두 손이 뒤통수와 어깨에 접촉하고 이현우의 팔이 허리를 감싸므로 포옹의 지지 관계가 성립한다. 하체와 발은 잘려 있지만 몸이 떠 있다고 볼 근거는 없다. 유리 조각과 물의 비행은 창밖 파도의 충격에서 시작되며, 아래쪽 조각들은 낙하 경로에 있다. 바닥에 고인 물은 바닥이 지지한다. 불가능한 부유는 없으나 물보라가 인물 윤곽 가까이 겹쳐, 아직 접촉하지 않은 순간을 명확하게 분리하지 못한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "기준 이미지의 인물과 조명을 완벽히 유지하면서, 뒤로 후진한 카메라 구도와 우측에서 밀려오는 물살의 공간감을 지시문대로 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 구도와 파도의 움직임을 따랐으나, 인물의 손 묘사가 어색하고 유리 파편의 질감이 다소 평면적으로 표현되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S37sh6_sel.png",
    "asset_id": "d5dfd10d-8a29-4ec6-af3d-07750a20bdc8",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:814924>",
    "asset_id": "c111939b-8620-4260-8164-459d42918c0a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a59-32fa-729c-afc5-9d4114e9a53c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S37sh6"
  }
 },
 "S38sh8::signage": {
  "fp": "b909947118d1766b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::30e49209e9b29fdd": {
  "subjects": [],
  "subject_text": "수몰된 인천 난민촌 7구역·표류 잔해 수면\n넓은 바닷물 수면 위로 철제 주거동 일부와 담벼락이 드러난 공간. 컨테이너 판자와 생활 잔해가 흩어져 떠 있고, 가장자리에 도심이 이어진다.",
  "identity": "canonical",
  "scope_id": "L53",
  "scope_role": "location_exterior",
  "scope_sha": "30df20591397d5f1"
 },
 "groupbg::family_floating_board": {
  "input_fingerprint": "d3e01ff4df6ce82b",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "family_floating_board",
    "tags": [
     "S38sh17",
     "S38sh8",
     "S63sh13"
    ]
   },
   "context_sig": "e715ccfe5d9bed1c"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야) / 수빈의 트럭 운전실: 현우의 차량과 나란히 주행하는 낡고 개방된 트럭 좌석이다. (특징: 지붕 없이 노출된 금속 프레임과 운전석의 수빈)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 저 멀리 수면 위에 컨테이너 판자 위에 누워있는 미연.\n- / INSERT(회상F.B): 앰버를... 하면서 숨을 거두기 직전의 엄마 모습 /\n\nTIME OF DAY (lock): dawn.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야) / 수빈의 트럭 운전실: 현우의 차량과 나란히 주행하는 낡고 개방된 트럭 좌석이다. (특징: 지붕 없이 노출된 금속 프레임과 운전석의 수빈)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 저 멀리 수면 위에 컨테이너 판자 위에 누워있는 미연.\n- / INSERT(회상F.B): 앰버를... 하면서 숨을 거두기 직전의 엄마 모습 /\n\nTIME OF DAY (lock): dawn.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_floating_board_369b22.png",
  "asset_id": "b8218b41-1568-4b21-8a79-84ea00652ec2",
  "input_asset_ids": [
   "1b8b2869-fb93-422b-ae03-adf5b196f3a8"
  ],
  "origin_tag": "S38sh8",
  "place_text": "On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.",
  "origin_inputs": {
   "place_text": "On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.",
   "time_of_day_en": "dawn",
   "conti_asset_id": "1b8b2869-fb93-422b-ae03-adf5b196f3a8"
  }
 },
 "S38sh8::bgfirst_bg": {
  "input_fingerprint": "e96f3f934913633c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 죽은 미연을 품에 꽉 껴안은 채 하늘을 향해 입을 크게 벌리고 절규하는 이현우의 상체.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside the platform at 이현우's seated shoulder height in the established three-quarter side view, tilting slightly upward rather than changing distance. His tightly embracing upper body occupies the center-right, his mouth open and gaze lifted toward the sky, while 미연's closed-eyed face and unsupported shoulders remain visible against him across the lower-left. Preserve a narrow margin of platform and sea so the embrace retains its physical isolation.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-board platform (Floating and supporting 이현우 and 미연) — Only a narrow portion of the supporting upper surface remains visible beneath them; used as Retains the precarious physical setting at the bottom edge; Sea (Surrounding the floating platform); used as Provides sparse peripheral space around the tightly held figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination and controlled contrast preserve the faces without theatrical emphasis, allowing grief to carry the image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 죽은 미연을 품에 꽉 껴안은 채 하늘을 향해 입을 크게 벌리고 절규하는 이현우의 상체.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside the platform at 이현우's seated shoulder height in the established three-quarter side view, tilting slightly upward rather than changing distance. His tightly embracing upper body occupies the center-right, his mouth open and gaze lifted toward the sky, while 미연's closed-eyed face and unsupported shoulders remain visible against him across the lower-left. Preserve a narrow margin of platform and sea so the embrace retains its physical isolation.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-board platform (Floating and supporting 이현우 and 미연) — Only a narrow portion of the supporting upper surface remains visible beneath them; used as Retains the precarious physical setting at the bottom edge; Sea (Surrounding the floating platform); used as Provides sparse peripheral space around the tightly held figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination and controlled contrast preserve the faces without theatrical emphasis, allowing grief to carry the image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh8__bgfirst_bg.png",
  "asset_id": "6879d4d5-23ab-4c18-bd7d-d97e8a244725",
  "input_asset_ids": [
   "1b8b2869-fb93-422b-ae03-adf5b196f3a8",
   "b8218b41-1568-4b21-8a79-84ea00652ec2"
  ]
 },
 "S38sh8": {
  "input_fingerprint": "d9d48ebe626f14ee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 죽은 미연을 품에 꽉 껴안은 채 하늘을 향해 입을 크게 벌리고 절규하는 이현우의 상체.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside the platform at 이현우's seated shoulder height in the established three-quarter side view, tilting slightly upward rather than changing distance. His tightly embracing upper body occupies the center-right, his mouth open and gaze lifted toward the sky, while 미연's closed-eyed face and unsupported shoulders remain visible against him across the lower-left. Preserve a narrow margin of platform and sea so the embrace retains its physical isolation.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-board platform (Floating and supporting 이현우 and 미연) — Only a narrow portion of the supporting upper surface remains visible beneath them; used as Retains the precarious physical setting at the bottom edge; Sea (Surrounding the floating platform); used as Provides sparse peripheral space around the tightly held figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination and controlled contrast preserve the faces without theatrical emphasis, allowing grief to carry the image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is held tightly against Hyunwoo's chest on the floating container panel, her eyes closed and her upper body enclosed by his arms. Her head's exact angle and the placement of her own arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, container panels and containers float on the sea above the inundated settlement. Blood remains on the floating panel at the abdominal-wound site. 미연: She is dead with her eyes closed on a floating container panel, soaked and bleeding from a puncture wound in her abdomen. Her earlier facial swelling remains. 이현우: He is soaked and crouched on the floating panel with both arms gathered into an embrace, crying intensely; his facial bruises and leg wound persist. The stiff contact card remains protected inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 죽은 미연을 품에 꽉 껴안은 채 하늘을 향해 입을 크게 벌리고 절규하는 이현우의 상체.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside the platform at 이현우's seated shoulder height in the established three-quarter side view, tilting slightly upward rather than changing distance. His tightly embracing upper body occupies the center-right, his mouth open and gaze lifted toward the sky, while 미연's closed-eyed face and unsupported shoulders remain visible against him across the lower-left. Preserve a narrow margin of platform and sea so the embrace retains its physical isolation.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-board platform (Floating and supporting 이현우 and 미연) — Only a narrow portion of the supporting upper surface remains visible beneath them; used as Retains the precarious physical setting at the bottom edge; Sea (Surrounding the floating platform); used as Provides sparse peripheral space around the tightly held figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination and controlled contrast preserve the faces without theatrical emphasis, allowing grief to carry the image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is held tightly against Hyunwoo's chest on the floating container panel, her eyes closed and her upper body enclosed by his arms. Her head's exact angle and the placement of her own arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, container panels and containers float on the sea above the inundated settlement. Blood remains on the floating panel at the abdominal-wound site. 미연: She is dead with her eyes closed on a floating container panel, soaked and bleeding from a puncture wound in her abdomen. Her earlier facial swelling remains. 이현우: He is soaked and crouched on the floating panel with both arms gathered into an embrace, crying intensely; his facial bruises and leg wound persist. The stiff contact card remains protected inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 죽은 미연을 품에 꽉 껴안은 채 하늘을 향해 입을 크게 벌리고 절규하는 이현우의 상체.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, surrounded by open water and drifting debris. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside the platform at 이현우's seated shoulder height in the established three-quarter side view, tilting slightly upward rather than changing distance. His tightly embracing upper body occupies the center-right, his mouth open and gaze lifted toward the sky, while 미연's closed-eyed face and unsupported shoulders remain visible against him across the lower-left. Preserve a narrow margin of platform and sea so the embrace retains its physical isolation.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Container-board platform (Floating and supporting 이현우 and 미연) — Only a narrow portion of the supporting upper surface remains visible beneath them; used as Retains the precarious physical setting at the bottom edge; Sea (Surrounding the floating platform); used as Provides sparse peripheral space around the tightly held figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination and controlled contrast preserve the faces without theatrical emphasis, allowing grief to carry the image.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is held tightly against Hyunwoo's chest on the floating container panel, her eyes closed and her upper body enclosed by his arms. Her head's exact angle and the placement of her own arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At dawn, container panels and containers float on the sea above the inundated settlement. Blood remains on the floating panel at the abdominal-wound site. 미연: She is dead with her eyes closed on a floating container panel, soaked and bleeding from a puncture wound in her abdomen. Her earlier facial swelling remains. 이현우: He is soaked and crouched on the floating panel with both arms gathered into an embrace, crying intensely; his facial bruises and leg wound persist. The stiff contact card remains protected inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh8__bgfirst_bg.png",
     "asset_id": "6879d4d5-23ab-4c18-bd7d-d97e8a244725",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S38sh8.png",
     "asset_id": "1b8b2869-fb93-422b-ae03-adf5b196f3a8",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1241442>",
     "asset_id": "d8a82347-dda6-4d03-8e8b-568b33ae1411",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_floating_board_369b22.png",
     "asset_id": "b8218b41-1568-4b21-8a79-84ea00652ec2",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1241442>",
     "asset_id": "d8a82347-dda6-4d03-8e8b-568b33ae1411",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 크게 벌린 입이 하늘을 향해 있음.",
    "built_space": "바다 위 컨테이너 패널 가장자리가 하단에 좁게 보임.",
    "entities": "이현우(인이어 무전기, 상처)와 미연(눈을 감은 얼굴)이 명확히 식별됨.",
    "hard_violations": [
     "[gemini-pro] 죽은 사람(미연)의 손이 근육 긴장을 유지한 채 이현우의 옷자락을 능동적으로 쥐고 있는 물리적 오류(시신의 중력 순응 규칙 위배)."
    ],
    "physics": "미연의 오른팔이 이현우의 팔 위에 얹혀진 채 손가락에 힘이 들어가 있어 죽은 시신의 물리적 상태와 모순됨."
   },
   {
    "label": "B",
    "direction": "이현우가 하늘을 향해 시선을 두고 입을 벌려 절규함.",
    "built_space": "일출 무렵의 바다와 컨테이너 패널 표면 일부가 프레임 하단에 정확히 배치됨.",
    "entities": "이현우(인이어 무전기, 젖은 흑발, 핏자국)와 미연(눈을 감은 시신, 복부 상처 부위 피)의 외형이 잘 묘사됨.",
    "hard_violations": [],
    "physics": "이현우가 미연을 단단히 안고 있으며, 미연의 몸은 근육 긴장 없이 완전히 축 늘어져 중력에 완벽히 순응함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "시신이 중력에 순응하는 디테일과 앵글, 캐릭터의 절규하는 표정 등 프롬프트의 요구사항을 완벽히 구현함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "죽은 미연의 손이 이현우의 팔을 능동적으로 쥐고 있는 물리적 오류(Hard Violation)가 발생하여 심각한 감점 요인이 됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 크게 벌린 입이 하늘을 향해 있음.",
        "built_space": "바다 위 컨테이너 패널 가장자리가 하단에 좁게 보임.",
        "entities": "이현우(인이어 무전기, 상처)와 미연(눈을 감은 얼굴)이 명확히 식별됨.",
        "hard_violations": [
         "죽은 사람(미연)의 손이 근육 긴장을 유지한 채 이현우의 옷자락을 능동적으로 쥐고 있는 물리적 오류(시신의 중력 순응 규칙 위배)."
        ],
        "physics": "미연의 오른팔이 이현우의 팔 위에 얹혀진 채 손가락에 힘이 들어가 있어 죽은 시신의 물리적 상태와 모순됨."
       },
       {
        "label": "B",
        "direction": "이현우가 하늘을 향해 시선을 두고 입을 벌려 절규함.",
        "built_space": "일출 무렵의 바다와 컨테이너 패널 표면 일부가 프레임 하단에 정확히 배치됨.",
        "entities": "이현우(인이어 무전기, 젖은 흑발, 핏자국)와 미연(눈을 감은 시신, 복부 상처 부위 피)의 외형이 잘 묘사됨.",
        "hard_violations": [],
        "physics": "이현우가 미연을 단단히 안고 있으며, 미연의 몸은 근육 긴장 없이 완전히 축 늘어져 중력에 완벽히 순응함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "시신이 중력에 순응하는 디테일과 앵글, 캐릭터의 절규하는 표정 등 프롬프트의 요구사항을 완벽히 구현함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "죽은 미연의 손이 이현우의 팔을 능동적으로 쥐고 있는 물리적 오류(Hard Violation)가 발생하여 심각한 감점 요인이 됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 크게 벌린 입이 하늘을 향해 있음.",
        "built_space": "바다 위 컨테이너 패널 가장자리가 하단에 좁게 보임.",
        "entities": "이현우(인이어 무전기, 상처)와 미연(눈을 감은 얼굴)이 명확히 식별됨.",
        "hard_violations": [
         "죽은 사람(미연)의 손이 근육 긴장을 유지한 채 이현우의 옷자락을 능동적으로 쥐고 있는 물리적 오류(시신의 중력 순응 규칙 위배)."
        ],
        "physics": "미연의 오른팔이 이현우의 팔 위에 얹혀진 채 손가락에 힘이 들어가 있어 죽은 시신의 물리적 상태와 모순됨."
       },
       {
        "label": "B",
        "direction": "이현우가 하늘을 향해 시선을 두고 입을 벌려 절규함.",
        "built_space": "일출 무렵의 바다와 컨테이너 패널 표면 일부가 프레임 하단에 정확히 배치됨.",
        "entities": "이현우(인이어 무전기, 젖은 흑발, 핏자국)와 미연(눈을 감은 시신, 복부 상처 부위 피)의 외형이 잘 묘사됨.",
        "hard_violations": [],
        "physics": "이현우가 미연을 단단히 안고 있으며, 미연의 몸은 근육 긴장 없이 완전히 축 늘어져 중력에 완벽히 순응함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "하늘을 보며 절규하는 현우와 왼쪽 아래에 안긴 미연을 더 크게 담아, 상체 중심 클로즈업과 꽉 끌어안는 순간을 더 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "위로 젖힌 얼굴과 측면 시점은 충실하지만, 인물이 더 작고 머리 위·주변 풍경의 여백이 넓어 지정된 밀착 클로즈업에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 턱을 들고 눈을 위쪽 하늘로 향하며 입을 크게 벌리고 있다. 절규의 방향이 지시와 맞는다. 미연은 눈을 감고 얼굴을 현우의 가슴 쪽으로 기댄다. 두 팔은 미연의 상체를 감싸며, 별도의 지향성 소품은 없다.",
        "built_space": "두 사람 아래에 녹슨 골판 금속 컨테이너 패널 하나가 있고, 상판과 테두리·모서리 결합부가 하단 양옆에 보인다. 뒤에는 바다, 떠 있는 컨테이너 두 개, 드럼통과 작은 잔해, 먼 도시 윤곽이 있어 장소 참조와 부합한다. 현우의 상체는 중앙 오른쪽, 미연의 얼굴은 그 왼쪽 아래에 놓인다. 어깨 높이 부근의 사선 시점으로 보이지만, 패널과 바다의 노출은 지시된 좁은 여백보다 넓다.",
        "entities": "등장인물은 현우와 미연 두 명뿐이다. 현우는 젊은 동아시아계 남성으로 짧고 헝클어진 검은 머리, 마른 체격, 어두운 오염된 셔츠, 귀의 소형 인이어 장치, 얼굴의 상처가 보인다. 미연은 중년 동아시아계 여성으로 검은 머리와 참조에 가까운 낡은 회갈색 겉옷을 입고 눈을 감고 있다. 두 사람의 젖은 질감과 혈흔은 보이지만 미연의 기존 얼굴 부기는 뚜렷하지 않다. 복부 관통상 자체는 포옹과 크롭 때문에 확인하기 어렵고, 다리 상처와 신발 속 카드는 화면 밖이다. 새벽 태양은 있으나 빛과 수면의 금빛 반사는 요구한 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "현우의 하체는 대부분 잘려 있지만 몸은 바로 아래 패널 위에 앉아 있는 배치로 읽힌다. 미연의 머리는 현우의 가슴과 목 옆에 기대고, 어깨와 몸통은 현우의 팔과 몸에 받쳐져 있다. 보이는 손들은 미연을 실제로 감싸 쥐며, 미연이 스스로 팔다리를 들어 올리는 모습은 없다. 패널은 수면에 놓여 부력으로 지지된다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "현우는 얼굴을 뒤로 젖혀 하늘 방향으로 입을 벌린다. 눈은 거의 감겨 있어 동공의 시선은 확인하기 어렵지만 머리와 절규의 방향은 명확히 위쪽이다. 미연은 눈을 감은 채 현우의 가슴에 얼굴을 기댄다. 현우의 양팔은 미연의 어깨와 몸통을 감싼다.",
        "built_space": "하단에는 두 사람을 받치는 컨테이너 패널 하나와 그 테두리·모서리 결합부가 보인다. 주변 바다에는 좌우의 컨테이너 두 개, 오른쪽 드럼통, 작은 부유 잔해가 있으며 먼 도시와 산의 배치도 참조에 가깝다. 현우는 중앙 오른쪽, 미연은 왼쪽 아래에 있고 사선 측면 시점도 잘 드러난다. 다만 인물 위의 하늘과 옆의 바다가 넓고 하단에 굽힌 무릎 일부까지 들어와, 상체를 밀착해 담으라는 프레이밍보다 느슨하다.",
        "entities": "현우와 미연 외의 사람은 없다. 현우는 검은 젖은 머리의 젊은 동아시아계 남성이며 어두운 낡은 셔츠, 얼굴의 찰과상, 소형 인이어 장치를 갖추고 있다. 미연은 검은 머리의 중년 동아시아계 여성으로 참조와 유사한 회갈색 겉옷을 입고 눈을 감고 있다. 옷의 젖음과 혈흔은 보이지만 미연의 얼굴 부기는 명확하지 않다. 복부의 정확한 상처 위치는 팔과 크롭에 가려지고, 다리 상처와 신발 속 카드는 확인되지 않는다. 새벽 배경은 맞지만 주황색 하늘과 강한 수면 반사는 절제된 조명 요구보다 강조되어 있다.",
        "hard_violations": [],
        "physics": "현우의 굽힌 무릎 일부와 패널 위의 낮은 자세가 보여 앉거나 웅크린 지지가 성립한다. 미연은 현우의 가슴에 머리를 기대고, 그의 두 팔이 어깨와 몸통을 받친다. 미연이 능동적으로 몸을 세우거나 손을 들어 보이는 부분은 없다. 현우의 손도 포옹 대상과 접촉한다. 패널과 주변 잔해는 수면에 떠 있으며, 지지 없는 공중 부양은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "하늘을 보며 절규하는 현우와 왼쪽 아래에 안긴 미연을 더 크게 담아, 상체 중심 클로즈업과 꽉 끌어안는 순간을 더 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "위로 젖힌 얼굴과 측면 시점은 충실하지만, 인물이 더 작고 머리 위·주변 풍경의 여백이 넓어 지정된 밀착 클로즈업에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 턱을 들고 눈을 위쪽 하늘로 향하며 입을 크게 벌리고 있다. 절규의 방향이 지시와 맞는다. 미연은 눈을 감고 얼굴을 현우의 가슴 쪽으로 기댄다. 두 팔은 미연의 상체를 감싸며, 별도의 지향성 소품은 없다.",
        "built_space": "두 사람 아래에 녹슨 골판 금속 컨테이너 패널 하나가 있고, 상판과 테두리·모서리 결합부가 하단 양옆에 보인다. 뒤에는 바다, 떠 있는 컨테이너 두 개, 드럼통과 작은 잔해, 먼 도시 윤곽이 있어 장소 참조와 부합한다. 현우의 상체는 중앙 오른쪽, 미연의 얼굴은 그 왼쪽 아래에 놓인다. 어깨 높이 부근의 사선 시점으로 보이지만, 패널과 바다의 노출은 지시된 좁은 여백보다 넓다.",
        "entities": "등장인물은 현우와 미연 두 명뿐이다. 현우는 젊은 동아시아계 남성으로 짧고 헝클어진 검은 머리, 마른 체격, 어두운 오염된 셔츠, 귀의 소형 인이어 장치, 얼굴의 상처가 보인다. 미연은 중년 동아시아계 여성으로 검은 머리와 참조에 가까운 낡은 회갈색 겉옷을 입고 눈을 감고 있다. 두 사람의 젖은 질감과 혈흔은 보이지만 미연의 기존 얼굴 부기는 뚜렷하지 않다. 복부 관통상 자체는 포옹과 크롭 때문에 확인하기 어렵고, 다리 상처와 신발 속 카드는 화면 밖이다. 새벽 태양은 있으나 빛과 수면의 금빛 반사는 요구한 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "현우의 하체는 대부분 잘려 있지만 몸은 바로 아래 패널 위에 앉아 있는 배치로 읽힌다. 미연의 머리는 현우의 가슴과 목 옆에 기대고, 어깨와 몸통은 현우의 팔과 몸에 받쳐져 있다. 보이는 손들은 미연을 실제로 감싸 쥐며, 미연이 스스로 팔다리를 들어 올리는 모습은 없다. 패널은 수면에 놓여 부력으로 지지된다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "현우는 얼굴을 뒤로 젖혀 하늘 방향으로 입을 벌린다. 눈은 거의 감겨 있어 동공의 시선은 확인하기 어렵지만 머리와 절규의 방향은 명확히 위쪽이다. 미연은 눈을 감은 채 현우의 가슴에 얼굴을 기댄다. 현우의 양팔은 미연의 어깨와 몸통을 감싼다.",
        "built_space": "하단에는 두 사람을 받치는 컨테이너 패널 하나와 그 테두리·모서리 결합부가 보인다. 주변 바다에는 좌우의 컨테이너 두 개, 오른쪽 드럼통, 작은 부유 잔해가 있으며 먼 도시와 산의 배치도 참조에 가깝다. 현우는 중앙 오른쪽, 미연은 왼쪽 아래에 있고 사선 측면 시점도 잘 드러난다. 다만 인물 위의 하늘과 옆의 바다가 넓고 하단에 굽힌 무릎 일부까지 들어와, 상체를 밀착해 담으라는 프레이밍보다 느슨하다.",
        "entities": "현우와 미연 외의 사람은 없다. 현우는 검은 젖은 머리의 젊은 동아시아계 남성이며 어두운 낡은 셔츠, 얼굴의 찰과상, 소형 인이어 장치를 갖추고 있다. 미연은 검은 머리의 중년 동아시아계 여성으로 참조와 유사한 회갈색 겉옷을 입고 눈을 감고 있다. 옷의 젖음과 혈흔은 보이지만 미연의 얼굴 부기는 명확하지 않다. 복부의 정확한 상처 위치는 팔과 크롭에 가려지고, 다리 상처와 신발 속 카드는 확인되지 않는다. 새벽 배경은 맞지만 주황색 하늘과 강한 수면 반사는 절제된 조명 요구보다 강조되어 있다.",
        "hard_violations": [],
        "physics": "현우의 굽힌 무릎 일부와 패널 위의 낮은 자세가 보여 앉거나 웅크린 지지가 성립한다. 미연은 현우의 가슴에 머리를 기대고, 그의 두 팔이 어깨와 몸통을 받친다. 미연이 능동적으로 몸을 세우거나 손을 들어 보이는 부분은 없다. 현우의 손도 포옹 대상과 접촉한다. 패널과 주변 잔해는 수면에 떠 있으며, 지지 없는 공중 부양은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.125,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 죽은 사람(미연)의 손이 근육 긴장을 유지한 채 이현우의 옷자락을 능동적으로 쥐고 있는 물리적 오류(시신의 중력 순응 규칙 위배)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "시신이 중력에 순응하는 디테일과 앵글, 캐릭터의 절규하는 표정 등 프롬프트의 요구사항을 완벽히 구현함."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "죽은 미연의 손이 이현우의 팔을 능동적으로 쥐고 있는 물리적 오류(Hard Violation)가 발생하여 심각한 감점 요인이 됨.  ★위반: [gemini-pro] 죽은 사람(미연)의 손이 근육 긴장을 유지한 채 이현우의 옷자락을 능동적으로 쥐고 있는 물리적 오류(시신의 중력 순응 규칙 위배)."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_family_floating_board_369b22.png",
    "asset_id": "b8218b41-1568-4b21-8a79-84ea00652ec2",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1241442>",
    "asset_id": "d8a82347-dda6-4d03-8e8b-568b33ae1411",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a5e-bd1b-7b38-bd5f-11ea64f6c96a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh8__bgfirst_bg.png",
   "bg_asset_id": "6879d4d5-23ab-4c18-bd7d-d97e8a244725",
   "bg_record_key": "S38sh8::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "family_floating_board",
   "groupbg_asset_id": "b8218b41-1568-4b21-8a79-84ea00652ec2"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S38sh10::signage": {
  "fp": "db2d8b0538eca082",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh10::bgfirst_bg": {
  "input_fingerprint": "908770f6b4c9b235",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 컨테이너 문 밖 판자 위에서 한쪽 팔을 뻗어 앰버의 손을 꽉 쥐고 몸쪽으로 끌어당기는 찰리의 굳건한 자세.\n\nLOCATION (lock): On a floating panel just outside an open container doorway, above the floodwater at dawn.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From outside the open doorway at 찰리's waist height, track gently alongside him and look slightly upward across his extended arm toward 앰버 at the threshold. Place 찰리 in three-quarter view on the left, braced backward into the pull, with 앰버 on the right pitching toward him; their clasped hands bridge the center without exaggerated foreground scale. Both direct their attention downward to the secure handhold, with the doorway defining 앰버's route out.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container doorway (Open as 앰버 climbs out) — The opening is seen obliquely, with its side and threshold framing 앰버; used as Separates the interior starting point from 찰리's supporting position outside; Exterior supporting platform (Supporting 찰리 outside the doorway) — A narrow upper surface is visible beneath his braced position; used as Makes the leverage of the assisted climb physically legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained dawn illumination maintains readable hands and precise hard-surface contours while keeping the rescue intimate rather than triumphant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 컨테이너 문 밖 판자 위에서 한쪽 팔을 뻗어 앰버의 손을 꽉 쥐고 몸쪽으로 끌어당기는 찰리의 굳건한 자세.\n\nLOCATION (lock): On a floating panel just outside an open container doorway, above the floodwater at dawn.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From outside the open doorway at 찰리's waist height, track gently alongside him and look slightly upward across his extended arm toward 앰버 at the threshold. Place 찰리 in three-quarter view on the left, braced backward into the pull, with 앰버 on the right pitching toward him; their clasped hands bridge the center without exaggerated foreground scale. Both direct their attention downward to the secure handhold, with the doorway defining 앰버's route out.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container doorway (Open as 앰버 climbs out) — The opening is seen obliquely, with its side and threshold framing 앰버; used as Separates the interior starting point from 찰리's supporting position outside; Exterior supporting platform (Supporting 찰리 outside the doorway) — A narrow upper surface is visible beneath his braced position; used as Makes the leverage of the assisted climb physically legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained dawn illumination maintains readable hands and precise hard-surface contours while keeping the rescue intimate rather than triumphant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh10__bgfirst_bg.png",
  "asset_id": "ed9aacf7-4d25-4105-a706-0e3472b33cbe",
  "input_asset_ids": [
   "c554d318-ee8c-4f02-97a0-f491c6957653",
   "d6290e1d-9c49-4e68-8488-f454a4d31ec4"
  ]
 },
 "S38sh10": {
  "input_fingerprint": "ae39b4f138d1dfda",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖 판자 위에서 한쪽 팔을 뻗어 앰버의 손을 꽉 쥐고 몸쪽으로 끌어당기는 찰리의 굳건한 자세.\n\nLOCATION (lock): On a floating panel just outside an open container doorway, above the floodwater at dawn. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From outside the open doorway at 찰리's waist height, track gently alongside him and look slightly upward across his extended arm toward 앰버 at the threshold. Place 찰리 in three-quarter view on the left, braced backward into the pull, with 앰버 on the right pitching toward him; their clasped hands bridge the center without exaggerated foreground scale. Both direct their attention downward to the secure handhold, with the doorway defining 앰버's route out.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container doorway (Open as 앰버 climbs out) — The opening is seen obliquely, with its side and threshold framing 앰버; used as Separates the interior starting point from 찰리's supporting position outside; Exterior supporting platform (Supporting 찰리 outside the doorway) — A narrow upper surface is visible beneath his braced position; used as Makes the leverage of the assisted climb physically legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained dawn illumination maintains readable hands and precise hard-surface contours while keeping the rescue intimate rather than triumphant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The floating container's door is open, with storage boxes still inside. Nearby floating container panels remain adrift in dawn light. 찰리: He has climbed onto another floating container and is braced with one arm extended. The old coat and hat remain his established disguise. 앰버: She is climbing out of the container with one hand extended, after spending the night inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖 판자 위에서 한쪽 팔을 뻗어 앰버의 손을 꽉 쥐고 몸쪽으로 끌어당기는 찰리의 굳건한 자세.\n\nLOCATION (lock): On a floating panel just outside an open container doorway, above the floodwater at dawn. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From outside the open doorway at 찰리's waist height, track gently alongside him and look slightly upward across his extended arm toward 앰버 at the threshold. Place 찰리 in three-quarter view on the left, braced backward into the pull, with 앰버 on the right pitching toward him; their clasped hands bridge the center without exaggerated foreground scale. Both direct their attention downward to the secure handhold, with the doorway defining 앰버's route out.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container doorway (Open as 앰버 climbs out) — The opening is seen obliquely, with its side and threshold framing 앰버; used as Separates the interior starting point from 찰리's supporting position outside; Exterior supporting platform (Supporting 찰리 outside the doorway) — A narrow upper surface is visible beneath his braced position; used as Makes the leverage of the assisted climb physically legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained dawn illumination maintains readable hands and precise hard-surface contours while keeping the rescue intimate rather than triumphant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The floating container's door is open, with storage boxes still inside. Nearby floating container panels remain adrift in dawn light. 찰리: He has climbed onto another floating container and is braced with one arm extended. The old coat and hat remain his established disguise. 앰버: She is climbing out of the container with one hand extended, after spending the night inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 열린 컨테이너 문 밖 판자 위에서 한쪽 팔을 뻗어 앰버의 손을 꽉 쥐고 몸쪽으로 끌어당기는 찰리의 굳건한 자세.\n\nLOCATION (lock): On a floating panel just outside an open container doorway, above the floodwater at dawn. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From outside the open doorway at 찰리's waist height, track gently alongside him and look slightly upward across his extended arm toward 앰버 at the threshold. Place 찰리 in three-quarter view on the left, braced backward into the pull, with 앰버 on the right pitching toward him; their clasped hands bridge the center without exaggerated foreground scale. Both direct their attention downward to the secure handhold, with the doorway defining 앰버's route out.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Container doorway (Open as 앰버 climbs out) — The opening is seen obliquely, with its side and threshold framing 앰버; used as Separates the interior starting point from 찰리's supporting position outside; Exterior supporting platform (Supporting 찰리 outside the doorway) — A narrow upper surface is visible beneath his braced position; used as Makes the leverage of the assisted climb physically legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained dawn illumination maintains readable hands and precise hard-surface contours while keeping the rescue intimate rather than triumphant.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The floating container's door is open, with storage boxes still inside. Nearby floating container panels remain adrift in dawn light. 찰리: He has climbed onto another floating container and is braced with one arm extended. The old coat and hat remain his established disguise. 앰버: She is climbing out of the container with one hand extended, after spending the night inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh10__bgfirst_bg.png",
     "asset_id": "ed9aacf7-4d25-4105-a706-0e3472b33cbe",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S38sh10.png",
     "asset_id": "c554d318-ee8c-4f02-97a0-f491c6957653",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_flooded_refugee_district_sel.png",
     "asset_id": "d6290e1d-9c49-4e68-8488-f454a4d31ec4",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "왼쪽 찰리가 오른팔을 오른쪽 앰버에게 뻗어 화면 중앙 부근에서 손을 잡는다. 앰버는 고개와 눈을 아래쪽 손잡이로 향하지만, 찰리의 얼굴은 손보다 높은 앰버의 얼굴 쪽을 향한다. 앰버의 상체는 찰리 쪽으로 기울었으나 찰리는 뒤로 당기기보다는 비교적 수직으로 버티고 있다.",
    "built_space": "오른쪽에 컨테이너 출입구 하나와 문턱 하나가 있고, 열린 문짝의 측면과 양쪽 세로 프레임이 보인다. 내부에는 적어도 세 개의 수납 상자가 보인다. 찰리는 문턱보다 낮은 외부 부유 패널 위에, 앰버는 문턱에 있다. 출입구와 외부 지지면의 연결은 이해되지만, 지지면이 좁게 보이는 미디엄 숏이 아니라 찰리의 전신과 넓은 바닥을 보여준다. 배경의 사장교와 건물군은 참조 장소에 대응하나 긴 콘크리트 방조제는 뚜렷하지 않다.",
    "entities": "인물은 찰리와 앰버 둘뿐이다. 찰리는 샌드 베이지 장갑, 흰 기계 얼굴, 주황색 눈 두 개, 선 모양 입, 파란 가슴 원자로와 낡은 모자·외투를 갖췄다. 짧은 다리와 육중한 팔은 참조 체형에 비교적 가깝다. 앰버는 금발의 어린 여자아이로, 방진 마스크와 때 묻은 카키 작업복, 공구 벨트를 착용했다. 마스크로 얼굴이 가려져 정확한 혼혈 외모와 얼굴 일치는 제한적으로만 확인된다. 홍수 수면과 떠다니는 패널·컨테이너, 내부 상자가 있으며 새벽 하늘이지만 직사광의 주황색은 요청한 절제된 조명보다 강하다.",
    "hard_violations": [],
    "physics": "찰리의 두 기계 발이 외부 패널에 닿아 있고, 벌린 다리가 체중을 지탱한다. 앰버의 앞쪽 부츠는 문턱에 놓였으며 다른 팔은 오른쪽 문틀 방향으로 뻗어 있으나 손의 접촉점은 가려져 있다. 맞잡은 손의 접촉은 보이고 무지지 공중부양은 없다. 다만 찰리의 상체가 충분히 뒤로 실리지 않아 자기 몸쪽으로 끌어당기는 힘은 약하게 읽힌다."
   },
   {
    "label": "A",
    "direction": "찰리는 왼쪽에서 오른팔을 뻗고, 앰버는 오른쪽 문턱에서 왼쪽으로 몸을 내민다. 두 사람의 손은 중앙 부근에서 맞물린다. 찰리의 숙인 얼굴과 앰버의 내려간 눈은 모두 손잡이 쪽을 향한다. 찰리의 몸통은 손에서 멀어지는 왼쪽으로 물러나 있어 앰버를 자기 쪽으로 당기는 방향이 더 명확하다.",
    "built_space": "오른쪽에 비스듬히 보이는 출입구 하나, 문턱 하나, 열린 문짝의 가장자리와 세로 프레임이 있다. 내부에는 검은 상자와 나무 상자 등 적어도 다섯 개의 수납 상자가 보인다. 찰리는 왼쪽의 별도 부유 컨테이너 상판에, 앰버는 오른쪽 문턱에 위치한다. 두 지지면 사이의 수면도 보인다. 배경의 긴 콘크리트 방조제와 도시 건물군은 장소 참조에 더 가깝다. 다만 양쪽 바닥과 거의 전신을 보여주어 지정된 미디엄 숏보다 넓다.",
    "entities": "찰리와 앰버 외의 인물은 없다. 찰리의 베이지 장갑, 흰 마스크형 얼굴, 빛나는 눈과 선 모양 입, 파란 원자로, 낡은 외투와 모자는 맞는다. 그러나 다리가 길고 앰버에 비해 몸집이 매우 커서 작고 땅딸막한 고릴라형이라는 참조 비율은 약해졌다. 앰버는 금발의 어린 여자아이이며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 마스크 때문에 정확한 얼굴 일치는 일부만 판단할 수 있다. 내부 보관 상자와 주변 부유 패널, 홍수 수면이 있으며 새벽을 나타내지만 따뜻한 역광은 다소 강하다.",
    "hard_violations": [],
    "physics": "찰리는 다리를 넓게 벌려 부유 상판을 딛고 있으며, 오른쪽에 보이는 발의 접지가 분명하고 왼쪽 앞발은 화면 하단에서 일부 잘린다. 뒤로 실린 몸통과 뻗은 팔이 당김의 반력을 만든다. 앰버는 한 부츠로 문턱을 딛고 다른 손바닥으로 오른쪽 문틀을 지지하며, 찰리의 손을 잡아 앞으로 체중을 옮긴다. 두 인물 모두 지지점이 확인되고, 떠 있는 패널과 컨테이너는 수면의 부력으로 지탱되는 배치다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리의 체형과 문턱의 지지는 잘 구현했지만, 전신으로 넓어진 구도와 앰버의 얼굴을 향한 찰리의 시선, 약한 뒤쪽 버팀이 지정된 구조 동작과 어긋난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "미디엄 숏보다 넓고 찰리가 지나치게 크지만, 맞잡은 손을 향한 시선과 뒤로 버티는 당김, 문턱에서 나오는 앰버의 동작 및 방조제 배경이 더 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 찰리가 오른팔을 오른쪽 앰버에게 뻗어 화면 중앙 부근에서 손을 잡는다. 앰버는 고개와 눈을 아래쪽 손잡이로 향하지만, 찰리의 얼굴은 손보다 높은 앰버의 얼굴 쪽을 향한다. 앰버의 상체는 찰리 쪽으로 기울었으나 찰리는 뒤로 당기기보다는 비교적 수직으로 버티고 있다.",
        "built_space": "오른쪽에 컨테이너 출입구 하나와 문턱 하나가 있고, 열린 문짝의 측면과 양쪽 세로 프레임이 보인다. 내부에는 적어도 세 개의 수납 상자가 보인다. 찰리는 문턱보다 낮은 외부 부유 패널 위에, 앰버는 문턱에 있다. 출입구와 외부 지지면의 연결은 이해되지만, 지지면이 좁게 보이는 미디엄 숏이 아니라 찰리의 전신과 넓은 바닥을 보여준다. 배경의 사장교와 건물군은 참조 장소에 대응하나 긴 콘크리트 방조제는 뚜렷하지 않다.",
        "entities": "인물은 찰리와 앰버 둘뿐이다. 찰리는 샌드 베이지 장갑, 흰 기계 얼굴, 주황색 눈 두 개, 선 모양 입, 파란 가슴 원자로와 낡은 모자·외투를 갖췄다. 짧은 다리와 육중한 팔은 참조 체형에 비교적 가깝다. 앰버는 금발의 어린 여자아이로, 방진 마스크와 때 묻은 카키 작업복, 공구 벨트를 착용했다. 마스크로 얼굴이 가려져 정확한 혼혈 외모와 얼굴 일치는 제한적으로만 확인된다. 홍수 수면과 떠다니는 패널·컨테이너, 내부 상자가 있으며 새벽 하늘이지만 직사광의 주황색은 요청한 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "찰리의 두 기계 발이 외부 패널에 닿아 있고, 벌린 다리가 체중을 지탱한다. 앰버의 앞쪽 부츠는 문턱에 놓였으며 다른 팔은 오른쪽 문틀 방향으로 뻗어 있으나 손의 접촉점은 가려져 있다. 맞잡은 손의 접촉은 보이고 무지지 공중부양은 없다. 다만 찰리의 상체가 충분히 뒤로 실리지 않아 자기 몸쪽으로 끌어당기는 힘은 약하게 읽힌다."
       },
       {
        "label": "B",
        "direction": "찰리는 왼쪽에서 오른팔을 뻗고, 앰버는 오른쪽 문턱에서 왼쪽으로 몸을 내민다. 두 사람의 손은 중앙 부근에서 맞물린다. 찰리의 숙인 얼굴과 앰버의 내려간 눈은 모두 손잡이 쪽을 향한다. 찰리의 몸통은 손에서 멀어지는 왼쪽으로 물러나 있어 앰버를 자기 쪽으로 당기는 방향이 더 명확하다.",
        "built_space": "오른쪽에 비스듬히 보이는 출입구 하나, 문턱 하나, 열린 문짝의 가장자리와 세로 프레임이 있다. 내부에는 검은 상자와 나무 상자 등 적어도 다섯 개의 수납 상자가 보인다. 찰리는 왼쪽의 별도 부유 컨테이너 상판에, 앰버는 오른쪽 문턱에 위치한다. 두 지지면 사이의 수면도 보인다. 배경의 긴 콘크리트 방조제와 도시 건물군은 장소 참조에 더 가깝다. 다만 양쪽 바닥과 거의 전신을 보여주어 지정된 미디엄 숏보다 넓다.",
        "entities": "찰리와 앰버 외의 인물은 없다. 찰리의 베이지 장갑, 흰 마스크형 얼굴, 빛나는 눈과 선 모양 입, 파란 원자로, 낡은 외투와 모자는 맞는다. 그러나 다리가 길고 앰버에 비해 몸집이 매우 커서 작고 땅딸막한 고릴라형이라는 참조 비율은 약해졌다. 앰버는 금발의 어린 여자아이이며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 마스크 때문에 정확한 얼굴 일치는 일부만 판단할 수 있다. 내부 보관 상자와 주변 부유 패널, 홍수 수면이 있으며 새벽을 나타내지만 따뜻한 역광은 다소 강하다.",
        "hard_violations": [],
        "physics": "찰리는 다리를 넓게 벌려 부유 상판을 딛고 있으며, 오른쪽에 보이는 발의 접지가 분명하고 왼쪽 앞발은 화면 하단에서 일부 잘린다. 뒤로 실린 몸통과 뻗은 팔이 당김의 반력을 만든다. 앰버는 한 부츠로 문턱을 딛고 다른 손바닥으로 오른쪽 문틀을 지지하며, 찰리의 손을 잡아 앞으로 체중을 옮긴다. 두 인물 모두 지지점이 확인되고, 떠 있는 패널과 컨테이너는 수면의 부력으로 지탱되는 배치다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리의 체형과 문턱의 지지는 잘 구현했지만, 전신으로 넓어진 구도와 앰버의 얼굴을 향한 찰리의 시선, 약한 뒤쪽 버팀이 지정된 구조 동작과 어긋난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 숏보다 넓고 찰리가 지나치게 크지만, 맞잡은 손을 향한 시선과 뒤로 버티는 당김, 문턱에서 나오는 앰버의 동작 및 방조제 배경이 더 충실하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 찰리가 오른팔을 오른쪽 앰버에게 뻗어 화면 중앙 부근에서 손을 잡는다. 앰버는 고개와 눈을 아래쪽 손잡이로 향하지만, 찰리의 얼굴은 손보다 높은 앰버의 얼굴 쪽을 향한다. 앰버의 상체는 찰리 쪽으로 기울었으나 찰리는 뒤로 당기기보다는 비교적 수직으로 버티고 있다.",
        "built_space": "오른쪽에 컨테이너 출입구 하나와 문턱 하나가 있고, 열린 문짝의 측면과 양쪽 세로 프레임이 보인다. 내부에는 적어도 세 개의 수납 상자가 보인다. 찰리는 문턱보다 낮은 외부 부유 패널 위에, 앰버는 문턱에 있다. 출입구와 외부 지지면의 연결은 이해되지만, 지지면이 좁게 보이는 미디엄 숏이 아니라 찰리의 전신과 넓은 바닥을 보여준다. 배경의 사장교와 건물군은 참조 장소에 대응하나 긴 콘크리트 방조제는 뚜렷하지 않다.",
        "entities": "인물은 찰리와 앰버 둘뿐이다. 찰리는 샌드 베이지 장갑, 흰 기계 얼굴, 주황색 눈 두 개, 선 모양 입, 파란 가슴 원자로와 낡은 모자·외투를 갖췄다. 짧은 다리와 육중한 팔은 참조 체형에 비교적 가깝다. 앰버는 금발의 어린 여자아이로, 방진 마스크와 때 묻은 카키 작업복, 공구 벨트를 착용했다. 마스크로 얼굴이 가려져 정확한 혼혈 외모와 얼굴 일치는 제한적으로만 확인된다. 홍수 수면과 떠다니는 패널·컨테이너, 내부 상자가 있으며 새벽 하늘이지만 직사광의 주황색은 요청한 절제된 조명보다 강하다.",
        "hard_violations": [],
        "physics": "찰리의 두 기계 발이 외부 패널에 닿아 있고, 벌린 다리가 체중을 지탱한다. 앰버의 앞쪽 부츠는 문턱에 놓였으며 다른 팔은 오른쪽 문틀 방향으로 뻗어 있으나 손의 접촉점은 가려져 있다. 맞잡은 손의 접촉은 보이고 무지지 공중부양은 없다. 다만 찰리의 상체가 충분히 뒤로 실리지 않아 자기 몸쪽으로 끌어당기는 힘은 약하게 읽힌다."
       },
       {
        "label": "A",
        "direction": "찰리는 왼쪽에서 오른팔을 뻗고, 앰버는 오른쪽 문턱에서 왼쪽으로 몸을 내민다. 두 사람의 손은 중앙 부근에서 맞물린다. 찰리의 숙인 얼굴과 앰버의 내려간 눈은 모두 손잡이 쪽을 향한다. 찰리의 몸통은 손에서 멀어지는 왼쪽으로 물러나 있어 앰버를 자기 쪽으로 당기는 방향이 더 명확하다.",
        "built_space": "오른쪽에 비스듬히 보이는 출입구 하나, 문턱 하나, 열린 문짝의 가장자리와 세로 프레임이 있다. 내부에는 검은 상자와 나무 상자 등 적어도 다섯 개의 수납 상자가 보인다. 찰리는 왼쪽의 별도 부유 컨테이너 상판에, 앰버는 오른쪽 문턱에 위치한다. 두 지지면 사이의 수면도 보인다. 배경의 긴 콘크리트 방조제와 도시 건물군은 장소 참조에 더 가깝다. 다만 양쪽 바닥과 거의 전신을 보여주어 지정된 미디엄 숏보다 넓다.",
        "entities": "찰리와 앰버 외의 인물은 없다. 찰리의 베이지 장갑, 흰 마스크형 얼굴, 빛나는 눈과 선 모양 입, 파란 원자로, 낡은 외투와 모자는 맞는다. 그러나 다리가 길고 앰버에 비해 몸집이 매우 커서 작고 땅딸막한 고릴라형이라는 참조 비율은 약해졌다. 앰버는 금발의 어린 여자아이이며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 마스크 때문에 정확한 얼굴 일치는 일부만 판단할 수 있다. 내부 보관 상자와 주변 부유 패널, 홍수 수면이 있으며 새벽을 나타내지만 따뜻한 역광은 다소 강하다.",
        "hard_violations": [],
        "physics": "찰리는 다리를 넓게 벌려 부유 상판을 딛고 있으며, 오른쪽에 보이는 발의 접지가 분명하고 왼쪽 앞발은 화면 하단에서 일부 잘린다. 뒤로 실린 몸통과 뻗은 팔이 당김의 반력을 만든다. 앰버는 한 부츠로 문턱을 딛고 다른 손바닥으로 오른쪽 문틀을 지지하며, 찰리의 손을 잡아 앞으로 체중을 옮긴다. 두 인물 모두 지지점이 확인되고, 떠 있는 패널과 컨테이너는 수면의 부력으로 지탱되는 배치다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 5,
   "A": 7
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "찰리의 체형과 문턱의 지지는 잘 구현했지만, 전신으로 넓어진 구도와 앰버의 얼굴을 향한 찰리의 시선, 약한 뒤쪽 버팀이 지정된 구조 동작과 어긋난다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "미디엄 숏보다 넓고 찰리가 지나치게 크지만, 맞잡은 손을 향한 시선과 뒤로 버티는 당김, 문턱에서 나오는 앰버의 동작 및 방조제 배경이 더 충실하다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_flooded_refugee_district_sel.png",
    "asset_id": "d6290e1d-9c49-4e68-8488-f454a4d31ec4",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a67-d7bd-7b7f-83b9-7e6f722f89a0",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh10__bgfirst_bg.png",
   "bg_asset_id": "ed9aacf7-4d25-4105-a706-0e3472b33cbe",
   "bg_record_key": "S38sh10::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S38sh17::signage": {
  "fp": "07abad9ec56d7a31",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh17": {
  "input_fingerprint": "f3836d74cd1d6533",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미연을 안고 우는 이현우와 앰버를 내려다보며 슬픈 눈빛으로 고개를 숙인 찰리의 굽은 상체.\n\nLOCATION (lock): On the drifting container wreckage where the family has gathered, above the flooded settlement's water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane rise to a stop just above 찰리's shoulder height on his front-quarter side, looking down at his bowed upper body in the upper-right of the composition. Keep 이현우 hunched around 미연 at lower-left and 앰버 folded toward her at lower-center, their figures partially cropped but their relationship readable across the supporting platform. 찰리's gaze falls toward the grieving siblings rather than the lens, while both siblings remain absorbed in 미연 below them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Floating platform beneath the family (Supporting the grieving family beside 찰리) — The upper surface is visible between the lower figures; used as Connects 찰리's elevated reaction with the family below in one physical space; Surrounding sea (Visible beyond the platform); used as Leaves a restrained peripheral margin around the family.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued dawn exposure and restrained contrast, letting bowed postures convey tenderness without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is recumbent on the floating container panel beside the crouching Hyunwoo, her eyes closed as if asleep. Her head rests at panel level, while the exact orientation of her torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating container surfaces and panels remain surrounded by seawater in dawn light. Blood from the abdominal wound remains on the supporting panel. 미연: She remains motionless and dead, with closed eyes, a bleeding abdominal puncture wound, and the earlier facial swelling. 이현우: He remains crouched and unresponsive with grief on the floating panel, retaining his facial bruises and leg wound. The stiff contact card remains inside his shoe. 앰버: She has reached the floating panel and is crying out in shock and grief. 찰리: He stands on the floating container surface with a sorrowful, lowered gaze, retaining the old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미연을 안고 우는 이현우와 앰버를 내려다보며 슬픈 눈빛으로 고개를 숙인 찰리의 굽은 상체.\n\nLOCATION (lock): On the drifting container wreckage where the family has gathered, above the flooded settlement's water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane rise to a stop just above 찰리's shoulder height on his front-quarter side, looking down at his bowed upper body in the upper-right of the composition. Keep 이현우 hunched around 미연 at lower-left and 앰버 folded toward her at lower-center, their figures partially cropped but their relationship readable across the supporting platform. 찰리's gaze falls toward the grieving siblings rather than the lens, while both siblings remain absorbed in 미연 below them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Floating platform beneath the family (Supporting the grieving family beside 찰리) — The upper surface is visible between the lower figures; used as Connects 찰리's elevated reaction with the family below in one physical space; Surrounding sea (Visible beyond the platform); used as Leaves a restrained peripheral margin around the family.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued dawn exposure and restrained contrast, letting bowed postures convey tenderness without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is recumbent on the floating container panel beside the crouching Hyunwoo, her eyes closed as if asleep. Her head rests at panel level, while the exact orientation of her torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating container surfaces and panels remain surrounded by seawater in dawn light. Blood from the abdominal wound remains on the supporting panel. 미연: She remains motionless and dead, with closed eyes, a bleeding abdominal puncture wound, and the earlier facial swelling. 이현우: He remains crouched and unresponsive with grief on the floating panel, retaining his facial bruises and leg wound. The stiff contact card remains inside his shoe. 앰버: She has reached the floating panel and is crying out in shock and grief. 찰리: He stands on the floating container surface with a sorrowful, lowered gaze, retaining the old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미연을 안고 우는 이현우와 앰버를 내려다보며 슬픈 눈빛으로 고개를 숙인 찰리의 굽은 상체.\n\nLOCATION (lock): On the drifting container wreckage where the family has gathered, above the flooded settlement's water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane rise to a stop just above 찰리's shoulder height on his front-quarter side, looking down at his bowed upper body in the upper-right of the composition. Keep 이현우 hunched around 미연 at lower-left and 앰버 folded toward her at lower-center, their figures partially cropped but their relationship readable across the supporting platform. 찰리's gaze falls toward the grieving siblings rather than the lens, while both siblings remain absorbed in 미연 below them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Floating platform beneath the family (Supporting the grieving family beside 찰리) — The upper surface is visible between the lower figures; used as Connects 찰리's elevated reaction with the family below in one physical space; Surrounding sea (Visible beyond the platform); used as Leaves a restrained peripheral margin around the family.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued dawn exposure and restrained contrast, letting bowed postures convey tenderness without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon's dead body is recumbent on the floating container panel beside the crouching Hyunwoo, her eyes closed as if asleep. Her head rests at panel level, while the exact orientation of her torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Floating container surfaces and panels remain surrounded by seawater in dawn light. Blood from the abdominal wound remains on the supporting panel. 미연: She remains motionless and dead, with closed eyes, a bleeding abdominal puncture wound, and the earlier facial swelling. 이현우: He remains crouched and unresponsive with grief on the floating panel, retaining his facial bruises and leg wound. The stiff contact card remains inside his shoe. 앰버: She has reached the floating panel and is crying out in shock and grief. 찰리: He stands on the floating container surface with a sorrowful, lowered gaze, retaining the old coat and hat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 가족을 내려다보고 있다. 이현우는 미연을 바라보지만, 앰버는 시선을 돌려 허공을 응시한다. 미연은 눈을 감고 누워 있다.",
    "built_space": "이전 샷에서 고정된 바다 위 잔해와 도시 스카이라인이 완전히 사라지고, 특징 없는 텅 빈 수면 위의 구조물로 묘사되었다.",
    "entities": "이현우는 이전 샷의 흙 묻은 셔츠가 아닌 깨끗한 남색 티셔츠를 입어 의상 고정 요건을 어겼다. 찰리는 흰색 평면 마스크 대신 인간의 코와 뼈대가 드러난 기괴한 얼굴을 하고 있다. 미연의 머리 밑에 프롬프트에 없는 베개가 추가되었다.",
    "hard_violations": [
     "[gpt-high] 미연의 머리를 앰버의 어깨 부근 높이의 쿠션에 놓고 상체를 비스듬히 세워, 머리가 패널 높이에 놓여야 한다는 고정 자세를 위반했다.",
     "[gpt-high] 지정되거나 참고에 없던 머리 받침 쿠션을 추가했다."
    ],
    "physics": "모든 인물이 구조물 바닥에 안정적으로 지지되어 있으며 부자연스럽게 떠 있는 요소는 없다."
   },
   {
    "label": "B",
    "direction": "찰리의 시선이 가족을 향하고 있다. 이현우와 앰버는 미연을 응시하고, 미연 역시 눈을 뜨고 앰버와 시선을 맞추고 있다.",
    "built_space": "이전 샷의 바다 위 컨테이너 잔해, 일출 조명, 멀리 보이는 스카이라인까지 정확히 일치하는 공간이 훌륭하게 구현되었다.",
    "entities": "찰리의 마스크와 빛나는 눈, 가슴의 원자로, 이현우의 낡은 셔츠 등 디자인과 의상 고정 요건을 정확히 지켰다. 다만 사망 상태여야 할 미연이 눈을 뜨고 미소 짓고 있으며, 가족들이 우는 대신 웃고 있어 상황 연출을 위반했다. 미연 위에 임의의 담요가 추가되었다.",
    "hard_violations": [
     "[gpt-high] 미연의 머리와 상체를 현우에게 기대어 높이 세우고 눈까지 뜨게 하여, 머리가 패널 높이에 놓인 사망자의 고정 자세를 위반했다.",
     "[gpt-high] 지정되거나 참고에 없던 큰 담요를 미연의 몸 위에 추가했다."
    ],
    "physics": "모든 인물이 컨테이너 잔해 위에 무게감을 가지고 자연스럽게 안착해 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "이전 샷의 배경과 조명, 이현우의 의상 및 찰리의 기계 외형을 매우 정확히 유지했으나, 우는 연기 대신 미소를 짓고 미연이 살아있는 상태로 묘사되어 감정선과 디테일에서 큰 감점이 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "미연이 눈을 감고 누워있으나, 이현우의 의상과 배경이 이전 샷의 고정 요건을 완전히 위반했으며 찰리의 마스크 형태도 심하게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 가족을 내려다보고 있다. 이현우는 미연을 바라보지만, 앰버는 시선을 돌려 허공을 응시한다. 미연은 눈을 감고 누워 있다.",
        "built_space": "이전 샷에서 고정된 바다 위 잔해와 도시 스카이라인이 완전히 사라지고, 특징 없는 텅 빈 수면 위의 구조물로 묘사되었다.",
        "entities": "이현우는 이전 샷의 흙 묻은 셔츠가 아닌 깨끗한 남색 티셔츠를 입어 의상 고정 요건을 어겼다. 찰리는 흰색 평면 마스크 대신 인간의 코와 뼈대가 드러난 기괴한 얼굴을 하고 있다. 미연의 머리 밑에 프롬프트에 없는 베개가 추가되었다.",
        "hard_violations": [],
        "physics": "모든 인물이 구조물 바닥에 안정적으로 지지되어 있으며 부자연스럽게 떠 있는 요소는 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 가족을 향하고 있다. 이현우와 앰버는 미연을 응시하고, 미연 역시 눈을 뜨고 앰버와 시선을 맞추고 있다.",
        "built_space": "이전 샷의 바다 위 컨테이너 잔해, 일출 조명, 멀리 보이는 스카이라인까지 정확히 일치하는 공간이 훌륭하게 구현되었다.",
        "entities": "찰리의 마스크와 빛나는 눈, 가슴의 원자로, 이현우의 낡은 셔츠 등 디자인과 의상 고정 요건을 정확히 지켰다. 다만 사망 상태여야 할 미연이 눈을 뜨고 미소 짓고 있으며, 가족들이 우는 대신 웃고 있어 상황 연출을 위반했다. 미연 위에 임의의 담요가 추가되었다.",
        "hard_violations": [],
        "physics": "모든 인물이 컨테이너 잔해 위에 무게감을 가지고 자연스럽게 안착해 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "이전 샷의 배경과 조명, 이현우의 의상 및 찰리의 기계 외형을 매우 정확히 유지했으나, 우는 연기 대신 미소를 짓고 미연이 살아있는 상태로 묘사되어 감정선과 디테일에서 큰 감점이 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "미연이 눈을 감고 누워있으나, 이현우의 의상과 배경이 이전 샷의 고정 요건을 완전히 위반했으며 찰리의 마스크 형태도 심하게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 가족을 내려다보고 있다. 이현우는 미연을 바라보지만, 앰버는 시선을 돌려 허공을 응시한다. 미연은 눈을 감고 누워 있다.",
        "built_space": "이전 샷에서 고정된 바다 위 잔해와 도시 스카이라인이 완전히 사라지고, 특징 없는 텅 빈 수면 위의 구조물로 묘사되었다.",
        "entities": "이현우는 이전 샷의 흙 묻은 셔츠가 아닌 깨끗한 남색 티셔츠를 입어 의상 고정 요건을 어겼다. 찰리는 흰색 평면 마스크 대신 인간의 코와 뼈대가 드러난 기괴한 얼굴을 하고 있다. 미연의 머리 밑에 프롬프트에 없는 베개가 추가되었다.",
        "hard_violations": [],
        "physics": "모든 인물이 구조물 바닥에 안정적으로 지지되어 있으며 부자연스럽게 떠 있는 요소는 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 가족을 향하고 있다. 이현우와 앰버는 미연을 응시하고, 미연 역시 눈을 뜨고 앰버와 시선을 맞추고 있다.",
        "built_space": "이전 샷의 바다 위 컨테이너 잔해, 일출 조명, 멀리 보이는 스카이라인까지 정확히 일치하는 공간이 훌륭하게 구현되었다.",
        "entities": "찰리의 마스크와 빛나는 눈, 가슴의 원자로, 이현우의 낡은 셔츠 등 디자인과 의상 고정 요건을 정확히 지켰다. 다만 사망 상태여야 할 미연이 눈을 뜨고 미소 짓고 있으며, 가족들이 우는 대신 웃고 있어 상황 연출을 위반했다. 미연 위에 임의의 담요가 추가되었다.",
        "hard_violations": [],
        "physics": "모든 인물이 컨테이너 잔해 위에 무게감을 가지고 자연스럽게 안착해 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "찰리의 외형과 현우의 의상은 더 가깝지만, 낮은 카메라와 눈을 뜨고 상체를 세운 미연이 지정된 내려다보기 구도와 사망 자세를 결정적으로 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "높은 전방 사선 시점과 절제된 노출은 더 충실하지만, 미연의 높은 머리 위치와 앰버의 정면 응시, 현우의 의상 변경 때문에 유효한 완성본은 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 숙인 얼굴은 왼쪽 아래 가족을 향하고, 현우와 앰버는 미연을 내려다본다. 그러나 미연은 눈을 뜨고 현우 쪽을 올려다보는 모습으로, 눈을 감은 사망 상태와 다르다. 무기나 이동 동작은 없다.",
        "built_space": "녹슨 골과 가장자리 철물이 있는 하나의 컨테이너 지지면에 네 인물이 모여 있다. 현우는 왼쪽, 앰버는 아래 중앙, 찰리는 오른쪽에 있어 상대 배치는 대체로 맞는다. 다만 카메라가 찰리의 어깨 위에서 내려다보기보다 가족의 높이에 가깝고, 수평선과 하늘 및 바다가 넓게 차지해 바다를 좁은 주변 여백으로 두라는 구도와 다르다.",
        "entities": "추가 인물 없이 현우·앰버·미연·찰리가 보인다. 현우는 검은 머리의 젊은 동아시아계 남성이며 이전 장면의 낡은 어두운 겉옷을 유지한다. 앰버는 금발 여자아이지만 참고의 남색 반소매 대신 긴 겉옷을 입었다. 미연은 중년 동아시아계 여성의 얼굴과 회갈색 의상을 유지하지만 살아 있는 표정이다. 찰리는 베이지 장갑, 흰 마스크, 둥근 발광 눈과 원형 가슴 장치가 참고와 비교적 가깝고 낡은 모자와 외투도 있다. 새 담요가 미연의 복부를 덮어 상처와 출혈 상태를 확인하기 어렵다.",
        "hard_violations": [
         "미연의 머리와 상체를 현우에게 기대어 높이 세우고 눈까지 뜨게 하여, 머리가 패널 높이에 놓인 사망자의 고정 자세를 위반했다.",
         "지정되거나 참고에 없던 큰 담요를 미연의 몸 위에 추가했다."
        ],
        "physics": "현우와 앰버는 패널에 앉거나 웅크려 체중을 싣고, 찰리는 발과 오른쪽 기계 손을 패널에 대고 몸을 굽힌다. 미연의 상체는 현우의 몸과 팔에 기대어 있어 공중에 떠 있는 것은 아니며, 손도 가족의 손과 몸 위에 놓여 있다. 다만 이러한 지지가 있더라도 머리를 패널에 내려놓으라는 사망 자세 조건은 충족하지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 아래쪽 가족을 향하고 현우도 아래를 보지만, 현우의 시선은 미연보다 앰버와 맞잡은 손 부근으로 읽힌다. 앰버는 미연에게 몸을 접지 않고 카메라 쪽을 정면으로 바라본다. 미연의 눈은 감겨 있다. 무기나 이동 동작은 없다.",
        "built_space": "바다에 둘러싸인 하나의 금속 패널 위에 가족과 찰리가 함께 있다. 높은 전방 사선에서 패널 윗면을 내려다보며 찰리를 오른쪽 위, 현우를 왼쪽 아래에 배치한 점은 요청에 더 가깝다. 그러나 앰버는 미연을 향해 숙이지 않고 현우 옆에 곧게 앉아 있으며, 미연의 상체가 중앙에 높게 놓인다. 패널은 이전 장면보다 평평하고 밝으며, 녹슨 골과 젖은 표면의 연속성이 약하다.",
        "entities": "등장 개체 수는 사람 셋과 기계 찰리 하나로 맞는다. 현우와 앰버는 참고와 가까운 젊은 동아시아계 남성과 금발 여자아이지만, 현우가 이전 장면의 오염된 겉옷 대신 깨끗한 남색 반소매를 입고 얼굴의 멍도 거의 보이지 않는다. 미연은 검은 머리의 중년 여성이고 눈을 감았으며 회갈색 겉옷은 유지한다. 찰리는 모자와 낡은 외투, 베이지 장갑을 갖췄으나 흰 얼굴에 입체적인 코와 눈구멍이 생겼고, 큰 원형 원자로가 작은 각진 청색 장치로 바뀌었다. 미연의 머리 아래에는 참고에 없는 쿠션이 있다.",
        "hard_violations": [
         "미연의 머리를 앰버의 어깨 부근 높이의 쿠션에 놓고 상체를 비스듬히 세워, 머리가 패널 높이에 놓여야 한다는 고정 자세를 위반했다.",
         "지정되거나 참고에 없던 머리 받침 쿠션을 추가했다."
        ],
        "physics": "현우와 앰버의 하체는 패널에 지지되고, 찰리의 보이는 발도 패널에 닿아 있다. 미연의 머리는 쿠션에, 몸통은 가족 옆과 패널에 걸쳐 기대며 손은 앞쪽의 다른 손과 몸 위에 놓인다. 명백하게 지지 없이 떠 있는 신체는 없지만, 머리 받침을 높인 연출은 요구된 완전한 와상 자세를 대신할 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "찰리의 외형과 현우의 의상은 더 가깝지만, 낮은 카메라와 눈을 뜨고 상체를 세운 미연이 지정된 내려다보기 구도와 사망 자세를 결정적으로 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "높은 전방 사선 시점과 절제된 노출은 더 충실하지만, 미연의 높은 머리 위치와 앰버의 정면 응시, 현우의 의상 변경 때문에 유효한 완성본은 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 숙인 얼굴은 왼쪽 아래 가족을 향하고, 현우와 앰버는 미연을 내려다본다. 그러나 미연은 눈을 뜨고 현우 쪽을 올려다보는 모습으로, 눈을 감은 사망 상태와 다르다. 무기나 이동 동작은 없다.",
        "built_space": "녹슨 골과 가장자리 철물이 있는 하나의 컨테이너 지지면에 네 인물이 모여 있다. 현우는 왼쪽, 앰버는 아래 중앙, 찰리는 오른쪽에 있어 상대 배치는 대체로 맞는다. 다만 카메라가 찰리의 어깨 위에서 내려다보기보다 가족의 높이에 가깝고, 수평선과 하늘 및 바다가 넓게 차지해 바다를 좁은 주변 여백으로 두라는 구도와 다르다.",
        "entities": "추가 인물 없이 현우·앰버·미연·찰리가 보인다. 현우는 검은 머리의 젊은 동아시아계 남성이며 이전 장면의 낡은 어두운 겉옷을 유지한다. 앰버는 금발 여자아이지만 참고의 남색 반소매 대신 긴 겉옷을 입었다. 미연은 중년 동아시아계 여성의 얼굴과 회갈색 의상을 유지하지만 살아 있는 표정이다. 찰리는 베이지 장갑, 흰 마스크, 둥근 발광 눈과 원형 가슴 장치가 참고와 비교적 가깝고 낡은 모자와 외투도 있다. 새 담요가 미연의 복부를 덮어 상처와 출혈 상태를 확인하기 어렵다.",
        "hard_violations": [
         "미연의 머리와 상체를 현우에게 기대어 높이 세우고 눈까지 뜨게 하여, 머리가 패널 높이에 놓인 사망자의 고정 자세를 위반했다.",
         "지정되거나 참고에 없던 큰 담요를 미연의 몸 위에 추가했다."
        ],
        "physics": "현우와 앰버는 패널에 앉거나 웅크려 체중을 싣고, 찰리는 발과 오른쪽 기계 손을 패널에 대고 몸을 굽힌다. 미연의 상체는 현우의 몸과 팔에 기대어 있어 공중에 떠 있는 것은 아니며, 손도 가족의 손과 몸 위에 놓여 있다. 다만 이러한 지지가 있더라도 머리를 패널에 내려놓으라는 사망 자세 조건은 충족하지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 아래쪽 가족을 향하고 현우도 아래를 보지만, 현우의 시선은 미연보다 앰버와 맞잡은 손 부근으로 읽힌다. 앰버는 미연에게 몸을 접지 않고 카메라 쪽을 정면으로 바라본다. 미연의 눈은 감겨 있다. 무기나 이동 동작은 없다.",
        "built_space": "바다에 둘러싸인 하나의 금속 패널 위에 가족과 찰리가 함께 있다. 높은 전방 사선에서 패널 윗면을 내려다보며 찰리를 오른쪽 위, 현우를 왼쪽 아래에 배치한 점은 요청에 더 가깝다. 그러나 앰버는 미연을 향해 숙이지 않고 현우 옆에 곧게 앉아 있으며, 미연의 상체가 중앙에 높게 놓인다. 패널은 이전 장면보다 평평하고 밝으며, 녹슨 골과 젖은 표면의 연속성이 약하다.",
        "entities": "등장 개체 수는 사람 셋과 기계 찰리 하나로 맞는다. 현우와 앰버는 참고와 가까운 젊은 동아시아계 남성과 금발 여자아이지만, 현우가 이전 장면의 오염된 겉옷 대신 깨끗한 남색 반소매를 입고 얼굴의 멍도 거의 보이지 않는다. 미연은 검은 머리의 중년 여성이고 눈을 감았으며 회갈색 겉옷은 유지한다. 찰리는 모자와 낡은 외투, 베이지 장갑을 갖췄으나 흰 얼굴에 입체적인 코와 눈구멍이 생겼고, 큰 원형 원자로가 작은 각진 청색 장치로 바뀌었다. 미연의 머리 아래에는 참고에 없는 쿠션이 있다.",
        "hard_violations": [
         "미연의 머리를 앰버의 어깨 부근 높이의 쿠션에 놓고 상체를 비스듬히 세워, 머리가 패널 높이에 놓여야 한다는 고정 자세를 위반했다.",
         "지정되거나 참고에 없던 머리 받침 쿠션을 추가했다."
        ],
        "physics": "현우와 앰버의 하체는 패널에 지지되고, 찰리의 보이는 발도 패널에 닿아 있다. 미연의 머리는 쿠션에, 몸통은 가족 옆과 패널에 걸쳐 기대며 손은 앞쪽의 다른 손과 몸 위에 놓인다. 명백하게 지지 없이 떠 있는 신체는 없지만, 머리 받침을 높인 연출은 요구된 완전한 와상 자세를 대신할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.417
   },
   "violations": {
    "B": [
     "[gpt-high] 미연의 머리와 상체를 현우에게 기대어 높이 세우고 눈까지 뜨게 하여, 머리가 패널 높이에 놓인 사망자의 고정 자세를 위반했다.",
     "[gpt-high] 지정되거나 참고에 없던 큰 담요를 미연의 몸 위에 추가했다."
    ],
    "A": [
     "[gpt-high] 미연의 머리를 앰버의 어깨 부근 높이의 쿠션에 놓고 상체를 비스듬히 세워, 머리가 패널 높이에 놓여야 한다는 고정 자세를 위반했다.",
     "[gpt-high] 지정되거나 참고에 없던 머리 받침 쿠션을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1417,
   "A": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "이전 샷의 배경과 조명, 이현우의 의상 및 찰리의 기계 외형을 매우 정확히 유지했으나, 우는 연기 대신 미소를 짓고 미연이 살아있는 상태로 묘사되어 감정선과 디테일에서 큰 감점이 있습니다.  ★위반: [gpt-high] 미연의 머리와 상체를 현우에게 기대어 높이 세우고 눈까지 뜨게 하여, 머리가 패널 높이에 놓인 사망자의 고정 자세를 위반했다. / [gpt-high] 지정되거나 참고에 없던 큰 담요를 미연의 몸 위에 추가했다."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "미연이 눈을 감고 누워있으나, 이현우의 의상과 배경이 이전 샷의 고정 요건을 완전히 위반했으며 찰리의 마스크 형태도 심하게 왜곡되었습니다.  ★위반: [gpt-high] 미연의 머리를 앰버의 어깨 부근 높이의 쿠션에 놓고 상체를 비스듬히 세워, 머리가 패널 높이에 놓여야 한다는 고정 자세를 위반했다. / [gpt-high] 지정되거나 참고에 없던 머리 받침 쿠션을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh8_sel.png",
    "asset_id": "c5910d68-60ae-4983-9d76-bf47827a68ec",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1241442>",
    "asset_id": "d8a82347-dda6-4d03-8e8b-568b33ae1411",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a71-0ae1-7cef-8ff3-0ceec26a7d66",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh8"
  },
  "staged_characters_added": [
   "C03",
   "C05",
   "C08"
  ]
 },
 "S39sh2::signage": {
  "fp": "c7233fb5ebed8967",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::flood_drift_field": {
  "input_fingerprint": "e2bb26e23f6388c6",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "flood_drift_field",
    "tags": [
     "S39sh2"
    ]
   },
   "context_sig": "a392f4891619846b"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On floating boards and rubbish amid the submerged refugee settlement in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 수몰된 난민촌의 풍경. 페드로와 구도환 등도 둥둥 떠다니고....\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On floating boards and rubbish amid the submerged refugee settlement in daylight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 수몰된 난민촌의 풍경. 페드로와 구도환 등도 둥둥 떠다니고....\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_drift_field_16f568.png",
  "asset_id": "345b088f-496d-415c-b40f-716d15d292b5",
  "input_asset_ids": [
   "eb4c2977-bc97-4497-bb58-0924d27f762c"
  ],
  "origin_tag": "S39sh2",
  "place_text": "On floating boards and rubbish amid the submerged refugee settlement in daylight.",
  "origin_inputs": {
   "place_text": "On floating boards and rubbish amid the submerged refugee settlement in daylight.",
   "time_of_day_en": "day",
   "conti_asset_id": "eb4c2977-bc97-4497-bb58-0924d27f762c"
  }
 },
 "S39sh2::bgfirst_bg": {
  "input_fingerprint": "2da1e0e92ae4aa1e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 판자와 쓰레기 더미 위에서 페드로와 구도환이 웅크린 채 표류하는 구도.\n\nLOCATION (lock): On floating boards and rubbish amid the submerged refugee settlement in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the descending approach, retain the broadcast drone's elevated three-quarter view from one side, with 페드로 and 구도환 fully visible near the center and surrounding water open on every side. 페드로 crouches more tightly on the left while 구도환 shifts his weight lower on the right, both looking down toward their precarious support rather than repeating an identical pose. Present the mediated aerial image cleanly, without invented interface graphics, as the descent settles before the lateral move toward the rescue area.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Drifting boards and rubbish (Floating beneath 페드로 and 구도환) — The supporting upper surfaces are seen diagonally from above; used as Forms a compact, irregular support beneath the two figures, leaving most of the frame to the water; Floodwater (Covering the refugee-camp area and surrounding the drifting pair); used as Makes the absence of secure ground immediately readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight exposure and controlled contrast retain the observational severity of the broadcast image without stylized surveillance tinting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 판자와 쓰레기 더미 위에서 페드로와 구도환이 웅크린 채 표류하는 구도.\n\nLOCATION (lock): On floating boards and rubbish amid the submerged refugee settlement in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the descending approach, retain the broadcast drone's elevated three-quarter view from one side, with 페드로 and 구도환 fully visible near the center and surrounding water open on every side. 페드로 crouches more tightly on the left while 구도환 shifts his weight lower on the right, both looking down toward their precarious support rather than repeating an identical pose. Present the mediated aerial image cleanly, without invented interface graphics, as the descent settles before the lateral move toward the rescue area.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Drifting boards and rubbish (Floating beneath 페드로 and 구도환) — The supporting upper surfaces are seen diagonally from above; used as Forms a compact, irregular support beneath the two figures, leaving most of the frame to the water; Floodwater (Covering the refugee-camp area and surrounding the drifting pair); used as Makes the absence of secure ground immediately readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight exposure and controlled contrast retain the observational severity of the broadcast image without stylized surveillance tinting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh2__bgfirst_bg.png",
  "asset_id": "6680de97-7641-423d-940d-c3ee3dc45066",
  "input_asset_ids": [
   "eb4c2977-bc97-4497-bb58-0924d27f762c",
   "345b088f-496d-415c-b40f-716d15d292b5"
  ]
 },
 "S39sh2": {
  "input_fingerprint": "e297ef7d759e6065",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 판자와 쓰레기 더미 위에서 페드로와 구도환이 웅크린 채 표류하는 구도.\n\nLOCATION (lock): On floating boards and rubbish amid the submerged refugee settlement in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the descending approach, retain the broadcast drone's elevated three-quarter view from one side, with 페드로 and 구도환 fully visible near the center and surrounding water open on every side. 페드로 crouches more tightly on the left while 구도환 shifts his weight lower on the right, both looking down toward their precarious support rather than repeating an identical pose. Present the mediated aerial image cleanly, without invented interface graphics, as the descent settles before the lateral move toward the rescue area.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Drifting boards and rubbish (Floating beneath 페드로 and 구도환) — The supporting upper surfaces are seen diagonally from above; used as Forms a compact, irregular support beneath the two figures, leaving most of the frame to the water; Floodwater (Covering the refugee-camp area and surrounding the drifting pair); used as Makes the absence of secure ground immediately readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight exposure and controlled contrast retain the observational severity of the broadcast image without stylized surveillance tinting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In daylight, Sector 7 remains completely submerged, with containers and debris floating across the former settlement. 페드로: He is afloat in the flooded settlement. 구도환: He is afloat in the flooded settlement; no active swimming or successful rescue is established here.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 판자와 쓰레기 더미 위에서 페드로와 구도환이 웅크린 채 표류하는 구도.\n\nLOCATION (lock): On floating boards and rubbish amid the submerged refugee settlement in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the descending approach, retain the broadcast drone's elevated three-quarter view from one side, with 페드로 and 구도환 fully visible near the center and surrounding water open on every side. 페드로 crouches more tightly on the left while 구도환 shifts his weight lower on the right, both looking down toward their precarious support rather than repeating an identical pose. Present the mediated aerial image cleanly, without invented interface graphics, as the descent settles before the lateral move toward the rescue area.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Drifting boards and rubbish (Floating beneath 페드로 and 구도환) — The supporting upper surfaces are seen diagonally from above; used as Forms a compact, irregular support beneath the two figures, leaving most of the frame to the water; Floodwater (Covering the refugee-camp area and surrounding the drifting pair); used as Makes the absence of secure ground immediately readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight exposure and controlled contrast retain the observational severity of the broadcast image without stylized surveillance tinting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In daylight, Sector 7 remains completely submerged, with containers and debris floating across the former settlement. 페드로: He is afloat in the flooded settlement. 구도환: He is afloat in the flooded settlement; no active swimming or successful rescue is established here.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 판자와 쓰레기 더미 위에서 페드로와 구도환이 웅크린 채 표류하는 구도.\n\nLOCATION (lock): On floating boards and rubbish amid the submerged refugee settlement in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the descending approach, retain the broadcast drone's elevated three-quarter view from one side, with 페드로 and 구도환 fully visible near the center and surrounding water open on every side. 페드로 crouches more tightly on the left while 구도환 shifts his weight lower on the right, both looking down toward their precarious support rather than repeating an identical pose. Present the mediated aerial image cleanly, without invented interface graphics, as the descent settles before the lateral move toward the rescue area.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Drifting boards and rubbish (Floating beneath 페드로 and 구도환) — The supporting upper surfaces are seen diagonally from above; used as Forms a compact, irregular support beneath the two figures, leaving most of the frame to the water; Floodwater (Covering the refugee-camp area and surrounding the drifting pair); used as Makes the absence of secure ground immediately readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight exposure and controlled contrast retain the observational severity of the broadcast image without stylized surveillance tinting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In daylight, Sector 7 remains completely submerged, with containers and debris floating across the former settlement. 페드로: He is afloat in the flooded settlement. 구도환: He is afloat in the flooded settlement; no active swimming or successful rescue is established here.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리) — wearing: 도둑질을 하며 입는 낡고 활동적인 집업 자켓과 통이 넓은 바지.; 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — wearing: 물이 빠지고 해진 갈색의 오래된 점퍼와 주름진 면바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh2__bgfirst_bg.png",
     "asset_id": "6680de97-7641-423d-940d-c3ee3dc45066",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S39sh2.png",
     "asset_id": "eb4c2977-bc97-4497-bb58-0924d27f762c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_drift_field_16f568.png",
     "asset_id": "345b088f-496d-415c-b40f-716d15d292b5",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761650>",
     "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:890748>",
     "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "페드로와 구도환 모두 시선을 아래로 향해 자신들이 딛고 있는 위태로운 잔해 더미를 바라보고 있음.",
    "built_space": "넓은 수면 중앙에 나무 판자와 쓰레기가 불규칙하게 뭉쳐진 부유물이 있으며, 카메라 앵글은 위에서 비스듬히 내려다보는 드론 뷰를 정확히 따름.",
    "entities": "왼쪽의 페드로는 비니와 녹색 자켓을 입고 팔을 모아 단단히 웅크리고 있으며, 오른쪽의 구도환은 갈색 자켓을 입고 한쪽 무릎을 낮춰 무게중심을 내린 모습으로 두 인물 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 사람 모두 물에 뜬 부유물 위에 손과 발을 딛고 자연스럽게 체중을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 시선이 아래쪽 잔해 더미를 향하고 있음.",
    "built_space": "수면 위로 잔해와 쓰레기가 뭉친 부유물이 위치하며, 프롬프트에서 요구한 높은 각도의 쿼터 뷰를 잘 보여줌.",
    "entities": "페드로(왼쪽)와 구도환(오른쪽)의 외형 및 복장 레퍼런스 일치도는 좋으나, 두 사람이 부유물을 짚고 있는 웅크린 자세가 서로 비슷하게 연출됨.",
    "hard_violations": [],
    "physics": "인물들의 손과 발이 부유물 위에 닿아 몸을 잘 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "페드로가 더 단단히 웅크리고 구도환이 무게 중심을 낮춘 비대칭적 자세 지시를 매우 충실하게 구현했으며, 레퍼런스와 앵글이 완벽히 조화됨."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "캐릭터와 배경의 기본적 묘사는 훌륭하나, 두 인물의 자세가 프롬프트가 요구한 차별화된 묘사(웅크림과 무게중심 낮춤)에 비해 다소 비슷하게 연출됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "페드로와 구도환 모두 시선을 아래로 향해 자신들이 딛고 있는 위태로운 잔해 더미를 바라보고 있음.",
        "built_space": "넓은 수면 중앙에 나무 판자와 쓰레기가 불규칙하게 뭉쳐진 부유물이 있으며, 카메라 앵글은 위에서 비스듬히 내려다보는 드론 뷰를 정확히 따름.",
        "entities": "왼쪽의 페드로는 비니와 녹색 자켓을 입고 팔을 모아 단단히 웅크리고 있으며, 오른쪽의 구도환은 갈색 자켓을 입고 한쪽 무릎을 낮춰 무게중심을 내린 모습으로 두 인물 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 물에 뜬 부유물 위에 손과 발을 딛고 자연스럽게 체중을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 아래쪽 잔해 더미를 향하고 있음.",
        "built_space": "수면 위로 잔해와 쓰레기가 뭉친 부유물이 위치하며, 프롬프트에서 요구한 높은 각도의 쿼터 뷰를 잘 보여줌.",
        "entities": "페드로(왼쪽)와 구도환(오른쪽)의 외형 및 복장 레퍼런스 일치도는 좋으나, 두 사람이 부유물을 짚고 있는 웅크린 자세가 서로 비슷하게 연출됨.",
        "hard_violations": [],
        "physics": "인물들의 손과 발이 부유물 위에 닿아 몸을 잘 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "페드로가 더 단단히 웅크리고 구도환이 무게 중심을 낮춘 비대칭적 자세 지시를 매우 충실하게 구현했으며, 레퍼런스와 앵글이 완벽히 조화됨."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "캐릭터와 배경의 기본적 묘사는 훌륭하나, 두 인물의 자세가 프롬프트가 요구한 차별화된 묘사(웅크림과 무게중심 낮춤)에 비해 다소 비슷하게 연출됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "페드로와 구도환 모두 시선을 아래로 향해 자신들이 딛고 있는 위태로운 잔해 더미를 바라보고 있음.",
        "built_space": "넓은 수면 중앙에 나무 판자와 쓰레기가 불규칙하게 뭉쳐진 부유물이 있으며, 카메라 앵글은 위에서 비스듬히 내려다보는 드론 뷰를 정확히 따름.",
        "entities": "왼쪽의 페드로는 비니와 녹색 자켓을 입고 팔을 모아 단단히 웅크리고 있으며, 오른쪽의 구도환은 갈색 자켓을 입고 한쪽 무릎을 낮춰 무게중심을 내린 모습으로 두 인물 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 사람 모두 물에 뜬 부유물 위에 손과 발을 딛고 자연스럽게 체중을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 아래쪽 잔해 더미를 향하고 있음.",
        "built_space": "수면 위로 잔해와 쓰레기가 뭉친 부유물이 위치하며, 프롬프트에서 요구한 높은 각도의 쿼터 뷰를 잘 보여줌.",
        "entities": "페드로(왼쪽)와 구도환(오른쪽)의 외형 및 복장 레퍼런스 일치도는 좋으나, 두 사람이 부유물을 짚고 있는 웅크린 자세가 서로 비슷하게 연출됨.",
        "hard_violations": [],
        "physics": "인물들의 손과 발이 부유물 위에 닿아 몸을 잘 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "더 높은 사선 시점과 작은 인물 크기가 방송 드론의 와이드숏에 더 충실하며, 좌우 배치와 발밑을 보는 시선도 맞지만 두 웅크린 자세의 차이는 다소 약하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "페드로의 단단히 접힌 자세와 구도환의 비대칭 하중 이동은 더 명확하지만, 인물이 커지고 시점이 낮아져 지정된 드론 와이드 구도에서는 A보다 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 페드로는 고개를 숙여 무릎 앞 판자를 보고, 오른쪽 구도환도 두 손 사이의 지지면을 내려다본다. 둘 다 카메라나 먼 구조물을 보지 않는다. 표류 방향을 특정할 단서는 없으며 수영이나 구조 동작은 없다.",
        "built_space": "두 사람은 화면 중앙 부근의 한 덩어리로 모인 판자·폐기물 위에 있고 사방에 물이 보인다. 가까운 주요 구조물은 왼쪽 가장자리의 파란 지붕 한 채, 오른쪽 가장자리의 밝은 지붕 한 채, 중경 중앙과 오른쪽 위의 상자형 구조물 각 하나, 중앙 오른쪽의 기울어진 어두운 지붕 하나다. 전경에는 누운 파란 드럼통 하나와 오른쪽 물 위의 선 드럼통 하나가 있어 참조 장소의 구성을 유지한다. 판자 윗면을 비스듬히 내려다보며, B보다 인물이 작고 내려다보는 각도가 크다. 다만 잔해 더미는 소형 지지대라기보다 전경에 넓게 펼쳐져 있다.",
        "entities": "사람은 남성 두 명뿐이다. 왼쪽은 젊은 얼굴에 검은 비니, 낡은 녹색 집업과 짙은 넓은 바지를 착용해 페드로 참조와 부합한다. 작은 얼굴과 숙인 고개 때문에 정확한 얼굴 동일성과 라틴계 혼혈 인상은 확정하기 어렵다. 오른쪽은 짧은 검은 머리의 중년 동아시아계 남성으로 갈색 점퍼와 올리브색 바지가 구도환 참조에 부합한다. 판자, 흰 폐패널, 천 뭉치, 드럼통, 부유 쓰레기와 침수된 정착지가 보인다. 자막이나 인터페이스는 없다.",
        "hard_violations": [],
        "physics": "페드로는 접은 두 다리와 판자 위 신발로 몸을 지탱하고 한 손을 지지면 가까이 둔다. 구도환은 굽힌 다리와 판자에 댄 양손으로 하중을 분산한다. 두 자세 모두 실제 쪼그려 앉기로 가능하며, 구도환은 팔을 더 벌려 버틴다. 겹친 목재와 통·폐자재는 수면에 잠겨 부력을 받는 것으로 읽히고, 사람이나 물체가 근거 없이 공중에 떠 있지는 않다."
       },
       {
        "label": "B",
        "direction": "페드로는 끌어안은 무릎과 신발 앞 판자 쪽으로 시선을 내리고, 구도환은 앞으로 짚은 손과 그 아래 폐패널 쪽을 본다. 두 시선 모두 불안정한 발판을 향한다. 구도환의 몸은 앞쪽으로 기울지만 물에 뛰어들거나 구조받는 행동은 아니다.",
        "built_space": "중앙 잔해 더미 위에서 페드로가 왼쪽, 구도환이 오른쪽을 차지하며 주변 네 방향에 물이 남는다. 왼쪽 파란 지붕 한 채, 오른쪽 가장자리의 밝은 지붕 한 채, 중경의 상자형 구조물 두 개와 기울어진 어두운 지붕 하나가 참조와 대응한다. 전경의 누운 파란 드럼통 하나와 오른쪽 수면의 선 드럼통 하나도 유지된다. 지지면 윗부분은 보이지만 A보다 인물이 크게 보이고 시점이 낮아, 높은 드론 사선 관찰보다 가까운 고각 촬영에 가깝다. 잔해 더미 역시 전경에 넓게 놓인다.",
        "entities": "추가 인물 없이 남성 두 명만 보인다. 왼쪽의 젊은 남성은 검은 비니, 녹색 집업, 짙은 넓은 바지로 페드로 참조에 부합하며, 얼굴은 숙여져 세부 동일성이나 혼혈 배경을 단정하기 어렵다. 오른쪽은 짧은 검은 머리의 중년 동아시아계 남성이고, 갈색 점퍼와 주름진 올리브색 바지가 구도환 참조와 맞는다. 목재, 흰 폐패널, 쓰레기, 천 뭉치, 드럼통과 침수 건축물이 실물 재질로 보인다. 화면 위에 추가된 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "페드로는 두 발을 판자에 놓고 무릎을 몸에 바짝 당겨 웅크린다. 구도환은 한쪽 다리를 더 낮게 접고 양손을 판자와 폐패널에 짚어 비대칭으로 체중을 옮긴다. 손·발·접힌 다리의 지지 관계가 보이며 두 사람의 자세 차이가 명확하다. 잔해는 수면과 접촉하고 목재와 통이 부력을 제공하는 것으로 읽힌다. 지지 없이 공중에 떠 있는 신체나 소품은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "더 높은 사선 시점과 작은 인물 크기가 방송 드론의 와이드숏에 더 충실하며, 좌우 배치와 발밑을 보는 시선도 맞지만 두 웅크린 자세의 차이는 다소 약하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "페드로의 단단히 접힌 자세와 구도환의 비대칭 하중 이동은 더 명확하지만, 인물이 커지고 시점이 낮아져 지정된 드론 와이드 구도에서는 A보다 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 페드로는 고개를 숙여 무릎 앞 판자를 보고, 오른쪽 구도환도 두 손 사이의 지지면을 내려다본다. 둘 다 카메라나 먼 구조물을 보지 않는다. 표류 방향을 특정할 단서는 없으며 수영이나 구조 동작은 없다.",
        "built_space": "두 사람은 화면 중앙 부근의 한 덩어리로 모인 판자·폐기물 위에 있고 사방에 물이 보인다. 가까운 주요 구조물은 왼쪽 가장자리의 파란 지붕 한 채, 오른쪽 가장자리의 밝은 지붕 한 채, 중경 중앙과 오른쪽 위의 상자형 구조물 각 하나, 중앙 오른쪽의 기울어진 어두운 지붕 하나다. 전경에는 누운 파란 드럼통 하나와 오른쪽 물 위의 선 드럼통 하나가 있어 참조 장소의 구성을 유지한다. 판자 윗면을 비스듬히 내려다보며, B보다 인물이 작고 내려다보는 각도가 크다. 다만 잔해 더미는 소형 지지대라기보다 전경에 넓게 펼쳐져 있다.",
        "entities": "사람은 남성 두 명뿐이다. 왼쪽은 젊은 얼굴에 검은 비니, 낡은 녹색 집업과 짙은 넓은 바지를 착용해 페드로 참조와 부합한다. 작은 얼굴과 숙인 고개 때문에 정확한 얼굴 동일성과 라틴계 혼혈 인상은 확정하기 어렵다. 오른쪽은 짧은 검은 머리의 중년 동아시아계 남성으로 갈색 점퍼와 올리브색 바지가 구도환 참조에 부합한다. 판자, 흰 폐패널, 천 뭉치, 드럼통, 부유 쓰레기와 침수된 정착지가 보인다. 자막이나 인터페이스는 없다.",
        "hard_violations": [],
        "physics": "페드로는 접은 두 다리와 판자 위 신발로 몸을 지탱하고 한 손을 지지면 가까이 둔다. 구도환은 굽힌 다리와 판자에 댄 양손으로 하중을 분산한다. 두 자세 모두 실제 쪼그려 앉기로 가능하며, 구도환은 팔을 더 벌려 버틴다. 겹친 목재와 통·폐자재는 수면에 잠겨 부력을 받는 것으로 읽히고, 사람이나 물체가 근거 없이 공중에 떠 있지는 않다."
       },
       {
        "label": "A",
        "direction": "페드로는 끌어안은 무릎과 신발 앞 판자 쪽으로 시선을 내리고, 구도환은 앞으로 짚은 손과 그 아래 폐패널 쪽을 본다. 두 시선 모두 불안정한 발판을 향한다. 구도환의 몸은 앞쪽으로 기울지만 물에 뛰어들거나 구조받는 행동은 아니다.",
        "built_space": "중앙 잔해 더미 위에서 페드로가 왼쪽, 구도환이 오른쪽을 차지하며 주변 네 방향에 물이 남는다. 왼쪽 파란 지붕 한 채, 오른쪽 가장자리의 밝은 지붕 한 채, 중경의 상자형 구조물 두 개와 기울어진 어두운 지붕 하나가 참조와 대응한다. 전경의 누운 파란 드럼통 하나와 오른쪽 수면의 선 드럼통 하나도 유지된다. 지지면 윗부분은 보이지만 A보다 인물이 크게 보이고 시점이 낮아, 높은 드론 사선 관찰보다 가까운 고각 촬영에 가깝다. 잔해 더미 역시 전경에 넓게 놓인다.",
        "entities": "추가 인물 없이 남성 두 명만 보인다. 왼쪽의 젊은 남성은 검은 비니, 녹색 집업, 짙은 넓은 바지로 페드로 참조에 부합하며, 얼굴은 숙여져 세부 동일성이나 혼혈 배경을 단정하기 어렵다. 오른쪽은 짧은 검은 머리의 중년 동아시아계 남성이고, 갈색 점퍼와 주름진 올리브색 바지가 구도환 참조와 맞는다. 목재, 흰 폐패널, 쓰레기, 천 뭉치, 드럼통과 침수 건축물이 실물 재질로 보인다. 화면 위에 추가된 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "페드로는 두 발을 판자에 놓고 무릎을 몸에 바짝 당겨 웅크린다. 구도환은 한쪽 다리를 더 낮게 접고 양손을 판자와 폐패널에 짚어 비대칭으로 체중을 옮긴다. 손·발·접힌 다리의 지지 관계가 보이며 두 사람의 자세 차이가 명확하다. 잔해는 수면과 접촉하고 목재와 통이 부력을 제공하는 것으로 읽힌다. 지지 없이 공중에 떠 있는 신체나 소품은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "페드로가 더 단단히 웅크리고 구도환이 무게 중심을 낮춘 비대칭적 자세 지시를 매우 충실하게 구현했으며, 레퍼런스와 앵글이 완벽히 조화됨."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "캐릭터와 배경의 기본적 묘사는 훌륭하나, 두 인물의 자세가 프롬프트가 요구한 차별화된 묘사(웅크림과 무게중심 낮춤)에 비해 다소 비슷하게 연출됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_drift_field_16f568.png",
    "asset_id": "345b088f-496d-415c-b40f-716d15d292b5",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761650>",
    "asset_id": "4c6f1bca-de84-4581-ac6e-c94c3b9f99f7",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:890748>",
    "asset_id": "f9a9c996-671b-428a-8a09-674293742f44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a8a-6864-7c61-ad88-31315e607e20",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh2__bgfirst_bg.png",
   "bg_asset_id": "6680de97-7641-423d-940d-c3ee3dc45066",
   "bg_record_key": "S39sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "flood_drift_field",
   "groupbg_asset_id": "345b088f-496d-415c-b40f-716d15d292b5"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S39sh8::signage": {
  "fp": "3f0b9188b93130bb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::flood_boat_search": {
  "input_fingerprint": "b85f42a95f0699e1",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "flood_boat_search",
    "tags": [
     "S39sh8"
    ]
   },
   "context_sig": "3ef567ba79e506a3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 고무보트를 타고 수몰된 난민촌을 가로지르는 민병대원들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n수몰된 인천 난민촌 7구역·표류 잔해 수면: 제방 붕괴 후 바다가 된 난민촌 위로 컨테이너와 시신이 떠다니는 수면부다. (특징: 물 위에 섬처럼 떠 있는 찌그러진 컨테이너 박스 판자들; 복부 출혈로 옷이 붉게 젖은 채 쓰러진 미연; 수면에 떠 있는 사체를 막대기로 뒤적이는 고무보트 위의 민병대원들; 상공에서 내려다보는 뉴스 중계용 드론 렌즈 시야)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 고무보트를 타고 수몰된 난민촌을 가로지르는 민병대원들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_boat_search_f3cb2b.png",
  "asset_id": "101c6aaa-cddf-45ea-a7fa-5510230aee38",
  "input_asset_ids": [
   "d52423f7-faeb-4459-8594-b420b34e12df"
  ],
  "origin_tag": "S39sh8",
  "place_text": "Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.",
  "origin_inputs": {
   "place_text": "Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.",
   "time_of_day_en": "day",
   "conti_asset_id": "d52423f7-faeb-4459-8594-b420b34e12df"
  }
 },
 "S39sh8::bgfirst_bg": {
  "input_fingerprint": "d3bb11ec36ab6529",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 막대기에 뒤집혀 하늘을 향해 드러난 창백한 시체의 얼굴을 빤히 내려다보는 민병대원의 상체.\n\nLOCATION (lock): Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inward move just above the water beside the body's head and outside the boat, looking obliquely upward toward the militia member's three-quarter upper body at upper-right. Keep the pale, upturned face in the lower-left foreground at believable scale, with the boat edge separating it from the member leaning down to inspect it. The member's gaze terminates on the face rather than the lens, and the probing stick remains a narrow diagonal connection between the two planes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubber boat (Carrying the militia member beside the floating body) — Its outer side and a limited portion of the upper edge face the camera; used as Establishes the height difference and separates the inspecting figure from the body; Probing stick (Held after turning the body to expose its face) — Runs obliquely from the member's hand toward the body; used as Links the inspection action across foreground and midground; Floodwater (Supporting the floating body beside the boat); used as Retains the low water-level context around the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled facial contrast preserve the corpse's stated pallor without sensational lighting or a borrowed searchlight effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 막대기에 뒤집혀 하늘을 향해 드러난 창백한 시체의 얼굴을 빤히 내려다보는 민병대원의 상체.\n\nLOCATION (lock): Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inward move just above the water beside the body's head and outside the boat, looking obliquely upward toward the militia member's three-quarter upper body at upper-right. Keep the pale, upturned face in the lower-left foreground at believable scale, with the boat edge separating it from the member leaning down to inspect it. The member's gaze terminates on the face rather than the lens, and the probing stick remains a narrow diagonal connection between the two planes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubber boat (Carrying the militia member beside the floating body) — Its outer side and a limited portion of the upper edge face the camera; used as Establishes the height difference and separates the inspecting figure from the body; Probing stick (Held after turning the body to expose its face) — Runs obliquely from the member's hand toward the body; used as Links the inspection action across foreground and midground; Floodwater (Supporting the floating body beside the boat); used as Retains the low water-level context around the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled facial contrast preserve the corpse's stated pallor without sensational lighting or a borrowed searchlight effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh8__bgfirst_bg.png",
  "asset_id": "7a335ad1-79ab-4053-b87b-bfafecf4ec28",
  "input_asset_ids": [
   "d52423f7-faeb-4459-8594-b420b34e12df",
   "101c6aaa-cddf-45ea-a7fa-5510230aee38"
  ]
 },
 "S39sh8": {
  "input_fingerprint": "1f452fe3b9b838b3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막대기에 뒤집혀 하늘을 향해 드러난 창백한 시체의 얼굴을 빤히 내려다보는 민병대원의 상체.\n\nLOCATION (lock): Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inward move just above the water beside the body's head and outside the boat, looking obliquely upward toward the militia member's three-quarter upper body at upper-right. Keep the pale, upturned face in the lower-left foreground at believable scale, with the boat edge separating it from the member leaning down to inspect it. The member's gaze terminates on the face rather than the lens, and the probing stick remains a narrow diagonal connection between the two planes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubber boat (Carrying the militia member beside the floating body) — Its outer side and a limited portion of the upper edge face the camera; used as Establishes the height difference and separates the inspecting figure from the body; Probing stick (Held after turning the body to expose its face) — Runs obliquely from the member's hand toward the body; used as Links the inspection action across foreground and midground; Floodwater (Supporting the floating body beside the boat); used as Retains the low water-level context around the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled facial contrast preserve the corpse's stated pallor without sensational lighting or a borrowed searchlight effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The corpse floats at the water's surface while a militiaman manipulates it with a pole to inspect its face. Its head is exposed for inspection, but the torso's orientation and the positions of the arms and legs are not specified in the scene text.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged in daylight, with floating debris and a rubber boat crossing the floodwater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막대기에 뒤집혀 하늘을 향해 드러난 창백한 시체의 얼굴을 빤히 내려다보는 민병대원의 상체.\n\nLOCATION (lock): Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inward move just above the water beside the body's head and outside the boat, looking obliquely upward toward the militia member's three-quarter upper body at upper-right. Keep the pale, upturned face in the lower-left foreground at believable scale, with the boat edge separating it from the member leaning down to inspect it. The member's gaze terminates on the face rather than the lens, and the probing stick remains a narrow diagonal connection between the two planes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubber boat (Carrying the militia member beside the floating body) — Its outer side and a limited portion of the upper edge face the camera; used as Establishes the height difference and separates the inspecting figure from the body; Probing stick (Held after turning the body to expose its face) — Runs obliquely from the member's hand toward the body; used as Links the inspection action across foreground and midground; Floodwater (Supporting the floating body beside the boat); used as Retains the low water-level context around the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled facial contrast preserve the corpse's stated pallor without sensational lighting or a borrowed searchlight effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The corpse floats at the water's surface while a militiaman manipulates it with a pole to inspect its face. Its head is exposed for inspection, but the torso's orientation and the positions of the arms and legs are not specified in the scene text.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged in daylight, with floating debris and a rubber boat crossing the floodwater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막대기에 뒤집혀 하늘을 향해 드러난 창백한 시체의 얼굴을 빤히 내려다보는 민병대원의 상체.\n\nLOCATION (lock): Aboard an open rubber boat on the flooded settlement's surface, beside a floating body being inspected. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the inward move just above the water beside the body's head and outside the boat, looking obliquely upward toward the militia member's three-quarter upper body at upper-right. Keep the pale, upturned face in the lower-left foreground at believable scale, with the boat edge separating it from the member leaning down to inspect it. The member's gaze terminates on the face rather than the lens, and the probing stick remains a narrow diagonal connection between the two planes.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rubber boat (Carrying the militia member beside the floating body) — Its outer side and a limited portion of the upper edge face the camera; used as Establishes the height difference and separates the inspecting figure from the body; Probing stick (Held after turning the body to expose its face) — Runs obliquely from the member's hand toward the body; used as Links the inspection action across foreground and midground; Floodwater (Supporting the floating body beside the boat); used as Retains the low water-level context around the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled facial contrast preserve the corpse's stated pallor without sensational lighting or a borrowed searchlight effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The corpse floats at the water's surface while a militiaman manipulates it with a pole to inspect its face. Its head is exposed for inspection, but the torso's orientation and the positions of the arms and legs are not specified in the scene text.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The settlement remains submerged in daylight, with floating debris and a rubber boat crossing the floodwater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 페드로 (라틴계 혼혈 남성, 10대 후반의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh8__bgfirst_bg.png",
     "asset_id": "7a335ad1-79ab-4053-b87b-bfafecf4ec28",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S39sh8.png",
     "asset_id": "d52423f7-faeb-4459-8594-b420b34e12df",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1302308>",
     "asset_id": "3e3943a9-b3e5-4e2c-ac87-2e35c5935244",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1278860>",
     "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_boat_search_f3cb2b.png",
     "asset_id": "101c6aaa-cddf-45ea-a7fa-5510230aee38",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1302308>",
     "asset_id": "3e3943a9-b3e5-4e2c-ac87-2e35c5935244",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1278860>",
     "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "민병대원의 시선이 시체의 얼굴을 정확히 향하고 있으며, 막대기도 시체 쪽을 가리킵니다.",
    "built_space": "수면 위 카메라 앵글과 보트를 사이에 둔 전경과 후경의 공간 배치가 요구사항과 일치합니다.",
    "entities": "구도환과 페드로의 얼굴 특징, 고무보트, 막대기가 모두 정확하게 묘사되었습니다.",
    "hard_violations": [],
    "physics": "시체는 물의 부력을 받아 자연스럽게 떠 있고, 인물은 보트에 체중을 싣고 막대기를 안정적으로 쥐고 있습니다."
   },
   {
    "label": "B",
    "direction": "민병대원의 시선이 시체의 얼굴이 아닌 허공을 향하고 있으며, 막대기는 시체의 목 부위를 향합니다.",
    "built_space": "카메라 앵글과 보트의 위치는 적절하나, 시체가 화면에 자리 잡은 비율과 위치가 다소 어색합니다.",
    "entities": "구도환의 외모는 일치하나, 시체의 얼굴 형태가 뭉개져 레퍼런스 인물을 알아보기 어렵습니다.",
    "hard_violations": [],
    "physics": "시체가 물에 떠 있고 인물이 막대기를 쥐고 있으나, 수면 위로 드러난 시체의 목과 머리 각도가 부자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 빤히 내려다보는 시선, 정확한 카메라 앵글, 그리고 두 인물의 레퍼런스를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "민병대원의 시선이 시체를 향하지 않으며, 시체의 얼굴 형태가 레퍼런스와 다르게 왜곡되어 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "민병대원의 시선이 시체의 얼굴을 정확히 향하고 있으며, 막대기도 시체 쪽을 가리킵니다.",
        "built_space": "수면 위 카메라 앵글과 보트를 사이에 둔 전경과 후경의 공간 배치가 요구사항과 일치합니다.",
        "entities": "구도환과 페드로의 얼굴 특징, 고무보트, 막대기가 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "시체는 물의 부력을 받아 자연스럽게 떠 있고, 인물은 보트에 체중을 싣고 막대기를 안정적으로 쥐고 있습니다."
       },
       {
        "label": "B",
        "direction": "민병대원의 시선이 시체의 얼굴이 아닌 허공을 향하고 있으며, 막대기는 시체의 목 부위를 향합니다.",
        "built_space": "카메라 앵글과 보트의 위치는 적절하나, 시체가 화면에 자리 잡은 비율과 위치가 다소 어색합니다.",
        "entities": "구도환의 외모는 일치하나, 시체의 얼굴 형태가 뭉개져 레퍼런스 인물을 알아보기 어렵습니다.",
        "hard_violations": [],
        "physics": "시체가 물에 떠 있고 인물이 막대기를 쥐고 있으나, 수면 위로 드러난 시체의 목과 머리 각도가 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 빤히 내려다보는 시선, 정확한 카메라 앵글, 그리고 두 인물의 레퍼런스를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "민병대원의 시선이 시체를 향하지 않으며, 시체의 얼굴 형태가 레퍼런스와 다르게 왜곡되어 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "민병대원의 시선이 시체의 얼굴을 정확히 향하고 있으며, 막대기도 시체 쪽을 가리킵니다.",
        "built_space": "수면 위 카메라 앵글과 보트를 사이에 둔 전경과 후경의 공간 배치가 요구사항과 일치합니다.",
        "entities": "구도환과 페드로의 얼굴 특징, 고무보트, 막대기가 모두 정확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "시체는 물의 부력을 받아 자연스럽게 떠 있고, 인물은 보트에 체중을 싣고 막대기를 안정적으로 쥐고 있습니다."
       },
       {
        "label": "B",
        "direction": "민병대원의 시선이 시체의 얼굴이 아닌 허공을 향하고 있으며, 막대기는 시체의 목 부위를 향합니다.",
        "built_space": "카메라 앵글과 보트의 위치는 적절하나, 시체가 화면에 자리 잡은 비율과 위치가 다소 어색합니다.",
        "entities": "구도환의 외모는 일치하나, 시체의 얼굴 형태가 뭉개져 레퍼런스 인물을 알아보기 어렵습니다.",
        "hard_violations": [],
        "physics": "시체가 물에 떠 있고 인물이 막대기를 쥐고 있으나, 수면 위로 드러난 시체의 목과 머리 각도가 부자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "수면 바로 위의 올려다보는 구도는 정확하지만, 민병대원의 시선이 시체 얼굴보다 카메라 쪽으로 향하고 시체의 얼굴·머리도 페드로 참조와 차이가 크다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴을 내려다보는 시선과 막대기의 연결, 보트 안팎의 배치 및 인물 식별이 더 충실하지만, 올려다보는 각도와 시체의 창백함은 다소 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "민병대원은 몸을 앞으로 기울였지만 눈은 비교적 정면인 카메라 부근을 향해 있어, 왼쪽 아래 시체 얼굴에 시선이 종착한다고 보기 어렵다. 장갑 낀 손의 막대기는 왼쪽 아래로 뻗어 시체의 턱·목 옆 수면에 닿는다. 시체는 눈을 감고 얼굴을 하늘로 향한다.",
        "built_space": "오른쪽에 고무보트 한 척의 외측 튜브와 제한된 상단이 보이고, 가까운 측면에는 밧줄이 통과하는 고정구 하나가 있다. 민병대원의 상체는 보트 가장자리 뒤 오른쪽 위에, 시체 얼굴은 보트 밖 왼쪽 아래에 있다. 카메라는 시체 머리 곁 수면에서 위를 올려다보며, 큰 전경 얼굴과 보트 측면이 깊이를 만든다. 배경의 침수 컨테이너, 전신주와 고가 구조물은 장소 참조의 주요 요소에 부합한다.",
        "entities": "중년 동아시아계로 보이는 남성 한 명과 시체 한 구가 보인다. 민병대원은 구도환 참조와 유사한 짧은 검은 머리와 남색 셔츠를 갖췄으나 전술조끼·무전기·장갑이 추가되어 있다. 시체는 창백한 피부와 감긴 눈을 갖지만 젖은 머리가 더 길고 얼굴 윤곽도 페드로 참조와 다르게 읽힌다. 고무보트, 탐침 막대기, 홍수와 부유 잔해가 보인다. 낮빛은 요청한 절제된 조명보다 다소 선명하다.",
        "hard_violations": [],
        "physics": "시체의 뒷머리와 목 아래가 수면에 잠겨 물의 부력으로 지지되며, 노출된 얼굴에 능동적으로 힘을 주는 자세는 보이지 않는다. 막대기는 장갑 낀 손이 확실히 쥐고 있다. 민병대원은 보트 안에서 가장자리 위로 상체를 내민 상태이고 하체 지지는 가려져 있다. 보트와 잔해는 물에 떠 있으며, 지지 없이 공중에 뜬 물체는 없다."
       },
       {
        "label": "B",
        "direction": "민병대원의 고개와 눈이 왼쪽 아래 시체 얼굴을 향해 내려가 있어 검사 대상이 명확하다. 손에 쥔 막대기는 같은 방향의 가는 대각선으로 뻗어 시체 목 옆 물속에 들어간다. 시체는 눈을 감고 얼굴을 위로 향한다.",
        "built_space": "고무보트 한 척이 오른쪽에 있고, 외측 튜브의 밧줄 고정구 하나와 상단 일부가 보인다. 민병대원은 보트 안쪽 오른쪽 위에서 몸을 숙이고, 시체는 가장자리 바깥 왼쪽 아래에 떠 있어 보트가 두 인물을 공간적으로 구분한다. 침수된 컨테이너 군, 왼쪽 전신주 열, 중앙 지붕과 오른쪽 고가 구조물은 장소 참조와 가깝다. 수면 가까운 시점이지만 A보다 올려다보는 각도가 약하고 시체의 상흉부까지 더 보인다.",
        "entities": "검사하는 중년 남성 한 명과 젊은 남성 시체 한 구만 보인다. 민병대원의 얼굴, 나이 인상, 짧은 검은 머리와 남색 셔츠는 구도환 참조에 가깝고 전술조끼가 추가되어 있다. 시체는 페드로 참조의 젊은 얼굴, 짧은 짙은 머리, 옅은 수염과 남색 상의에 더 가깝지만 머리의 곱슬기가 강하고 피부의 창백함은 약하다. 고무보트, 나무 막대기, 홍수와 부유 잔해가 모두 식별된다.",
        "hard_violations": [],
        "physics": "시체의 뒷머리와 몸통이 물에 잠긴 채 부력으로 받쳐지고, 드러난 목과 얼굴은 누운 몸에 자연스럽게 이어진다. 들어 올린 팔다리나 근육으로 유지하는 자세는 없다. 막대기는 맨손으로 확실히 잡혀 있고 끝은 물속에 잠긴다. 민병대원은 보트 가장자리에 상체를 가까이 대고 숙여 무게를 지지하는 자세이며, 보트와 잔해도 수면의 지지를 받는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "수면 바로 위의 올려다보는 구도는 정확하지만, 민병대원의 시선이 시체 얼굴보다 카메라 쪽으로 향하고 시체의 얼굴·머리도 페드로 참조와 차이가 크다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴을 내려다보는 시선과 막대기의 연결, 보트 안팎의 배치 및 인물 식별이 더 충실하지만, 올려다보는 각도와 시체의 창백함은 다소 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "민병대원은 몸을 앞으로 기울였지만 눈은 비교적 정면인 카메라 부근을 향해 있어, 왼쪽 아래 시체 얼굴에 시선이 종착한다고 보기 어렵다. 장갑 낀 손의 막대기는 왼쪽 아래로 뻗어 시체의 턱·목 옆 수면에 닿는다. 시체는 눈을 감고 얼굴을 하늘로 향한다.",
        "built_space": "오른쪽에 고무보트 한 척의 외측 튜브와 제한된 상단이 보이고, 가까운 측면에는 밧줄이 통과하는 고정구 하나가 있다. 민병대원의 상체는 보트 가장자리 뒤 오른쪽 위에, 시체 얼굴은 보트 밖 왼쪽 아래에 있다. 카메라는 시체 머리 곁 수면에서 위를 올려다보며, 큰 전경 얼굴과 보트 측면이 깊이를 만든다. 배경의 침수 컨테이너, 전신주와 고가 구조물은 장소 참조의 주요 요소에 부합한다.",
        "entities": "중년 동아시아계로 보이는 남성 한 명과 시체 한 구가 보인다. 민병대원은 구도환 참조와 유사한 짧은 검은 머리와 남색 셔츠를 갖췄으나 전술조끼·무전기·장갑이 추가되어 있다. 시체는 창백한 피부와 감긴 눈을 갖지만 젖은 머리가 더 길고 얼굴 윤곽도 페드로 참조와 다르게 읽힌다. 고무보트, 탐침 막대기, 홍수와 부유 잔해가 보인다. 낮빛은 요청한 절제된 조명보다 다소 선명하다.",
        "hard_violations": [],
        "physics": "시체의 뒷머리와 목 아래가 수면에 잠겨 물의 부력으로 지지되며, 노출된 얼굴에 능동적으로 힘을 주는 자세는 보이지 않는다. 막대기는 장갑 낀 손이 확실히 쥐고 있다. 민병대원은 보트 안에서 가장자리 위로 상체를 내민 상태이고 하체 지지는 가려져 있다. 보트와 잔해는 물에 떠 있으며, 지지 없이 공중에 뜬 물체는 없다."
       },
       {
        "label": "A",
        "direction": "민병대원의 고개와 눈이 왼쪽 아래 시체 얼굴을 향해 내려가 있어 검사 대상이 명확하다. 손에 쥔 막대기는 같은 방향의 가는 대각선으로 뻗어 시체 목 옆 물속에 들어간다. 시체는 눈을 감고 얼굴을 위로 향한다.",
        "built_space": "고무보트 한 척이 오른쪽에 있고, 외측 튜브의 밧줄 고정구 하나와 상단 일부가 보인다. 민병대원은 보트 안쪽 오른쪽 위에서 몸을 숙이고, 시체는 가장자리 바깥 왼쪽 아래에 떠 있어 보트가 두 인물을 공간적으로 구분한다. 침수된 컨테이너 군, 왼쪽 전신주 열, 중앙 지붕과 오른쪽 고가 구조물은 장소 참조와 가깝다. 수면 가까운 시점이지만 A보다 올려다보는 각도가 약하고 시체의 상흉부까지 더 보인다.",
        "entities": "검사하는 중년 남성 한 명과 젊은 남성 시체 한 구만 보인다. 민병대원의 얼굴, 나이 인상, 짧은 검은 머리와 남색 셔츠는 구도환 참조에 가깝고 전술조끼가 추가되어 있다. 시체는 페드로 참조의 젊은 얼굴, 짧은 짙은 머리, 옅은 수염과 남색 상의에 더 가깝지만 머리의 곱슬기가 강하고 피부의 창백함은 약하다. 고무보트, 나무 막대기, 홍수와 부유 잔해가 모두 식별된다.",
        "hard_violations": [],
        "physics": "시체의 뒷머리와 몸통이 물에 잠긴 채 부력으로 받쳐지고, 드러난 목과 얼굴은 누운 몸에 자연스럽게 이어진다. 들어 올린 팔다리나 근육으로 유지하는 자세는 없다. 막대기는 맨손으로 확실히 잡혀 있고 끝은 물속에 잠긴다. 민병대원은 보트 가장자리에 상체를 가까이 대고 숙여 무게를 지지하는 자세이며, 보트와 잔해도 수면의 지지를 받는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.321
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.321
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 빤히 내려다보는 시선, 정확한 카메라 앵글, 그리고 두 인물의 레퍼런스를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "민병대원의 시선이 시체를 향하지 않으며, 시체의 얼굴 형태가 레퍼런스와 다르게 왜곡되어 묘사되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flood_boat_search_f3cb2b.png",
    "asset_id": "101c6aaa-cddf-45ea-a7fa-5510230aee38",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1302308>",
    "asset_id": "3e3943a9-b3e5-4e2c-ac87-2e35c5935244",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 페드로: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1278860>",
    "asset_id": "7489692f-8267-42f4-a12f-be8c5c9013a8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a93-8d95-7571-9f37-19b45589717c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S39sh8__bgfirst_bg.png",
   "bg_asset_id": "7a335ad1-79ab-4053-b87b-bfafecf4ec28",
   "bg_record_key": "S39sh8::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "flood_boat_search",
   "groupbg_asset_id": "101c6aaa-cddf-45ea-a7fa-5510230aee38"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S40sh2::signage": {
  "fp": "5ccf3d49613c90dd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::c3c90e59582f7e94": {
  "subjects": [],
  "subject_text": "인천 난민촌 외곽 도심 거리·무인점포 앞\n주거 밀집지 외곽에서 도심으로 이어지는 거리. 도로 건너편에 무인점포의 정면과 출입문이 보인다.",
  "identity": "canonical",
  "scope_id": "L54",
  "scope_role": "location_exterior",
  "scope_sha": "c1355dabe9787e22"
 },
 "S40sh2::bgfirst_bg": {
  "input_fingerprint": "ae5cdf48d9c7c462",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 제자리에 멈춰 서서 바닷물에 잠긴 난민촌 방향으로 몸을 틀어 뒤돌아보는 앰버의 전신.\n\nLOCATION (lock): On an exposed street at the refugee settlement's outer edge, looking back toward the inundated district.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the tracking camera behind and to one side of 앰버 at below-shoulder height, keeping her full body center-right and preserving the group's established travel axis toward screen-right. Capture her halted weight on the rear foot, torso twisting back toward screen-left, with her tear-covered face partly visible in profile as she looks toward the refugee camp beyond the frame. Leave generous space on that backward-looking side and keep the companions farther along the route outside this framing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp outskirts (The group is passing through toward the city) — The route continues diagonally forward toward screen-right; used as Leaves opposing spaces for onward escape and 앰버's backward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and subdued contrast leave the tears readable without isolating 앰버 with an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 제자리에 멈춰 서서 바닷물에 잠긴 난민촌 방향으로 몸을 틀어 뒤돌아보는 앰버의 전신.\n\nLOCATION (lock): On an exposed street at the refugee settlement's outer edge, looking back toward the inundated district.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the tracking camera behind and to one side of 앰버 at below-shoulder height, keeping her full body center-right and preserving the group's established travel axis toward screen-right. Capture her halted weight on the rear foot, torso twisting back toward screen-left, with her tear-covered face partly visible in profile as she looks toward the refugee camp beyond the frame. Leave generous space on that backward-looking side and keep the companions farther along the route outside this framing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp outskirts (The group is passing through toward the city) — The route continues diagonally forward toward screen-right; used as Leaves opposing spaces for onward escape and 앰버's backward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and subdued contrast leave the tears readable without isolating 앰버 with an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S40sh2__bgfirst_bg.png",
  "asset_id": "64e4b41f-ed27-45ea-ab5e-6e1d35b87c75",
  "input_asset_ids": [
   "331384f4-cbcd-4583-b617-e981c6377aa6",
   "a3eab2e7-9787-4314-b6d8-717f050d2d82"
  ]
 },
 "S40sh2": {
  "input_fingerprint": "28813e99e29f4be1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 멈춰 서서 바닷물에 잠긴 난민촌 방향으로 몸을 틀어 뒤돌아보는 앰버의 전신.\n\nLOCATION (lock): On an exposed street at the refugee settlement's outer edge, looking back toward the inundated district. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the tracking camera behind and to one side of 앰버 at below-shoulder height, keeping her full body center-right and preserving the group's established travel axis toward screen-right. Capture her halted weight on the rear foot, torso twisting back toward screen-left, with her tear-covered face partly visible in profile as she looks toward the refugee camp beyond the frame. Leave generous space on that backward-looking side and keep the companions farther along the route outside this framing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp outskirts (The group is passing through toward the city) — The route continues diagonally forward toward screen-right; used as Leaves opposing spaces for onward escape and 앰버's backward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and subdued contrast leave the tears readable without isolating 앰버 with an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The refugee settlement remains flooded behind the route toward the city in daylight. 앰버: She has stopped from exhaustion, turning back repeatedly with her face covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 멈춰 서서 바닷물에 잠긴 난민촌 방향으로 몸을 틀어 뒤돌아보는 앰버의 전신.\n\nLOCATION (lock): On an exposed street at the refugee settlement's outer edge, looking back toward the inundated district. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the tracking camera behind and to one side of 앰버 at below-shoulder height, keeping her full body center-right and preserving the group's established travel axis toward screen-right. Capture her halted weight on the rear foot, torso twisting back toward screen-left, with her tear-covered face partly visible in profile as she looks toward the refugee camp beyond the frame. Leave generous space on that backward-looking side and keep the companions farther along the route outside this framing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp outskirts (The group is passing through toward the city) — The route continues diagonally forward toward screen-right; used as Leaves opposing spaces for onward escape and 앰버's backward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and subdued contrast leave the tears readable without isolating 앰버 with an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The refugee settlement remains flooded behind the route toward the city in daylight. 앰버: She has stopped from exhaustion, turning back repeatedly with her face covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 제자리에 멈춰 서서 바닷물에 잠긴 난민촌 방향으로 몸을 틀어 뒤돌아보는 앰버의 전신.\n\nLOCATION (lock): On an exposed street at the refugee settlement's outer edge, looking back toward the inundated district. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the tracking camera behind and to one side of 앰버 at below-shoulder height, keeping her full body center-right and preserving the group's established travel axis toward screen-right. Capture her halted weight on the rear foot, torso twisting back toward screen-left, with her tear-covered face partly visible in profile as she looks toward the refugee camp beyond the frame. Leave generous space on that backward-looking side and keep the companions farther along the route outside this framing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Refugee-camp outskirts (The group is passing through toward the city) — The route continues diagonally forward toward screen-right; used as Leaves opposing spaces for onward escape and 앰버's backward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and subdued contrast leave the tears readable without isolating 앰버 with an artificial lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The refugee settlement remains flooded behind the route toward the city in daylight. 앰버: She has stopped from exhaustion, turning back repeatedly with her face covered in tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S40sh2__bgfirst_bg.png",
     "asset_id": "64e4b41f-ed27-45ea-ab5e-6e1d35b87c75",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S40sh2.png",
     "asset_id": "331384f4-cbcd-4583-b617-e981c6377aa6",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L54B01.png",
     "asset_id": "a3eab2e7-9787-4314-b6d8-717f050d2d82",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선은 화면 왼쪽의 물에 잠긴 지역을 향하고 있음.",
    "built_space": "화면 우측으로 대각선으로 뻗은 아스팔트 도로와 좌측에 수몰된 건물들이 배치되어 있으나, 레퍼런스 사진의 상점 건물은 존재하지 않음.",
    "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 지정된 앰버의 특징과 일치함.",
    "hard_violations": [
     "[gpt-high] 정확한 촬영 장소에 없는 연속 수변 방호벽과 침수 주택지를 배치하고 도로 배치를 새로 만들어, 고정된 상점 앞 거리 대신 별개의 수변 도로를 구현했습니다."
    ],
    "physics": "앰버의 두 발이 아스팔트 지면에 닿아 몸을 안정적으로 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "앰버의 시선은 화면 왼쪽을 향하고 있으나 특정한 목표물(수몰 구역)이 보이지 않음.",
    "built_space": "위치 레퍼런스와 동일한 상점 건물, 가로 방향의 차도와 보도블록이 평면적으로 배치되어 있으며 수몰된 구역이 없음.",
    "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 앰버의 지정된 특징과 일치함.",
    "hard_violations": [],
    "physics": "앰버의 두 발이 보도블록 위에 닿아 몸을 안정적으로 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 각도, 대각선 구도, 수몰된 지역을 바라보는 자세 등 프레이밍 지침(Priority 2)을 훌륭하게 구현했으나 위치 레퍼런스의 상점 건축물(Priority 3)이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "위치 레퍼런스의 배경은 일치하나, 지시된 대각선 구도와 인물의 비틀린 자세를 무시하고 레퍼런스 사진의 평면적 구도를 그대로 복사하여 핵심 프레이밍(Priority 2)을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 화면 왼쪽의 물에 잠긴 지역을 향하고 있음.",
        "built_space": "화면 우측으로 대각선으로 뻗은 아스팔트 도로와 좌측에 수몰된 건물들이 배치되어 있으나, 레퍼런스 사진의 상점 건물은 존재하지 않음.",
        "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 지정된 앰버의 특징과 일치함.",
        "hard_violations": [],
        "physics": "앰버의 두 발이 아스팔트 지면에 닿아 몸을 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "앰버의 시선은 화면 왼쪽을 향하고 있으나 특정한 목표물(수몰 구역)이 보이지 않음.",
        "built_space": "위치 레퍼런스와 동일한 상점 건물, 가로 방향의 차도와 보도블록이 평면적으로 배치되어 있으며 수몰된 구역이 없음.",
        "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 앰버의 지정된 특징과 일치함.",
        "hard_violations": [],
        "physics": "앰버의 두 발이 보도블록 위에 닿아 몸을 안정적으로 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 각도, 대각선 구도, 수몰된 지역을 바라보는 자세 등 프레이밍 지침(Priority 2)을 훌륭하게 구현했으나 위치 레퍼런스의 상점 건축물(Priority 3)이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "위치 레퍼런스의 배경은 일치하나, 지시된 대각선 구도와 인물의 비틀린 자세를 무시하고 레퍼런스 사진의 평면적 구도를 그대로 복사하여 핵심 프레이밍(Priority 2)을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 화면 왼쪽의 물에 잠긴 지역을 향하고 있음.",
        "built_space": "화면 우측으로 대각선으로 뻗은 아스팔트 도로와 좌측에 수몰된 건물들이 배치되어 있으나, 레퍼런스 사진의 상점 건물은 존재하지 않음.",
        "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 지정된 앰버의 특징과 일치함.",
        "hard_violations": [],
        "physics": "앰버의 두 발이 아스팔트 지면에 닿아 몸을 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "앰버의 시선은 화면 왼쪽을 향하고 있으나 특정한 목표물(수몰 구역)이 보이지 않음.",
        "built_space": "위치 레퍼런스와 동일한 상점 건물, 가로 방향의 차도와 보도블록이 평면적으로 배치되어 있으며 수몰된 구역이 없음.",
        "entities": "금발, 카키색 작업복, 공구 벨트, 방진 마스크 등 앰버의 지정된 특징과 일치함.",
        "hard_violations": [],
        "physics": "앰버의 두 발이 보도블록 위에 닿아 몸을 안정적으로 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정 장소와 중앙 오른쪽 전신 배치는 충실하지만, 뒤쪽 사선 카메라가 아니라 앞쪽에서 촬영되어 멈춘 채 몸을 비틀어 뒤돌아보는 핵심 동작이 빠졌습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "침수지와 오른쪽으로 이어지는 탈출로는 보이지만, 지정 장소를 새로운 수변 도로로 바꾸었고 뒤돌아보는 동작도 구현하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 왼쪽 바깥을 향하지만 그곳의 난민촌은 보이지 않습니다. 가슴과 골반도 대체로 카메라와 왼쪽을 향하여, 오른쪽으로 이동하다 상체만 왼쪽으로 돌린 관계가 읽히지 않습니다. 도로는 화면을 가로질러 놓여 있어 오른쪽 전방으로 이어지는 이동 축도 약합니다.",
        "built_space": "맞은편에 넓은 상점 진열창 한 면, 그 오른쪽 출입구 한 곳과 작은 경사판 하나, 외벽의 실외기 하나와 전기함 하나가 보입니다. 왼쪽 골목, 전신주, 차도와 앞쪽 점자블록 보도까지 장소 사진의 주요 구성이 유지됩니다. 앰버는 앞쪽 보도에 서 있습니다. 전신은 중앙 오른쪽에 있고 왼쪽 여백도 넓지만, 작업복 앞면이 보이는 카메라 위치는 요구된 뒤쪽 사선 위치가 아닙니다. 침수 상태는 화면에서 확인되지 않습니다.",
        "entities": "사람은 금발의 밝은 피부를 가진 어린 여자아이 한 명뿐이며, 약 10세의 체격과 얼굴은 참조에 대체로 부합합니다. 혼혈 배경 자체는 외관만으로 확정할 수 없습니다. 기름때 묻은 카키 작업복, 가죽 공구 벨트와 파우치, 검은 부츠가 보입니다. 정교한 방진 마스크는 참조처럼 목에 걸려 있습니다. 볼에는 가는 눈물 자국이 보이지만 얼굴 전체가 눈물로 뒤덮인 상태는 약하게 표현되었습니다. 동행인이나 추가 인물은 없습니다.",
        "hard_violations": [],
        "physics": "두 부츠의 밑창이 보도에 닿아 몸을 지지하며 부유하거나 불가능한 관절은 보이지 않습니다. 마스크는 목끈으로, 공구와 파우치는 허리 벨트로 지지됩니다. 다만 발을 비교적 나란히 둔 직립 자세라 이동 중 뒷발에 체중을 실어 멈춘 순간보다 정적인 서 있는 자세에 가깝습니다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 왼쪽 침수 주택지 방향을 향합니다. 도로는 오른쪽 먼 배경으로 이어져 요구한 탈출 방향은 읽힙니다. 그러나 가슴과 골반은 카메라 쪽을 향하고 머리만 왼쪽으로 돌아가 있어, 오른쪽 진행 중 몸을 비틀어 뒤돌아보는 동작은 드러나지 않습니다.",
        "built_space": "앰버는 보도가 아니라 차도 위에 서 있습니다. 왼쪽에는 여러 콘크리트 지주와 금속 난간으로 이어진 수변 방호벽이 있고, 그 너머로 침수된 저층 주택들과 지붕들이 펼쳐집니다. 오른쪽에는 전신주가 줄지어 선 도로가 먼 도시 쪽으로 이어집니다. 지정 장소의 큰 상점 진열창, 오른쪽 출입구, 외벽 실외기와 전기함은 보이지 않으며, 단순한 촬영 각도 변화가 아니라 다른 도로 구조로 교체되었습니다. 중앙 오른쪽 전신과 왼쪽 여백은 맞지만 카메라는 인물 앞쪽입니다.",
        "entities": "금발의 밝은 피부를 가진 어린 여자아이 한 명이며 연령과 체격은 참조에 대체로 맞습니다. 혼혈 배경은 외관만으로 확정하기 어렵습니다. 오염된 카키 작업복, 가죽 공구 벨트와 파우치, 부츠, 목에 걸린 기계식 방진 마스크가 보입니다. 볼을 따라 흐르는 눈물이 확인됩니다. 추가 인물은 없고, 물에 잠긴 주거지는 보이지만 난민촌이라는 용도와 바닷물 여부는 외관만으로 확인되지 않습니다.",
        "hard_violations": [
         "정확한 촬영 장소에 없는 연속 수변 방호벽과 침수 주택지를 배치하고 도로 배치를 새로 만들어, 고정된 상점 앞 거리 대신 별개의 수변 도로를 구현했습니다."
        ],
        "physics": "양쪽 부츠가 아스팔트에 닿아 몸을 지지하고 있으며, 마스크와 허리 장비도 각각 목끈과 벨트에 연결됩니다. 수면 위 지붕과 기둥은 아래쪽이 물에 잠긴 구조물로 읽혀 부유한다고 볼 근거는 없습니다. 인물의 자세는 물리적으로 가능하지만 양발로 선 정적인 자세라 뒷발에 체중이 걸린 급정지와 상체 비틀림은 분명하지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지정 장소와 중앙 오른쪽 전신 배치는 충실하지만, 뒤쪽 사선 카메라가 아니라 앞쪽에서 촬영되어 멈춘 채 몸을 비틀어 뒤돌아보는 핵심 동작이 빠졌습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "침수지와 오른쪽으로 이어지는 탈출로는 보이지만, 지정 장소를 새로운 수변 도로로 바꾸었고 뒤돌아보는 동작도 구현하지 못했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 왼쪽 바깥을 향하지만 그곳의 난민촌은 보이지 않습니다. 가슴과 골반도 대체로 카메라와 왼쪽을 향하여, 오른쪽으로 이동하다 상체만 왼쪽으로 돌린 관계가 읽히지 않습니다. 도로는 화면을 가로질러 놓여 있어 오른쪽 전방으로 이어지는 이동 축도 약합니다.",
        "built_space": "맞은편에 넓은 상점 진열창 한 면, 그 오른쪽 출입구 한 곳과 작은 경사판 하나, 외벽의 실외기 하나와 전기함 하나가 보입니다. 왼쪽 골목, 전신주, 차도와 앞쪽 점자블록 보도까지 장소 사진의 주요 구성이 유지됩니다. 앰버는 앞쪽 보도에 서 있습니다. 전신은 중앙 오른쪽에 있고 왼쪽 여백도 넓지만, 작업복 앞면이 보이는 카메라 위치는 요구된 뒤쪽 사선 위치가 아닙니다. 침수 상태는 화면에서 확인되지 않습니다.",
        "entities": "사람은 금발의 밝은 피부를 가진 어린 여자아이 한 명뿐이며, 약 10세의 체격과 얼굴은 참조에 대체로 부합합니다. 혼혈 배경 자체는 외관만으로 확정할 수 없습니다. 기름때 묻은 카키 작업복, 가죽 공구 벨트와 파우치, 검은 부츠가 보입니다. 정교한 방진 마스크는 참조처럼 목에 걸려 있습니다. 볼에는 가는 눈물 자국이 보이지만 얼굴 전체가 눈물로 뒤덮인 상태는 약하게 표현되었습니다. 동행인이나 추가 인물은 없습니다.",
        "hard_violations": [],
        "physics": "두 부츠의 밑창이 보도에 닿아 몸을 지지하며 부유하거나 불가능한 관절은 보이지 않습니다. 마스크는 목끈으로, 공구와 파우치는 허리 벨트로 지지됩니다. 다만 발을 비교적 나란히 둔 직립 자세라 이동 중 뒷발에 체중을 실어 멈춘 순간보다 정적인 서 있는 자세에 가깝습니다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 왼쪽 침수 주택지 방향을 향합니다. 도로는 오른쪽 먼 배경으로 이어져 요구한 탈출 방향은 읽힙니다. 그러나 가슴과 골반은 카메라 쪽을 향하고 머리만 왼쪽으로 돌아가 있어, 오른쪽 진행 중 몸을 비틀어 뒤돌아보는 동작은 드러나지 않습니다.",
        "built_space": "앰버는 보도가 아니라 차도 위에 서 있습니다. 왼쪽에는 여러 콘크리트 지주와 금속 난간으로 이어진 수변 방호벽이 있고, 그 너머로 침수된 저층 주택들과 지붕들이 펼쳐집니다. 오른쪽에는 전신주가 줄지어 선 도로가 먼 도시 쪽으로 이어집니다. 지정 장소의 큰 상점 진열창, 오른쪽 출입구, 외벽 실외기와 전기함은 보이지 않으며, 단순한 촬영 각도 변화가 아니라 다른 도로 구조로 교체되었습니다. 중앙 오른쪽 전신과 왼쪽 여백은 맞지만 카메라는 인물 앞쪽입니다.",
        "entities": "금발의 밝은 피부를 가진 어린 여자아이 한 명이며 연령과 체격은 참조에 대체로 맞습니다. 혼혈 배경은 외관만으로 확정하기 어렵습니다. 오염된 카키 작업복, 가죽 공구 벨트와 파우치, 부츠, 목에 걸린 기계식 방진 마스크가 보입니다. 볼을 따라 흐르는 눈물이 확인됩니다. 추가 인물은 없고, 물에 잠긴 주거지는 보이지만 난민촌이라는 용도와 바닷물 여부는 외관만으로 확인되지 않습니다.",
        "hard_violations": [
         "정확한 촬영 장소에 없는 연속 수변 방호벽과 침수 주택지를 배치하고 도로 배치를 새로 만들어, 고정된 상점 앞 거리 대신 별개의 수변 도로를 구현했습니다."
        ],
        "physics": "양쪽 부츠가 아스팔트에 닿아 몸을 지지하고 있으며, 마스크와 허리 장비도 각각 목끈과 벨트에 연결됩니다. 수면 위 지붕과 기둥은 아래쪽이 물에 잠긴 구조물로 읽혀 부유한다고 볼 근거는 없습니다. 인물의 자세는 물리적으로 가능하지만 양발로 선 정적인 자세라 뒷발에 체중이 걸린 급정지와 상체 비틀림은 분명하지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.4,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.15,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[gpt-high] 정확한 촬영 장소에 없는 연속 수변 방호벽과 침수 주택지를 배치하고 도로 배치를 새로 만들어, 고정된 상점 앞 거리 대신 별개의 수변 도로를 구현했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1150,
   "B": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1150,
    "verdict_ko": "카메라 각도, 대각선 구도, 수몰된 지역을 바라보는 자세 등 프레이밍 지침(Priority 2)을 훌륭하게 구현했으나 위치 레퍼런스의 상점 건축물(Priority 3)이 누락되었습니다.  ★위반: [gpt-high] 정확한 촬영 장소에 없는 연속 수변 방호벽과 침수 주택지를 배치하고 도로 배치를 새로 만들어, 고정된 상점 앞 거리 대신 별개의 수변 도로를 구현했습니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "위치 레퍼런스의 배경은 일치하나, 지시된 대각선 구도와 인물의 비틀린 자세를 무시하고 레퍼런스 사진의 평면적 구도를 그대로 복사하여 핵심 프레이밍(Priority 2)을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L54B01.png",
    "asset_id": "a3eab2e7-9787-4314-b6d8-717f050d2d82",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0a9c-07ed-759d-b784-dffb34fc5a1a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S40sh2__bgfirst_bg.png",
   "bg_asset_id": "64e4b41f-ed27-45ea-ab5e-6e1d35b87c75",
   "bg_record_key": "S40sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S40sh5::signage": {
  "fp": "106e5018e3e0aef9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh5": {
  "input_fingerprint": "d1179d5632345e06",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 앰버의 팔뚝을 꽉 쥐고 앞을 향해 강하게 끌어당겨 앰버의 상체가 앞으로 쏠린 찰나.\n\nLOCATION (lock): On the street leading away from the flooded refugee settlement toward the city outskirts. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 앰버's near side at forearm height, finish the close approach outside the space between the siblings, looking obliquely across the grip and slightly upward toward her torso. 이현우's hand enters from screen-right and clamps her forearm near center, while enough of her upper body remains on the left to show it pitching sharply toward their screen-right escape route. Her head and attention are drawn toward that route beyond the crop; keep 이현우's face outside the image so the grip and resulting displacement carry the moment.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Outskirts route (The siblings are resuming their escape toward the city) — Continues beyond the right side of the frame; used as Remains a reduced contextual margin behind the arm and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding daylight exposure and controlled contrast, emphasizing the physical pull through shape rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the outer-city street surfaces, nearby buildings, and daylight appearance from the reference. Exclude floating containers and debris as street furnishings; those belong to the flooded settlement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight escape route leads away from the flooded settlement toward the city. 앰버: She remains exhausted and tear-streaked, her body pulled forward from the stopped position. 이현우: He is moving forward with one arm held back in a pulling posture; his facial bruises and untreated leg wound persist. The stiff contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 앰버의 팔뚝을 꽉 쥐고 앞을 향해 강하게 끌어당겨 앰버의 상체가 앞으로 쏠린 찰나.\n\nLOCATION (lock): On the street leading away from the flooded refugee settlement toward the city outskirts. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 앰버's near side at forearm height, finish the close approach outside the space between the siblings, looking obliquely across the grip and slightly upward toward her torso. 이현우's hand enters from screen-right and clamps her forearm near center, while enough of her upper body remains on the left to show it pitching sharply toward their screen-right escape route. Her head and attention are drawn toward that route beyond the crop; keep 이현우's face outside the image so the grip and resulting displacement carry the moment.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Outskirts route (The siblings are resuming their escape toward the city) — Continues beyond the right side of the frame; used as Remains a reduced contextual margin behind the arm and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding daylight exposure and controlled contrast, emphasizing the physical pull through shape rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the outer-city street surfaces, nearby buildings, and daylight appearance from the reference. Exclude floating containers and debris as street furnishings; those belong to the flooded settlement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight escape route leads away from the flooded settlement toward the city. 앰버: She remains exhausted and tear-streaked, her body pulled forward from the stopped position. 이현우: He is moving forward with one arm held back in a pulling posture; his facial bruises and untreated leg wound persist. The stiff contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 앰버의 팔뚝을 꽉 쥐고 앞을 향해 강하게 끌어당겨 앰버의 상체가 앞으로 쏠린 찰나.\n\nLOCATION (lock): On the street leading away from the flooded refugee settlement toward the city outskirts. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 앰버's near side at forearm height, finish the close approach outside the space between the siblings, looking obliquely across the grip and slightly upward toward her torso. 이현우's hand enters from screen-right and clamps her forearm near center, while enough of her upper body remains on the left to show it pitching sharply toward their screen-right escape route. Her head and attention are drawn toward that route beyond the crop; keep 이현우's face outside the image so the grip and resulting displacement carry the moment.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Outskirts route (The siblings are resuming their escape toward the city) — Continues beyond the right side of the frame; used as Remains a reduced contextual margin behind the arm and torso.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the preceding daylight exposure and controlled contrast, emphasizing the physical pull through shape rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the outer-city street surfaces, nearby buildings, and daylight appearance from the reference. Exclude floating containers and debris as street furnishings; those belong to the flooded settlement.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight escape route leads away from the flooded settlement toward the city. 앰버: She remains exhausted and tear-streaked, her body pulled forward from the stopped position. 이현우: He is moving forward with one arm held back in a pulling posture; his facial bruises and untreated leg wound persist. The stiff contact card remains inside his shoe.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선과 몸의 방향이 화면 우측 탈출로를 향해 있으며, 우측에서 들어온 손이 그녀를 같은 방향으로 당기고 있습니다.",
    "built_space": "아웃포커싱된 배경에 참조 이미지의 거리와 건물 외곽이 올바른 원근감으로 축소되어 배치되어 있습니다.",
    "entities": "앰버(금발, 작업복, 목에 건 마스크)와 이현우의 거친 손과 어두운 셔츠 소매가 지침대로 묘사되었습니다.",
    "hard_violations": [],
    "physics": "앰버의 상체가 앞(화면 우측)으로 강하게 쏠려 있으며, 우측에서 들어온 손에 의해 당겨지는 힘과 무게 중심의 이동이 명확히 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "앰버의 시선이 탈출로가 아닌 화면 좌측의 이현우를 향하고 있습니다.",
    "built_space": "참조 이미지의 거리와 상점 입구가 화면 전체에 넓게 나타납니다.",
    "entities": "앰버와 함께 프레임에서 제외되어야 할 이현우의 얼굴과 상체가 온전히 나타나 있습니다.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥에 서서 정적인 자세를 취하고 있으며, 상체가 앞으로 쏠리거나 강하게 끌어당겨지는 힘이 작용하지 않습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "클로즈업 프레이밍과 이현우의 얼굴을 배제하라는 카메라 지시를 정확히 따랐으며 앰버가 앞으로 쏠리는 역동적인 찰나를 훌륭하게 연출했으나, 팔뚝이 아닌 손을 잡고 있는 점이 유일한 아쉬움입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 얼굴을 화면에서 제외하라는 지침을 완전히 무시하고 와이드한 구도로 렌더링했으며, 강하게 끌어당겨지는 물리적 액션이 전혀 표현되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선과 몸의 방향이 화면 우측 탈출로를 향해 있으며, 우측에서 들어온 손이 그녀를 같은 방향으로 당기고 있습니다.",
        "built_space": "아웃포커싱된 배경에 참조 이미지의 거리와 건물 외곽이 올바른 원근감으로 축소되어 배치되어 있습니다.",
        "entities": "앰버(금발, 작업복, 목에 건 마스크)와 이현우의 거친 손과 어두운 셔츠 소매가 지침대로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "앰버의 상체가 앞(화면 우측)으로 강하게 쏠려 있으며, 우측에서 들어온 손에 의해 당겨지는 힘과 무게 중심의 이동이 명확히 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선이 탈출로가 아닌 화면 좌측의 이현우를 향하고 있습니다.",
        "built_space": "참조 이미지의 거리와 상점 입구가 화면 전체에 넓게 나타납니다.",
        "entities": "앰버와 함께 프레임에서 제외되어야 할 이현우의 얼굴과 상체가 온전히 나타나 있습니다.",
        "hard_violations": [],
        "physics": "두 사람 모두 바닥에 서서 정적인 자세를 취하고 있으며, 상체가 앞으로 쏠리거나 강하게 끌어당겨지는 힘이 작용하지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "클로즈업 프레이밍과 이현우의 얼굴을 배제하라는 카메라 지시를 정확히 따랐으며 앰버가 앞으로 쏠리는 역동적인 찰나를 훌륭하게 연출했으나, 팔뚝이 아닌 손을 잡고 있는 점이 유일한 아쉬움입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이현우의 얼굴을 화면에서 제외하라는 지침을 완전히 무시하고 와이드한 구도로 렌더링했으며, 강하게 끌어당겨지는 물리적 액션이 전혀 표현되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선과 몸의 방향이 화면 우측 탈출로를 향해 있으며, 우측에서 들어온 손이 그녀를 같은 방향으로 당기고 있습니다.",
        "built_space": "아웃포커싱된 배경에 참조 이미지의 거리와 건물 외곽이 올바른 원근감으로 축소되어 배치되어 있습니다.",
        "entities": "앰버(금발, 작업복, 목에 건 마스크)와 이현우의 거친 손과 어두운 셔츠 소매가 지침대로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "앰버의 상체가 앞(화면 우측)으로 강하게 쏠려 있으며, 우측에서 들어온 손에 의해 당겨지는 힘과 무게 중심의 이동이 명확히 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선이 탈출로가 아닌 화면 좌측의 이현우를 향하고 있습니다.",
        "built_space": "참조 이미지의 거리와 상점 입구가 화면 전체에 넓게 나타납니다.",
        "entities": "앰버와 함께 프레임에서 제외되어야 할 이현우의 얼굴과 상체가 온전히 나타나 있습니다.",
        "hard_violations": [],
        "physics": "두 사람 모두 바닥에 서서 정적인 자세를 취하고 있으며, 상체가 앞으로 쏠리거나 강하게 끌어당겨지는 힘이 작용하지 않습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장소와 의상은 잘 이어지지만, 서로 바라보며 멈춘 중간 거리 구도여서 얼굴을 제외한 팔뚝 클로즈업과 강한 전방 당김이라는 핵심 지시를 놓쳤다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽에서 들어온 손의 강한 움켜쥠과 왼쪽 앰버의 전방 쏠림을 근접 구도로 구현했으나, 카메라가 다소 내려다보고 배경도 요구보다 넓게 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 화면 오른쪽의 이현우 얼굴을 올려다보고, 이현우는 왼쪽 아래의 앰버를 바라본다. 두 사람의 주의가 화면 밖 오른쪽 탈출로가 아니라 서로에게 향한다. 이현우의 손은 오른쪽에서 앰버의 손목 부근을 잡지만 앰버의 상체는 거의 수직이며, 오른쪽으로 급히 끌리는 방향성이 없다.",
        "built_space": "뒤쪽에 큰 상점 유리창 하나와 그 오른쪽 출입구 하나, 오른쪽 벽의 실외기 하나가 보인다. 왼쪽 골목과 전신주, 앞쪽 도로·경계석·보도·노란 점자블록도 이전 사진과 잘 이어진다. 두 사람은 가까운 보도 쪽에 서 있으나, 상점 전경까지 크게 드러나는 구도여서 배경을 좁은 여백으로 남기라는 지시와 다르다.",
        "entities": "금발의 어린 여자아이와 짧은 검은 머리의 젊은 동아시아계 남성으로 보이며 참고 인물의 외형에 대체로 부합한다. 앰버의 카키색 작업복, 목에 걸린 방진 마스크, 가죽 공구 벨트가 유지된다. 이현우의 어두운 낡은 셔츠도 맞지만, 제외해야 할 얼굴과 몸통이 크게 보이고 얼굴의 멍은 뚜렷하지 않다. 앰버의 눈물 자국과 탈진한 표정도 약하다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 손가락이 앰버의 손목 가까이를 실제로 감싸고 있어 접촉 자체는 가능하다. 두 몸통은 화면 아래로 이어지며 발은 잘려 있어 접지 상태는 확인할 수 없지만 공중에 뜬 모습은 아니다. 마스크는 목의 끈에, 공구와 주머니는 허리 벨트에 지지된다. 다만 느슨하게 처진 앰버의 손과 곧게 선 상체에는 강한 당김에 따른 변위나 체중 이동이 나타나지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 소매와 손이 오른쪽에서 들어와 중앙 아래의 앰버 팔뚝 끝을 감싼다. 앰버의 팔은 오른쪽으로 뻗고 어깨와 상체도 그쪽으로 기울며, 시선은 화면 밖 오른쪽 앞을 향한다. 머리와 몸이 탈출 방향으로 이끌리는 관계가 읽히고 이현우의 얼굴은 보이지 않는다.",
        "built_space": "아스팔트 도로 한 줄기와 왼쪽 경계석·보도, 낮은 담장과 건물들이 뒤로 이어진다. 여러 전신주가 원근에 따라 작아지며 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 이전 사진과 비슷한 외곽 거리 재료와 낮 시간대이지만, 같은 상점의 창과 출입구는 확인되지 않아 정확한 장소 연속성은 제한적이다. 도로가 상당한 배경 면적을 차지하며 카메라는 약간 아래를 내려다보는 인상이다.",
        "entities": "앰버는 참고와 부합하는 금발의 어린 여자아이로, 젖은 눈과 볼의 눈물 흔적이 보인다. 때 묻은 카키색 작업복, 목에 걸린 금속성 방진 마스크, 가죽 공구 벨트가 유지된다. 이현우는 거친 손과 어두운 셔츠 소매만 보이며, 얼굴·다리·신발 속 카드가 잘린 것은 지시에 맞는다. 손만으로 정확한 나이나 민족적 배경을 확정할 수는 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 엄지와 나머지 손가락이 앰버의 손목에 가까운 팔뚝을 위아래로 감싸 직접 지지하고 당긴다. 앰버의 팔과 어깨가 연결된 채 긴장되고 상체가 오른쪽 앞으로 기울어, 잡아당기는 힘에 반응하는 자세로 가능하다. 굽힌 다리는 화면 아래로 이어지고 발의 접지는 대부분 잘렸으나 몸이 떠 있다는 징후는 없다. 마스크는 목끈으로, 공구 주머니는 벨트로 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소와 의상은 잘 이어지지만, 서로 바라보며 멈춘 중간 거리 구도여서 얼굴을 제외한 팔뚝 클로즈업과 강한 전방 당김이라는 핵심 지시를 놓쳤다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽에서 들어온 손의 강한 움켜쥠과 왼쪽 앰버의 전방 쏠림을 근접 구도로 구현했으나, 카메라가 다소 내려다보고 배경도 요구보다 넓게 남는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 화면 오른쪽의 이현우 얼굴을 올려다보고, 이현우는 왼쪽 아래의 앰버를 바라본다. 두 사람의 주의가 화면 밖 오른쪽 탈출로가 아니라 서로에게 향한다. 이현우의 손은 오른쪽에서 앰버의 손목 부근을 잡지만 앰버의 상체는 거의 수직이며, 오른쪽으로 급히 끌리는 방향성이 없다.",
        "built_space": "뒤쪽에 큰 상점 유리창 하나와 그 오른쪽 출입구 하나, 오른쪽 벽의 실외기 하나가 보인다. 왼쪽 골목과 전신주, 앞쪽 도로·경계석·보도·노란 점자블록도 이전 사진과 잘 이어진다. 두 사람은 가까운 보도 쪽에 서 있으나, 상점 전경까지 크게 드러나는 구도여서 배경을 좁은 여백으로 남기라는 지시와 다르다.",
        "entities": "금발의 어린 여자아이와 짧은 검은 머리의 젊은 동아시아계 남성으로 보이며 참고 인물의 외형에 대체로 부합한다. 앰버의 카키색 작업복, 목에 걸린 방진 마스크, 가죽 공구 벨트가 유지된다. 이현우의 어두운 낡은 셔츠도 맞지만, 제외해야 할 얼굴과 몸통이 크게 보이고 얼굴의 멍은 뚜렷하지 않다. 앰버의 눈물 자국과 탈진한 표정도 약하다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 손가락이 앰버의 손목 가까이를 실제로 감싸고 있어 접촉 자체는 가능하다. 두 몸통은 화면 아래로 이어지며 발은 잘려 있어 접지 상태는 확인할 수 없지만 공중에 뜬 모습은 아니다. 마스크는 목의 끈에, 공구와 주머니는 허리 벨트에 지지된다. 다만 느슨하게 처진 앰버의 손과 곧게 선 상체에는 강한 당김에 따른 변위나 체중 이동이 나타나지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 소매와 손이 오른쪽에서 들어와 중앙 아래의 앰버 팔뚝 끝을 감싼다. 앰버의 팔은 오른쪽으로 뻗고 어깨와 상체도 그쪽으로 기울며, 시선은 화면 밖 오른쪽 앞을 향한다. 머리와 몸이 탈출 방향으로 이끌리는 관계가 읽히고 이현우의 얼굴은 보이지 않는다.",
        "built_space": "아스팔트 도로 한 줄기와 왼쪽 경계석·보도, 낮은 담장과 건물들이 뒤로 이어진다. 여러 전신주가 원근에 따라 작아지며 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 이전 사진과 비슷한 외곽 거리 재료와 낮 시간대이지만, 같은 상점의 창과 출입구는 확인되지 않아 정확한 장소 연속성은 제한적이다. 도로가 상당한 배경 면적을 차지하며 카메라는 약간 아래를 내려다보는 인상이다.",
        "entities": "앰버는 참고와 부합하는 금발의 어린 여자아이로, 젖은 눈과 볼의 눈물 흔적이 보인다. 때 묻은 카키색 작업복, 목에 걸린 금속성 방진 마스크, 가죽 공구 벨트가 유지된다. 이현우는 거친 손과 어두운 셔츠 소매만 보이며, 얼굴·다리·신발 속 카드가 잘린 것은 지시에 맞는다. 손만으로 정확한 나이나 민족적 배경을 확정할 수는 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 엄지와 나머지 손가락이 앰버의 손목에 가까운 팔뚝을 위아래로 감싸 직접 지지하고 당긴다. 앰버의 팔과 어깨가 연결된 채 긴장되고 상체가 오른쪽 앞으로 기울어, 잡아당기는 힘에 반응하는 자세로 가능하다. 굽힌 다리는 화면 아래로 이어지고 발의 접지는 대부분 잘렸으나 몸이 떠 있다는 징후는 없다. 마스크는 목끈으로, 공구 주머니는 벨트로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.804
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 804
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "클로즈업 프레이밍과 이현우의 얼굴을 배제하라는 카메라 지시를 정확히 따랐으며 앰버가 앞으로 쏠리는 역동적인 찰나를 훌륭하게 연출했으나, 팔뚝이 아닌 손을 잡고 있는 점이 유일한 아쉬움입니다."
   },
   {
    "label": "B",
    "score": 804,
    "verdict_ko": "이현우의 얼굴을 화면에서 제외하라는 지침을 완전히 무시하고 와이드한 구도로 렌더링했으며, 강하게 끌어당겨지는 물리적 액션이 전혀 표현되지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S40sh2_sel.png",
    "asset_id": "815b8518-3480-474f-b145-957651d25f63",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0aa5-70ad-73b6-8afd-5cebae5a5f19",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S40sh2"
  }
 },
 "S40sh7::signage": {
  "fp": "1fe517835294dd51",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh7": {
  "input_fingerprint": "1cb40bfa90b38d66",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 인적 없는 삭막한 도심 거리 건너편에 유리창이 온전한 무인 점포가 보이는 시점 구도.\n\nLOCATION (lock): On a deserted city-outskirts street opposite an unattended shop with intact front glazing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the pan at 이현우's eye height from the group's side of the street, aligning the unobstructed view with his look while keeping his shoulder and all other figures outside the frame. Look obliquely across the empty roadway toward the unattended shop in the upper-right, its intact front windows readable but the storefront occupying less than two-fifths of the image. The intervening road fills the lower field, making the distance to their possible destination clear without moving across the street.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unattended shop across the street in the upper-right of the frame, background; Empty intervening roadway in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Unattended shop (Visible across the street) — Its street-facing frontage is seen obliquely from the opposite side; used as Serves as the destination of 이현우's look in the upper-right background; Shop windows (Intact) — The street-facing glazing is visible; interior contents are not resolved; used as Identifies an intact storefront without adding reflections or specific merchandise; Intervening roadway (Empty, with no police visible); used as Separates the viewpoint from the shop across the lower and middle frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the empty street and intact storefront legible without inventing illuminated signage or interior light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An unmanned shop stands across the daylight city street. No damage to its entrance or glazing has yet been established.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 인적 없는 삭막한 도심 거리 건너편에 유리창이 온전한 무인 점포가 보이는 시점 구도.\n\nLOCATION (lock): On a deserted city-outskirts street opposite an unattended shop with intact front glazing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the pan at 이현우's eye height from the group's side of the street, aligning the unobstructed view with his look while keeping his shoulder and all other figures outside the frame. Look obliquely across the empty roadway toward the unattended shop in the upper-right, its intact front windows readable but the storefront occupying less than two-fifths of the image. The intervening road fills the lower field, making the distance to their possible destination clear without moving across the street.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unattended shop across the street in the upper-right of the frame, background; Empty intervening roadway in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Unattended shop (Visible across the street) — Its street-facing frontage is seen obliquely from the opposite side; used as Serves as the destination of 이현우's look in the upper-right background; Shop windows (Intact) — The street-facing glazing is visible; interior contents are not resolved; used as Identifies an intact storefront without adding reflections or specific merchandise; Intervening roadway (Empty, with no police visible); used as Separates the viewpoint from the shop across the lower and middle frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the empty street and intact storefront legible without inventing illuminated signage or interior light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An unmanned shop stands across the daylight city street. No damage to its entrance or glazing has yet been established.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 인적 없는 삭막한 도심 거리 건너편에 유리창이 온전한 무인 점포가 보이는 시점 구도.\n\nLOCATION (lock): On a deserted city-outskirts street opposite an unattended shop with intact front glazing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the pan at 이현우's eye height from the group's side of the street, aligning the unobstructed view with his look while keeping his shoulder and all other figures outside the frame. Look obliquely across the empty roadway toward the unattended shop in the upper-right, its intact front windows readable but the storefront occupying less than two-fifths of the image. The intervening road fills the lower field, making the distance to their possible destination clear without moving across the street.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unattended shop across the street in the upper-right of the frame, background; Empty intervening roadway in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Unattended shop (Visible across the street) — Its street-facing frontage is seen obliquely from the opposite side; used as Serves as the destination of 이현우's look in the upper-right background; Shop windows (Intact) — The street-facing glazing is visible; interior contents are not resolved; used as Identifies an intact storefront without adding reflections or specific merchandise; Intervening roadway (Empty, with no police visible); used as Separates the viewpoint from the shop across the lower and middle frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and restrained contrast keep the empty street and intact storefront legible without inventing illuminated signage or interior light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An unmanned shop stands across the daylight city street. No damage to its entrance or glazing has yet been established.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
    "built_space": "우측에 무인 점포가 위치함. LOCATION 사진의 주택가 좁은 도로 대신 STRUCTURE 사진의 넓은 도로를 배경으로 잘못 렌더링함.",
    "entities": "무인 점포(지정된 외관과 일치하나 내부 진열품이 명확히 보임), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
    "hard_violations": [],
    "physics": "도로 차선, 가로등, 점포 건물이 모두 지면에 안정적으로 배치됨."
   },
   {
    "label": "B",
    "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
    "built_space": "우측에 무인 점포가 위치함. A와 동일하게 LOCATION 배경 지침을 위반함. 도로 좌측 차선 구조에 명백한 오류가 있음.",
    "entities": "무인 점포(외관 일치, 내부 진열품 선명함), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 도로 구조 (좌측 점선 차선이 사선으로 꺾여 가장자리 실선과 비정상적으로 병합됨)"
    ],
    "physics": "도로의 점선 차선이 원근법과 물리적 구조를 무시하고 잘못된 방향으로 합쳐짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "LOCATION 사진의 배경을 무시하고 내부 상품이 선명하게 묘사된 한계가 있으나, 물리적 결함이 없어 차선 오류가 있는 B보다 우수합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일한 배경 지침 위반이 있으며, 도로 점선이 실선과 비정상적으로 합쳐지는 물리적 구조 오류가 있어 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
        "built_space": "우측에 무인 점포가 위치함. LOCATION 사진의 주택가 좁은 도로 대신 STRUCTURE 사진의 넓은 도로를 배경으로 잘못 렌더링함.",
        "entities": "무인 점포(지정된 외관과 일치하나 내부 진열품이 명확히 보임), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
        "hard_violations": [],
        "physics": "도로 차선, 가로등, 점포 건물이 모두 지면에 안정적으로 배치됨."
       },
       {
        "label": "B",
        "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
        "built_space": "우측에 무인 점포가 위치함. A와 동일하게 LOCATION 배경 지침을 위반함. 도로 좌측 차선 구조에 명백한 오류가 있음.",
        "entities": "무인 점포(외관 일치, 내부 진열품 선명함), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
        "hard_violations": [
         "물리적으로 불가능한 도로 구조 (좌측 점선 차선이 사선으로 꺾여 가장자리 실선과 비정상적으로 병합됨)"
        ],
        "physics": "도로의 점선 차선이 원근법과 물리적 구조를 무시하고 잘못된 방향으로 합쳐짐."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "LOCATION 사진의 배경을 무시하고 내부 상품이 선명하게 묘사된 한계가 있으나, 물리적 결함이 없어 차선 오류가 있는 B보다 우수합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일한 배경 지침 위반이 있으며, 도로 점선이 실선과 비정상적으로 합쳐지는 물리적 구조 오류가 있어 감점되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
        "built_space": "우측에 무인 점포가 위치함. LOCATION 사진의 주택가 좁은 도로 대신 STRUCTURE 사진의 넓은 도로를 배경으로 잘못 렌더링함.",
        "entities": "무인 점포(지정된 외관과 일치하나 내부 진열품이 명확히 보임), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
        "hard_violations": [],
        "physics": "도로 차선, 가로등, 점포 건물이 모두 지면에 안정적으로 배치됨."
       },
       {
        "label": "B",
        "direction": "시선은 빈 도로 너머 우측 점포를 비스듬히 향함.",
        "built_space": "우측에 무인 점포가 위치함. A와 동일하게 LOCATION 배경 지침을 위반함. 도로 좌측 차선 구조에 명백한 오류가 있음.",
        "entities": "무인 점포(외관 일치, 내부 진열품 선명함), 텅 빈 도로(배경 불일치), 인물 없음(지침 준수).",
        "hard_violations": [
         "물리적으로 불가능한 도로 구조 (좌측 점선 차선이 사선으로 꺾여 가장자리 실선과 비정상적으로 병합됨)"
        ],
        "physics": "도로의 점선 차선이 원근법과 물리적 구조를 무시하고 잘못된 방향으로 합쳐짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "점포를 오른쪽 위에 작게 두고 빈 도로를 하단에 넓게 배치해 구도는 더 충실하지만, 지정된 동네 거리 대신 구조 참고사진의 도로 주변까지 복제했고 실내 상품도 드러난다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "점포의 외형과 온전한 유리는 잘 유지했지만, 점포가 오른쪽 중앙까지 내려오고 구조 참고사진의 구도와 주변 환경을 거의 그대로 따라 지정 장소와 독립적인 촬영 구도 요구를 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사람과 무기, 이동하는 물체는 없다. 카메라는 도로 반대편 점포 정면과 오른쪽 측벽을 비스듬히 바라본다. 시선의 목적지인 점포는 오른쪽 위에 있고, 도로의 소실점은 왼쪽에 있어 요구한 대각선 관찰 관계가 읽힌다.",
        "built_space": "단층 점포 한 채에 파란 간판 띠 하나, 중앙 유리 출입구 한 곳, 그 왼쪽 진열창 구역 하나와 오른쪽 두 구획의 진열창, 오른쪽 측벽 창 하나와 계량기함 하나가 보인다. 흰 외벽의 녹과 금속 창틀은 구조 참고사진에 가깝다. 점포 전면은 화면의 5분의 2보다 작고 하단 대부분은 빈 차도다. 그러나 주변은 방음벽, 관목 띠, 연속 가로등, 가드레일과 먼 산으로 구성되어, 장소 참고사진의 낮은 주택과 왼쪽 골목, 전봇대 및 가까운 맞은편 보도 환경이 아니다. 불가능한 반사는 보이지 않는다.",
        "entities": "무인 점포, 파손되지 않은 전면 유리와 출입구, 빈 아스팔트 도로가 모두 보인다. 사람, 얼굴, 어깨, 경찰, 차량은 없다. 구조 참고사진의 낡은 흰색 점포와 파란색·흰색·붉은색 간판 띠를 유지한다. 다만 유리 너머 선반의 병과 포장 상품이 구분되어 실내 내용물을 해상하지 말라는 요구에 어긋난다. 낮이지만 뚜렷한 햇빛과 그림자는 요청한 절제된 광량보다 강하다. 새로 읽히는 문구나 화면 위 자막은 없다.",
        "hard_violations": [],
        "physics": "점포는 맞은편 보도에 놓이고, 가로등과 울타리는 지면에 고정되어 있다. 간판과 계량기함은 건물에 부착되어 있으며 실내 상품은 선반 위에 놓여 있다. 공중에 떠 있는 인물이나 지지 없는 물체, 물리적으로 불가능한 동작은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "인물의 시선이나 겨냥하는 물체, 이동체는 없다. 카메라는 빈 도로 건너 오른쪽 점포를 사선으로 보며 정면과 오른쪽 측벽을 함께 담는다. 다만 점포의 창과 출입구가 오른쪽 중앙 높이에 놓여, 오른쪽 위 배경을 목적지로 삼는 요구는 A보다 약하다.",
        "built_space": "단층 점포 한 채, 파란 간판 띠 하나, 중앙 유리 출입구 한 곳, 왼쪽 진열창 구역 하나, 오른쪽 두 구획의 진열창, 측벽 창 하나와 계량기함 하나가 보인다. 외형과 개구부 배치는 구조 참고사진에 매우 가깝다. 전면 면적은 화면의 5분의 2 미만이고 빈 도로가 전경을 차지한다. 그러나 방음벽과 관목, 길게 늘어선 가로등, 가드레일, 산과 아파트까지 구조 참고사진의 환경을 가져와 장소 참고사진의 주택가 골목과 보도 배치를 대체했다. 구조 참고사진의 카메라 구도도 거의 반복한다. 광학적으로 불가능한 반사는 확인되지 않는다.",
        "entities": "온전한 전면 유리를 가진 무인 점포와 빈 차도가 있고, 사람이나 경찰 및 차량은 없다. 녹슨 흰 외벽과 삼색 간판 띠는 구조 정체성에 부합한다. 창 너머 상품 진열과 여러 포장 물체가 뚜렷해 내부를 식별하지 않도록 하라는 요구에는 맞지 않는다. 낮의 자연광이지만 가로등 그림자와 맑은 하늘이 강조된다. 추가 자막이나 명확한 새 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "건물과 가로등, 방음벽 및 가드레일은 각각 지면과 기둥으로 지지된다. 점포의 간판은 지붕 가장자리에 부착되어 있고 실내 물체는 선반에 놓여 있다. 움직이는 신체나 부유 물체는 없으며, 지지와 중력에 어긋나는 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "점포를 오른쪽 위에 작게 두고 빈 도로를 하단에 넓게 배치해 구도는 더 충실하지만, 지정된 동네 거리 대신 구조 참고사진의 도로 주변까지 복제했고 실내 상품도 드러난다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "점포의 외형과 온전한 유리는 잘 유지했지만, 점포가 오른쪽 중앙까지 내려오고 구조 참고사진의 구도와 주변 환경을 거의 그대로 따라 지정 장소와 독립적인 촬영 구도 요구를 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "사람과 무기, 이동하는 물체는 없다. 카메라는 도로 반대편 점포 정면과 오른쪽 측벽을 비스듬히 바라본다. 시선의 목적지인 점포는 오른쪽 위에 있고, 도로의 소실점은 왼쪽에 있어 요구한 대각선 관찰 관계가 읽힌다.",
        "built_space": "단층 점포 한 채에 파란 간판 띠 하나, 중앙 유리 출입구 한 곳, 그 왼쪽 진열창 구역 하나와 오른쪽 두 구획의 진열창, 오른쪽 측벽 창 하나와 계량기함 하나가 보인다. 흰 외벽의 녹과 금속 창틀은 구조 참고사진에 가깝다. 점포 전면은 화면의 5분의 2보다 작고 하단 대부분은 빈 차도다. 그러나 주변은 방음벽, 관목 띠, 연속 가로등, 가드레일과 먼 산으로 구성되어, 장소 참고사진의 낮은 주택과 왼쪽 골목, 전봇대 및 가까운 맞은편 보도 환경이 아니다. 불가능한 반사는 보이지 않는다.",
        "entities": "무인 점포, 파손되지 않은 전면 유리와 출입구, 빈 아스팔트 도로가 모두 보인다. 사람, 얼굴, 어깨, 경찰, 차량은 없다. 구조 참고사진의 낡은 흰색 점포와 파란색·흰색·붉은색 간판 띠를 유지한다. 다만 유리 너머 선반의 병과 포장 상품이 구분되어 실내 내용물을 해상하지 말라는 요구에 어긋난다. 낮이지만 뚜렷한 햇빛과 그림자는 요청한 절제된 광량보다 강하다. 새로 읽히는 문구나 화면 위 자막은 없다.",
        "hard_violations": [],
        "physics": "점포는 맞은편 보도에 놓이고, 가로등과 울타리는 지면에 고정되어 있다. 간판과 계량기함은 건물에 부착되어 있으며 실내 상품은 선반 위에 놓여 있다. 공중에 떠 있는 인물이나 지지 없는 물체, 물리적으로 불가능한 동작은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "인물의 시선이나 겨냥하는 물체, 이동체는 없다. 카메라는 빈 도로 건너 오른쪽 점포를 사선으로 보며 정면과 오른쪽 측벽을 함께 담는다. 다만 점포의 창과 출입구가 오른쪽 중앙 높이에 놓여, 오른쪽 위 배경을 목적지로 삼는 요구는 A보다 약하다.",
        "built_space": "단층 점포 한 채, 파란 간판 띠 하나, 중앙 유리 출입구 한 곳, 왼쪽 진열창 구역 하나, 오른쪽 두 구획의 진열창, 측벽 창 하나와 계량기함 하나가 보인다. 외형과 개구부 배치는 구조 참고사진에 매우 가깝다. 전면 면적은 화면의 5분의 2 미만이고 빈 도로가 전경을 차지한다. 그러나 방음벽과 관목, 길게 늘어선 가로등, 가드레일, 산과 아파트까지 구조 참고사진의 환경을 가져와 장소 참고사진의 주택가 골목과 보도 배치를 대체했다. 구조 참고사진의 카메라 구도도 거의 반복한다. 광학적으로 불가능한 반사는 확인되지 않는다.",
        "entities": "온전한 전면 유리를 가진 무인 점포와 빈 차도가 있고, 사람이나 경찰 및 차량은 없다. 녹슨 흰 외벽과 삼색 간판 띠는 구조 정체성에 부합한다. 창 너머 상품 진열과 여러 포장 물체가 뚜렷해 내부를 식별하지 않도록 하라는 요구에는 맞지 않는다. 낮의 자연광이지만 가로등 그림자와 맑은 하늘이 강조된다. 추가 자막이나 명확한 새 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "건물과 가로등, 방음벽 및 가드레일은 각각 지면과 기둥으로 지지된다. 점포의 간판은 지붕 가장자리에 부착되어 있고 실내 물체는 선반에 놓여 있다. 움직이는 신체나 부유 물체는 없으며, 지지와 중력에 어긋나는 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.8,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 도로 구조 (좌측 점선 차선이 사선으로 꺾여 가장자리 실선과 비정상적으로 병합됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1800,
   "B": 1350
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1800,
    "verdict_ko": "LOCATION 사진의 배경을 무시하고 내부 상품이 선명하게 묘사된 한계가 있으나, 물리적 결함이 없어 차선 오류가 있는 B보다 우수합니다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "A와 동일한 배경 지침 위반이 있으며, 도로 점선이 실선과 비정상적으로 합쳐지는 물리적 구조 오류가 있어 감점되었습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 도로 구조 (좌측 점선 차선이 사선으로 꺾여 가장자리 실선과 비정상적으로 병합됨)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L54B01.png",
    "asset_id": "a3eab2e7-9787-4314-b6d8-717f050d2d82",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_unattended_store_sel.png",
    "asset_id": "bdf4f3ec-d87e-45b4-93f6-17d8c1eaf331",
    "role": "structure_seed_look"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0abc-4d07-78b7-b9c7-20a86edee543",
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S41sh18::signage": {
  "fp": "513bfa18df466c9e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9a7dc0a1d8aafbd2": {
  "subjects": [],
  "subject_text": "무인점포 내부\n신발과 가방, 식품, 생활용품이 진열된 무인 매장. 진열대 사이 통로와 별도 의약품 코너, 현금인출기가 있다.",
  "identity": "canonical",
  "scope_id": "L55",
  "scope_role": "location_interior",
  "scope_sha": "9747386951ff3bd7"
 },
 "S41sh18::bgfirst_bg": {
  "input_fingerprint": "db57b0d735926a06",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 금속 주먹이 현금인출기의 전면 패널을 정통으로 꿰뚫고 들어간 파괴적인 찰나.\n\nLOCATION (lock): At the cash machine inside the unattended shop, beside the retail aisles. Daylight enters through the intact storefront glazing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside 찰리's punching arm, slightly below the impact and looking obliquely upward across the ATM front, without crossing the strike axis. His forearm runs from lower-left into the embedded metal fist near center-right, with the breached panel occupying no more than two-fifths of the frame and a soft margin of the shop display retaining context. Exclude his head and torso, concentrating on the completed forward extension and contact point before withdrawal or falling money begins.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: ATM front (Pierced by 찰리's fist, which remains embedded) — The front and a narrow adjoining side are viewed obliquely from below; used as Locates the impact without enlarging the machine beyond believable scale; Shop display (Present within the shop) — Only an indistinct edge remains beyond the punching arm; used as Provides a small contextual margin around the isolated action detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the shop and controlled highlights articulate the metal fist and breached panel without invented sparks or dramatic light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 금속 주먹이 현금인출기의 전면 패널을 정통으로 꿰뚫고 들어간 파괴적인 찰나.\n\nLOCATION (lock): At the cash machine inside the unattended shop, beside the retail aisles. Daylight enters through the intact storefront glazing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside 찰리's punching arm, slightly below the impact and looking obliquely upward across the ATM front, without crossing the strike axis. His forearm runs from lower-left into the embedded metal fist near center-right, with the breached panel occupying no more than two-fifths of the frame and a soft margin of the shop display retaining context. Exclude his head and torso, concentrating on the completed forward extension and contact point before withdrawal or falling money begins.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: ATM front (Pierced by 찰리's fist, which remains embedded) — The front and a narrow adjoining side are viewed obliquely from below; used as Locates the impact without enlarging the machine beyond believable scale; Shop display (Present within the shop) — Only an indistinct edge remains beyond the punching arm; used as Provides a small contextual margin around the isolated action detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the shop and controlled highlights articulate the metal fist and breached panel without invented sparks or dramatic light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S41sh18__bgfirst_bg.png",
  "asset_id": "55d50b19-ec48-4a42-9b3b-967f4c27b608",
  "input_asset_ids": [
   "9a6bb787-0846-4533-8eca-f856ce454482",
   "76b5f00a-f9af-40d4-81fa-1eaf208efccc"
  ]
 },
 "S41sh18": {
  "input_fingerprint": "4db60d5d8068754e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기의 전면 패널을 정통으로 꿰뚫고 들어간 파괴적인 찰나.\n\nLOCATION (lock): At the cash machine inside the unattended shop, beside the retail aisles. Daylight enters through the intact storefront glazing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside 찰리's punching arm, slightly below the impact and looking obliquely upward across the ATM front, without crossing the strike axis. His forearm runs from lower-left into the embedded metal fist near center-right, with the breached panel occupying no more than two-fifths of the frame and a soft margin of the shop display retaining context. Exclude his head and torso, concentrating on the completed forward extension and contact point before withdrawal or falling money begins.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: ATM front (Pierced by 찰리's fist, which remains embedded) — The front and a narrow adjoining side are viewed obliquely from below; used as Locates the impact without enlarging the machine beyond believable scale; Shop display (Present within the shop) — Only an indistinct edge remains beyond the punching arm; used as Provides a small contextual margin around the isolated action detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the shop and controlled highlights articulate the metal fist and breached panel without invented sparks or dramatic light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM now has a fresh hole punched through its front; cash has not yet begun pouring out at this impact. The collected duffel and backpack contain supplies including a disposable phone, a map, food, drinks, and the remaining COPD medicine. 찰리: He stands at the ATM with his fist embedded in its front. The blanket has not yet been handed over.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기의 전면 패널을 정통으로 꿰뚫고 들어간 파괴적인 찰나.\n\nLOCATION (lock): At the cash machine inside the unattended shop, beside the retail aisles. Daylight enters through the intact storefront glazing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside 찰리's punching arm, slightly below the impact and looking obliquely upward across the ATM front, without crossing the strike axis. His forearm runs from lower-left into the embedded metal fist near center-right, with the breached panel occupying no more than two-fifths of the frame and a soft margin of the shop display retaining context. Exclude his head and torso, concentrating on the completed forward extension and contact point before withdrawal or falling money begins.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: ATM front (Pierced by 찰리's fist, which remains embedded) — The front and a narrow adjoining side are viewed obliquely from below; used as Locates the impact without enlarging the machine beyond believable scale; Shop display (Present within the shop) — Only an indistinct edge remains beyond the punching arm; used as Provides a small contextual margin around the isolated action detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the shop and controlled highlights articulate the metal fist and breached panel without invented sparks or dramatic light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM now has a fresh hole punched through its front; cash has not yet begun pouring out at this impact. The collected duffel and backpack contain supplies including a disposable phone, a map, food, drinks, and the remaining COPD medicine. 찰리: He stands at the ATM with his fist embedded in its front. The blanket has not yet been handed over.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 금속 주먹이 현금인출기의 전면 패널을 정통으로 꿰뚫고 들어간 파괴적인 찰나.\n\nLOCATION (lock): At the cash machine inside the unattended shop, beside the retail aisles. Daylight enters through the intact storefront glazing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold beside 찰리's punching arm, slightly below the impact and looking obliquely upward across the ATM front, without crossing the strike axis. His forearm runs from lower-left into the embedded metal fist near center-right, with the breached panel occupying no more than two-fifths of the frame and a soft margin of the shop display retaining context. Exclude his head and torso, concentrating on the completed forward extension and contact point before withdrawal or falling money begins.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: ATM front (Pierced by 찰리's fist, which remains embedded) — The front and a narrow adjoining side are viewed obliquely from below; used as Locates the impact without enlarging the machine beyond believable scale; Shop display (Present within the shop) — Only an indistinct edge remains beyond the punching arm; used as Provides a small contextual margin around the isolated action detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient illumination appropriate to the shop and controlled highlights articulate the metal fist and breached panel without invented sparks or dramatic light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM now has a fresh hole punched through its front; cash has not yet begun pouring out at this impact. The collected duffel and backpack contain supplies including a disposable phone, a map, food, drinks, and the remaining COPD medicine. 찰리: He stands at the ATM with his fist embedded in its front. The blanket has not yet been handed over.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S41sh18__bgfirst_bg.png",
     "asset_id": "55d50b19-ec48-4a42-9b3b-967f4c27b608",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S41sh18.png",
     "asset_id": "9a6bb787-0846-4533-8eca-f856ce454482",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L55B01.png",
     "asset_id": "76b5f00a-f9af-40d4-81fa-1eaf208efccc",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "기계 팔이 화면의 왼쪽 중간에서 오른쪽으로 뻗어 ATM 전면에 타격을 가하고 있습니다.",
    "built_space": "편의점 내부로, 우측에 ATM 기기가 있고 좌측 배경에는 매장 진열대와 입구의 채광이 확인됩니다.",
    "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
    "hard_violations": [],
    "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
   },
   {
    "label": "B",
    "direction": "기계 팔이 화면의 왼쪽 아래에서 오른쪽 위를 향해 뻗어 ATM 전면을 꿰뚫고 있습니다.",
    "built_space": "편의점 내부로, 우측에 ATM 기기가 위치하며 좌측 배경으로 포커스 아웃된 진열대가 보입니다.",
    "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
    "hard_violations": [],
    "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 카메라 앵글(약간 아래에서 올려다보는 구도)과 팔의 방향(왼쪽 아래에서 시작해 오른쪽 중앙으로 향함)을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "기계 팔이 왼쪽 중간에서 시작해 수평에 가깝게 뻗어 있어, 프롬프트가 요구한 구도와 팔의 방향 지시를 따르지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "기계 팔이 화면의 왼쪽 중간에서 오른쪽으로 뻗어 ATM 전면에 타격을 가하고 있습니다.",
        "built_space": "편의점 내부로, 우측에 ATM 기기가 있고 좌측 배경에는 매장 진열대와 입구의 채광이 확인됩니다.",
        "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
        "hard_violations": [],
        "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
       },
       {
        "label": "B",
        "direction": "기계 팔이 화면의 왼쪽 아래에서 오른쪽 위를 향해 뻗어 ATM 전면을 꿰뚫고 있습니다.",
        "built_space": "편의점 내부로, 우측에 ATM 기기가 위치하며 좌측 배경으로 포커스 아웃된 진열대가 보입니다.",
        "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
        "hard_violations": [],
        "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 카메라 앵글(약간 아래에서 올려다보는 구도)과 팔의 방향(왼쪽 아래에서 시작해 오른쪽 중앙으로 향함)을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "기계 팔이 왼쪽 중간에서 시작해 수평에 가깝게 뻗어 있어, 프롬프트가 요구한 구도와 팔의 방향 지시를 따르지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "기계 팔이 화면의 왼쪽 중간에서 오른쪽으로 뻗어 ATM 전면에 타격을 가하고 있습니다.",
        "built_space": "편의점 내부로, 우측에 ATM 기기가 있고 좌측 배경에는 매장 진열대와 입구의 채광이 확인됩니다.",
        "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
        "hard_violations": [],
        "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
       },
       {
        "label": "B",
        "direction": "기계 팔이 화면의 왼쪽 아래에서 오른쪽 위를 향해 뻗어 ATM 전면을 꿰뚫고 있습니다.",
        "built_space": "편의점 내부로, 우측에 ATM 기기가 위치하며 좌측 배경으로 포커스 아웃된 진열대가 보입니다.",
        "entities": "찰리의 샌드 베이지색 기계 장갑판 팔과 주먹, 파손된 ATM 전면 패널이 보입니다.",
        "hard_violations": [],
        "physics": "기계 팔은 화면 밖의 본체에 의해 지탱되며 물리적으로 ATM에 단단히 박혀 있는 상태입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "주먹이 전면을 관통한 접점을 더 밀착해 보여주고 배경도 더 흐리지만, 전완의 좌하단 진입과 몸통 제외, 최소한의 매장 여백이라는 구도 지시는 충족하지 못한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "관통 동작과 장소는 맞지만 팔이 좌상단에서 내려오고 가슴 원자로까지 노출되며, ATM 상부와 매장을 넓게 보여줘 지정된 접점 인서트에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "금속 전완이 화면 왼쪽 중상단에서 오른쪽 중앙의 ATM 구멍으로 이어지고, 접힌 손가락과 주먹 앞부분이 전면 패널 안에 박혀 있다. 타격 대상은 정확히 ATM 전면이다. 다만 지정된 좌하단에서 중앙 오른쪽으로 올라가는 전완 방향은 아니다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "오른쪽에 ATM 한 대의 전면과 좁은 측면이 보이며 화면·키패드·카드 투입부가 위에, 파손부가 아래에 있다. 왼쪽 배경에는 상품 진열대 두 구간과 온전한 유리 출입구, 천장 조명이 보인다. 회색 ATM과 생활용품 진열은 장소 참조와 부합한다. 카메라는 접점 아래에서 비스듬히 올려다보지만, ATM이 화면 오른쪽 절반가량을 차지하고 매장도 단순한 흐릿한 가장자리보다 넓게 남는다. 머리는 없지만 왼쪽 가장자리에 몸체 장갑 일부가 걸친다.",
        "entities": "보이는 팔과 주먹은 샌드 베이지 장갑판, 검은 기계 관절, 마모된 금속 표면을 갖춰 찰리의 비인간 기계 신체와 맞는다. ATM 전면에는 새로 벌어진 구멍과 내부 부품이 보인다. 현금, 불꽃, 다른 인물은 없다. 얼굴과 가슴 표식은 판독할 수 없으며, 가방과 의약품 등은 이 인서트에서 확인 대상이 아니다.",
        "hard_violations": [],
        "physics": "주먹은 손목과 전완의 기계 관절에 연결되어 있고 팔은 화면 밖 몸체 쪽으로 이어진다. 주먹이 패널 구멍에 계속 맞물려 있어 타격 후 회수 전 상태로 읽힌다. 찢어진 판재는 가장자리에 붙어 있으며, 지지 없이 떠 있는 물체는 보이지 않는다. 발이 잘려 있어 지상 지지는 확인할 수 없지만 팔 자체가 공중에 분리된 상태는 아니다."
       },
       {
        "label": "B",
        "direction": "팔이 화면 좌상단의 어깨 쪽에서 우하향해 중앙 오른쪽의 ATM 전면에 박힌 주먹으로 이어진다. 타격 축은 실제 파손 구멍에 도달하지만, 요구한 좌하단 진입 방향과 반대다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "오른쪽에는 ATM 한 대의 상단 간판, 화면, 조작부, 파손된 하부 전면과 측면이 함께 보인다. 왼쪽에는 긴 중앙 상품 진열대, 출입구 옆 신발·가방 진열, 뒤쪽 냉장 진열 일부와 온전한 유리 출입구가 보인다. 장소의 기본 배치와 재질은 참조와 잘 대응한다. 그러나 낮은 시점에서 기계 상부와 바닥·천장까지 넓게 담아 접점 인서트보다 장면 설명 구도가 되고, 진열대도 작은 흐릿한 여백에 그치지 않는다. 왼쪽에는 어깨와 가슴 일부가 노출된다.",
        "entities": "베이지 장갑과 원형 관절의 거대한 기계 팔은 찰리와 맞으며, 왼쪽 가장자리에 푸른 가슴 원자로 일부도 보인다. 이는 정체성에는 맞지만 몸통을 제외하라는 지시에는 어긋난다. ATM과 전면의 관통 구멍은 명확하고 상단의 은행공용 ATM 표기도 장소 참조와 대응한다. 다른 사람, 쏟아지는 현금, 불꽃은 없다. 가방과 내용물은 지정된 상세 프레임 밖 항목이다.",
        "hard_violations": [],
        "physics": "어깨에서 팔 관절과 손목을 거쳐 주먹까지 기계적으로 연결되어 있고, 주먹 앞부분은 찢어진 패널 내부에 걸려 있다. 판재 조각들은 구멍 둘레에 붙어 있어 무지지 부유가 보이지 않는다. 팔을 전진시켜 전면을 뚫은 상태로 성립하지만, 완성된 전완과 접점만 고립시키기보다 상완과 몸통까지 보여준다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "주먹이 전면을 관통한 접점을 더 밀착해 보여주고 배경도 더 흐리지만, 전완의 좌하단 진입과 몸통 제외, 최소한의 매장 여백이라는 구도 지시는 충족하지 못한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "관통 동작과 장소는 맞지만 팔이 좌상단에서 내려오고 가슴 원자로까지 노출되며, ATM 상부와 매장을 넓게 보여줘 지정된 접점 인서트에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "금속 전완이 화면 왼쪽 중상단에서 오른쪽 중앙의 ATM 구멍으로 이어지고, 접힌 손가락과 주먹 앞부분이 전면 패널 안에 박혀 있다. 타격 대상은 정확히 ATM 전면이다. 다만 지정된 좌하단에서 중앙 오른쪽으로 올라가는 전완 방향은 아니다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "오른쪽에 ATM 한 대의 전면과 좁은 측면이 보이며 화면·키패드·카드 투입부가 위에, 파손부가 아래에 있다. 왼쪽 배경에는 상품 진열대 두 구간과 온전한 유리 출입구, 천장 조명이 보인다. 회색 ATM과 생활용품 진열은 장소 참조와 부합한다. 카메라는 접점 아래에서 비스듬히 올려다보지만, ATM이 화면 오른쪽 절반가량을 차지하고 매장도 단순한 흐릿한 가장자리보다 넓게 남는다. 머리는 없지만 왼쪽 가장자리에 몸체 장갑 일부가 걸친다.",
        "entities": "보이는 팔과 주먹은 샌드 베이지 장갑판, 검은 기계 관절, 마모된 금속 표면을 갖춰 찰리의 비인간 기계 신체와 맞는다. ATM 전면에는 새로 벌어진 구멍과 내부 부품이 보인다. 현금, 불꽃, 다른 인물은 없다. 얼굴과 가슴 표식은 판독할 수 없으며, 가방과 의약품 등은 이 인서트에서 확인 대상이 아니다.",
        "hard_violations": [],
        "physics": "주먹은 손목과 전완의 기계 관절에 연결되어 있고 팔은 화면 밖 몸체 쪽으로 이어진다. 주먹이 패널 구멍에 계속 맞물려 있어 타격 후 회수 전 상태로 읽힌다. 찢어진 판재는 가장자리에 붙어 있으며, 지지 없이 떠 있는 물체는 보이지 않는다. 발이 잘려 있어 지상 지지는 확인할 수 없지만 팔 자체가 공중에 분리된 상태는 아니다."
       },
       {
        "label": "A",
        "direction": "팔이 화면 좌상단의 어깨 쪽에서 우하향해 중앙 오른쪽의 ATM 전면에 박힌 주먹으로 이어진다. 타격 축은 실제 파손 구멍에 도달하지만, 요구한 좌하단 진입 방향과 반대다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "오른쪽에는 ATM 한 대의 상단 간판, 화면, 조작부, 파손된 하부 전면과 측면이 함께 보인다. 왼쪽에는 긴 중앙 상품 진열대, 출입구 옆 신발·가방 진열, 뒤쪽 냉장 진열 일부와 온전한 유리 출입구가 보인다. 장소의 기본 배치와 재질은 참조와 잘 대응한다. 그러나 낮은 시점에서 기계 상부와 바닥·천장까지 넓게 담아 접점 인서트보다 장면 설명 구도가 되고, 진열대도 작은 흐릿한 여백에 그치지 않는다. 왼쪽에는 어깨와 가슴 일부가 노출된다.",
        "entities": "베이지 장갑과 원형 관절의 거대한 기계 팔은 찰리와 맞으며, 왼쪽 가장자리에 푸른 가슴 원자로 일부도 보인다. 이는 정체성에는 맞지만 몸통을 제외하라는 지시에는 어긋난다. ATM과 전면의 관통 구멍은 명확하고 상단의 은행공용 ATM 표기도 장소 참조와 대응한다. 다른 사람, 쏟아지는 현금, 불꽃은 없다. 가방과 내용물은 지정된 상세 프레임 밖 항목이다.",
        "hard_violations": [],
        "physics": "어깨에서 팔 관절과 손목을 거쳐 주먹까지 기계적으로 연결되어 있고, 주먹 앞부분은 찢어진 패널 내부에 걸려 있다. 판재 조각들은 구멍 둘레에 붙어 있어 무지지 부유가 보이지 않는다. 팔을 전진시켜 전면을 뚫은 상태로 성립하지만, 완성된 전완과 접점만 고립시키기보다 상완과 몸통까지 보여준다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.292,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.292,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1292
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 카메라 앵글(약간 아래에서 올려다보는 구도)과 팔의 방향(왼쪽 아래에서 시작해 오른쪽 중앙으로 향함)을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1292,
    "verdict_ko": "기계 팔이 왼쪽 중간에서 시작해 수평에 가깝게 뻗어 있어, 프롬프트가 요구한 구도와 팔의 방향 지시를 따르지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L55B01.png",
    "asset_id": "76b5f00a-f9af-40d4-81fa-1eaf208efccc",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ac2-e375-73b4-b648-47219248375c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S41sh18__bgfirst_bg.png",
   "bg_asset_id": "55d50b19-ec48-4a42-9b3b-967f4c27b608",
   "bg_record_key": "S41sh18::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S41sh33::signage": {
  "fp": "a1444ad5e705bdf1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh33": {
  "input_fingerprint": "e63bcff5bbc38941",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 진열대 뒤에서 달려나온 앰버가 기쁜 표정으로 라울의 품에 와락 안겨 있는 전신 구도.\n\nLOCATION (lock): In the aisle beside the display shelves inside the unattended shop, lit by daylight through the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track at lower-torso height on the established side of 앰버's approach, leaving both figures unobstructed from head to foot. Place 앰버 center-left with her back three-quarter to the camera as her forward momentum closes into the embrace, while the shorter, ponytailed 라울 on center-right shifts his weight to receive her. 앰버 closes her eyes against him and 라울 looks down toward their embrace, with the display edge at far left preserving the direction she has just run from.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Shop display (The hiding place 앰버 has just left) — Its side edge is visible at the far left, clear of both figures; used as Marks the origin of her approach without obstructing the full-body embrace; Shop entrance (Opened by 라울 on arrival) — Visible obliquely beyond the reunion area; used as Retains the spatial context of 라울's arrival in the background.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient illumination and restrained contrast unchanged, allowing bodily contact and relieved expressions to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM retains its punched-open front, and cash has been collected into a bag. The packed supplies and COPD medicine remain gathered for departure. 앰버: She has emerged from behind the display shelving, still wearing the packed bag. 라울: He is newly inside the shop, with his short stature and ponytail established at the entrance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 진열대 뒤에서 달려나온 앰버가 기쁜 표정으로 라울의 품에 와락 안겨 있는 전신 구도.\n\nLOCATION (lock): In the aisle beside the display shelves inside the unattended shop, lit by daylight through the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track at lower-torso height on the established side of 앰버's approach, leaving both figures unobstructed from head to foot. Place 앰버 center-left with her back three-quarter to the camera as her forward momentum closes into the embrace, while the shorter, ponytailed 라울 on center-right shifts his weight to receive her. 앰버 closes her eyes against him and 라울 looks down toward their embrace, with the display edge at far left preserving the direction she has just run from.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Shop display (The hiding place 앰버 has just left) — Its side edge is visible at the far left, clear of both figures; used as Marks the origin of her approach without obstructing the full-body embrace; Shop entrance (Opened by 라울 on arrival) — Visible obliquely beyond the reunion area; used as Retains the spatial context of 라울's arrival in the background.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient illumination and restrained contrast unchanged, allowing bodily contact and relieved expressions to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM retains its punched-open front, and cash has been collected into a bag. The packed supplies and COPD medicine remain gathered for departure. 앰버: She has emerged from behind the display shelving, still wearing the packed bag. 라울: He is newly inside the shop, with his short stature and ponytail established at the entrance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 진열대 뒤에서 달려나온 앰버가 기쁜 표정으로 라울의 품에 와락 안겨 있는 전신 구도.\n\nLOCATION (lock): In the aisle beside the display shelves inside the unattended shop, lit by daylight through the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track at lower-torso height on the established side of 앰버's approach, leaving both figures unobstructed from head to foot. Place 앰버 center-left with her back three-quarter to the camera as her forward momentum closes into the embrace, while the shorter, ponytailed 라울 on center-right shifts his weight to receive her. 앰버 closes her eyes against him and 라울 looks down toward their embrace, with the display edge at far left preserving the direction she has just run from.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Shop display (The hiding place 앰버 has just left) — Its side edge is visible at the far left, clear of both figures; used as Marks the origin of her approach without obstructing the full-body embrace; Shop entrance (Opened by 라울 on arrival) — Visible obliquely beyond the reunion area; used as Retains the spatial context of 라울's arrival in the background.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the shop's neutral ambient illumination and restrained contrast unchanged, allowing bodily contact and relieved expressions to supply the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM retains its punched-open front, and cash has been collected into a bag. The packed supplies and COPD medicine remain gathered for departure. 앰버: She has emerged from behind the display shelving, still wearing the packed bag. 라울: He is newly inside the shop, with his short stature and ponytail established at the entrance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 라울에게 안겨 얼굴을 묻고 눈을 감고 있으며, 라울은 시선을 아래로 향해 앰버를 내려다봄.",
    "built_space": "우측에 부서진 ATM, 좌측 가장자리에 진열대, 후경에 매장 출입문이 레퍼런스와 일치하는 구도로 배치됨.",
    "entities": "앰버(작업복, 공구 벨트, 백팩 일치 / 방진 마스크 누락), 라울(복장 일치). 두 인물 모두 지시된 나이대와 피부색을 반영했으나, 라울이 앰버보다 작아야 한다는 설정을 무시하고 더 크게 묘사됨.",
    "hard_violations": [],
    "physics": "앰버의 한쪽 다리가 뒤로 들려 있고, 두 발로 선 라울이 앰버의 체중을 자연스럽게 지탱하며 안고 있음."
   },
   {
    "label": "B",
    "direction": "앰버는 눈을 감은 채 라울을 끌어안고 있고, 라울은 고개를 숙여 앰버를 바라봄.",
    "built_space": "화면 우측의 파손된 ATM, 좌측의 진열대, 뒤쪽의 유리문 등 지정된 공간 요소들이 올바른 위치에 있음.",
    "entities": "앰버(백팩, 작업복 착용 / 방진 마스크 누락), 라울(복장 일치). A와 마찬가지로 라울이 앰버보다 확연히 크게 묘사되어 인물 설정을 위반함.",
    "hard_violations": [],
    "physics": "앰버가 한 발을 든 상태로 라울에게 안겨 있으며, 라울이 지면에 두 발을 딛고 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 후보 모두 라울이 앰버보다 작아야 한다는 설정과 앰버의 하이테크 마스크를 누락했지만, A가 ATM 파손부 형태를 레퍼런스에 더 가깝게 구현하여 근소하게 우세합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프롬프트에 명시된 신장 조건(라울이 더 작음)과 앰버의 마스크 착용 지시를 모두 위반했으며, 앰버의 오른손 묘사가 다소 뭉개져 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 라울에게 안겨 얼굴을 묻고 눈을 감고 있으며, 라울은 시선을 아래로 향해 앰버를 내려다봄.",
        "built_space": "우측에 부서진 ATM, 좌측 가장자리에 진열대, 후경에 매장 출입문이 레퍼런스와 일치하는 구도로 배치됨.",
        "entities": "앰버(작업복, 공구 벨트, 백팩 일치 / 방진 마스크 누락), 라울(복장 일치). 두 인물 모두 지시된 나이대와 피부색을 반영했으나, 라울이 앰버보다 작아야 한다는 설정을 무시하고 더 크게 묘사됨.",
        "hard_violations": [],
        "physics": "앰버의 한쪽 다리가 뒤로 들려 있고, 두 발로 선 라울이 앰버의 체중을 자연스럽게 지탱하며 안고 있음."
       },
       {
        "label": "B",
        "direction": "앰버는 눈을 감은 채 라울을 끌어안고 있고, 라울은 고개를 숙여 앰버를 바라봄.",
        "built_space": "화면 우측의 파손된 ATM, 좌측의 진열대, 뒤쪽의 유리문 등 지정된 공간 요소들이 올바른 위치에 있음.",
        "entities": "앰버(백팩, 작업복 착용 / 방진 마스크 누락), 라울(복장 일치). A와 마찬가지로 라울이 앰버보다 확연히 크게 묘사되어 인물 설정을 위반함.",
        "hard_violations": [],
        "physics": "앰버가 한 발을 든 상태로 라울에게 안겨 있으며, 라울이 지면에 두 발을 딛고 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 후보 모두 라울이 앰버보다 작아야 한다는 설정과 앰버의 하이테크 마스크를 누락했지만, A가 ATM 파손부 형태를 레퍼런스에 더 가깝게 구현하여 근소하게 우세합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프롬프트에 명시된 신장 조건(라울이 더 작음)과 앰버의 마스크 착용 지시를 모두 위반했으며, 앰버의 오른손 묘사가 다소 뭉개져 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 라울에게 안겨 얼굴을 묻고 눈을 감고 있으며, 라울은 시선을 아래로 향해 앰버를 내려다봄.",
        "built_space": "우측에 부서진 ATM, 좌측 가장자리에 진열대, 후경에 매장 출입문이 레퍼런스와 일치하는 구도로 배치됨.",
        "entities": "앰버(작업복, 공구 벨트, 백팩 일치 / 방진 마스크 누락), 라울(복장 일치). 두 인물 모두 지시된 나이대와 피부색을 반영했으나, 라울이 앰버보다 작아야 한다는 설정을 무시하고 더 크게 묘사됨.",
        "hard_violations": [],
        "physics": "앰버의 한쪽 다리가 뒤로 들려 있고, 두 발로 선 라울이 앰버의 체중을 자연스럽게 지탱하며 안고 있음."
       },
       {
        "label": "B",
        "direction": "앰버는 눈을 감은 채 라울을 끌어안고 있고, 라울은 고개를 숙여 앰버를 바라봄.",
        "built_space": "화면 우측의 파손된 ATM, 좌측의 진열대, 뒤쪽의 유리문 등 지정된 공간 요소들이 올바른 위치에 있음.",
        "entities": "앰버(백팩, 작업복 착용 / 방진 마스크 누락), 라울(복장 일치). A와 마찬가지로 라울이 앰버보다 확연히 크게 묘사되어 인물 설정을 위반함.",
        "hard_violations": [],
        "physics": "앰버가 한 발을 든 상태로 라울에게 안겨 있으며, 라울이 지면에 두 발을 딛고 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전신 포옹과 시선, 매장 연속성은 맞지만 앰버의 등 사선 구도가 덜 분명하고, 더 작아야 할 라울이 오히려 크게 보인다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앰버의 등 사선 방향과 앞으로 쏠리는 포옹 동작이 지시에 더 가깝지만, 라울의 작은 체구 조건은 여전히 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 왼쪽 앰버가 오른쪽 라울에게 몸을 기울여 안기며 눈을 감고 웃는다. 라울은 앰버의 얼굴과 두 사람의 접촉 부위를 내려다본다. 앰버의 뒤로 든 발과 왼쪽 진열대가 왼쪽에서 오른쪽으로 접근한 방향을 뒷받침한다. 다만 앰버는 지정된 등 사선보다 옆모습이 더 많이 드러난다.",
        "built_space": "왼쪽 가장자리에 전경 진열대 한 줄, 후방에 상품 진열대, 왼쪽 뒤에 유리 출입구 한 곳, 오른쪽에 전면이 파손된 ATM 한 대가 보인다. 두 아이는 그 사이 통로에 서 있고 머리부터 신발까지 진열대에 가리지 않는다. 금속 설비, 상품과 휴지 진열, 밝은 바닥 및 주간 채광은 이전 장면과 대체로 이어진다. 전신 와이드와 낮은 시점도 대체로 맞는다.",
        "entities": "사람은 두 아이뿐이다. 앰버는 밝은 피부와 금발, 어린 여자아이 외형이며 때 묻은 카키 작업복, 가죽 공구 벨트, 부푼 배낭과 부츠를 갖췄다. 라울은 짙은 피부, 묶은 곱슬머리, 어린 남자아이 외형에 낡은 티셔츠와 반바지를 착용했다. 그러나 라울이 앰버보다 키와 체격이 커 보여 명시된 상대적 크기와 다르다. 앰버의 얼굴에는 마스크가 없고 목 부분은 포옹에 가려 하이테크 방진 마스크를 확인할 수 없다. 현금과 약품은 보이지 않아 배낭 내용물을 검증할 수 없다.",
        "hard_violations": [],
        "physics": "앰버는 앞쪽 부츠 한 짝을 바닥에 딛고 반대쪽 다리를 뒤로 굽혀 들었다. 라울은 양발을 벌려 바닥에 딛고 팔로 앰버의 몸을 받는다. 접지한 발과 서로 감싼 팔이 기울어진 몸을 지지하므로 부유하는 자세가 아니다. 배낭은 어깨끈으로, 공구는 허리 벨트로 지지된다."
       },
       {
        "label": "B",
        "direction": "앰버는 중앙 왼쪽에서 등을 사선으로 보이며 오른쪽 라울의 품으로 기울어져 있다. 눈을 감고 웃으며 얼굴을 라울에게 붙이고, 라울은 포옹 부위를 내려다본다. 왼쪽 뒤로 뻗은 다리와 오른쪽으로 향한 상체가 진열대 쪽에서 달려와 안긴 진행 방향을 보여준다.",
        "built_space": "왼쪽 끝의 전경 진열대 한 줄은 두 사람을 가리지 않는다. 후방 상품 진열대와 왼쪽 뒤의 비스듬한 유리 출입구 한 곳, 오른쪽의 파손된 ATM 한 대가 통로를 둘러싼다. 두 아이의 머리와 발이 모두 포함된 낮은 전신 와이드이며, 앰버의 등 방향이 A보다 명확하다. 이전 장면의 금속 ATM, 진열 상품, 바닥 재질과 중성적인 주간 조명을 대체로 유지한다.",
        "entities": "앰버와 라울에 해당하는 두 아이만 보인다. 앰버의 금발과 밝은 피부, 카키 작업복, 공구 벨트, 배낭은 참조의 주요 특징에 부합한다. 라울의 짙은 피부와 묶은 곱슬머리, 낡은 티셔츠와 반바지도 부합한다. 두 얼굴은 어린아이로 읽히지만 라울이 더 크고 높게 보여 작은 체구라는 조건과 어긋난다. 앰버의 얼굴은 마스크로 덮이지 않았고 가려진 목 부분에서도 마스크 존재를 확인할 수 없다. 현금과 COPD 약품은 별도로 식별되지 않으며 배낭 내부는 보이지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 앞쪽 부츠는 바닥에 확실히 닿고 뒤쪽 다리는 뒤로 뻗어 발끝이 바닥 가까이에 있다. 라울은 벌린 양발로 서서 앰버의 등과 몸통을 감싸 지지한다. 앰버의 전방 기울기는 접지한 발과 포옹의 접촉으로 성립한다. 배낭과 공구 벨트도 몸에 부착되어 있으며 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전신 포옹과 시선, 매장 연속성은 맞지만 앰버의 등 사선 구도가 덜 분명하고, 더 작아야 할 라울이 오히려 크게 보인다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 등 사선 방향과 앞으로 쏠리는 포옹 동작이 지시에 더 가깝지만, 라울의 작은 체구 조건은 여전히 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 왼쪽 앰버가 오른쪽 라울에게 몸을 기울여 안기며 눈을 감고 웃는다. 라울은 앰버의 얼굴과 두 사람의 접촉 부위를 내려다본다. 앰버의 뒤로 든 발과 왼쪽 진열대가 왼쪽에서 오른쪽으로 접근한 방향을 뒷받침한다. 다만 앰버는 지정된 등 사선보다 옆모습이 더 많이 드러난다.",
        "built_space": "왼쪽 가장자리에 전경 진열대 한 줄, 후방에 상품 진열대, 왼쪽 뒤에 유리 출입구 한 곳, 오른쪽에 전면이 파손된 ATM 한 대가 보인다. 두 아이는 그 사이 통로에 서 있고 머리부터 신발까지 진열대에 가리지 않는다. 금속 설비, 상품과 휴지 진열, 밝은 바닥 및 주간 채광은 이전 장면과 대체로 이어진다. 전신 와이드와 낮은 시점도 대체로 맞는다.",
        "entities": "사람은 두 아이뿐이다. 앰버는 밝은 피부와 금발, 어린 여자아이 외형이며 때 묻은 카키 작업복, 가죽 공구 벨트, 부푼 배낭과 부츠를 갖췄다. 라울은 짙은 피부, 묶은 곱슬머리, 어린 남자아이 외형에 낡은 티셔츠와 반바지를 착용했다. 그러나 라울이 앰버보다 키와 체격이 커 보여 명시된 상대적 크기와 다르다. 앰버의 얼굴에는 마스크가 없고 목 부분은 포옹에 가려 하이테크 방진 마스크를 확인할 수 없다. 현금과 약품은 보이지 않아 배낭 내용물을 검증할 수 없다.",
        "hard_violations": [],
        "physics": "앰버는 앞쪽 부츠 한 짝을 바닥에 딛고 반대쪽 다리를 뒤로 굽혀 들었다. 라울은 양발을 벌려 바닥에 딛고 팔로 앰버의 몸을 받는다. 접지한 발과 서로 감싼 팔이 기울어진 몸을 지지하므로 부유하는 자세가 아니다. 배낭은 어깨끈으로, 공구는 허리 벨트로 지지된다."
       },
       {
        "label": "A",
        "direction": "앰버는 중앙 왼쪽에서 등을 사선으로 보이며 오른쪽 라울의 품으로 기울어져 있다. 눈을 감고 웃으며 얼굴을 라울에게 붙이고, 라울은 포옹 부위를 내려다본다. 왼쪽 뒤로 뻗은 다리와 오른쪽으로 향한 상체가 진열대 쪽에서 달려와 안긴 진행 방향을 보여준다.",
        "built_space": "왼쪽 끝의 전경 진열대 한 줄은 두 사람을 가리지 않는다. 후방 상품 진열대와 왼쪽 뒤의 비스듬한 유리 출입구 한 곳, 오른쪽의 파손된 ATM 한 대가 통로를 둘러싼다. 두 아이의 머리와 발이 모두 포함된 낮은 전신 와이드이며, 앰버의 등 방향이 A보다 명확하다. 이전 장면의 금속 ATM, 진열 상품, 바닥 재질과 중성적인 주간 조명을 대체로 유지한다.",
        "entities": "앰버와 라울에 해당하는 두 아이만 보인다. 앰버의 금발과 밝은 피부, 카키 작업복, 공구 벨트, 배낭은 참조의 주요 특징에 부합한다. 라울의 짙은 피부와 묶은 곱슬머리, 낡은 티셔츠와 반바지도 부합한다. 두 얼굴은 어린아이로 읽히지만 라울이 더 크고 높게 보여 작은 체구라는 조건과 어긋난다. 앰버의 얼굴은 마스크로 덮이지 않았고 가려진 목 부분에서도 마스크 존재를 확인할 수 없다. 현금과 COPD 약품은 별도로 식별되지 않으며 배낭 내부는 보이지 않는다.",
        "hard_violations": [],
        "physics": "앰버의 앞쪽 부츠는 바닥에 확실히 닿고 뒤쪽 다리는 뒤로 뻗어 발끝이 바닥 가까이에 있다. 라울은 벌린 양발로 서서 앰버의 등과 몸통을 감싸 지지한다. 앰버의 전방 기울기는 접지한 발과 포옹의 접촉으로 성립한다. 배낭과 공구 벨트도 몸에 부착되어 있으며 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.657
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.657
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1657
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "두 후보 모두 라울이 앰버보다 작아야 한다는 설정과 앰버의 하이테크 마스크를 누락했지만, A가 ATM 파손부 형태를 레퍼런스에 더 가깝게 구현하여 근소하게 우세합니다."
   },
   {
    "label": "B",
    "score": 1657,
    "verdict_ko": "프롬프트에 명시된 신장 조건(라울이 더 작음)과 앰버의 마스크 착용 지시를 모두 위반했으며, 앰버의 오른손 묘사가 다소 뭉개져 어색합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S41sh18_sel.png",
    "asset_id": "d20f21a4-7a1b-48f9-9bb3-09d135a89a43",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0aca-15c7-76c0-9d25-fde389398be4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh18"
  }
 },
 "S41sh36::signage": {
  "fp": "b528dc2d7ef15824",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh36": {
  "input_fingerprint": "1caa8f47698d4d0f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울의 옅은 미소가 싹 사라지고 굳어진 얼굴에 동공이 미세하게 흔들리는 순간의 클로즈업.\n\nLOCATION (lock): In the reunion spot beside the unattended shop's display shelves, with daylight from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach from the reunion side at 라울's eye level, just outside 앰버's shoulder line, and ease to a stop on his shallow three-quarter face. His face occupies the central-right portion of the frame, with a narrow, soft fragment of 앰버's shoulder at the left edge; he remains absorbed in her face just outside the crop as his smile disappears and his pupils shift. Let camera distance be the only emphasized change, retaining the established eyeline, subject positions, and exposure.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Store display shelving (Present behind the reunion, largely outside the narrow focus plane) — Only an oblique fragment of the shelving remains visible behind 라울; used as A subdued spatial reference that keeps the facial close-up situated inside the store.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the minute eye movements without dramatizing the loss through a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 라울, 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the store shelves, remaining merchandise, and daytime interior lighting from the reference. Exclude the rushing reunion movement as a repeated event and any exterior street fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punched open, and the collected cash, bags, food, disposable phone, map, and medicine are ready to be taken away. 라울: He remains inside the shop with his ponytail, his expression now subdued. 앰버: She still wears the packed bag and presses her lips together as her expression turns sorrowful.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울의 옅은 미소가 싹 사라지고 굳어진 얼굴에 동공이 미세하게 흔들리는 순간의 클로즈업.\n\nLOCATION (lock): In the reunion spot beside the unattended shop's display shelves, with daylight from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach from the reunion side at 라울's eye level, just outside 앰버's shoulder line, and ease to a stop on his shallow three-quarter face. His face occupies the central-right portion of the frame, with a narrow, soft fragment of 앰버's shoulder at the left edge; he remains absorbed in her face just outside the crop as his smile disappears and his pupils shift. Let camera distance be the only emphasized change, retaining the established eyeline, subject positions, and exposure.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Store display shelving (Present behind the reunion, largely outside the narrow focus plane) — Only an oblique fragment of the shelving remains visible behind 라울; used as A subdued spatial reference that keeps the facial close-up situated inside the store.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the minute eye movements without dramatizing the loss through a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 라울, 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the store shelves, remaining merchandise, and daytime interior lighting from the reference. Exclude the rushing reunion movement as a repeated event and any exterior street fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punched open, and the collected cash, bags, food, disposable phone, map, and medicine are ready to be taken away. 라울: He remains inside the shop with his ponytail, his expression now subdued. 앰버: She still wears the packed bag and presses her lips together as her expression turns sorrowful.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 라울의 옅은 미소가 싹 사라지고 굳어진 얼굴에 동공이 미세하게 흔들리는 순간의 클로즈업.\n\nLOCATION (lock): In the reunion spot beside the unattended shop's display shelves, with daylight from the storefront. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach from the reunion side at 라울's eye level, just outside 앰버's shoulder line, and ease to a stop on his shallow three-quarter face. His face occupies the central-right portion of the frame, with a narrow, soft fragment of 앰버's shoulder at the left edge; he remains absorbed in her face just outside the crop as his smile disappears and his pupils shift. Let camera distance be the only emphasized change, retaining the established eyeline, subject positions, and exposure.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Store display shelving (Present behind the reunion, largely outside the narrow focus plane) — Only an oblique fragment of the shelving remains visible behind 라울; used as A subdued spatial reference that keeps the facial close-up situated inside the store.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the minute eye movements without dramatizing the loss through a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 라울, 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the store shelves, remaining merchandise, and daytime interior lighting from the reference. Exclude the rushing reunion movement as a repeated event and any exterior street fixtures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ATM remains punched open, and the collected cash, bags, food, disposable phone, map, and medicine are ready to be taken away. 라울: He remains inside the shop with his ponytail, his expression now subdued. 앰버: She still wears the packed bag and presses her lips together as her expression turns sorrowful.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
    "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
    "entities": "라울(흑인 혼혈 소년, 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습).",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
    "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
    "entities": "라울(흑인 혼혈 소년, 옷깃이 찢어진 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습, 백팩 끈 착용).",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 구도와 굳은 표정을 잘 살렸으며, 앰버의 백팩 끈과 라울의 셔츠 찢어짐 등 레퍼런스 디테일을 정확히 반영했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 시선 처리는 훌륭하나, 라울의 셔츠 디테일과 앰버의 백팩 끈 등 세부 요소의 묘사가 다소 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
        "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
        "entities": "라울(흑인 혼혈 소년, 옷깃이 찢어진 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습, 백팩 끈 착용).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       },
       {
        "label": "A",
        "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
        "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
        "entities": "라울(흑인 혼혈 소년, 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 구도와 굳은 표정을 잘 살렸으며, 앰버의 백팩 끈과 라울의 셔츠 찢어짐 등 레퍼런스 디테일을 정확히 반영했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 시선 처리는 훌륭하나, 라울의 셔츠 디테일과 앰버의 백팩 끈 등 세부 요소의 묘사가 다소 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
        "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
        "entities": "라울(흑인 혼혈 소년, 옷깃이 찢어진 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습, 백팩 끈 착용).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       },
       {
        "label": "A",
        "direction": "라울의 시선이 화면 왼쪽 밖 앰버의 얼굴을 향함.",
        "built_space": "상점 내부. 배경 왼쪽에 흐릿한 유리문, 오른쪽에 진열대 일부가 보임.",
        "entities": "라울(흑인 혼혈 소년, 낡은 녹색 셔츠, 꽁지머리), 앰버(금발 소녀, 뒷모습).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "눈높이의 얕은 사선 얼굴과 앰버를 향한 시선은 맞지만, 앰버의 머리·어깨와 진열대가 너무 넓게 들어오고 라울의 표정은 굳어지는 순간보다 차분한 관찰에 가깝다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "더 밀착된 얼굴 크기와 긴장된 눈매가 미소가 사라진 순간에 더 가깝고 진열대도 오른쪽으로 제한되지만, 앰버를 좁은 어깨 조각만 남기라는 구도 지시는 여전히 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울은 얼굴을 화면 왼쪽으로 조금 돌리고 두 눈으로 바로 앞 앰버의 얼굴을 바라본다. 렌즈를 응시하지 않는다. 앰버도 라울 쪽으로 돌아서 있으나 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 겨누거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤에 유리 출입구 일부가 있고, 중앙과 오른쪽 뒤에는 상품이 놓인 진열대가 이어진다. 오른쪽에는 선반의 수평 경계가 세 줄 정도, 중앙에는 두 줄 정도 보이며 동일 설비의 중복이라고 볼 근거는 없다. 두 아이는 출입구와 진열대 앞에서 마주한다. 낮빛과 매장 재질은 참조와 부합하지만 진열대가 단지 비스듬한 작은 조각으로 남기에는 노출 면적이 크다. 문제 될 반사상은 없다.",
        "entities": "라울은 참조와 닮은 약 10세의 갈색 피부 남자아이로, 뒤로 묶은 검은 곱슬머리와 얼룩지고 해진 녹색 티셔츠를 유지한다. 라틴계·흑인 혼혈이라는 구체적 계통은 외모만으로 확정할 수 없으나 참조 정체성과 충돌하지 않는다. 앰버는 금발과 밝은 피부, 이전 장면의 흙빛 옷과 검은 배낭 끈이 보인다. 얼굴 대부분이 가려져 얼굴 정체성이나 입술 표정은 판정할 수 없다. 두 아이 외 인물은 없다. 현금인출기, 현금, 식량, 전화기, 지도, 약과 하의는 구도 밖이므로 유지 여부를 평가하지 않는다. 라울의 미소는 없지만 동공의 미세한 흔들림 자체는 정지 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 아이의 머리는 목과 몸통에 자연스럽게 연결되고, 앰버의 배낭 끈은 어깨에 걸쳐 있다. 하체와 발은 화면 밖이므로 접지 상태는 보이지 않지만 공중에 떠 있다는 징후도 없다. 진열 상품은 선반 위에 놓여 있으며 지지 없는 물체나 불가능한 자세는 없다."
       },
       {
        "label": "B",
        "direction": "라울의 얼굴은 왼쪽으로 얕게 돌아가 있고, 두 눈은 왼쪽의 앰버 얼굴 쪽을 약간 올려다본다. 시선이 렌즈로 빠지지 않으며 눈매에 경계와 긴장이 보인다. 앰버는 라울을 향해 있지만 눈은 가려져 있다. 무기나 지향성 소품, 이동 중인 물체는 없다.",
        "built_space": "왼쪽 뒤에는 유리 출입구 한 구역과 그 앞 바닥 일부가, 오른쪽 뒤에는 상단 상자·봉지 상품 두 단·하단 대형 포장품으로 이어지는 진열대 한 구역이 보인다. 참조의 출입구와 진열대 관계 및 하단 포장품이 유지된다. 진열대는 흐릿한 오른쪽 배경으로 한정되어 A보다 요구에 가깝지만 여전히 작은 조각 이상이다. 카메라는 라울을 아주 조금 내려다보는 인상이 있어 정확한 눈높이 조건에는 약간 못 미친다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "라울은 참조와 닮은 어린 갈색 피부 남자아이이며 검은 곱슬 꽁지머리, 푸른 머리끈, 얼룩지고 구멍 난 녹색 티셔츠가 보인다. 약 10세의 외형과 참조의 얼굴 특징을 유지한다. 앰버는 금발, 밝은 피부, 흙빛 옷과 배낭 끈으로 이전 장면의 모습을 잇지만 얼굴이 가려져 구체적 얼굴 특징과 슬픈 입술 연기는 확인할 수 없다. 추가 인물은 없다. 화면 밖 현금인출기와 수거 물품, 하의는 평가 대상에서 제외한다. 라울은 입꼬리가 내려가고 눈 주위가 긴장되어 미소가 사라진 순간으로 읽히며, 홍채와 동공은 정상적인 사람 눈이다.",
        "hard_violations": [],
        "physics": "라울의 기울어진 머리는 목과 어깨가 자연스럽게 받치며, 앰버의 옷과 배낭 끈도 몸에 걸쳐 있다. 발은 잘려 있지만 몸이 떠 있거나 지지가 끊어진 모습은 없다. 배경 상품은 선반이 받친다. 정지한 대면 자세로 가능한 장면이며 해부학적·물리적 불가능성은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "눈높이의 얕은 사선 얼굴과 앰버를 향한 시선은 맞지만, 앰버의 머리·어깨와 진열대가 너무 넓게 들어오고 라울의 표정은 굳어지는 순간보다 차분한 관찰에 가깝다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "더 밀착된 얼굴 크기와 긴장된 눈매가 미소가 사라진 순간에 더 가깝고 진열대도 오른쪽으로 제한되지만, 앰버를 좁은 어깨 조각만 남기라는 구도 지시는 여전히 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "라울은 얼굴을 화면 왼쪽으로 조금 돌리고 두 눈으로 바로 앞 앰버의 얼굴을 바라본다. 렌즈를 응시하지 않는다. 앰버도 라울 쪽으로 돌아서 있으나 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 겨누거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤에 유리 출입구 일부가 있고, 중앙과 오른쪽 뒤에는 상품이 놓인 진열대가 이어진다. 오른쪽에는 선반의 수평 경계가 세 줄 정도, 중앙에는 두 줄 정도 보이며 동일 설비의 중복이라고 볼 근거는 없다. 두 아이는 출입구와 진열대 앞에서 마주한다. 낮빛과 매장 재질은 참조와 부합하지만 진열대가 단지 비스듬한 작은 조각으로 남기에는 노출 면적이 크다. 문제 될 반사상은 없다.",
        "entities": "라울은 참조와 닮은 약 10세의 갈색 피부 남자아이로, 뒤로 묶은 검은 곱슬머리와 얼룩지고 해진 녹색 티셔츠를 유지한다. 라틴계·흑인 혼혈이라는 구체적 계통은 외모만으로 확정할 수 없으나 참조 정체성과 충돌하지 않는다. 앰버는 금발과 밝은 피부, 이전 장면의 흙빛 옷과 검은 배낭 끈이 보인다. 얼굴 대부분이 가려져 얼굴 정체성이나 입술 표정은 판정할 수 없다. 두 아이 외 인물은 없다. 현금인출기, 현금, 식량, 전화기, 지도, 약과 하의는 구도 밖이므로 유지 여부를 평가하지 않는다. 라울의 미소는 없지만 동공의 미세한 흔들림 자체는 정지 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 아이의 머리는 목과 몸통에 자연스럽게 연결되고, 앰버의 배낭 끈은 어깨에 걸쳐 있다. 하체와 발은 화면 밖이므로 접지 상태는 보이지 않지만 공중에 떠 있다는 징후도 없다. 진열 상품은 선반 위에 놓여 있으며 지지 없는 물체나 불가능한 자세는 없다."
       },
       {
        "label": "A",
        "direction": "라울의 얼굴은 왼쪽으로 얕게 돌아가 있고, 두 눈은 왼쪽의 앰버 얼굴 쪽을 약간 올려다본다. 시선이 렌즈로 빠지지 않으며 눈매에 경계와 긴장이 보인다. 앰버는 라울을 향해 있지만 눈은 가려져 있다. 무기나 지향성 소품, 이동 중인 물체는 없다.",
        "built_space": "왼쪽 뒤에는 유리 출입구 한 구역과 그 앞 바닥 일부가, 오른쪽 뒤에는 상단 상자·봉지 상품 두 단·하단 대형 포장품으로 이어지는 진열대 한 구역이 보인다. 참조의 출입구와 진열대 관계 및 하단 포장품이 유지된다. 진열대는 흐릿한 오른쪽 배경으로 한정되어 A보다 요구에 가깝지만 여전히 작은 조각 이상이다. 카메라는 라울을 아주 조금 내려다보는 인상이 있어 정확한 눈높이 조건에는 약간 못 미친다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "라울은 참조와 닮은 어린 갈색 피부 남자아이이며 검은 곱슬 꽁지머리, 푸른 머리끈, 얼룩지고 구멍 난 녹색 티셔츠가 보인다. 약 10세의 외형과 참조의 얼굴 특징을 유지한다. 앰버는 금발, 밝은 피부, 흙빛 옷과 배낭 끈으로 이전 장면의 모습을 잇지만 얼굴이 가려져 구체적 얼굴 특징과 슬픈 입술 연기는 확인할 수 없다. 추가 인물은 없다. 화면 밖 현금인출기와 수거 물품, 하의는 평가 대상에서 제외한다. 라울은 입꼬리가 내려가고 눈 주위가 긴장되어 미소가 사라진 순간으로 읽히며, 홍채와 동공은 정상적인 사람 눈이다.",
        "hard_violations": [],
        "physics": "라울의 기울어진 머리는 목과 어깨가 자연스럽게 받치며, 앰버의 옷과 배낭 끈도 몸에 걸쳐 있다. 발은 잘려 있지만 몸이 떠 있거나 지지가 끊어진 모습은 없다. 배경 상품은 선반이 받친다. 정지한 대면 자세로 가능한 장면이며 해부학적·물리적 불가능성은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1875,
   "A": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "지정된 클로즈업 구도와 굳은 표정을 잘 살렸으며, 앰버의 백팩 끈과 라울의 셔츠 찢어짐 등 레퍼런스 디테일을 정확히 반영했습니다."
   },
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "구도와 시선 처리는 훌륭하나, 라울의 셔츠 디테일과 앰버의 백팩 끈 등 세부 요소의 묘사가 다소 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 라울, 앰버 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S41sh33_sel.png",
    "asset_id": "bf818815-6dea-43fd-9eac-483fc320f722",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0acf-4586-7b59-a044-1d84853ebf9f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh33"
  },
  "staged_characters_added": [
   "C05"
  ]
 },
 "S42sh2::signage": {
  "fp": "22440dfbde9fe9c2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh2::bgfirst_bg": {
  "input_fingerprint": "a139f0549c6ceb5f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반쯤 열린 보디백 사이로 창백하게 굳어 죽어 있는 구도환의 평온한 얼굴이 드러난 구도.\n\nLOCATION (lock): At the muddy outdoor entrance to the flooded refugee settlement, beside a body bag laid out for identification at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane at its initial high position beside 구도환's head, looking steeply downward into the half-open body bag rather than adopting 박철진's literal viewpoint. Place the pale, motionless face diagonally across the center, with the parted bag edges forming narrow borders and occupying less than two-fifths of the image together. Pause before the rise, keeping 박철진 outside the frame so the revelation rests entirely on 구도환's peaceful stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Body bag opening (Half-open, revealing 구도환's face) — The parted upper edges face the downward-looking camera on either side of his head; used as A narrow enclosing frame around the revealed face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination keeps the pale face legible with restrained contrast and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반쯤 열린 보디백 사이로 창백하게 굳어 죽어 있는 구도환의 평온한 얼굴이 드러난 구도.\n\nLOCATION (lock): At the muddy outdoor entrance to the flooded refugee settlement, beside a body bag laid out for identification at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane at its initial high position beside 구도환's head, looking steeply downward into the half-open body bag rather than adopting 박철진's literal viewpoint. Place the pale, motionless face diagonally across the center, with the parted bag edges forming narrow borders and occupying less than two-fifths of the image together. Pause before the rise, keeping 박철진 outside the frame so the revelation rests entirely on 구도환's peaceful stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Body bag opening (Half-open, revealing 구도환's face) — The parted upper edges face the downward-looking camera on either side of his head; used as A narrow enclosing frame around the revealed face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination keeps the pale face legible with restrained contrast and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S42sh2__bgfirst_bg.png",
  "asset_id": "ab156aa6-e228-44f1-bc2f-ba5782ec7024",
  "input_asset_ids": [
   "8e843507-bc8f-4860-8b5a-579ba2cc1239",
   "cf1ff170-e216-45ef-956e-1c0eea9b4407"
  ]
 },
 "S42sh2": {
  "input_fingerprint": "03a1242360905a5c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 보디백 사이로 창백하게 굳어 죽어 있는 구도환의 평온한 얼굴이 드러난 구도.\n\nLOCATION (lock): At the muddy outdoor entrance to the flooded refugee settlement, beside a body bag laid out for identification at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane at its initial high position beside 구도환's head, looking steeply downward into the half-open body bag rather than adopting 박철진's literal viewpoint. Place the pale, motionless face diagonally across the center, with the parted bag edges forming narrow borders and occupying less than two-fifths of the image together. Pause before the rise, keeping 박철진 outside the frame so the revelation rests entirely on 구도환's peaceful stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Body bag opening (Half-open, revealing 구도환's face) — The parted upper edges face the downward-looking camera on either side of his head; used as A narrow enclosing frame around the revealed face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination keeps the pale face legible with restrained contrast and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Gu Dohwan's dead body is enclosed in a body bag whose opened zipper exposes his face. The bag conceals his torso and limbs, and the scene text does not specify the body's underlying orientation or support.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the refugee-settlement entrance at night, a body bag lies with its zipper opened far enough to expose the face inside. 구도환: His dead body is enclosed in the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 보디백 사이로 창백하게 굳어 죽어 있는 구도환의 평온한 얼굴이 드러난 구도.\n\nLOCATION (lock): At the muddy outdoor entrance to the flooded refugee settlement, beside a body bag laid out for identification at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane at its initial high position beside 구도환's head, looking steeply downward into the half-open body bag rather than adopting 박철진's literal viewpoint. Place the pale, motionless face diagonally across the center, with the parted bag edges forming narrow borders and occupying less than two-fifths of the image together. Pause before the rise, keeping 박철진 outside the frame so the revelation rests entirely on 구도환's peaceful stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Body bag opening (Half-open, revealing 구도환's face) — The parted upper edges face the downward-looking camera on either side of his head; used as A narrow enclosing frame around the revealed face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination keeps the pale face legible with restrained contrast and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Gu Dohwan's dead body is enclosed in a body bag whose opened zipper exposes his face. The bag conceals his torso and limbs, and the scene text does not specify the body's underlying orientation or support.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the refugee-settlement entrance at night, a body bag lies with its zipper opened far enough to expose the face inside. 구도환: His dead body is enclosed in the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 보디백 사이로 창백하게 굳어 죽어 있는 구도환의 평온한 얼굴이 드러난 구도.\n\nLOCATION (lock): At the muddy outdoor entrance to the flooded refugee settlement, beside a body bag laid out for identification at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane at its initial high position beside 구도환's head, looking steeply downward into the half-open body bag rather than adopting 박철진's literal viewpoint. Place the pale, motionless face diagonally across the center, with the parted bag edges forming narrow borders and occupying less than two-fifths of the image together. Pause before the rise, keeping 박철진 outside the frame so the revelation rests entirely on 구도환's peaceful stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Body bag opening (Half-open, revealing 구도환's face) — The parted upper edges face the downward-looking camera on either side of his head; used as A narrow enclosing frame around the revealed face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination keeps the pale face legible with restrained contrast and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Gu Dohwan's dead body is enclosed in a body bag whose opened zipper exposes his face. The bag conceals his torso and limbs, and the scene text does not specify the body's underlying orientation or support.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the refugee-settlement entrance at night, a body bag lies with its zipper opened far enough to expose the face inside. 구도환: His dead body is enclosed in the opened body bag.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 구도환 (한국인 남성, 50대의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S42sh2__bgfirst_bg.png",
     "asset_id": "ab156aa6-e228-44f1-bc2f-ba5782ec7024",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S42sh2.png",
     "asset_id": "8e843507-bc8f-4860-8b5a-579ba2cc1239",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1281254>",
     "asset_id": "ef44ca91-a8e5-4b86-9566-53391731bd57",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camp_gate_b39179.png",
     "asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1281254>",
     "asset_id": "ef44ca91-a8e5-4b86-9566-53391731bd57",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 보디백 안에 있는 구도환의 얼굴을 비스듬히 아래로 내려다보고 있음.",
    "built_space": "진흙 바닥 위에 놓인 보디백이 널찍하게 열려 있어 프레임 모서리에 진흙이 보임.",
    "entities": "눈을 감은 구도환(갈색 재킷, 올리브색 티셔츠), 검은색 보디백이 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "신체는 바닥에 중력에 맞게 자연스럽게 누워 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 보디백 안에 평온하게 누워 있는 구도환의 얼굴을 직각에 가깝게 내려다보고 있음.",
    "built_space": "보디백 외부의 배경은 우측 가장자리에 아주 좁게만 보임.",
    "entities": "구도환의 얼굴 및 복장, 검은색 보디백의 질감이 지침과 레퍼런스에 잘 부합함.",
    "hard_violations": [],
    "physics": "시신이 바닥에 완전히 밀착되어 지지를 받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지침에 명시된 클로즈업 스케일과 보디백 가장자리가 좁은 테두리를 형성한다는 프레임 요구 사항을 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "보디백이 너무 넓게 열려 있고 주변 진흙 바닥이 프레임의 상당 부분을 차지하여, 명시된 프레이밍 비율을 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 보디백 안에 있는 구도환의 얼굴을 비스듬히 아래로 내려다보고 있음.",
        "built_space": "진흙 바닥 위에 놓인 보디백이 널찍하게 열려 있어 프레임 모서리에 진흙이 보임.",
        "entities": "눈을 감은 구도환(갈색 재킷, 올리브색 티셔츠), 검은색 보디백이 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "신체는 바닥에 중력에 맞게 자연스럽게 누워 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 보디백 안에 평온하게 누워 있는 구도환의 얼굴을 직각에 가깝게 내려다보고 있음.",
        "built_space": "보디백 외부의 배경은 우측 가장자리에 아주 좁게만 보임.",
        "entities": "구도환의 얼굴 및 복장, 검은색 보디백의 질감이 지침과 레퍼런스에 잘 부합함.",
        "hard_violations": [],
        "physics": "시신이 바닥에 완전히 밀착되어 지지를 받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지침에 명시된 클로즈업 스케일과 보디백 가장자리가 좁은 테두리를 형성한다는 프레임 요구 사항을 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "보디백이 너무 넓게 열려 있고 주변 진흙 바닥이 프레임의 상당 부분을 차지하여, 명시된 프레이밍 비율을 벗어났습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 보디백 안에 있는 구도환의 얼굴을 비스듬히 아래로 내려다보고 있음.",
        "built_space": "진흙 바닥 위에 놓인 보디백이 널찍하게 열려 있어 프레임 모서리에 진흙이 보임.",
        "entities": "눈을 감은 구도환(갈색 재킷, 올리브색 티셔츠), 검은색 보디백이 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "신체는 바닥에 중력에 맞게 자연스럽게 누워 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 보디백 안에 평온하게 누워 있는 구도환의 얼굴을 직각에 가깝게 내려다보고 있음.",
        "built_space": "보디백 외부의 배경은 우측 가장자리에 아주 좁게만 보임.",
        "entities": "구도환의 얼굴 및 복장, 검은색 보디백의 질감이 지침과 레퍼런스에 잘 부합함.",
        "hard_violations": [],
        "physics": "시신이 바닥에 완전히 밀착되어 지지를 받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "중앙을 대각선으로 채운 얼굴 클로즈업과 참조 인물의 외모가 더 충실하지만, 보디백 양옆이 지정된 좁은 테두리보다 넓고 가슴 일부까지 드러난다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "평온한 시신과 젖은 야간 지면은 맞지만, 얼굴보다 보디백과 상반신을 넓게 보여 주어 지정된 얼굴 중심 클로즈업에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈이 감겨 있어 응시 대상은 없다. 얼굴은 위쪽 카메라를 향하고, 정수리에서 턱으로 이어지는 축이 화면 오른쪽 위에서 왼쪽 아래로 기울어 중앙을 대각선으로 차지한다. 카메라는 열린 보디백 안을 가파르게 내려다본다.",
        "built_space": "보디백 하나의 열린 지퍼 양쪽이 머리를 둘러싸고 있다. 가장자리 밖으로 젖은 지면이 조금 보이며, 문·울타리 등 고정 시설은 클로즈업 밖이라 장소의 정확한 건축적 일치는 확인할 수 없다. 양옆 덮개는 합계가 화면의 약 2/5에 이르거나 넘어 보여, 요구한 좁은 테두리에는 미달한다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명으로, 한국인 50대 구도환의 참조 얼굴과 머리 모양에 가깝다. 창백한 피부, 감긴 눈과 이완된 입이 평온한 죽음의 표현에 부합한다. 갈색 재킷과 올리브색 셔츠도 참조와 맞지만, 얼굴뿐 아니라 목과 윗가슴까지 노출된다. 검은 지퍼식 보디백은 실제 방수 소재로 보이며, 다른 사람이나 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수가 보디백 내부의 주름진 바닥에 놓여 있고 목도 몸통으로 자연스럽게 이어진다. 머리를 스스로 들거나 공중에 떠 있는 징후는 없다. 보디백은 지면에 놓여 있으며 젖은 덮개도 아래로 처져 있다. 팔다리는 프레임에 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 눈이 감겨 있어 응시 대상은 없다. 얼굴은 위를 향하며 머리 축은 화면 왼쪽 위에서 오른쪽 아래로 기울어 있다. 카메라는 열린 보디백을 내려다보지만, 얼굴이 A보다 작고 주변 개구부와 가슴이 더 많이 포함된다.",
        "built_space": "젖은 진흙 지면 위에 보디백 하나가 놓여 있고 열린 지퍼가 머리와 상부 가슴 둘레에 넓은 개구부를 만든다. 고정 시설은 보이지 않아 참조 입구의 정확한 위치는 확인할 수 없다. 보디백 덮개가 화면의 상당 부분을 차지해 양쪽 합계 2/5 미만의 좁은 테두리 조건을 충족하지 못한다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명이며 참조와 대체로 비슷하지만, 얼굴 윤곽과 머리 정돈은 A보다 차이가 보인다. 창백한 얼굴과 감긴 눈, 긴장 없는 표정은 요구한 평온한 시신에 맞는다. 갈색 재킷과 올리브색 셔츠가 보이고, 가슴과 어깨까지 상당히 드러나 몸통을 감춘 상태보다 개방 범위가 넓다. 검은 보디백 외에 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 몸통은 지면에 놓인 보디백 내부에 받쳐져 있다. 머리나 목이 능동적으로 들려 있다는 뚜렷한 징후는 없다. 열린 덮개는 지면과 내부 신체 위에 접혀 놓이고, 물방울과 지면의 반사도 자연스럽다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "중앙을 대각선으로 채운 얼굴 클로즈업과 참조 인물의 외모가 더 충실하지만, 보디백 양옆이 지정된 좁은 테두리보다 넓고 가슴 일부까지 드러난다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "평온한 시신과 젖은 야간 지면은 맞지만, 얼굴보다 보디백과 상반신을 넓게 보여 주어 지정된 얼굴 중심 클로즈업에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈이 감겨 있어 응시 대상은 없다. 얼굴은 위쪽 카메라를 향하고, 정수리에서 턱으로 이어지는 축이 화면 오른쪽 위에서 왼쪽 아래로 기울어 중앙을 대각선으로 차지한다. 카메라는 열린 보디백 안을 가파르게 내려다본다.",
        "built_space": "보디백 하나의 열린 지퍼 양쪽이 머리를 둘러싸고 있다. 가장자리 밖으로 젖은 지면이 조금 보이며, 문·울타리 등 고정 시설은 클로즈업 밖이라 장소의 정확한 건축적 일치는 확인할 수 없다. 양옆 덮개는 합계가 화면의 약 2/5에 이르거나 넘어 보여, 요구한 좁은 테두리에는 미달한다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명으로, 한국인 50대 구도환의 참조 얼굴과 머리 모양에 가깝다. 창백한 피부, 감긴 눈과 이완된 입이 평온한 죽음의 표현에 부합한다. 갈색 재킷과 올리브색 셔츠도 참조와 맞지만, 얼굴뿐 아니라 목과 윗가슴까지 노출된다. 검은 지퍼식 보디백은 실제 방수 소재로 보이며, 다른 사람이나 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수가 보디백 내부의 주름진 바닥에 놓여 있고 목도 몸통으로 자연스럽게 이어진다. 머리를 스스로 들거나 공중에 떠 있는 징후는 없다. 보디백은 지면에 놓여 있으며 젖은 덮개도 아래로 처져 있다. 팔다리는 프레임에 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 눈이 감겨 있어 응시 대상은 없다. 얼굴은 위를 향하며 머리 축은 화면 왼쪽 위에서 오른쪽 아래로 기울어 있다. 카메라는 열린 보디백을 내려다보지만, 얼굴이 A보다 작고 주변 개구부와 가슴이 더 많이 포함된다.",
        "built_space": "젖은 진흙 지면 위에 보디백 하나가 놓여 있고 열린 지퍼가 머리와 상부 가슴 둘레에 넓은 개구부를 만든다. 고정 시설은 보이지 않아 참조 입구의 정확한 위치는 확인할 수 없다. 보디백 덮개가 화면의 상당 부분을 차지해 양쪽 합계 2/5 미만의 좁은 테두리 조건을 충족하지 못한다.",
        "entities": "짧은 검은 머리의 중년 동아시아계 남성 한 명이며 참조와 대체로 비슷하지만, 얼굴 윤곽과 머리 정돈은 A보다 차이가 보인다. 창백한 얼굴과 감긴 눈, 긴장 없는 표정은 요구한 평온한 시신에 맞는다. 갈색 재킷과 올리브색 셔츠가 보이고, 가슴과 어깨까지 상당히 드러나 몸통을 감춘 상태보다 개방 범위가 넓다. 검은 보디백 외에 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 몸통은 지면에 놓인 보디백 내부에 받쳐져 있다. 머리나 목이 능동적으로 들려 있다는 뚜렷한 징후는 없다. 열린 덮개는 지면과 내부 신체 위에 접혀 놓이고, 물방울과 지면의 반사도 자연스럽다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지침에 명시된 클로즈업 스케일과 보디백 가장자리가 좁은 테두리를 형성한다는 프레임 요구 사항을 충실히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "보디백이 너무 넓게 열려 있고 주변 진흙 바닥이 프레임의 상당 부분을 차지하여, 명시된 프레이밍 비율을 벗어났습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camp_gate_b39179.png",
    "asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 구도환: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1281254>",
    "asset_id": "ef44ca91-a8e5-4b86-9566-53391731bd57",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ad4-cc59-7ee8-9a5b-370f34629134",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S42sh2__bgfirst_bg.png",
   "bg_asset_id": "ab156aa6-e228-44f1-bc2f-ba5782ec7024",
   "bg_record_key": "S42sh2::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "camp_gate",
   "groupbg_asset_id": "cf1ff170-e216-45ef-956e-1c0eea9b4407"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S42sh18::signage": {
  "fp": "7577023cf376b984",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh18": {
  "input_fingerprint": "7379d42833387037",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차 문을 꽉 잡은 박철진이 열린 창문 너머 차 안의 윤성찬을 향해 교활한 야심이 담긴 미소를 짓는 옆얼굴.\n\nLOCATION (lock): On the muddy ground at the refugee settlement's entrance, immediately outside the luxury sedan's open door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain outside the sedan on the established side, clear of the open door's swing, with a low, slightly upward view that runs obliquely through the open window. Frame 박철진's near-profile at left and his gripping hand below it, while 윤성찬 sits farther right inside the car; 박철진 directs his calculating smile toward 윤성찬, whose attention is fixed on the man obstructing his departure. Ease the lateral track to hold their positions, making 박철진's altered expression and eyeline the emphasis rather than adding a new camera move.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sedan door and open window (Door open and held by 박철진; window open) — The exterior side is seen obliquely, with the window opening exposing 윤성찬 inside; used as Separates exterior leverage from the seated man's confined space without obscuring the hand or smile; Sedan passenger compartment (Occupied by 윤성찬) — Seen diagonally through the open window from the established exterior side; used as Provides depth and a scale reference behind 박철진.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination holds both faces and the gripping hand readable without introducing an additional light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan is stopped at the muddy settlement entrance with its door open. The opened body bag remains at the inspection site. 윤성찬: He is getting back into the sedan with his cane, whose tip remains muddy. He retains the handkerchief taken out on arrival. 박철진: He stands beside the sedan, holding its open door.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차 문을 꽉 잡은 박철진이 열린 창문 너머 차 안의 윤성찬을 향해 교활한 야심이 담긴 미소를 짓는 옆얼굴.\n\nLOCATION (lock): On the muddy ground at the refugee settlement's entrance, immediately outside the luxury sedan's open door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain outside the sedan on the established side, clear of the open door's swing, with a low, slightly upward view that runs obliquely through the open window. Frame 박철진's near-profile at left and his gripping hand below it, while 윤성찬 sits farther right inside the car; 박철진 directs his calculating smile toward 윤성찬, whose attention is fixed on the man obstructing his departure. Ease the lateral track to hold their positions, making 박철진's altered expression and eyeline the emphasis rather than adding a new camera move.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sedan door and open window (Door open and held by 박철진; window open) — The exterior side is seen obliquely, with the window opening exposing 윤성찬 inside; used as Separates exterior leverage from the seated man's confined space without obscuring the hand or smile; Sedan passenger compartment (Occupied by 윤성찬) — Seen diagonally through the open window from the established exterior side; used as Provides depth and a scale reference behind 박철진.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination holds both faces and the gripping hand readable without introducing an additional light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan is stopped at the muddy settlement entrance with its door open. The opened body bag remains at the inspection site. 윤성찬: He is getting back into the sedan with his cane, whose tip remains muddy. He retains the handkerchief taken out on arrival. 박철진: He stands beside the sedan, holding its open door.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 차 문을 꽉 잡은 박철진이 열린 창문 너머 차 안의 윤성찬을 향해 교활한 야심이 담긴 미소를 짓는 옆얼굴.\n\nLOCATION (lock): On the muddy ground at the refugee settlement's entrance, immediately outside the luxury sedan's open door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain outside the sedan on the established side, clear of the open door's swing, with a low, slightly upward view that runs obliquely through the open window. Frame 박철진's near-profile at left and his gripping hand below it, while 윤성찬 sits farther right inside the car; 박철진 directs his calculating smile toward 윤성찬, whose attention is fixed on the man obstructing his departure. Ease the lateral track to hold their positions, making 박철진's altered expression and eyeline the emphasis rather than adding a new camera move.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sedan door and open window (Door open and held by 박철진; window open) — The exterior side is seen obliquely, with the window opening exposing 윤성찬 inside; used as Separates exterior leverage from the seated man's confined space without obscuring the hand or smile; Sedan passenger compartment (Occupied by 윤성찬) — Seen diagonally through the open window from the established exterior side; used as Provides depth and a scale reference behind 박철진.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination holds both faces and the gripping hand readable without introducing an additional light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The luxury sedan is stopped at the muddy settlement entrance with its door open. The opened body bag remains at the inspection site. 윤성찬: He is getting back into the sedan with his cane, whose tip remains muddy. He retains the handkerchief taken out on arrival. 박철진: He stands beside the sedan, holding its open door.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음.; 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박철진이 차 안의 윤성찬을 향해 시선을 던지며, 윤성찬 역시 그를 응시함.",
    "built_space": "야간의 진흙탕 배경. 열린 세단 문을 기준으로 밖에는 박철진, 안에는 윤성찬이 위치함.",
    "entities": "박철진(전투복, 완장)과 윤성찬(정장)의 외형이 참조와 일치. 지팡이와 손수건 소품 모두 존재함.",
    "hard_violations": [
     "[gpt-high] 배경의 울타리와 건물 앞에 여러 인물을 추가하여 등장인물 제한을 위반했다."
    ],
    "physics": "박철진이 열린 차 문틀을 단단히 쥐고 있으며, 윤성찬은 좌석에 체중을 싣고 안정적으로 앉아 있음."
   },
   {
    "label": "B",
    "direction": "박철진과 윤성찬이 서로 정확히 시선을 맞추고 있음.",
    "built_space": "어두운 야간의 진흙탕 배경과 열린 차 문이 올바르게 배치됨.",
    "entities": "두 인물의 인상착의는 참조와 일치하나, 윤성찬의 손수건이 묘사되지 않음.",
    "hard_violations": [
     "[gpt-high] 박철진과 윤성찬만 보여야 하는 장면에 배경 인물들을 추가했다."
    ],
    "physics": "차 문을 잡은 박철진의 손과 내부에 앉은 윤성찬의 자세에 무리가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 앵글과 인물의 시선 교환을 완벽히 구현했으며, 윤성찬이 지팡이와 손수건을 모두 쥐고 있는 세부 지시사항까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 조명은 우수하나, 프롬프트에 명시된 윤성찬의 손수건이 누락되어 A에 비해 정확도가 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진이 차 안의 윤성찬을 향해 시선을 던지며, 윤성찬 역시 그를 응시함.",
        "built_space": "야간의 진흙탕 배경. 열린 세단 문을 기준으로 밖에는 박철진, 안에는 윤성찬이 위치함.",
        "entities": "박철진(전투복, 완장)과 윤성찬(정장)의 외형이 참조와 일치. 지팡이와 손수건 소품 모두 존재함.",
        "hard_violations": [],
        "physics": "박철진이 열린 차 문틀을 단단히 쥐고 있으며, 윤성찬은 좌석에 체중을 싣고 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "박철진과 윤성찬이 서로 정확히 시선을 맞추고 있음.",
        "built_space": "어두운 야간의 진흙탕 배경과 열린 차 문이 올바르게 배치됨.",
        "entities": "두 인물의 인상착의는 참조와 일치하나, 윤성찬의 손수건이 묘사되지 않음.",
        "hard_violations": [],
        "physics": "차 문을 잡은 박철진의 손과 내부에 앉은 윤성찬의 자세에 무리가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 앵글과 인물의 시선 교환을 완벽히 구현했으며, 윤성찬이 지팡이와 손수건을 모두 쥐고 있는 세부 지시사항까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 조명은 우수하나, 프롬프트에 명시된 윤성찬의 손수건이 누락되어 A에 비해 정확도가 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박철진이 차 안의 윤성찬을 향해 시선을 던지며, 윤성찬 역시 그를 응시함.",
        "built_space": "야간의 진흙탕 배경. 열린 세단 문을 기준으로 밖에는 박철진, 안에는 윤성찬이 위치함.",
        "entities": "박철진(전투복, 완장)과 윤성찬(정장)의 외형이 참조와 일치. 지팡이와 손수건 소품 모두 존재함.",
        "hard_violations": [],
        "physics": "박철진이 열린 차 문틀을 단단히 쥐고 있으며, 윤성찬은 좌석에 체중을 싣고 안정적으로 앉아 있음."
       },
       {
        "label": "B",
        "direction": "박철진과 윤성찬이 서로 정확히 시선을 맞추고 있음.",
        "built_space": "어두운 야간의 진흙탕 배경과 열린 차 문이 올바르게 배치됨.",
        "entities": "두 인물의 인상착의는 참조와 일치하나, 윤성찬의 손수건이 묘사되지 않음.",
        "hard_violations": [],
        "physics": "차 문을 잡은 박철진의 손과 내부에 앉은 윤성찬의 자세에 무리가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "옆얼굴의 미소와 서로 향하는 시선은 맞지만, 지정된 미디엄 숏보다 인물 크롭이 타이트하고 배경에 허용되지 않은 사람들이 추가되어 탈락 사유가 있습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 박철진의 상체와 문을 쥔 손, 오른쪽 실내의 윤성찬을 담은 미디엄 구도가 더 충실하지만, 배경의 여러 추가 인물 때문에 그대로 사용할 수는 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 박철진은 고개를 오른쪽 아래로 기울여 차 안의 윤성찬을 바라보며 입꼬리를 올린다. 윤성찬도 왼쪽 위의 박철진 얼굴을 올려다본다. 두 사람의 시선 대상은 지시와 일치한다.",
        "built_space": "전경에 바깥으로 열린 차량 문 하나가 비스듬히 걸쳐 있고, 박철진의 손은 그 윗부분을 잡는다. 오른쪽에는 차체의 출입구와 뒷좌석 공간, 좌석 등받이와 머리받침이 보인다. 열린 창 쪽을 통한 외부 시점은 성립하지만, 박철진의 몸통과 윤성찬의 가슴 아래가 많이 잘려 비교적 타이트하다. 배경에는 진흙길과 임시 건축물, 여러 조명이 보이지만 이전 참고 사진의 좁은 시야만으로 이 시설들의 동일성을 확인할 수 없다.",
        "entities": "주요 인물 두 명은 중년 한국인 남성과 고령 한국인 남성으로 읽힌다. 박철진의 얼굴, 검은 머리, 남색 모자와 전투복, 붉은 완장은 참고와 대체로 맞는다. 윤성찬의 정돈된 회색 머리, 흰 콧수염, 주름과 검은 재킷·푸른 셔츠·넥타이도 참고에 가깝다. 오른쪽 아래에 굽은 손잡이 지팡이가 일부 보인다. 손수건과 지팡이 끝, 검시 장소의 시신 가방은 화면에서 확인되지 않는다. 배경 중앙에는 두 명을 비롯해 추가 사람 형체들이 보인다.",
        "hard_violations": [
         "박철진과 윤성찬만 보여야 하는 장면에 배경 인물들을 추가했다."
        ],
        "physics": "박철진의 손가락은 문 상단을 감싸며, 손목과 소매도 그의 팔로 자연스럽게 이어진다. 열린 문은 차량에 연결되어 있고 손으로 붙들려 있다. 윤성찬은 좌석에 앉아 상체를 박철진 쪽으로 돌린 자세로 읽힌다. 두 사람의 발과 지팡이 하단은 프레임 밖이므로 접지나 지팡이의 정확한 지지점은 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다."
       },
       {
        "label": "B",
        "direction": "박철진은 오른쪽 아래 실내의 윤성찬을 향해 미소를 짓고, 윤성찬은 왼쪽 위 박철진에게 시선을 고정한다. 박철진의 얼굴은 완전한 측면보다는 비스듬한 옆얼굴이며, 두 사람의 상호 응시는 지시대로 연결된다.",
        "built_space": "전경에 열린 차량 문 하나와 내려간 창의 상단 경계가 있고, 그 뒤 오른쪽으로 차체 출입구와 실내 좌석이 이어진다. 박철진은 차 밖 왼쪽에서 문을 붙들고, 윤성찬은 오른쪽 좌석 안에 자리한다. 왼쪽 상체와 그 아래 손을 함께 담는 미디엄 구도가 A보다 분명하다. 실내에는 검은 좌석 등받이와 머리받침 하나가 뚜렷하다. 배경의 울타리, 임시 건물과 진흙 바닥은 보이지만 이전 참고 사진으로 시설 배치의 일치 여부까지 검증할 수는 없다.",
        "entities": "박철진의 중년 남성 얼굴과 체격, 짧은 검은 머리, 남색 모자·전투복과 붉은 완장이 참고에 가깝다. 윤성찬도 고령 남성의 회색 머리와 흰 콧수염, 얼굴 특징, 정장 차림을 유지한다. 그의 몸 앞에 지팡이 손잡이와 옅은 손수건이 보이며, 진흙 묻은 지팡이 끝은 프레임 밖이다. 시신 가방은 보이지 않는다. 울타리와 건물 앞에는 허용된 두 사람 외에 여러 사람이 서 있다.",
        "hard_violations": [
         "배경의 울타리와 건물 앞에 여러 인물을 추가하여 등장인물 제한을 위반했다."
        ],
        "physics": "박철진의 손은 문 윗부분을 감싸고 엄지와 손가락의 접촉이 보이며, 팔은 그의 어깨와 자연스럽게 연결된다. 앞으로 기울인 상체는 서서 문을 붙드는 동작으로 가능하다. 윤성찬은 좌석의 지지를 받으며 고개를 돌리고 있다. 지팡이와 손수건은 몸 앞과 무릎 쪽에 놓인 것으로 읽히지만 하단과 손의 접촉은 가려져 있다. 명백히 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "옆얼굴의 미소와 서로 향하는 시선은 맞지만, 지정된 미디엄 숏보다 인물 크롭이 타이트하고 배경에 허용되지 않은 사람들이 추가되어 탈락 사유가 있습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 박철진의 상체와 문을 쥔 손, 오른쪽 실내의 윤성찬을 담은 미디엄 구도가 더 충실하지만, 배경의 여러 추가 인물 때문에 그대로 사용할 수는 없습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 박철진은 고개를 오른쪽 아래로 기울여 차 안의 윤성찬을 바라보며 입꼬리를 올린다. 윤성찬도 왼쪽 위의 박철진 얼굴을 올려다본다. 두 사람의 시선 대상은 지시와 일치한다.",
        "built_space": "전경에 바깥으로 열린 차량 문 하나가 비스듬히 걸쳐 있고, 박철진의 손은 그 윗부분을 잡는다. 오른쪽에는 차체의 출입구와 뒷좌석 공간, 좌석 등받이와 머리받침이 보인다. 열린 창 쪽을 통한 외부 시점은 성립하지만, 박철진의 몸통과 윤성찬의 가슴 아래가 많이 잘려 비교적 타이트하다. 배경에는 진흙길과 임시 건축물, 여러 조명이 보이지만 이전 참고 사진의 좁은 시야만으로 이 시설들의 동일성을 확인할 수 없다.",
        "entities": "주요 인물 두 명은 중년 한국인 남성과 고령 한국인 남성으로 읽힌다. 박철진의 얼굴, 검은 머리, 남색 모자와 전투복, 붉은 완장은 참고와 대체로 맞는다. 윤성찬의 정돈된 회색 머리, 흰 콧수염, 주름과 검은 재킷·푸른 셔츠·넥타이도 참고에 가깝다. 오른쪽 아래에 굽은 손잡이 지팡이가 일부 보인다. 손수건과 지팡이 끝, 검시 장소의 시신 가방은 화면에서 확인되지 않는다. 배경 중앙에는 두 명을 비롯해 추가 사람 형체들이 보인다.",
        "hard_violations": [
         "박철진과 윤성찬만 보여야 하는 장면에 배경 인물들을 추가했다."
        ],
        "physics": "박철진의 손가락은 문 상단을 감싸며, 손목과 소매도 그의 팔로 자연스럽게 이어진다. 열린 문은 차량에 연결되어 있고 손으로 붙들려 있다. 윤성찬은 좌석에 앉아 상체를 박철진 쪽으로 돌린 자세로 읽힌다. 두 사람의 발과 지팡이 하단은 프레임 밖이므로 접지나 지팡이의 정확한 지지점은 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없다."
       },
       {
        "label": "A",
        "direction": "박철진은 오른쪽 아래 실내의 윤성찬을 향해 미소를 짓고, 윤성찬은 왼쪽 위 박철진에게 시선을 고정한다. 박철진의 얼굴은 완전한 측면보다는 비스듬한 옆얼굴이며, 두 사람의 상호 응시는 지시대로 연결된다.",
        "built_space": "전경에 열린 차량 문 하나와 내려간 창의 상단 경계가 있고, 그 뒤 오른쪽으로 차체 출입구와 실내 좌석이 이어진다. 박철진은 차 밖 왼쪽에서 문을 붙들고, 윤성찬은 오른쪽 좌석 안에 자리한다. 왼쪽 상체와 그 아래 손을 함께 담는 미디엄 구도가 A보다 분명하다. 실내에는 검은 좌석 등받이와 머리받침 하나가 뚜렷하다. 배경의 울타리, 임시 건물과 진흙 바닥은 보이지만 이전 참고 사진으로 시설 배치의 일치 여부까지 검증할 수는 없다.",
        "entities": "박철진의 중년 남성 얼굴과 체격, 짧은 검은 머리, 남색 모자·전투복과 붉은 완장이 참고에 가깝다. 윤성찬도 고령 남성의 회색 머리와 흰 콧수염, 얼굴 특징, 정장 차림을 유지한다. 그의 몸 앞에 지팡이 손잡이와 옅은 손수건이 보이며, 진흙 묻은 지팡이 끝은 프레임 밖이다. 시신 가방은 보이지 않는다. 울타리와 건물 앞에는 허용된 두 사람 외에 여러 사람이 서 있다.",
        "hard_violations": [
         "배경의 울타리와 건물 앞에 여러 인물을 추가하여 등장인물 제한을 위반했다."
        ],
        "physics": "박철진의 손은 문 윗부분을 감싸고 엄지와 손가락의 접촉이 보이며, 팔은 그의 어깨와 자연스럽게 연결된다. 앞으로 기울인 상체는 서서 문을 붙드는 동작으로 가능하다. 윤성찬은 좌석의 지지를 받으며 고개를 돌리고 있다. 지팡이와 손수건은 몸 앞과 무릎 쪽에 놓인 것으로 읽히지만 하단과 손의 접촉은 가려져 있다. 명백히 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gpt-high] 박철진과 윤성찬만 보여야 하는 장면에 배경 인물들을 추가했다."
    ],
    "A": [
     "[gpt-high] 배경의 울타리와 건물 앞에 여러 인물을 추가하여 등장인물 제한을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 앵글과 인물의 시선 교환을 완벽히 구현했으며, 윤성찬이 지팡이와 손수건을 모두 쥐고 있는 세부 지시사항까지 정확하게 반영했습니다.  ★위반: [gpt-high] 배경의 울타리와 건물 앞에 여러 인물을 추가하여 등장인물 제한을 위반했다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "전반적인 구도와 조명은 우수하나, 프롬프트에 명시된 윤성찬의 손수건이 누락되어 A에 비해 정확도가 떨어집니다.  ★위반: [gpt-high] 박철진과 윤성찬만 보여야 하는 장면에 배경 인물들을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S42sh2_sel.png",
    "asset_id": "d9de811e-a41a-45d3-8d2b-86b87b13bcda",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1456670>",
    "asset_id": "cb968520-53a5-4b55-8260-e357fc83e9f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0adb-8017-704f-882c-1d57317ad578",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh2"
  },
  "staged_characters_added": [
   "C17"
  ]
 },
 "S42sh27::signage": {
  "fp": "1a4b947a2a0e7379",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh27": {
  "input_fingerprint": "1e29348a195411c5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수하가 건넨 무전기를 귀에 댄 채 짙은 눈썹이 꿈틀하며 뻣뻣하게 굳어 있는 박철진의 찰나.\n\nLOCATION (lock): At the nighttime roadside checkpoint area at the refugee settlement's muddy entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inward move slightly above 박철진's eye line at the established oblique angle, tilting gently downward so his brow and the radio at his ear remain visible together. Place his face just left of center and the radio along its right boundary, with the subordinate's completed handoff outside the crop; his attention is on the call, his eyes lowered toward the ground beyond the lower frame as his brow twitches and his jaw stiffens. Emphasize only the final reduction in camera distance, not a simultaneous lighting or eyeline shift.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Handheld radio (Received from the subordinate and held against 박철진's ear) — Its side is visible beside his cheek without covering his brow or mouth; used as A small contextual element linking the frozen expression to the incoming call.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued nighttime exposure, using controlled facial contrast to reveal the brow's small movement and the sudden rigidity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sedan has departed, and its door was closed before departure. The opened body bag remains at the nighttime inspection site. 박철진: He remains at the entrance after the sedan's departure, his earlier smile gone and his face now tense.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수하가 건넨 무전기를 귀에 댄 채 짙은 눈썹이 꿈틀하며 뻣뻣하게 굳어 있는 박철진의 찰나.\n\nLOCATION (lock): At the nighttime roadside checkpoint area at the refugee settlement's muddy entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inward move slightly above 박철진's eye line at the established oblique angle, tilting gently downward so his brow and the radio at his ear remain visible together. Place his face just left of center and the radio along its right boundary, with the subordinate's completed handoff outside the crop; his attention is on the call, his eyes lowered toward the ground beyond the lower frame as his brow twitches and his jaw stiffens. Emphasize only the final reduction in camera distance, not a simultaneous lighting or eyeline shift.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Handheld radio (Received from the subordinate and held against 박철진's ear) — Its side is visible beside his cheek without covering his brow or mouth; used as A small contextual element linking the frozen expression to the incoming call.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued nighttime exposure, using controlled facial contrast to reveal the brow's small movement and the sudden rigidity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sedan has departed, and its door was closed before departure. The opened body bag remains at the nighttime inspection site. 박철진: He remains at the entrance after the sedan's departure, his earlier smile gone and his face now tense.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수하가 건넨 무전기를 귀에 댄 채 짙은 눈썹이 꿈틀하며 뻣뻣하게 굳어 있는 박철진의 찰나.\n\nLOCATION (lock): At the nighttime roadside checkpoint area at the refugee settlement's muddy entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the inward move slightly above 박철진's eye line at the established oblique angle, tilting gently downward so his brow and the radio at his ear remain visible together. Place his face just left of center and the radio along its right boundary, with the subordinate's completed handoff outside the crop; his attention is on the call, his eyes lowered toward the ground beyond the lower frame as his brow twitches and his jaw stiffens. Emphasize only the final reduction in camera distance, not a simultaneous lighting or eyeline shift.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Handheld radio (Received from the subordinate and held against 박철진's ear) — Its side is visible beside his cheek without covering his brow or mouth; used as A small contextual element linking the frozen expression to the incoming call.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the subdued nighttime exposure, using controlled facial contrast to reveal the brow's small movement and the sudden rigidity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sedan has departed, and its door was closed before departure. The opened body bag remains at the nighttime inspection site. 박철진: He remains at the entrance after the sedan's departure, his earlier smile gone and his face now tense.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 하단 밖을 향하고 있음.",
    "built_space": "야간 진흙탕 검문소 배경이 보이며, 우측 근경에 차량의 지붕이 위치함.",
    "entities": "인물의 외형과 무전기 소품은 레퍼런스와 일치하나, 출발해야 할 세단이 프레임에 남아있음.",
    "hard_violations": [
     "[gpt-high] 박철진 외에는 등장시키지 말라는 지시와 달리 배경에 여러 사람을 배치했다.",
     "[gpt-high] 이미 출발한 세단의 지붕과 창틀을 오른쪽 전경에 크게 남겨 장면의 명시적 상태를 위반했다."
    ],
    "physics": "손이 무전기를 쥐고 귀에 안정적으로 밀착시키고 있음."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 하단 밖을 향하고 있음.",
    "built_space": "야간 진흙탕 검문소 배경이며, 우측 하단에 차량 일부가 위치함.",
    "entities": "인물의 외형은 일치하나 무전기에 레퍼런스에 없는 구조물이 묘사됨. 출발해야 할 세단이 남아있음.",
    "hard_violations": [
     "[gpt-high] 박철진만 보여야 하는 장면에 여러 배경 인물의 몸을 추가했다.",
     "[gpt-high] 이미 출발해 없어야 할 세단의 젖은 차체 일부가 오른쪽 아래에 남아 있다."
    ],
    "physics": "손이 무전기를 쥐고 귀에 안정적으로 대고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "무전기의 형태와 질감을 레퍼런스에 가깝게 잘 구현하였으나, 프롬프트에서 이미 출발했다고 명시된 세단이 화면 우측에 그대로 남아있는 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지시된 시선과 구도는 따랐으나 무전기에 불필요한 얇은 안테나가 추가되는 등 소품 구현이 부정확하며, 출발한 세단이 여전히 화면에 존재하여 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 하단 밖을 향하고 있음.",
        "built_space": "야간 진흙탕 검문소 배경이 보이며, 우측 근경에 차량의 지붕이 위치함.",
        "entities": "인물의 외형과 무전기 소품은 레퍼런스와 일치하나, 출발해야 할 세단이 프레임에 남아있음.",
        "hard_violations": [],
        "physics": "손이 무전기를 쥐고 귀에 안정적으로 밀착시키고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 하단 밖을 향하고 있음.",
        "built_space": "야간 진흙탕 검문소 배경이며, 우측 하단에 차량 일부가 위치함.",
        "entities": "인물의 외형은 일치하나 무전기에 레퍼런스에 없는 구조물이 묘사됨. 출발해야 할 세단이 남아있음.",
        "hard_violations": [],
        "physics": "손이 무전기를 쥐고 귀에 안정적으로 대고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "무전기의 형태와 질감을 레퍼런스에 가깝게 잘 구현하였으나, 프롬프트에서 이미 출발했다고 명시된 세단이 화면 우측에 그대로 남아있는 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "지시된 시선과 구도는 따랐으나 무전기에 불필요한 얇은 안테나가 추가되는 등 소품 구현이 부정확하며, 출발한 세단이 여전히 화면에 존재하여 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 프레임 하단 밖을 향하고 있음.",
        "built_space": "야간 진흙탕 검문소 배경이 보이며, 우측 근경에 차량의 지붕이 위치함.",
        "entities": "인물의 외형과 무전기 소품은 레퍼런스와 일치하나, 출발해야 할 세단이 프레임에 남아있음.",
        "hard_violations": [],
        "physics": "손이 무전기를 쥐고 귀에 안정적으로 밀착시키고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 프레임 하단 밖을 향하고 있음.",
        "built_space": "야간 진흙탕 검문소 배경이며, 우측 하단에 차량 일부가 위치함.",
        "entities": "인물의 외형은 일치하나 무전기에 레퍼런스에 없는 구조물이 묘사됨. 출발해야 할 세단이 남아있음.",
        "hard_violations": [],
        "physics": "손이 무전기를 쥐고 귀에 안정적으로 대고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 중심의 밀착 구도와 내려간 시선은 더 충실하지만, 금지된 배경 인물들과 출발했어야 할 세단 일부가 남아 있어 사용 가능한 장면은 아니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "귀에 무전기를 댄 긴장은 구현했으나, 배경 인물들을 그대로 남기고 출발한 세단을 화면 오른쪽에 크게 재등장시켜 장면 연속성을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박철진의 두 눈은 카메라가 아니라 화면 아래쪽의 지면 방향을 향한다. 고개는 오른쪽으로 비스듬히 숙였고 미간에 힘이 들어가 있다. 무전기는 화면 오른쪽 귀에 세워져 있으며 안테나는 위를 향한다. 옆면과 격자 면 일부가 보이고 눈썹과 입을 가리지 않는다. 다만 상단 모서리가 관자놀이 쪽에 닿아 있어 귀와 스피커의 정확한 접촉은 확인하기 어렵다.",
        "built_space": "왼쪽 위에 두 발광부가 달린 조명 기둥 하나, 뒤쪽에 철망 울타리와 여러 지주, 오른쪽에 박공지붕 구조물과 따뜻한 조명이 보인다. 진흙 바닥의 물웅덩이에는 조명 반사가 있어 이전 장소의 재질과 야간 분위기는 이어진다. 배경에는 여러 사람이 서 있고 오른쪽 아래에는 젖은 검은 차체 일부가 들어온다. 얼굴은 중앙보다 왼쪽, 무전기는 얼굴의 오른쪽 경계에 놓인 밀착 클로즈업이다.",
        "entities": "주인공은 참조와 유사한 중년 동아시아계 남성으로, 짧은 검은 머리 일부와 남색 챙모자, 거친 남색 전투복, 붉은 완장 일부가 보인다. 얼굴과 손의 나이 및 피부 질감도 대체로 일치한다. 무전기는 검은 휴대형 장비이지만 참조의 은색 전면 격자와 표시창 구성은 확인되지 않아 정확한 소품 일치는 부족하다. 수하의 손은 없지만, 허용되지 않은 배경 사람들은 분명히 보인다. 시신 가방은 이 클로즈업에서 확인되지 않는다.",
        "hard_violations": [
         "박철진만 보여야 하는 장면에 여러 배경 인물의 몸을 추가했다.",
         "이미 출발해 없어야 할 세단의 젖은 차체 일부가 오른쪽 아래에 남아 있다."
        ],
        "physics": "박철진 자신의 손이 무전기 몸체를 감싸 쥐고 있으며 손목과 소매가 자연스럽게 연결된다. 장비는 손으로 지지되고 귀 옆에 유지되어 공중에 떠 있지 않다. 고개를 숙인 목과 어깨의 연결도 가능하며, 하체는 클로즈업 밖이므로 발의 지지 상태는 판단할 수 없다."
       },
       {
        "label": "B",
        "direction": "박철진의 눈은 화면 아래 지면 쪽을 향하며 렌즈를 보지 않는다. 미간이 찌푸려지고 입이 닫혀 있어 통화를 듣다가 굳은 표정으로 읽힌다. 무전기는 화면 오른쪽 귀에 밀착되어 안테나가 위로 향하고 눈썹과 입은 드러난다. 카메라에는 넓은 격자 면과 측면 일부가 보이지만, 귀를 향한 스피커 면은 명확하게 확인되지 않는다.",
        "built_space": "왼쪽 위에 두 발광부가 달린 조명 기둥 하나, 뒤에 철망 울타리와 여러 지주, 중앙 오른쪽에 박공지붕 구조물 및 따뜻한 조명이 있다. 젖은 진흙 바닥과 물웅덩이는 참조 장소와 유사하다. 그러나 배경 여러 곳에 사람이 서 있고 오른쪽 전경에는 세단의 지붕과 창틀이 크게 들어온다. 얼굴은 중앙 왼쪽에 있지만 A보다 작고 상체와 차량이 더 많이 보여 얼굴 중심의 최종 접근 구도에서 멀어진다.",
        "entities": "중년 동아시아계 남성의 얼굴, 검은 머리 일부, 남색 챙모자, 남색 전투복과 붉은 완장은 참조 박철진과 대체로 일치한다. 무전기를 쥔 손도 주인공의 나이와 체격에 어울린다. 무전기는 참조보다 단순한 검은 격자형 외관으로 보이며 은색 전면판과 표시창의 일치는 확인되지 않는다. 수하의 전달 동작은 프레임 밖이지만 배경에 금지된 사람들이 남아 있고, 출발한 세단도 명백하게 보인다. 시신 가방은 프레임에서 확인되지 않는다.",
        "hard_violations": [
         "박철진 외에는 등장시키지 말라는 지시와 달리 배경에 여러 사람을 배치했다.",
         "이미 출발한 세단의 지붕과 창틀을 오른쪽 전경에 크게 남겨 장면의 명시적 상태를 위반했다."
        ],
        "physics": "손가락이 무전기를 감싸고 손바닥이 하단을 받쳐 장비를 확실하게 지지한다. 손목은 전투복 소매로 이어지며 귀에 장비를 대는 팔의 굽힘도 물리적으로 가능하다. 고개와 어깨에도 불가능한 연결은 없다. 하체는 보이지 않으며, 보이는 범위에서 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴 중심의 밀착 구도와 내려간 시선은 더 충실하지만, 금지된 배경 인물들과 출발했어야 할 세단 일부가 남아 있어 사용 가능한 장면은 아니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "귀에 무전기를 댄 긴장은 구현했으나, 배경 인물들을 그대로 남기고 출발한 세단을 화면 오른쪽에 크게 재등장시켜 장면 연속성을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박철진의 두 눈은 카메라가 아니라 화면 아래쪽의 지면 방향을 향한다. 고개는 오른쪽으로 비스듬히 숙였고 미간에 힘이 들어가 있다. 무전기는 화면 오른쪽 귀에 세워져 있으며 안테나는 위를 향한다. 옆면과 격자 면 일부가 보이고 눈썹과 입을 가리지 않는다. 다만 상단 모서리가 관자놀이 쪽에 닿아 있어 귀와 스피커의 정확한 접촉은 확인하기 어렵다.",
        "built_space": "왼쪽 위에 두 발광부가 달린 조명 기둥 하나, 뒤쪽에 철망 울타리와 여러 지주, 오른쪽에 박공지붕 구조물과 따뜻한 조명이 보인다. 진흙 바닥의 물웅덩이에는 조명 반사가 있어 이전 장소의 재질과 야간 분위기는 이어진다. 배경에는 여러 사람이 서 있고 오른쪽 아래에는 젖은 검은 차체 일부가 들어온다. 얼굴은 중앙보다 왼쪽, 무전기는 얼굴의 오른쪽 경계에 놓인 밀착 클로즈업이다.",
        "entities": "주인공은 참조와 유사한 중년 동아시아계 남성으로, 짧은 검은 머리 일부와 남색 챙모자, 거친 남색 전투복, 붉은 완장 일부가 보인다. 얼굴과 손의 나이 및 피부 질감도 대체로 일치한다. 무전기는 검은 휴대형 장비이지만 참조의 은색 전면 격자와 표시창 구성은 확인되지 않아 정확한 소품 일치는 부족하다. 수하의 손은 없지만, 허용되지 않은 배경 사람들은 분명히 보인다. 시신 가방은 이 클로즈업에서 확인되지 않는다.",
        "hard_violations": [
         "박철진만 보여야 하는 장면에 여러 배경 인물의 몸을 추가했다.",
         "이미 출발해 없어야 할 세단의 젖은 차체 일부가 오른쪽 아래에 남아 있다."
        ],
        "physics": "박철진 자신의 손이 무전기 몸체를 감싸 쥐고 있으며 손목과 소매가 자연스럽게 연결된다. 장비는 손으로 지지되고 귀 옆에 유지되어 공중에 떠 있지 않다. 고개를 숙인 목과 어깨의 연결도 가능하며, 하체는 클로즈업 밖이므로 발의 지지 상태는 판단할 수 없다."
       },
       {
        "label": "A",
        "direction": "박철진의 눈은 화면 아래 지면 쪽을 향하며 렌즈를 보지 않는다. 미간이 찌푸려지고 입이 닫혀 있어 통화를 듣다가 굳은 표정으로 읽힌다. 무전기는 화면 오른쪽 귀에 밀착되어 안테나가 위로 향하고 눈썹과 입은 드러난다. 카메라에는 넓은 격자 면과 측면 일부가 보이지만, 귀를 향한 스피커 면은 명확하게 확인되지 않는다.",
        "built_space": "왼쪽 위에 두 발광부가 달린 조명 기둥 하나, 뒤에 철망 울타리와 여러 지주, 중앙 오른쪽에 박공지붕 구조물 및 따뜻한 조명이 있다. 젖은 진흙 바닥과 물웅덩이는 참조 장소와 유사하다. 그러나 배경 여러 곳에 사람이 서 있고 오른쪽 전경에는 세단의 지붕과 창틀이 크게 들어온다. 얼굴은 중앙 왼쪽에 있지만 A보다 작고 상체와 차량이 더 많이 보여 얼굴 중심의 최종 접근 구도에서 멀어진다.",
        "entities": "중년 동아시아계 남성의 얼굴, 검은 머리 일부, 남색 챙모자, 남색 전투복과 붉은 완장은 참조 박철진과 대체로 일치한다. 무전기를 쥔 손도 주인공의 나이와 체격에 어울린다. 무전기는 참조보다 단순한 검은 격자형 외관으로 보이며 은색 전면판과 표시창의 일치는 확인되지 않는다. 수하의 전달 동작은 프레임 밖이지만 배경에 금지된 사람들이 남아 있고, 출발한 세단도 명백하게 보인다. 시신 가방은 프레임에서 확인되지 않는다.",
        "hard_violations": [
         "박철진 외에는 등장시키지 말라는 지시와 달리 배경에 여러 사람을 배치했다.",
         "이미 출발한 세단의 지붕과 창틀을 오른쪽 전경에 크게 남겨 장면의 명시적 상태를 위반했다."
        ],
        "physics": "손가락이 무전기를 감싸고 손바닥이 하단을 받쳐 장비를 확실하게 지지한다. 손목은 전투복 소매로 이어지며 귀에 장비를 대는 팔의 굽힘도 물리적으로 가능하다. 고개와 어깨에도 불가능한 연결은 없다. 하체는 보이지 않으며, 보이는 범위에서 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.464
   },
   "violations": {
    "B": [
     "[gpt-high] 박철진만 보여야 하는 장면에 여러 배경 인물의 몸을 추가했다.",
     "[gpt-high] 이미 출발해 없어야 할 세단의 젖은 차체 일부가 오른쪽 아래에 남아 있다."
    ],
    "A": [
     "[gpt-high] 박철진 외에는 등장시키지 말라는 지시와 달리 배경에 여러 사람을 배치했다.",
     "[gpt-high] 이미 출발한 세단의 지붕과 창틀을 오른쪽 전경에 크게 남겨 장면의 명시적 상태를 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1417,
   "B": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "무전기의 형태와 질감을 레퍼런스에 가깝게 잘 구현하였으나, 프롬프트에서 이미 출발했다고 명시된 세단이 화면 우측에 그대로 남아있는 점이 아쉽습니다.  ★위반: [gpt-high] 박철진 외에는 등장시키지 말라는 지시와 달리 배경에 여러 사람을 배치했다. / [gpt-high] 이미 출발한 세단의 지붕과 창틀을 오른쪽 전경에 크게 남겨 장면의 명시적 상태를 위반했다."
   },
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "지시된 시선과 구도는 따랐으나 무전기에 불필요한 얇은 안테나가 추가되는 등 소품 구현이 부정확하며, 출발한 세단이 여전히 화면에 존재하여 감점되었습니다.  ★위반: [gpt-high] 박철진만 보여야 하는 장면에 여러 배경 인물의 몸을 추가했다. / [gpt-high] 이미 출발해 없어야 할 세단의 젖은 차체 일부가 오른쪽 아래에 남아 있다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S42sh18_sel.png",
    "asset_id": "efd50d88-85cf-47c6-891d-8cb22c70c99b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 무전기: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1027091>",
    "asset_id": "68be9f54-1b7a-4963-ad00-d3aea327c7a6",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ae0-b028-7f7d-9a64-d556fb2df348",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh18"
  }
 },
 "S43sh8::signage": {
  "fp": "a06946bae5f7a847",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh8": {
  "input_fingerprint": "f432133f6b85354b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전화기를 든 채 허공을 향해 허리를 90도로 굽히고 극도로 비굴하게 고개를 숙인 박철진의 전신.\n\nLOCATION (lock): Inside the militia commander's office, at the telephone conversation area. Ordinary office lighting supports the ongoing call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the retreat along the established oblique line in 박철진's office, holding a high, downward-looking view that includes his lowered head, receiver hand, and both feet. Place his deeply bowed body slightly right of center with empty space before him, making clear that he is bending to an authority present only on the telephone; his eyes point toward the floor beneath the bow. Let the widening distance alone expose the full humiliation as his waist reaches a right-angle bend.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the call in 박철진's trembling hand) — Seen in profile beside his lowered head; used as A small narrative link between the deep bow and the unseen authority; Office floor (Visible beneath his feet and lowered head); used as Establishes the full-body bow and the empty space to which it is addressed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient office illumination remains steady and restrained, allowing posture rather than a lighting effect to express submission.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife previously thrown into the militia-office wall remains embedded there; no removal is established. 박철진: He holds the telephone during the call, with trembling hands and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전화기를 든 채 허공을 향해 허리를 90도로 굽히고 극도로 비굴하게 고개를 숙인 박철진의 전신.\n\nLOCATION (lock): Inside the militia commander's office, at the telephone conversation area. Ordinary office lighting supports the ongoing call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the retreat along the established oblique line in 박철진's office, holding a high, downward-looking view that includes his lowered head, receiver hand, and both feet. Place his deeply bowed body slightly right of center with empty space before him, making clear that he is bending to an authority present only on the telephone; his eyes point toward the floor beneath the bow. Let the widening distance alone expose the full humiliation as his waist reaches a right-angle bend.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the call in 박철진's trembling hand) — Seen in profile beside his lowered head; used as A small narrative link between the deep bow and the unseen authority; Office floor (Visible beneath his feet and lowered head); used as Establishes the full-body bow and the empty space to which it is addressed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient office illumination remains steady and restrained, allowing posture rather than a lighting effect to express submission.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife previously thrown into the militia-office wall remains embedded there; no removal is established. 박철진: He holds the telephone during the call, with trembling hands and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전화기를 든 채 허공을 향해 허리를 90도로 굽히고 극도로 비굴하게 고개를 숙인 박철진의 전신.\n\nLOCATION (lock): Inside the militia commander's office, at the telephone conversation area. Ordinary office lighting supports the ongoing call. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the retreat along the established oblique line in 박철진's office, holding a high, downward-looking view that includes his lowered head, receiver hand, and both feet. Place his deeply bowed body slightly right of center with empty space before him, making clear that he is bending to an authority present only on the telephone; his eyes point toward the floor beneath the bow. Let the widening distance alone expose the full humiliation as his waist reaches a right-angle bend.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the call in 박철진's trembling hand) — Seen in profile beside his lowered head; used as A small narrative link between the deep bow and the unseen authority; Office floor (Visible beneath his feet and lowered head); used as Establishes the full-body bow and the empty space to which it is addressed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient office illumination remains steady and restrained, allowing posture rather than a lighting effect to express submission.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The knife previously thrown into the militia-office wall remains embedded there; no removal is established. 박철진: He holds the telephone during the call, with trembling hands and a submissively bowed posture.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "오른쪽 벽면(화이트보드)을 향해 몸을 굽히고 시선은 바닥을 향함.",
    "built_space": "야간 사무실. 왼쪽 책상과 오른쪽 벽 배치는 맞으나, 인물 앞의 빈 공간이 확보되지 않음.",
    "entities": "박철진이 등장하고 노인은 없음. 왼쪽 팔의 완장 등 복장은 참조 이미지와 일치함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 전화선 구조 (책상에서 온 선이 수화기가 아닌 인물의 허리/등에 연결됨)"
    ],
    "physics": "바닥에 서서 수화기를 들고 있으나, 전화기 본체와 연결된 선이 인물의 몸을 관통하거나 허리에 붙어 있음."
   },
   {
    "label": "B",
    "direction": "왼쪽 책상과 빈 공간을 향해 몸을 굽히고 있으며 시선은 바닥을 향함.",
    "built_space": "야간 사무실. 인물이 중앙 우측에 위치하고 카메라 하향 각도 및 전방 빈 공간 지시가 올바르게 적용됨.",
    "entities": "박철진이 등장하며 노인은 없음. 단, 참조 이미지와 달리 붉은 완장이 오른팔에 잘못 그려짐.",
    "hard_violations": [],
    "physics": "바닥을 안정적으로 딛고 서 있으며, 전화선이 수화기 하단에서 책상 쪽으로 자연스럽게 이어짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "완장 위치가 반대로 된 오류가 있으나, 빈 공간을 향한 인사 자세와 자연스러운 전화선 연결을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전화선이 몸에 연결되는 물리적 오류가 있으며, 전방에 빈 공간을 두라는 지시와 달리 벽을 향해 절하고 있음."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 책상과 빈 공간을 향해 몸을 굽히고 있으며 시선은 바닥을 향함.",
        "built_space": "야간 사무실. 인물이 중앙 우측에 위치하고 카메라 하향 각도 및 전방 빈 공간 지시가 올바르게 적용됨.",
        "entities": "박철진이 등장하며 노인은 없음. 단, 참조 이미지와 달리 붉은 완장이 오른팔에 잘못 그려짐.",
        "hard_violations": [],
        "physics": "바닥을 안정적으로 딛고 서 있으며, 전화선이 수화기 하단에서 책상 쪽으로 자연스럽게 이어짐."
       },
       {
        "label": "A",
        "direction": "오른쪽 벽면(화이트보드)을 향해 몸을 굽히고 시선은 바닥을 향함.",
        "built_space": "야간 사무실. 왼쪽 책상과 오른쪽 벽 배치는 맞으나, 인물 앞의 빈 공간이 확보되지 않음.",
        "entities": "박철진이 등장하고 노인은 없음. 왼쪽 팔의 완장 등 복장은 참조 이미지와 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 전화선 구조 (책상에서 온 선이 수화기가 아닌 인물의 허리/등에 연결됨)"
        ],
        "physics": "바닥에 서서 수화기를 들고 있으나, 전화기 본체와 연결된 선이 인물의 몸을 관통하거나 허리에 붙어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "완장 위치가 반대로 된 오류가 있으나, 빈 공간을 향한 인사 자세와 자연스러운 전화선 연결을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전화선이 몸에 연결되는 물리적 오류가 있으며, 전방에 빈 공간을 두라는 지시와 달리 벽을 향해 절하고 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 책상과 빈 공간을 향해 몸을 굽히고 있으며 시선은 바닥을 향함.",
        "built_space": "야간 사무실. 인물이 중앙 우측에 위치하고 카메라 하향 각도 및 전방 빈 공간 지시가 올바르게 적용됨.",
        "entities": "박철진이 등장하며 노인은 없음. 단, 참조 이미지와 달리 붉은 완장이 오른팔에 잘못 그려짐.",
        "hard_violations": [],
        "physics": "바닥을 안정적으로 딛고 서 있으며, 전화선이 수화기 하단에서 책상 쪽으로 자연스럽게 이어짐."
       },
       {
        "label": "A",
        "direction": "오른쪽 벽면(화이트보드)을 향해 몸을 굽히고 시선은 바닥을 향함.",
        "built_space": "야간 사무실. 왼쪽 책상과 오른쪽 벽 배치는 맞으나, 인물 앞의 빈 공간이 확보되지 않음.",
        "entities": "박철진이 등장하고 노인은 없음. 왼쪽 팔의 완장 등 복장은 참조 이미지와 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 전화선 구조 (책상에서 온 선이 수화기가 아닌 인물의 허리/등에 연결됨)"
        ],
        "physics": "바닥에 서서 수화기를 들고 있으나, 전화기 본체와 연결된 선이 인물의 몸을 관통하거나 허리에 붙어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "높은 사선 와이드숏에서 인물을 중앙 오른쪽에 놓고, 빈 바닥을 향한 거의 직각의 굴욕적인 절과 수화기를 든 손·양발을 모두 보여 주어 지시를 가장 충실하게 구현했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "야간 사무실과 통화 중 전신 인사는 유지하지만, 허리가 직각에 못 미치고 인물이 중앙 왼쪽에 있으며 내려다보는 각도도 약해 핵심 구도와 굴욕의 정도가 덜 맞는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "몸과 숙인 머리는 화면 왼쪽 아래의 비어 있는 바닥을 향한다. 눈은 모자와 고개에 가려 직접 확인하기 어렵지만 얼굴 방향은 절 아래 바닥을 향한다. 앞에 상대 인물은 없고, 검은 수화기는 숙인 머리 옆에서 귀와 입을 향해 잡혀 있다. 벽에 박힌 칼은 손잡이가 왼쪽으로 돌출되어 있다.",
        "built_space": "뒤쪽 창가에 책상 한 개와 검은 사무용 의자 한 개, 깃발 한 개, 책상등 한 개가 있다. 뒤쪽 오른편에는 높은 서류 선반 두 열과 그 아래·옆의 낮은 수납장들이 있고, 창 아래에는 낮은 수납장 두 개가 보인다. 오른쪽 벽에는 지도 한 개, 일정판 한 개, 박힌 칼 한 자루가 유지된다. 인물은 책상 오른쪽의 열린 바닥에 서 있으며 앞쪽 바닥이 넓게 비어 있다. 높은 사선 시점에서 머리부터 양발까지 보이고, 바닥의 빛 반사도 광택 있는 표면에서 가능한 형태다.",
        "entities": "등장인물은 박철진에 해당하는 성인 남성 한 명뿐이다. 얼굴과 머리카락이 대부분 가려 정확한 연령·얼굴 일치는 판단하기 어렵지만, 보이는 피부와 체격은 참조와 양립한다. 남색 전투복, 같은 계열의 모자, 흰 별 모양이 있는 붉은 완장, 허리 주머니와 검은 전투화가 참조를 따른다. 검은 유선 수화기와 책상 위 전화기, 벽에 남은 칼이 보인다. 창밖은 야간이며, 실내 조명은 다소 어둡지만 과장된 효과 없이 유지된다.",
        "hard_violations": [],
        "physics": "두 전투화가 바닥에 닿아 체중을 지탱하고, 다리를 세운 채 고관절에서 상체를 거의 수평으로 접고 있다. 수화기는 머리 옆 손에 실제로 쥐어져 있으며 전화선은 책상 쪽으로 이어진다. 주머니는 허리띠에 매달리고 칼은 벽에 박혀 지지된다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "몸과 얼굴은 화면 오른쪽 아래의 빈 바닥을 향하며 다른 사람에게 절하는 장면은 아니다. 눈 자체는 작고 가려져 확인하기 어렵지만 고개 방향은 바닥을 향한다. 한 손의 수화기는 숙인 얼굴 옆에서 귀와 입 쪽으로 향한다. 벽의 칼 손잡이는 왼쪽으로 돌출되어 있다.",
        "built_space": "창가의 책상 한 개, 검은 의자 한 개, 깃발 한 개, 책상등 한 개와 뒤쪽의 높은 서류 선반 두 열이 보인다. 선반 아래와 지도 아래에 낮은 수납장들이 이어지고, 오른쪽 벽에는 지도 한 개, 일정판 한 개, 박힌 칼 한 자루가 있다. 참조의 주요 배치와 회색 바닥은 유지된다. 인물은 책상 앞 열린 바닥의 중앙 왼쪽에 있으며 오른쪽으로 빈 공간을 둔다. 전신은 들어오지만 내려다보는 높이가 A보다 낮고, 지정된 중앙 오른쪽 배치와 다르다. 바닥 반사는 가능한 범위다.",
        "entities": "성인 남성 한 명만 있으며, 드러난 옆얼굴과 체격은 한국인 중년 남성 박철진의 참조와 대체로 부합한다. 짧은 머리는 모자에 대부분 가려져 있다. 남색 전투복과 모자, 붉은 별 완장, 허리 주머니, 검은 전투화가 유지된다. 한 손에 검은 유선 수화기를 들고 있으며 벽의 칼도 남아 있다. 창밖은 밤이다. 다른 인물이나 사진 위에 얹힌 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발이 바닥을 지지하고 상체를 앞으로 굽힌 자세로, 신체 지지는 자연스럽다. 다만 상체가 수평까지 내려가지 않아 요구한 허리의 90도 굽힘에는 못 미친다. 한 손이 수화기를 잡고 다른 손은 허벅지 앞에 내려와 있다. 전화선 일부는 몸에 가려지지만 책상 쪽 연결이 보이며, 허리 장비와 벽의 칼에도 각각 지지점이 있다. 공중에 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "높은 사선 와이드숏에서 인물을 중앙 오른쪽에 놓고, 빈 바닥을 향한 거의 직각의 굴욕적인 절과 수화기를 든 손·양발을 모두 보여 주어 지시를 가장 충실하게 구현했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "야간 사무실과 통화 중 전신 인사는 유지하지만, 허리가 직각에 못 미치고 인물이 중앙 왼쪽에 있으며 내려다보는 각도도 약해 핵심 구도와 굴욕의 정도가 덜 맞는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "몸과 숙인 머리는 화면 왼쪽 아래의 비어 있는 바닥을 향한다. 눈은 모자와 고개에 가려 직접 확인하기 어렵지만 얼굴 방향은 절 아래 바닥을 향한다. 앞에 상대 인물은 없고, 검은 수화기는 숙인 머리 옆에서 귀와 입을 향해 잡혀 있다. 벽에 박힌 칼은 손잡이가 왼쪽으로 돌출되어 있다.",
        "built_space": "뒤쪽 창가에 책상 한 개와 검은 사무용 의자 한 개, 깃발 한 개, 책상등 한 개가 있다. 뒤쪽 오른편에는 높은 서류 선반 두 열과 그 아래·옆의 낮은 수납장들이 있고, 창 아래에는 낮은 수납장 두 개가 보인다. 오른쪽 벽에는 지도 한 개, 일정판 한 개, 박힌 칼 한 자루가 유지된다. 인물은 책상 오른쪽의 열린 바닥에 서 있으며 앞쪽 바닥이 넓게 비어 있다. 높은 사선 시점에서 머리부터 양발까지 보이고, 바닥의 빛 반사도 광택 있는 표면에서 가능한 형태다.",
        "entities": "등장인물은 박철진에 해당하는 성인 남성 한 명뿐이다. 얼굴과 머리카락이 대부분 가려 정확한 연령·얼굴 일치는 판단하기 어렵지만, 보이는 피부와 체격은 참조와 양립한다. 남색 전투복, 같은 계열의 모자, 흰 별 모양이 있는 붉은 완장, 허리 주머니와 검은 전투화가 참조를 따른다. 검은 유선 수화기와 책상 위 전화기, 벽에 남은 칼이 보인다. 창밖은 야간이며, 실내 조명은 다소 어둡지만 과장된 효과 없이 유지된다.",
        "hard_violations": [],
        "physics": "두 전투화가 바닥에 닿아 체중을 지탱하고, 다리를 세운 채 고관절에서 상체를 거의 수평으로 접고 있다. 수화기는 머리 옆 손에 실제로 쥐어져 있으며 전화선은 책상 쪽으로 이어진다. 주머니는 허리띠에 매달리고 칼은 벽에 박혀 지지된다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "몸과 얼굴은 화면 오른쪽 아래의 빈 바닥을 향하며 다른 사람에게 절하는 장면은 아니다. 눈 자체는 작고 가려져 확인하기 어렵지만 고개 방향은 바닥을 향한다. 한 손의 수화기는 숙인 얼굴 옆에서 귀와 입 쪽으로 향한다. 벽의 칼 손잡이는 왼쪽으로 돌출되어 있다.",
        "built_space": "창가의 책상 한 개, 검은 의자 한 개, 깃발 한 개, 책상등 한 개와 뒤쪽의 높은 서류 선반 두 열이 보인다. 선반 아래와 지도 아래에 낮은 수납장들이 이어지고, 오른쪽 벽에는 지도 한 개, 일정판 한 개, 박힌 칼 한 자루가 있다. 참조의 주요 배치와 회색 바닥은 유지된다. 인물은 책상 앞 열린 바닥의 중앙 왼쪽에 있으며 오른쪽으로 빈 공간을 둔다. 전신은 들어오지만 내려다보는 높이가 A보다 낮고, 지정된 중앙 오른쪽 배치와 다르다. 바닥 반사는 가능한 범위다.",
        "entities": "성인 남성 한 명만 있으며, 드러난 옆얼굴과 체격은 한국인 중년 남성 박철진의 참조와 대체로 부합한다. 짧은 머리는 모자에 대부분 가려져 있다. 남색 전투복과 모자, 붉은 별 완장, 허리 주머니, 검은 전투화가 유지된다. 한 손에 검은 유선 수화기를 들고 있으며 벽의 칼도 남아 있다. 창밖은 밤이다. 다른 인물이나 사진 위에 얹힌 자막은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 발이 바닥을 지지하고 상체를 앞으로 굽힌 자세로, 신체 지지는 자연스럽다. 다만 상체가 수평까지 내려가지 않아 요구한 허리의 90도 굽힘에는 못 미친다. 한 손이 수화기를 잡고 다른 손은 허벅지 앞에 내려와 있다. 전화선 일부는 몸에 가려지지만 책상 쪽 연결이 보이며, 허리 장비와 벽의 칼에도 각각 지지점이 있다. 공중에 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.206,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.956,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 전화선 구조 (책상에서 온 선이 수화기가 아닌 인물의 허리/등에 연결됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 956
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "완장 위치가 반대로 된 오류가 있으나, 빈 공간을 향한 인사 자세와 자연스러운 전화선 연결을 충실히 구현함."
   },
   {
    "label": "A",
    "score": 956,
    "verdict_ko": "전화선이 몸에 연결되는 물리적 오류가 있으며, 전방에 빈 공간을 두라는 지시와 달리 벽을 향해 절하고 있음.  ★위반: [gemini-pro] 물리적으로 불가능한 전화선 구조 (책상에서 온 선이 수화기가 아닌 인물의 허리/등에 연결됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S18sh8_sel.png",
    "asset_id": "02c3226d-6c8c-434a-872d-b7d5f2b65f30",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ae5-c992-70f2-9da7-7eb5c2a38dcb",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh8"
  }
 },
 "S43sh9::signage": {
  "fp": "12548097640e727c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::0ebb21869cc59adf": {
  "subjects": [],
  "subject_text": "국방장관 집무실\n집무용 책상과 의자, 대면 좌석이 배치된 사무실. 책상에는 통화용 전화기가 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L56",
  "scope_role": "location_interior",
  "scope_sha": "83c3f7d02f34f011"
 },
 "S43sh9::bgfirst_bg": {
  "input_fingerprint": "14bd11f09c1dfadd",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수화기에 대고 손가락을 강하게 뻗은 채 오만한 기세로 입을 벌려 명령하는 자세의 국방장관의 역동적인 상반신.\n\nLOCATION (lock): At the telephone position inside the defense minister's private office, under office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly into 국방장관's office and hold the prescribed static, low oblique view, looking upward from below his seated shoulder line. His upper body sits right of center, while the receiver and forcefully extended finger occupy the central space at natural scale; his open mouth addresses the call and his attention stays on the receiver rather than the lens. Keep the pointing hand lateral to his face so the command reads through the relationship between hand, mouth, and telephone without exaggerated foreshortening.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the command) — The receiver is seen obliquely near his mouth, with the extended finger directed toward its lower end; used as Makes the target of the commanding gesture readable without showing the remote listener.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient office illumination retains facial and hand detail without adding an unsupported source or theatrical color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수화기에 대고 손가락을 강하게 뻗은 채 오만한 기세로 입을 벌려 명령하는 자세의 국방장관의 역동적인 상반신.\n\nLOCATION (lock): At the telephone position inside the defense minister's private office, under office lighting.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly into 국방장관's office and hold the prescribed static, low oblique view, looking upward from below his seated shoulder line. His upper body sits right of center, while the receiver and forcefully extended finger occupy the central space at natural scale; his open mouth addresses the call and his attention stays on the receiver rather than the lens. Keep the pointing hand lateral to his face so the command reads through the relationship between hand, mouth, and telephone without exaggerated foreshortening.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the command) — The receiver is seen obliquely near his mouth, with the extended finger directed toward its lower end; used as Makes the target of the commanding gesture readable without showing the remote listener.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient office illumination retains facial and hand detail without adding an unsupported source or theatrical color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh9__bgfirst_bg.png",
  "asset_id": "7e32544a-841b-4aff-9a20-47b6b1a6c072",
  "input_asset_ids": [
   "621e897f-6865-4cff-ad78-84a56b41b8ce",
   "42b2d8cf-e604-41d4-92ef-f04fa7aa7624"
  ]
 },
 "S43sh9": {
  "input_fingerprint": "4a7a9d557a45af02",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수화기에 대고 손가락을 강하게 뻗은 채 오만한 기세로 입을 벌려 명령하는 자세의 국방장관의 역동적인 상반신.\n\nLOCATION (lock): At the telephone position inside the defense minister's private office, under office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly into 국방장관's office and hold the prescribed static, low oblique view, looking upward from below his seated shoulder line. His upper body sits right of center, while the receiver and forcefully extended finger occupy the central space at natural scale; his open mouth addresses the call and his attention stays on the receiver rather than the lens. Keep the pointing hand lateral to his face so the command reads through the relationship between hand, mouth, and telephone without exaggerated foreshortening.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the command) — The receiver is seen obliquely near his mouth, with the extended finger directed toward its lower end; used as Makes the target of the commanding gesture readable without showing the remote listener.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient office illumination retains facial and hand detail without adding an unsupported source or theatrical color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. 국방장관: He remains in his office on the telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수화기에 대고 손가락을 강하게 뻗은 채 오만한 기세로 입을 벌려 명령하는 자세의 국방장관의 역동적인 상반신.\n\nLOCATION (lock): At the telephone position inside the defense minister's private office, under office lighting. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly into 국방장관's office and hold the prescribed static, low oblique view, looking upward from below his seated shoulder line. His upper body sits right of center, while the receiver and forcefully extended finger occupy the central space at natural scale; his open mouth addresses the call and his attention stays on the receiver rather than the lens. Keep the pointing hand lateral to his face so the command reads through the relationship between hand, mouth, and telephone without exaggerated foreshortening.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the command) — The receiver is seen obliquely near his mouth, with the extended finger directed toward its lower end; used as Makes the target of the commanding gesture readable without showing the remote listener.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient office illumination retains facial and hand detail without adding an unsupported source or theatrical color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. 국방장관: He remains in his office on the telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수화기에 대고 손가락을 강하게 뻗은 채 오만한 기세로 입을 벌려 명령하는 자세의 국방장관의 역동적인 상반신.\n\nLOCATION (lock): At the telephone position inside the defense minister's private office, under office lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut explicitly into 국방장관's office and hold the prescribed static, low oblique view, looking upward from below his seated shoulder line. His upper body sits right of center, while the receiver and forcefully extended finger occupy the central space at natural scale; his open mouth addresses the call and his attention stays on the receiver rather than the lens. Keep the pointing hand lateral to his face so the command reads through the relationship between hand, mouth, and telephone without exaggerated foreshortening.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Telephone receiver (Held during the command) — The receiver is seen obliquely near his mouth, with the extended finger directed toward its lower end; used as Makes the target of the commanding gesture readable without showing the remote listener.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient office illumination retains facial and hand detail without adding an unsupported source or theatrical color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. 국방장관: He remains in his office on the telephone.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 국방장관 right now, so 국방장관's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 국방장관: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh9__bgfirst_bg.png",
     "asset_id": "7e32544a-841b-4aff-9a20-47b6b1a6c072",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S43sh9.png",
     "asset_id": "621e897f-6865-4cff-ad78-84a56b41b8ce",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:805558>",
     "asset_id": "2ee9d902-fac0-42dc-986b-7bd16edba439",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L56B01.png",
     "asset_id": "42b2d8cf-e604-41d4-92ef-f04fa7aa7624",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:805558>",
     "asset_id": "2ee9d902-fac0-42dc-986b-7bd16edba439",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 수화기 쪽을 향하고 있으며, 왼손 검지손가락이 입가에 있는 수화기의 하단(송화기)을 정확하게 가리키고 있음.",
    "built_space": "국방장관의 집무실 책상, 뒤편의 태극기와 부대기, 중앙의 액자, 그리고 창밖의 야경이 레퍼런스 사진의 구조에 맞게 올바른 위치에 배치되어 있음.",
    "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "오른손은 수화기를 자연스럽게 쥐고 귀에 대고 있으며, 왼손은 수화기를 향해 지시하는 자세를 안정적으로 취하고 있음. 무게중심과 의자에 앉은 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선은 앞이나 아래를 향해 있으며, 왼손 검지손가락이 수화기가 아닌 책상 위 전화기 본체를 가리키고 있음.",
    "built_space": "집무실 책상과 뒤편의 태극기, 창밖의 야경은 보이나, 우측의 부대기가 생략되었고 액자의 위치가 어긋나 있음.",
    "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "오른손으로 수화기를 쥐고 있는 모습과 의자에 앉은 자세는 자연스러우나, 손가락의 방향이 지시된 동작과 다름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 카메라 앵글과 인물의 화면 배치, 특히 손가락이 수화기 하단을 정확히 향하고 있는 점이 프롬프트와 완벽하게 일치합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 배경의 분위기는 좋으나, 뻗은 손가락이 수화기의 하단이 아닌 책상 위의 전화기 본체를 향하고 있어 프롬프트의 지시를 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 수화기 쪽을 향하고 있으며, 왼손 검지손가락이 입가에 있는 수화기의 하단(송화기)을 정확하게 가리키고 있음.",
        "built_space": "국방장관의 집무실 책상, 뒤편의 태극기와 부대기, 중앙의 액자, 그리고 창밖의 야경이 레퍼런스 사진의 구조에 맞게 올바른 위치에 배치되어 있음.",
        "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "오른손은 수화기를 자연스럽게 쥐고 귀에 대고 있으며, 왼손은 수화기를 향해 지시하는 자세를 안정적으로 취하고 있음. 무게중심과 의자에 앉은 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 앞이나 아래를 향해 있으며, 왼손 검지손가락이 수화기가 아닌 책상 위 전화기 본체를 가리키고 있음.",
        "built_space": "집무실 책상과 뒤편의 태극기, 창밖의 야경은 보이나, 우측의 부대기가 생략되었고 액자의 위치가 어긋나 있음.",
        "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "오른손으로 수화기를 쥐고 있는 모습과 의자에 앉은 자세는 자연스러우나, 손가락의 방향이 지시된 동작과 다름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시된 카메라 앵글과 인물의 화면 배치, 특히 손가락이 수화기 하단을 정확히 향하고 있는 점이 프롬프트와 완벽하게 일치합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 배경의 분위기는 좋으나, 뻗은 손가락이 수화기의 하단이 아닌 책상 위의 전화기 본체를 향하고 있어 프롬프트의 지시를 벗어났습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 수화기 쪽을 향하고 있으며, 왼손 검지손가락이 입가에 있는 수화기의 하단(송화기)을 정확하게 가리키고 있음.",
        "built_space": "국방장관의 집무실 책상, 뒤편의 태극기와 부대기, 중앙의 액자, 그리고 창밖의 야경이 레퍼런스 사진의 구조에 맞게 올바른 위치에 배치되어 있음.",
        "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "오른손은 수화기를 자연스럽게 쥐고 귀에 대고 있으며, 왼손은 수화기를 향해 지시하는 자세를 안정적으로 취하고 있음. 무게중심과 의자에 앉은 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 앞이나 아래를 향해 있으며, 왼손 검지손가락이 수화기가 아닌 책상 위 전화기 본체를 가리키고 있음.",
        "built_space": "집무실 책상과 뒤편의 태극기, 창밖의 야경은 보이나, 우측의 부대기가 생략되었고 액자의 위치가 어긋나 있음.",
        "entities": "국방장관(한국인 남성, 단정한 짧은 검은 머리, 네이비 정장, 화이트 셔츠)이 캐릭터 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "오른손으로 수화기를 쥐고 있는 모습과 의자에 앉은 자세는 자연스러우나, 손가락의 방향이 지시된 동작과 다름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "밤의 집무실과 낮은 사선 미디엄 숏은 맞지만, 검지가 수화기 하단이 아닌 그 아래 전화선 쪽을 향해 핵심 명령 제스처의 표적이 빗나간다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙의 검지가 수화기 하단을 정확히 겨냥하고, 오른쪽 상반신·벌린 입·수화기에 집중한 시선을 낮은 사선 미디엄 숏으로 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 렌즈가 아닌 수화기 쪽 아래를 바라보며 입을 벌리고 통화한다. 뻗은 검지는 화면 왼쪽을 향하지만 수화기 하단보다 낮아, 직접 가리키는 대상은 아래로 늘어진 전화선 부근이다. 손은 얼굴 옆보다 가슴 앞쪽에 놓여 있다.",
        "built_space": "책상 하나 뒤의 검은 집무 의자 하나에 앉아 있다. 왼쪽 창과 블라인드, 달력 하나, 태극기 하나, 뒤쪽 산수화 액자 하나와 낮은 목재 수납장, 여러 화분이 보인다. 책상에는 전화기 본체 하나와 필기구 받침, 서류가 있다. 참조의 재료와 주요 배치는 유지되며 오른쪽 책장과 군기는 화면 밖이다. 어깨 아래에서 올려다보는 사선 구도이고 상반신은 중앙 오른쪽에 있다. 창밖은 밤이며 불가능한 반사는 보이지 않는다.",
        "entities": "성숙한 동아시아계 남성 한 명으로, 참조 인물과 유사한 얼굴·짧고 정돈된 검은 머리·체격을 보인다. 짙은 네이비 정장, 흰 셔츠, 파란 넥타이와 흰 포켓치프가 참조에 부합한다. 검은 유선 수화기는 본인의 손에 들려 있고 입 가까이에 있다. 다른 사람이나 덧씌운 문자는 없다. 민병대 사무실의 칼은 이 집무실 장면에서 확인할 대상이 아니다.",
        "hard_violations": [],
        "physics": "몸은 집무 의자에 앉아 있고 등 뒤에 등받이가 보인다. 한 손이 수화기를 확실히 감싸 쥐며, 반대쪽 팔과 손은 팔꿈치를 굽힌 지시 동작으로 자연스럽게 이어진다. 전화선은 수화기에서 본체 쪽으로 늘어지고 본체·서류·필기구는 책상에 지지된다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "시선은 렌즈가 아닌 얼굴 옆 수화기를 향하고, 벌린 입은 수화기 하단 가까이에서 명령하는 관계를 이룬다. 반대 손의 검지가 왼쪽 위로 뻗어 수화기의 아래쪽 송화부를 직접 겨냥한다. 손은 얼굴 옆에 놓이며 과장된 원근 단축 없이 손끝과 표적의 관계가 분명하다.",
        "built_space": "책상 하나와 그 뒤 검은 집무 의자 하나가 보인다. 왼쪽 창·블라인드와 달력 하나, 태극기 하나, 뒤쪽 산수화 액자 하나, 오른쪽 군기 하나와 목재 책장 하나가 참조의 상대적 배치에 맞는다. 천장 조명 패널 두 개가 일부 보이며 실내 조명의 근거가 있다. 전화기 본체 하나는 책상 왼쪽, 서류 더미는 오른쪽에 놓인다. 낮은 사선 미디엄 숏에서 상반신은 중앙 오른쪽, 수화기와 검지는 중앙을 차지한다. 창밖은 밤이다.",
        "entities": "인물은 한 명이며 성숙한 동아시아계 남성의 얼굴, 정돈된 짧은 검은 머리와 체격이 국방장관 참조와 유사하다. 네이비 정장, 흰 셔츠, 파란 넥타이, 넥타이핀과 포켓치프도 일치한다. 검은 유선 수화기를 본인의 손으로 들고 있으며 아래 송화부가 입을 향한다. 전화기 본체는 사용자를 향해 기울어져 카메라에는 뒷면이 보인다. 추가 인물이나 그래픽 문자는 없다.",
        "hard_violations": [],
        "physics": "남자는 의자에 앉아 등받이를 등 뒤에 두고 상체를 앞으로 기울인다. 한 손의 손가락이 수화기 손잡이를 감싸 지지하고, 다른 팔은 굽힌 상태에서 검지만 뻗어 표적을 가리킨다. 수화기에서 내려온 코일선은 중력 방향으로 늘어져 책상 위 본체와 연결된다. 전화기·책·받침은 모두 책상에 놓여 있으며 부유하거나 해부학적으로 불가능한 부분은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "밤의 집무실과 낮은 사선 미디엄 숏은 맞지만, 검지가 수화기 하단이 아닌 그 아래 전화선 쪽을 향해 핵심 명령 제스처의 표적이 빗나간다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙의 검지가 수화기 하단을 정확히 겨냥하고, 오른쪽 상반신·벌린 입·수화기에 집중한 시선을 낮은 사선 미디엄 숏으로 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 렌즈가 아닌 수화기 쪽 아래를 바라보며 입을 벌리고 통화한다. 뻗은 검지는 화면 왼쪽을 향하지만 수화기 하단보다 낮아, 직접 가리키는 대상은 아래로 늘어진 전화선 부근이다. 손은 얼굴 옆보다 가슴 앞쪽에 놓여 있다.",
        "built_space": "책상 하나 뒤의 검은 집무 의자 하나에 앉아 있다. 왼쪽 창과 블라인드, 달력 하나, 태극기 하나, 뒤쪽 산수화 액자 하나와 낮은 목재 수납장, 여러 화분이 보인다. 책상에는 전화기 본체 하나와 필기구 받침, 서류가 있다. 참조의 재료와 주요 배치는 유지되며 오른쪽 책장과 군기는 화면 밖이다. 어깨 아래에서 올려다보는 사선 구도이고 상반신은 중앙 오른쪽에 있다. 창밖은 밤이며 불가능한 반사는 보이지 않는다.",
        "entities": "성숙한 동아시아계 남성 한 명으로, 참조 인물과 유사한 얼굴·짧고 정돈된 검은 머리·체격을 보인다. 짙은 네이비 정장, 흰 셔츠, 파란 넥타이와 흰 포켓치프가 참조에 부합한다. 검은 유선 수화기는 본인의 손에 들려 있고 입 가까이에 있다. 다른 사람이나 덧씌운 문자는 없다. 민병대 사무실의 칼은 이 집무실 장면에서 확인할 대상이 아니다.",
        "hard_violations": [],
        "physics": "몸은 집무 의자에 앉아 있고 등 뒤에 등받이가 보인다. 한 손이 수화기를 확실히 감싸 쥐며, 반대쪽 팔과 손은 팔꿈치를 굽힌 지시 동작으로 자연스럽게 이어진다. 전화선은 수화기에서 본체 쪽으로 늘어지고 본체·서류·필기구는 책상에 지지된다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "시선은 렌즈가 아닌 얼굴 옆 수화기를 향하고, 벌린 입은 수화기 하단 가까이에서 명령하는 관계를 이룬다. 반대 손의 검지가 왼쪽 위로 뻗어 수화기의 아래쪽 송화부를 직접 겨냥한다. 손은 얼굴 옆에 놓이며 과장된 원근 단축 없이 손끝과 표적의 관계가 분명하다.",
        "built_space": "책상 하나와 그 뒤 검은 집무 의자 하나가 보인다. 왼쪽 창·블라인드와 달력 하나, 태극기 하나, 뒤쪽 산수화 액자 하나, 오른쪽 군기 하나와 목재 책장 하나가 참조의 상대적 배치에 맞는다. 천장 조명 패널 두 개가 일부 보이며 실내 조명의 근거가 있다. 전화기 본체 하나는 책상 왼쪽, 서류 더미는 오른쪽에 놓인다. 낮은 사선 미디엄 숏에서 상반신은 중앙 오른쪽, 수화기와 검지는 중앙을 차지한다. 창밖은 밤이다.",
        "entities": "인물은 한 명이며 성숙한 동아시아계 남성의 얼굴, 정돈된 짧은 검은 머리와 체격이 국방장관 참조와 유사하다. 네이비 정장, 흰 셔츠, 파란 넥타이, 넥타이핀과 포켓치프도 일치한다. 검은 유선 수화기를 본인의 손으로 들고 있으며 아래 송화부가 입을 향한다. 전화기 본체는 사용자를 향해 기울어져 카메라에는 뒷면이 보인다. 추가 인물이나 그래픽 문자는 없다.",
        "hard_violations": [],
        "physics": "남자는 의자에 앉아 등받이를 등 뒤에 두고 상체를 앞으로 기울인다. 한 손의 손가락이 수화기 손잡이를 감싸 지지하고, 다른 팔은 굽힌 상태에서 검지만 뻗어 표적을 가리킨다. 수화기에서 내려온 코일선은 중력 방향으로 늘어져 책상 위 본체와 연결된다. 전화기·책·받침은 모두 책상에 놓여 있으며 부유하거나 해부학적으로 불가능한 부분은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.378
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.378
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1378
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 카메라 앵글과 인물의 화면 배치, 특히 손가락이 수화기 하단을 정확히 향하고 있는 점이 프롬프트와 완벽하게 일치합니다."
   },
   {
    "label": "B",
    "score": 1378,
    "verdict_ko": "인물과 배경의 분위기는 좋으나, 뻗은 손가락이 수화기의 하단이 아닌 책상 위의 전화기 본체를 향하고 있어 프롬프트의 지시를 벗어났습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L56B01.png",
    "asset_id": "42b2d8cf-e604-41d4-92ef-f04fa7aa7624",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:805558>",
    "asset_id": "2ee9d902-fac0-42dc-986b-7bd16edba439",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0aeb-19f5-76da-b00f-5d49e6e98261",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh9__bgfirst_bg.png",
   "bg_asset_id": "7e32544a-841b-4aff-9a20-47b6b1a6c072",
   "bg_record_key": "S43sh9::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S43sh14::signage": {
  "fp": "b469a5e2ca71c879",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh14": {
  "input_fingerprint": "c71a08fa8edb53ba",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 책상을 짚은 채 수하 1을 향해 목에 핏대를 세우고 고함을 치듯 입을 크게 벌린 박철진의 광기 어린 얼굴.\n\nLOCATION (lock): Behind the desk in the militia commander's office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the close endpoint from the subordinate's side of the desk, retaining the low lens height and slight upward tilt beneath 박철진's forward-leaning face. His face occupies the central-right area, with his open mouth and taut neck clearly readable and narrow fragments of his planted hands and the desktop along the bottom edge; he shouts toward 박철진의 수하 1 just off-screen left, never into the lens. Stop before the smile, emphasizing only the approach in distance while the subordinate remains outside the facial crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Braced by both of 박철진's hands) — A shallow strip of the top and near edge is visible beneath his leaning body; used as Provides physical evidence of his aggressive weight shift without competing with his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established office illumination unchanged, with controlled contrast preserving the strained neck and open mouth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the militia office's furnishings, desk, telephone, and established lighting from the reference. Exclude the ongoing telephone-call moment and all furnishings from the minister's separate office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. The telephone call has ended. 박철진: He remains in his office after ending the call, now upright and visibly assertive rather than bowed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 책상을 짚은 채 수하 1을 향해 목에 핏대를 세우고 고함을 치듯 입을 크게 벌린 박철진의 광기 어린 얼굴.\n\nLOCATION (lock): Behind the desk in the militia commander's office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the close endpoint from the subordinate's side of the desk, retaining the low lens height and slight upward tilt beneath 박철진's forward-leaning face. His face occupies the central-right area, with his open mouth and taut neck clearly readable and narrow fragments of his planted hands and the desktop along the bottom edge; he shouts toward 박철진의 수하 1 just off-screen left, never into the lens. Stop before the smile, emphasizing only the approach in distance while the subordinate remains outside the facial crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Braced by both of 박철진's hands) — A shallow strip of the top and near edge is visible beneath his leaning body; used as Provides physical evidence of his aggressive weight shift without competing with his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established office illumination unchanged, with controlled contrast preserving the strained neck and open mouth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the militia office's furnishings, desk, telephone, and established lighting from the reference. Exclude the ongoing telephone-call moment and all furnishings from the minister's separate office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. The telephone call has ended. 박철진: He remains in his office after ending the call, now upright and visibly assertive rather than bowed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 책상을 짚은 채 수하 1을 향해 목에 핏대를 세우고 고함을 치듯 입을 크게 벌린 박철진의 광기 어린 얼굴.\n\nLOCATION (lock): Behind the desk in the militia commander's office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the close endpoint from the subordinate's side of the desk, retaining the low lens height and slight upward tilt beneath 박철진's forward-leaning face. His face occupies the central-right area, with his open mouth and taut neck clearly readable and narrow fragments of his planted hands and the desktop along the bottom edge; he shouts toward 박철진의 수하 1 just off-screen left, never into the lens. Stop before the smile, emphasizing only the approach in distance while the subordinate remains outside the facial crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Desk (Braced by both of 박철진's hands) — A shallow strip of the top and near edge is visible beneath his leaning body; used as Provides physical evidence of his aggressive weight shift without competing with his face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established office illumination unchanged, with controlled contrast preserving the strained neck and open mouth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the militia office's furnishings, desk, telephone, and established lighting from the reference. Exclude the ongoing telephone-call moment and all furnishings from the minister's separate office.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia-office wall retains the previously embedded knife. The telephone call has ended. 박철진: He remains in his office after ending the call, now upright and visibly assertive rather than bowed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽 밖의 수하를 향해 강렬하게 시선을 두고 고함을 치고 있음.",
    "built_space": "야간의 집무실 내부로, 책상, 지도, 깃발, 창밖 야경이 적절히 배치되었으며 벽에 꽂힌 칼과 책상 위 전화기가 정확히 존재함.",
    "entities": "박철진의 얼굴, 모자, 짙은 남색 전투복과 붉은 완장이 레퍼런스와 일치함.",
    "hard_violations": [
     "[gpt-high] 완전히 화면 밖에 있어야 할 수하의 어깨와 상체 일부를 왼쪽 전경에 포함했다."
    ],
    "physics": "체중을 실어 두 손으로 책상을 단단히 짚고 있는 자세가 화면 하단에 물리적으로 자연스럽게 표현됨."
   },
   {
    "label": "B",
    "direction": "화면 왼쪽 밖을 향해 시선을 던지며 소리치고 있음.",
    "built_space": "야간의 집무실로 책상과 지도는 보이나, 유지되어야 할 벽면의 칼과 책상 위 전화기가 프레임 내에 보이지 않음.",
    "entities": "박철진의 외형과 복장, 완장 디테일이 레퍼런스와 일치함.",
    "hard_violations": [
     "[gpt-high] 수하를 화면 밖에 두고 박철진만 보여야 하는데, 왼쪽 아래 전경에 상대 인물의 어깨와 상체 일부가 노출되어 있다."
    ],
    "physics": "상체를 숙이고 있으나 프레임 하단에 한 손만 지탱하고 있어 두 손으로 짚었다는 제시문을 온전히 충족하지 못함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 손으로 책상을 짚은 자세와 로우 앵글 구도를 정확히 구현했으며, 벽의 칼과 책상 위 전화기 등 이전 샷의 디테일을 충실하게 유지했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴 표정과 기본 구도는 좋으나, 두 손으로 책상을 짚는 지시와 달리 한 손만 보이며 배경에 있어야 할 칼이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 밖의 수하를 향해 강렬하게 시선을 두고 고함을 치고 있음.",
        "built_space": "야간의 집무실 내부로, 책상, 지도, 깃발, 창밖 야경이 적절히 배치되었으며 벽에 꽂힌 칼과 책상 위 전화기가 정확히 존재함.",
        "entities": "박철진의 얼굴, 모자, 짙은 남색 전투복과 붉은 완장이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "체중을 실어 두 손으로 책상을 단단히 짚고 있는 자세가 화면 하단에 물리적으로 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽 밖을 향해 시선을 던지며 소리치고 있음.",
        "built_space": "야간의 집무실로 책상과 지도는 보이나, 유지되어야 할 벽면의 칼과 책상 위 전화기가 프레임 내에 보이지 않음.",
        "entities": "박철진의 외형과 복장, 완장 디테일이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "상체를 숙이고 있으나 프레임 하단에 한 손만 지탱하고 있어 두 손으로 짚었다는 제시문을 온전히 충족하지 못함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 손으로 책상을 짚은 자세와 로우 앵글 구도를 정확히 구현했으며, 벽의 칼과 책상 위 전화기 등 이전 샷의 디테일을 충실하게 유지했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴 표정과 기본 구도는 좋으나, 두 손으로 책상을 짚는 지시와 달리 한 손만 보이며 배경에 있어야 할 칼이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 밖의 수하를 향해 강렬하게 시선을 두고 고함을 치고 있음.",
        "built_space": "야간의 집무실 내부로, 책상, 지도, 깃발, 창밖 야경이 적절히 배치되었으며 벽에 꽂힌 칼과 책상 위 전화기가 정확히 존재함.",
        "entities": "박철진의 얼굴, 모자, 짙은 남색 전투복과 붉은 완장이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "체중을 실어 두 손으로 책상을 단단히 짚고 있는 자세가 화면 하단에 물리적으로 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽 밖을 향해 시선을 던지며 소리치고 있음.",
        "built_space": "야간의 집무실로 책상과 지도는 보이나, 유지되어야 할 벽면의 칼과 책상 위 전화기가 프레임 내에 보이지 않음.",
        "entities": "박철진의 외형과 복장, 완장 디테일이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "상체를 숙이고 있으나 프레임 하단에 한 손만 지탱하고 있어 두 손으로 짚었다는 제시문을 온전히 충족하지 못함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "얼굴의 중앙 오른쪽 배치와 화면 밖 왼쪽을 향한 고함은 더 정확하지만, 허리까지 드러나는 넓은 구도와 왼쪽 전경의 수하 신체 노출이 지시를 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "양손의 책상 접촉과 벽에 박힌 칼은 잘 보존했으나, 얼굴이 중앙 왼쪽에 놓이고 구도가 지나치게 넓으며 제외해야 할 수하의 어깨까지 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "눈과 벌린 입은 렌즈가 아니라 화면 왼쪽의 상대를 향한다. 수하에게 고함치는 방향은 맞지만, 그 상대의 어깨로 읽히는 흐릿한 신체 일부가 왼쪽 아래에 들어와 완전히 화면 밖에 있지는 않다. 무기나 이동하는 물체는 보이지 않는다.",
        "built_space": "왼쪽에 야경이 보이는 연속 창, 깃발 하나와 켜진 탁상등 하나, 중앙 뒤에 서류 책장, 오른쪽 벽에 한반도 지도 하나가 보인다. 하단 전경의 책상 면에 인물이 손을 짚고 있으며, 탁상등과 서류가 놓인 면은 왼쪽 뒤에도 드러난다. 사무실의 재료와 야간 조명은 대체로 이어지지만, 얼굴 클로즈업보다 배경과 허리까지 포함한 상반신 구도에 가깝다. 얼굴은 요구한 중앙 오른쪽에 있다. 벽의 칼은 이 크롭에서 확인되지 않는다.",
        "entities": "중년 동아시아계 남성의 얼굴, 검은 머리 일부, 남색 모자와 전투복, 흰 문양이 있는 붉은 완장이 참조와 대체로 부합한다. 크게 벌린 입과 긴장된 목이 보이고 웃음은 없다. 두 손은 하단에서 일부 보인다. 전화기는 명확히 식별되지 않으며 통화 동작도 없다. 왼쪽 전경에는 박철진 외 인물의 어깨로 읽히는 부분이 있다.",
        "hard_violations": [
         "수하를 화면 밖에 두고 박철진만 보여야 하는데, 왼쪽 아래 전경에 상대 인물의 어깨와 상체 일부가 노출되어 있다."
        ],
        "physics": "화면 왼쪽 손바닥과 손가락이 책상에 닿고, 오른쪽 손 일부도 하단 가장자리에서 책상에 이어진다. 팔을 통해 기울인 상체의 무게를 받는 자세는 가능하다. 하체가 잘렸다는 이유로 부유한다고 볼 근거는 없다. 모자와 완장은 몸에 착용되어 있으며 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "시선과 얼굴은 화면 왼쪽의 상대를 향하고, 렌즈를 정면으로 응시하지 않는다. 입을 크게 벌려 상대에게 고함치는 순간으로 읽힌다. 다만 왼쪽 아래에 상대의 어깨가 들어온다. 오른쪽 벽의 칼은 손잡이가 실내로 돌출되고 날이 벽에 들어간 상태다.",
        "built_space": "왼쪽의 야경 창과 깃발 하나, 탁상등 하나, 뒤쪽 서류 책장, 벽의 지도 하나와 박힌 칼 하나가 보인다. 전경 책상 위에는 전화기가 있고, 박철진은 양손을 벌려 책상을 짚는다. 참조 사무실의 주요 요소는 이어지지만, 넓은 상반신과 팔·손·전화기까지 보여 주어 지정된 얼굴 클로즈업에 도달하지 못했다. 얼굴 중심도 요구된 중앙 오른쪽이 아니라 중앙 왼쪽이다.",
        "entities": "중년 동아시아계 남성의 얼굴과 검은 머리 일부, 남색 모자 및 전투복, 붉은 완장이 참조와 대체로 맞는다. 벌어진 입, 정상적인 눈, 긴장된 목 근육이 보인다. 양손과 전화기, 벽에 박힌 칼이 식별되며 전화기를 들고 있지는 않다. 수하는 보이지 않아야 하지만 왼쪽 전경에 어깨와 상체 일부가 나타난다.",
        "hard_violations": [
         "완전히 화면 밖에 있어야 할 수하의 어깨와 상체 일부를 왼쪽 전경에 포함했다."
        ],
        "physics": "양손의 손가락과 손바닥이 책상 면에 닿아 전방으로 기울인 몸을 받친다. 팔과 어깨의 연결 및 체중 이동은 현실적으로 가능하다. 전화기는 책상 위에 놓여 있고 칼은 벽에 박혀 지지된다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "얼굴의 중앙 오른쪽 배치와 화면 밖 왼쪽을 향한 고함은 더 정확하지만, 허리까지 드러나는 넓은 구도와 왼쪽 전경의 수하 신체 노출이 지시를 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "양손의 책상 접촉과 벽에 박힌 칼은 잘 보존했으나, 얼굴이 중앙 왼쪽에 놓이고 구도가 지나치게 넓으며 제외해야 할 수하의 어깨까지 보인다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "눈과 벌린 입은 렌즈가 아니라 화면 왼쪽의 상대를 향한다. 수하에게 고함치는 방향은 맞지만, 그 상대의 어깨로 읽히는 흐릿한 신체 일부가 왼쪽 아래에 들어와 완전히 화면 밖에 있지는 않다. 무기나 이동하는 물체는 보이지 않는다.",
        "built_space": "왼쪽에 야경이 보이는 연속 창, 깃발 하나와 켜진 탁상등 하나, 중앙 뒤에 서류 책장, 오른쪽 벽에 한반도 지도 하나가 보인다. 하단 전경의 책상 면에 인물이 손을 짚고 있으며, 탁상등과 서류가 놓인 면은 왼쪽 뒤에도 드러난다. 사무실의 재료와 야간 조명은 대체로 이어지지만, 얼굴 클로즈업보다 배경과 허리까지 포함한 상반신 구도에 가깝다. 얼굴은 요구한 중앙 오른쪽에 있다. 벽의 칼은 이 크롭에서 확인되지 않는다.",
        "entities": "중년 동아시아계 남성의 얼굴, 검은 머리 일부, 남색 모자와 전투복, 흰 문양이 있는 붉은 완장이 참조와 대체로 부합한다. 크게 벌린 입과 긴장된 목이 보이고 웃음은 없다. 두 손은 하단에서 일부 보인다. 전화기는 명확히 식별되지 않으며 통화 동작도 없다. 왼쪽 전경에는 박철진 외 인물의 어깨로 읽히는 부분이 있다.",
        "hard_violations": [
         "수하를 화면 밖에 두고 박철진만 보여야 하는데, 왼쪽 아래 전경에 상대 인물의 어깨와 상체 일부가 노출되어 있다."
        ],
        "physics": "화면 왼쪽 손바닥과 손가락이 책상에 닿고, 오른쪽 손 일부도 하단 가장자리에서 책상에 이어진다. 팔을 통해 기울인 상체의 무게를 받는 자세는 가능하다. 하체가 잘렸다는 이유로 부유한다고 볼 근거는 없다. 모자와 완장은 몸에 착용되어 있으며 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "시선과 얼굴은 화면 왼쪽의 상대를 향하고, 렌즈를 정면으로 응시하지 않는다. 입을 크게 벌려 상대에게 고함치는 순간으로 읽힌다. 다만 왼쪽 아래에 상대의 어깨가 들어온다. 오른쪽 벽의 칼은 손잡이가 실내로 돌출되고 날이 벽에 들어간 상태다.",
        "built_space": "왼쪽의 야경 창과 깃발 하나, 탁상등 하나, 뒤쪽 서류 책장, 벽의 지도 하나와 박힌 칼 하나가 보인다. 전경 책상 위에는 전화기가 있고, 박철진은 양손을 벌려 책상을 짚는다. 참조 사무실의 주요 요소는 이어지지만, 넓은 상반신과 팔·손·전화기까지 보여 주어 지정된 얼굴 클로즈업에 도달하지 못했다. 얼굴 중심도 요구된 중앙 오른쪽이 아니라 중앙 왼쪽이다.",
        "entities": "중년 동아시아계 남성의 얼굴과 검은 머리 일부, 남색 모자 및 전투복, 붉은 완장이 참조와 대체로 맞는다. 벌어진 입, 정상적인 눈, 긴장된 목 근육이 보인다. 양손과 전화기, 벽에 박힌 칼이 식별되며 전화기를 들고 있지는 않다. 수하는 보이지 않아야 하지만 왼쪽 전경에 어깨와 상체 일부가 나타난다.",
        "hard_violations": [
         "완전히 화면 밖에 있어야 할 수하의 어깨와 상체 일부를 왼쪽 전경에 포함했다."
        ],
        "physics": "양손의 손가락과 손바닥이 책상 면에 닿아 전방으로 기울인 몸을 받친다. 팔과 어깨의 연결 및 체중 이동은 현실적으로 가능하다. 전화기는 책상 위에 놓여 있고 칼은 벽에 박혀 지지된다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.464
   },
   "violations": {
    "B": [
     "[gpt-high] 수하를 화면 밖에 두고 박철진만 보여야 하는데, 왼쪽 아래 전경에 상대 인물의 어깨와 상체 일부가 노출되어 있다."
    ],
    "A": [
     "[gpt-high] 완전히 화면 밖에 있어야 할 수하의 어깨와 상체 일부를 왼쪽 전경에 포함했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1464
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "두 손으로 책상을 짚은 자세와 로우 앵글 구도를 정확히 구현했으며, 벽의 칼과 책상 위 전화기 등 이전 샷의 디테일을 충실하게 유지했습니다.  ★위반: [gpt-high] 완전히 화면 밖에 있어야 할 수하의 어깨와 상체 일부를 왼쪽 전경에 포함했다."
   },
   {
    "label": "B",
    "score": 1464,
    "verdict_ko": "얼굴 표정과 기본 구도는 좋으나, 두 손으로 책상을 짚는 지시와 달리 한 손만 보이며 배경에 있어야 할 칼이 누락되었습니다.  ★위반: [gpt-high] 수하를 화면 밖에 두고 박철진만 보여야 하는데, 왼쪽 아래 전경에 상대 인물의 어깨와 상체 일부가 노출되어 있다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh8_sel.png",
    "asset_id": "f665ac08-4cc1-481e-a440-ee2ff6e4add8",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0af1-3adc-7a07-93d7-f816437a7a28",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh8"
  }
 },
 "S44sh4::signage": {
  "fp": "f0fe126f7bea7ba3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh4": {
  "input_fingerprint": "d9cae56597687be6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위로 활짝 열린 셔터 안쪽 짙은 어둠 속에 감춰져 있던 낡은 소형 캠핑카의 앞범퍼와 둥근 실루엣이 드러난 전경.\n\nLOCATION (lock): Inside the storage bay beside an abandoned factory, immediately behind the raised rolling shutter. The parked camper emerges from the otherwise dark warehouse interior. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just outside the fully raised shutter at the flow's low, off-axis position, looking slightly upward at the camper's front three-quarter outline. Place the front bumper near the lower center and let the rounded vehicle occupy roughly one-third of the image, surrounded by the warehouse opening and dark interior so its small scale remains credible. Keep the shot static and unoccupied, completing the reveal before moving along the visible side toward the rear.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Warehouse shutter and opening (Shutter fully raised) — The opening faces the camera obliquely, with the raised shutter above the camper; used as Architectural framing and a scale reference around the vehicle; Small camper (Old and parked inside the warehouse) — Front bumper and one side are visible in a front three-quarter view; used as The revealed subject, held below two-fifths of the frame area.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness while retaining enough subdued tonal separation to distinguish the bumper and rounded silhouette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, revealing a small camper in the dark interior. Its rear door has not yet been opened at this point.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위로 활짝 열린 셔터 안쪽 짙은 어둠 속에 감춰져 있던 낡은 소형 캠핑카의 앞범퍼와 둥근 실루엣이 드러난 전경.\n\nLOCATION (lock): Inside the storage bay beside an abandoned factory, immediately behind the raised rolling shutter. The parked camper emerges from the otherwise dark warehouse interior. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just outside the fully raised shutter at the flow's low, off-axis position, looking slightly upward at the camper's front three-quarter outline. Place the front bumper near the lower center and let the rounded vehicle occupy roughly one-third of the image, surrounded by the warehouse opening and dark interior so its small scale remains credible. Keep the shot static and unoccupied, completing the reveal before moving along the visible side toward the rear.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Warehouse shutter and opening (Shutter fully raised) — The opening faces the camera obliquely, with the raised shutter above the camper; used as Architectural framing and a scale reference around the vehicle; Small camper (Old and parked inside the warehouse) — Front bumper and one side are visible in a front three-quarter view; used as The revealed subject, held below two-fifths of the frame area.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness while retaining enough subdued tonal separation to distinguish the bumper and rounded silhouette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, revealing a small camper in the dark interior. Its rear door has not yet been opened at this point.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 위로 활짝 열린 셔터 안쪽 짙은 어둠 속에 감춰져 있던 낡은 소형 캠핑카의 앞범퍼와 둥근 실루엣이 드러난 전경.\n\nLOCATION (lock): Inside the storage bay beside an abandoned factory, immediately behind the raised rolling shutter. The parked camper emerges from the otherwise dark warehouse interior. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold just outside the fully raised shutter at the flow's low, off-axis position, looking slightly upward at the camper's front three-quarter outline. Place the front bumper near the lower center and let the rounded vehicle occupy roughly one-third of the image, surrounded by the warehouse opening and dark interior so its small scale remains credible. Keep the shot static and unoccupied, completing the reveal before moving along the visible side toward the rear.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Warehouse shutter and opening (Shutter fully raised) — The opening faces the camera obliquely, with the raised shutter above the camper; used as Architectural framing and a scale reference around the vehicle; Small camper (Old and parked inside the warehouse) — Front bumper and one side are visible in a front three-quarter view; used as The revealed subject, held below two-fifths of the frame area.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the warehouse darkness while retaining enough subdued tonal separation to distinguish the bumper and rounded silhouette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter is raised, revealing a small camper in the dark interior. Its rear door has not yet been opened at this point.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 창고 밖 로우 앵글에서 내부 중앙의 캠핑카 전면부를 비스듬히 올려다보고 있습니다.",
    "built_space": "롤링 셔터가 위로 완전히 말려 올라간 창고 입구가 프레임을 형성하며, 내부는 어둡고 바닥에 물기가 있습니다.",
    "entities": "프롬프트에 명시된 낡은 소형 캠핑카와 수평으로 말려 올라간 셔터가 정확히 존재하며 인물은 없습니다.",
    "hard_violations": [
     "[gpt-high] 명시된 장소 요소만 구성하라는 제한에도 불구하고 팔레트 운반 장비, 선반 위 적재물과 여러 용기 등 별도 소품을 추가했다."
    ],
    "physics": "캠핑카가 창고 바닥에 안정적으로 정차되어 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라가 창고 밖 로우 앵글에서 안쪽의 캠핑카를 향해 배치되어 있습니다.",
    "built_space": "창고 입구가 열려 있으나, 상단 셔터 라인이 사선으로 기울어져 있어 일반적인 롤링 셔터 구조와 어긋납니다.",
    "entities": "낡은 소형 캠핑카가 존재하며 인물은 없으나, 셔터의 형태가 왜곡되어 있습니다.",
    "hard_violations": [
     "[gpt-high] 장소와 차량 외에는 발명하지 말라는 제한에도 불구하고 드럼통, 기대 놓은 판재, 사다리와 전경 적재물 등 독립적인 소품을 추가했다."
    ],
    "physics": "캠핑카의 바퀴가 지면에 닿아 무게감을 유지하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 지시한 카메라 구도, 완전히 열린 롤링 셔터의 수평적 형태, 그리고 캠핑카가 화면의 약 3분의 1을 차지하는 비율을 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캠핑카의 전반적인 배치와 분위기는 부합하나, 위로 말려 올라가야 할 셔터 상단이 비스듬하게 기울어져 구조적 현실성이 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 창고 밖 로우 앵글에서 내부 중앙의 캠핑카 전면부를 비스듬히 올려다보고 있습니다.",
        "built_space": "롤링 셔터가 위로 완전히 말려 올라간 창고 입구가 프레임을 형성하며, 내부는 어둡고 바닥에 물기가 있습니다.",
        "entities": "프롬프트에 명시된 낡은 소형 캠핑카와 수평으로 말려 올라간 셔터가 정확히 존재하며 인물은 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카가 창고 바닥에 안정적으로 정차되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라가 창고 밖 로우 앵글에서 안쪽의 캠핑카를 향해 배치되어 있습니다.",
        "built_space": "창고 입구가 열려 있으나, 상단 셔터 라인이 사선으로 기울어져 있어 일반적인 롤링 셔터 구조와 어긋납니다.",
        "entities": "낡은 소형 캠핑카가 존재하며 인물은 없으나, 셔터의 형태가 왜곡되어 있습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면에 닿아 무게감을 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 지시한 카메라 구도, 완전히 열린 롤링 셔터의 수평적 형태, 그리고 캠핑카가 화면의 약 3분의 1을 차지하는 비율을 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캠핑카의 전반적인 배치와 분위기는 부합하나, 위로 말려 올라가야 할 셔터 상단이 비스듬하게 기울어져 구조적 현실성이 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 창고 밖 로우 앵글에서 내부 중앙의 캠핑카 전면부를 비스듬히 올려다보고 있습니다.",
        "built_space": "롤링 셔터가 위로 완전히 말려 올라간 창고 입구가 프레임을 형성하며, 내부는 어둡고 바닥에 물기가 있습니다.",
        "entities": "프롬프트에 명시된 낡은 소형 캠핑카와 수평으로 말려 올라간 셔터가 정확히 존재하며 인물은 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카가 창고 바닥에 안정적으로 정차되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라가 창고 밖 로우 앵글에서 안쪽의 캠핑카를 향해 배치되어 있습니다.",
        "built_space": "창고 입구가 열려 있으나, 상단 셔터 라인이 사선으로 기울어져 있어 일반적인 롤링 셔터 구조와 어긋납니다.",
        "entities": "낡은 소형 캠핑카가 존재하며 인물은 없으나, 셔터의 형태가 왜곡되어 있습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면에 닿아 무게감을 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "낮은 사선의 야간 전경은 맞지만 캠핑카가 요구한 약 3분의 1보다 지나치게 작으며, 드럼통·사다리·적재물 등 금지된 추가 소품이 보인다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "하단 중앙의 앞범퍼와 둥근 전면·측면을 더 적절한 크기로 드러내지만, 운반 장비와 적재물 등 명시되지 않은 소품을 발명하여 완전한 준수에는 실패한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 앞면은 카메라 쪽에서 화면 왼쪽으로 조금 틀어져 있고, 보이는 측면은 오른쪽 뒤로 이어진다. 카메라는 개구부 밖의 낮은 위치에서 위쪽을 바라보며 전면과 한쪽 측면을 함께 담는다. 사람의 시선이나 이동은 없다.",
        "built_space": "큰 창고 개구부 하나와 그 상단에 올라간 롤링 셔터 하나, 양쪽 문기둥이 보인다. 차량은 문턱 뒤 실내 바닥에 주차되어 있다. 천장에는 매달린 등 하나가 보이고, 왼쪽에는 드럼통과 기대 놓은 판재, 오른쪽 안쪽에는 선반과 사다리, 오른쪽 전경에는 적재물이 있다. 젖은 바닥의 희미한 반사는 이 낮은 시점에서 가능하다.",
        "entities": "낡은 밝은색 소형 캠핑카 한 대가 있으며 둥근 상부 돌출부, 원형 전조등, 앞범퍼와 한쪽 측면이 식별된다. 사람이나 얼굴은 없다. 후면 문은 보이지 않아 열림 여부를 확인할 수 없고, 눈에 보이는 문은 닫혀 있다. 차량은 화면 너비의 약 3분의 1이지만 실제 점유 면적은 요구한 약 3분의 1보다 훨씬 작다. 별도의 참조 이미지는 없다.",
        "hard_violations": [
         "장소와 차량 외에는 발명하지 말라는 제한에도 불구하고 드럼통, 기대 놓은 판재, 사다리와 전경 적재물 등 독립적인 소품을 추가했다."
        ],
        "physics": "차량은 타이어로 바닥에 지지되어 있으며 움직이는 흔적이 없다. 드럼통과 적재물은 바닥에 놓여 있고 판재는 벽에 기대어 있다. 셔터는 양옆 문틀과 상부 구조에 연결되어 있다. 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "캠핑카 앞면은 카메라를 향하면서 화면 왼쪽으로 조금 틀어져 있고, 한쪽 측면은 오른쪽 뒤로 뻗는다. 낮은 외부 카메라가 개구부를 비스듬히 올려다보므로 요구한 전면 사분의 삼 시점에 맞는다. 앞범퍼는 하단 중앙 가까이에 있다. 사람이나 이동 방향을 판단할 대상은 없다.",
        "built_space": "창고 개구부 하나, 상단으로 올라간 롤링 셔터 하나와 양쪽 문기둥이 보인다. 캠핑카는 문턱 바로 뒤 실내에 서 있다. 왼쪽 기둥에는 전기함 하나와 배관이 있고 양쪽 입구에는 보호기둥이 하나씩 있다. 왼쪽 내부에는 선반, 여러 용기와 운반 장비가 들어차 있다. 전경 물웅덩이의 약한 반사는 카메라 위치와 모순되지 않는다.",
        "entities": "오래되고 얼룩진 밝은색 캠핑카 한 대이며 둥근 상부 윤곽, 앞범퍼, 원형 전조등과 한쪽 측면이 보인다. 차량은 화면 오른쪽에 조금 치우치지만 A보다 요구한 주제 크기에 가까우며 전체 면적의 5분의 2 미만이다. 사람이나 얼굴은 없고 전면 유리 안에는 커튼이 보인다. 후면 문은 가려져 있어 상태를 단정할 수 없다. 읽을 수 있는 추가 문구는 보이지 않는다.",
        "hard_violations": [
         "명시된 장소 요소만 구성하라는 제한에도 불구하고 팔레트 운반 장비, 선반 위 적재물과 여러 용기 등 별도 소품을 추가했다."
        ],
        "physics": "차량의 앞뒤 바퀴가 실내 바닥을 딛고 차체를 지지한다. 커튼은 전면 유리 안쪽에서 자연스럽게 늘어져 있다. 운반 장비와 용기는 바닥 또는 선반에 놓여 있고 전기함은 기둥에 부착되어 있다. 셔터는 문틀 상부에 지지되며, 지지 없이 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "낮은 사선의 야간 전경은 맞지만 캠핑카가 요구한 약 3분의 1보다 지나치게 작으며, 드럼통·사다리·적재물 등 금지된 추가 소품이 보인다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "하단 중앙의 앞범퍼와 둥근 전면·측면을 더 적절한 크기로 드러내지만, 운반 장비와 적재물 등 명시되지 않은 소품을 발명하여 완전한 준수에는 실패한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 앞면은 카메라 쪽에서 화면 왼쪽으로 조금 틀어져 있고, 보이는 측면은 오른쪽 뒤로 이어진다. 카메라는 개구부 밖의 낮은 위치에서 위쪽을 바라보며 전면과 한쪽 측면을 함께 담는다. 사람의 시선이나 이동은 없다.",
        "built_space": "큰 창고 개구부 하나와 그 상단에 올라간 롤링 셔터 하나, 양쪽 문기둥이 보인다. 차량은 문턱 뒤 실내 바닥에 주차되어 있다. 천장에는 매달린 등 하나가 보이고, 왼쪽에는 드럼통과 기대 놓은 판재, 오른쪽 안쪽에는 선반과 사다리, 오른쪽 전경에는 적재물이 있다. 젖은 바닥의 희미한 반사는 이 낮은 시점에서 가능하다.",
        "entities": "낡은 밝은색 소형 캠핑카 한 대가 있으며 둥근 상부 돌출부, 원형 전조등, 앞범퍼와 한쪽 측면이 식별된다. 사람이나 얼굴은 없다. 후면 문은 보이지 않아 열림 여부를 확인할 수 없고, 눈에 보이는 문은 닫혀 있다. 차량은 화면 너비의 약 3분의 1이지만 실제 점유 면적은 요구한 약 3분의 1보다 훨씬 작다. 별도의 참조 이미지는 없다.",
        "hard_violations": [
         "장소와 차량 외에는 발명하지 말라는 제한에도 불구하고 드럼통, 기대 놓은 판재, 사다리와 전경 적재물 등 독립적인 소품을 추가했다."
        ],
        "physics": "차량은 타이어로 바닥에 지지되어 있으며 움직이는 흔적이 없다. 드럼통과 적재물은 바닥에 놓여 있고 판재는 벽에 기대어 있다. 셔터는 양옆 문틀과 상부 구조에 연결되어 있다. 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "캠핑카 앞면은 카메라를 향하면서 화면 왼쪽으로 조금 틀어져 있고, 한쪽 측면은 오른쪽 뒤로 뻗는다. 낮은 외부 카메라가 개구부를 비스듬히 올려다보므로 요구한 전면 사분의 삼 시점에 맞는다. 앞범퍼는 하단 중앙 가까이에 있다. 사람이나 이동 방향을 판단할 대상은 없다.",
        "built_space": "창고 개구부 하나, 상단으로 올라간 롤링 셔터 하나와 양쪽 문기둥이 보인다. 캠핑카는 문턱 바로 뒤 실내에 서 있다. 왼쪽 기둥에는 전기함 하나와 배관이 있고 양쪽 입구에는 보호기둥이 하나씩 있다. 왼쪽 내부에는 선반, 여러 용기와 운반 장비가 들어차 있다. 전경 물웅덩이의 약한 반사는 카메라 위치와 모순되지 않는다.",
        "entities": "오래되고 얼룩진 밝은색 캠핑카 한 대이며 둥근 상부 윤곽, 앞범퍼, 원형 전조등과 한쪽 측면이 보인다. 차량은 화면 오른쪽에 조금 치우치지만 A보다 요구한 주제 크기에 가까우며 전체 면적의 5분의 2 미만이다. 사람이나 얼굴은 없고 전면 유리 안에는 커튼이 보인다. 후면 문은 가려져 있어 상태를 단정할 수 없다. 읽을 수 있는 추가 문구는 보이지 않는다.",
        "hard_violations": [
         "명시된 장소 요소만 구성하라는 제한에도 불구하고 팔레트 운반 장비, 선반 위 적재물과 여러 용기 등 별도 소품을 추가했다."
        ],
        "physics": "차량의 앞뒤 바퀴가 실내 바닥을 딛고 차체를 지지한다. 커튼은 전면 유리 안쪽에서 자연스럽게 늘어져 있다. 운반 장비와 용기는 바닥 또는 선반에 놓여 있고 전기함은 기둥에 부착되어 있다. 셔터는 문틀 상부에 지지되며, 지지 없이 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.464
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.214
   },
   "violations": {
    "B": [
     "[gpt-high] 장소와 차량 외에는 발명하지 말라는 제한에도 불구하고 드럼통, 기대 놓은 판재, 사다리와 전경 적재물 등 독립적인 소품을 추가했다."
    ],
    "A": [
     "[gpt-high] 명시된 장소 요소만 구성하라는 제한에도 불구하고 팔레트 운반 장비, 선반 위 적재물과 여러 용기 등 별도 소품을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1214
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 지시한 카메라 구도, 완전히 열린 롤링 셔터의 수평적 형태, 그리고 캠핑카가 화면의 약 3분의 1을 차지하는 비율을 정확하게 구현했습니다.  ★위반: [gpt-high] 명시된 장소 요소만 구성하라는 제한에도 불구하고 팔레트 운반 장비, 선반 위 적재물과 여러 용기 등 별도 소품을 추가했다."
   },
   {
    "label": "B",
    "score": 1214,
    "verdict_ko": "캠핑카의 전반적인 배치와 분위기는 부합하나, 위로 말려 올라가야 할 셔터 상단이 비스듬하게 기울어져 구조적 현실성이 떨어집니다.  ★위반: [gpt-high] 장소와 차량 외에는 발명하지 말라는 제한에도 불구하고 드럼통, 기대 놓은 판재, 사다리와 전경 적재물 등 독립적인 소품을 추가했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0af5-e495-753e-9773-16adc205bb8f",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S44sh7::signage": {
  "fp": "68bd2c25cea4712a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh7": {
  "input_fingerprint": "2f8a9966e157298d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 뒷문 안으로 성인 네 명이 거뜬히 쉴 수 있는 널찍한 침대와 아늑한 생활 공간이 훤히 들여다보이는 구도.\n\nLOCATION (lock): Inside the small camper's rear living compartment, viewed through its open rear door while parked in the warehouse. The bed and roomy resting area are visible against the surrounding darkness. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward dolly just short of the open rear threshold, offset to one side and angled diagonally downward into the camper. Keep the doorway edges peripheral, place the bed across the lower-left portion at less than two-fifths of the image, and preserve the adjoining living space deeper at right so the capacity for four adults reads through the layout rather than lens exaggeration. No people enter the crop, and camera distance is the only emphasized change as the approach settles.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper bed in the lower-left of the frame, midground; Adjoining camper living space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Rear doorway (Rear door fully open) — Seen obliquely from just outside the threshold, opening onto the bed and living area; used as A peripheral frame establishing that the interior is being viewed directly from outside; Camper bed (Visible inside the camper) — The sleeping surface is viewed diagonally from above and extends inward from the lower-left region; used as Primary interior feature and a natural scale reference; Camper living space (Open to view through the rear door, with room for four occupants) — Interior depth continues beyond and alongside the bed; used as Preserves the relationship between sleeping accommodation and usable interior space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained, setting-appropriate ambient illumination that makes the interior layout readable without inventing a visible fixture or a warm color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter remains raised, and the camper's rear door is now open. A bed and an interior roomy enough for four are visible inside.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 뒷문 안으로 성인 네 명이 거뜬히 쉴 수 있는 널찍한 침대와 아늑한 생활 공간이 훤히 들여다보이는 구도.\n\nLOCATION (lock): Inside the small camper's rear living compartment, viewed through its open rear door while parked in the warehouse. The bed and roomy resting area are visible against the surrounding darkness. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward dolly just short of the open rear threshold, offset to one side and angled diagonally downward into the camper. Keep the doorway edges peripheral, place the bed across the lower-left portion at less than two-fifths of the image, and preserve the adjoining living space deeper at right so the capacity for four adults reads through the layout rather than lens exaggeration. No people enter the crop, and camera distance is the only emphasized change as the approach settles.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper bed in the lower-left of the frame, midground; Adjoining camper living space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Rear doorway (Rear door fully open) — Seen obliquely from just outside the threshold, opening onto the bed and living area; used as A peripheral frame establishing that the interior is being viewed directly from outside; Camper bed (Visible inside the camper) — The sleeping surface is viewed diagonally from above and extends inward from the lower-left region; used as Primary interior feature and a natural scale reference; Camper living space (Open to view through the rear door, with room for four occupants) — Interior depth continues beyond and alongside the bed; used as Preserves the relationship between sleeping accommodation and usable interior space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained, setting-appropriate ambient illumination that makes the interior layout readable without inventing a visible fixture or a warm color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter remains raised, and the camper's rear door is now open. A bed and an interior roomy enough for four are visible inside.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 뒷문 안으로 성인 네 명이 거뜬히 쉴 수 있는 널찍한 침대와 아늑한 생활 공간이 훤히 들여다보이는 구도.\n\nLOCATION (lock): Inside the small camper's rear living compartment, viewed through its open rear door while parked in the warehouse. The bed and roomy resting area are visible against the surrounding darkness. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward dolly just short of the open rear threshold, offset to one side and angled diagonally downward into the camper. Keep the doorway edges peripheral, place the bed across the lower-left portion at less than two-fifths of the image, and preserve the adjoining living space deeper at right so the capacity for four adults reads through the layout rather than lens exaggeration. No people enter the crop, and camera distance is the only emphasized change as the approach settles.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper bed in the lower-left of the frame, midground; Adjoining camper living space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Rear doorway (Rear door fully open) — Seen obliquely from just outside the threshold, opening onto the bed and living area; used as A peripheral frame establishing that the interior is being viewed directly from outside; Camper bed (Visible inside the camper) — The sleeping surface is viewed diagonally from above and extends inward from the lower-left region; used as Primary interior feature and a natural scale reference; Camper living space (Open to view through the rear door, with room for four occupants) — Interior depth continues beyond and alongside the bed; used as Preserves the relationship between sleeping accommodation and usable interior space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained, setting-appropriate ambient illumination that makes the interior layout readable without inventing a visible fixture or a warm color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse shutter remains raised, and the camper's rear door is now open. A bed and an interior roomy enough for four are visible inside.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
    "built_space": "캠핑카 내부. 좌측 하단 및 중경에 침대가 있고, 우측 후경에 4인이 앉을 수 있는 테이블과 소파 공간이 깊이 있게 배치됨. 프레임 가장자리에 문틀이 위치함.",
    "entities": "캠핑카 문틀, 넓은 침대, 거실 공간(소파, 테이블). 프롬프트 지시대로 사람은 없음.",
    "hard_violations": [
     "[gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 여러 가방과 조리 용기 등 미지정 물체를 다수 추가했다.",
     "[gpt-high] 허용된 원문이나 참조가 없는 문자성 표식을 왼쪽 창 하단에 추가했다."
    ],
    "physics": "모든 가구와 사물이 중력에 맞게 바닥과 표면에 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
    "built_space": "캠핑카 내부. 좌측에 침대, 우측 후경에 거실 공간 배치. 좌측 가장자리에 캠핑카 외부의 후미등이 포함되어 문틀보다 바깥쪽 시점이 강조됨.",
    "entities": "차량 후미등, 캠핑카 문틀, 침대, 거실 공간. 천장에 켜진 조명 기구가 포함됨. 사람은 없음.",
    "hard_violations": [
     "[gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 컵·용기와 주방 기기 등 미지정 물체를 다수 추가했다.",
     "[gpt-high] 가시적인 조명 기구를 만들지 말라는 명시적 조건과 달리 상부에 원형 조명 기구를 추가했다."
    ],
    "physics": "모든 가구와 사물이 물리 법칙에 어긋남 없이 배치됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 카메라 구도와 조명 제한(눈에 띄는 조명 기구 배제)을 잘 준수하며, 요구된 캠핑카 내부의 깊이감과 공간 배치를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 공간 배치는 맞으나, 프레임 좌측에 차량 외부(후미등)가 과하게 노출되었고 프롬프트에서 금지한 켜진 조명 기구가 임의로 추가되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
        "built_space": "캠핑카 내부. 좌측 하단 및 중경에 침대가 있고, 우측 후경에 4인이 앉을 수 있는 테이블과 소파 공간이 깊이 있게 배치됨. 프레임 가장자리에 문틀이 위치함.",
        "entities": "캠핑카 문틀, 넓은 침대, 거실 공간(소파, 테이블). 프롬프트 지시대로 사람은 없음.",
        "hard_violations": [],
        "physics": "모든 가구와 사물이 중력에 맞게 바닥과 표면에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
        "built_space": "캠핑카 내부. 좌측에 침대, 우측 후경에 거실 공간 배치. 좌측 가장자리에 캠핑카 외부의 후미등이 포함되어 문틀보다 바깥쪽 시점이 강조됨.",
        "entities": "차량 후미등, 캠핑카 문틀, 침대, 거실 공간. 천장에 켜진 조명 기구가 포함됨. 사람은 없음.",
        "hard_violations": [],
        "physics": "모든 가구와 사물이 물리 법칙에 어긋남 없이 배치됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 카메라 구도와 조명 제한(눈에 띄는 조명 기구 배제)을 잘 준수하며, 요구된 캠핑카 내부의 깊이감과 공간 배치를 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전반적인 공간 배치는 맞으나, 프레임 좌측에 차량 외부(후미등)가 과하게 노출되었고 프롬프트에서 금지한 켜진 조명 기구가 임의로 추가되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
        "built_space": "캠핑카 내부. 좌측 하단 및 중경에 침대가 있고, 우측 후경에 4인이 앉을 수 있는 테이블과 소파 공간이 깊이 있게 배치됨. 프레임 가장자리에 문틀이 위치함.",
        "entities": "캠핑카 문틀, 넓은 침대, 거실 공간(소파, 테이블). 프롬프트 지시대로 사람은 없음.",
        "hard_violations": [],
        "physics": "모든 가구와 사물이 중력에 맞게 바닥과 표면에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 열린 뒷문 밖에서 캠핑카 내부 안쪽을 향하고 있음.",
        "built_space": "캠핑카 내부. 좌측에 침대, 우측 후경에 거실 공간 배치. 좌측 가장자리에 캠핑카 외부의 후미등이 포함되어 문틀보다 바깥쪽 시점이 강조됨.",
        "entities": "차량 후미등, 캠핑카 문틀, 침대, 거실 공간. 천장에 켜진 조명 기구가 포함됨. 사람은 없음.",
        "hard_violations": [],
        "physics": "모든 가구와 사물이 물리 법칙에 어긋남 없이 배치됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "문턱 밖에서 왼쪽 침대와 오른쪽 안쪽 생활 공간을 보는 배치는 맞지만, 침대가 더 크게 전경을 차지하고 금지된 가시 조명과 미지정 장식·소품을 추가했다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "침대의 화면 점유가 더 작고 오른쪽 생활 공간까지의 동선이 더 명료해 상대적으로 가깝지만, 미지정 사진·짐·주방 소품과 창문의 문자 표식을 만들어 엄격한 조건에는 실패했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이나 시선, 겨냥하는 물체는 없다. 카메라는 열린 후면 출입구 바로 밖에서 실내를 비스듬히 내려다본다. 침대는 왼쪽 아래에서 안쪽으로 이어지고, 오른쪽 통로는 뒤편 테이블과 좌석을 향한다.",
        "built_space": "침대 한 개, 후면 출입구 한 곳, 왼쪽 측면 창 한 개와 안쪽 가로 창 한 개가 보인다. 안쪽에는 받침 기둥이 있는 테이블 한 개와 이를 둘러싼 벤치 좌석이 있고, 통로 양쪽에 수납장과 주방 설비가 있다. 문틀은 양쪽 가장자리에 있지만 침대는 중경보다 전경에 강하게 걸쳐 화면의 약 4할을 차지한다. 침대 옆 통로와 뒤편 휴식 공간의 연결은 가능하다. 상부에는 원형 조명 두 곳이 보인다. 창고 셔터는 화면 밖이라 상태를 확인할 수 없다.",
        "entities": "사람이나 얼굴은 없으며 침대와 생활 공간은 실제 캠퍼 내부로 읽힌다. 다만 침대 자체는 통상적인 두 사람용에 가까워 성인 네 명이 여유 있게 쉴 규모가 확실하지 않다. 침구 외에 벽 사진 여러 장, 수납 주머니, 컵과 용기, 러그, 주방 기기 등을 추가했다. 어두운 외부는 밤과 부합하지만 창고라는 장소를 입증하는 구조는 거의 보이지 않는다.",
        "hard_violations": [
         "새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 컵·용기와 주방 기기 등 미지정 물체를 다수 추가했다.",
         "가시적인 조명 기구를 만들지 말라는 명시적 조건과 달리 상부에 원형 조명 기구를 추가했다."
        ],
        "physics": "매트리스는 목재 침대 받침 위에 놓이고 이불은 매트리스 가장자리를 따라 처진다. 테이블은 바닥에 닿는 기둥으로 지지되며 컵과 용기는 상판에 놓여 있다. 좌석과 수납장은 바닥에 닿고 러그도 바닥 위에 있다. 떠 있거나 지지되지 않은 물체, 불가능한 신체나 반사는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "인물의 시선이나 무기, 이동 중인 물체는 없다. 문턱 바깥의 카메라가 왼쪽 침대의 윗면과 오른쪽 통로를 비스듬히 내려다본다. 통로는 화면 가운데 오른쪽에서 안쪽 테이블과 벤치 공간으로 이어진다.",
        "built_space": "침대 한 개와 열린 후면 출입구 한 곳, 왼쪽 측면 창 한 개 및 안쪽 가로 창 한 개가 보인다. 안쪽 테이블은 한 개이며 주변에 꺾여 이어지는 벤치 좌석이 있다. 왼쪽 침대 위 선반과 커튼, 통로 양쪽 수납·주방 설비가 배치되어 있다. 양쪽 문 가장자리는 주변부에 남고 침대는 화면의 약 3분의 1에서 4할 미만을 차지한다. 침대는 여전히 전경까지 내려오지만 A보다 오른쪽 생활 공간과 바닥 동선이 뚜렷하다. 창고 셔터는 보이지 않는다.",
        "entities": "사람과 얼굴은 없고 침대, 열린 후면 출입구, 안쪽 휴식 공간은 식별된다. 뒤편 좌석은 여러 명이 앉을 수 있어 보이지만 침대는 성인 네 명이 함께 여유 있게 쉬기에는 좁아 보인다. 벽 사진 여러 장, 상부 가방들, 통로의 여행 가방, 조리 용기와 주방 설비 등 미지정 물체가 있다. 왼쪽 창 하단에는 문구를 확정하기 어려운 문자성 표식도 보인다. 외부는 어둡지만 창고 위치는 확증하기 어렵다.",
        "hard_violations": [
         "새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 여러 가방과 조리 용기 등 미지정 물체를 다수 추가했다.",
         "허용된 원문이나 참조가 없는 문자성 표식을 왼쪽 창 하단에 추가했다."
        ],
        "physics": "침대와 매트리스는 목재 받침에 지지되고 침구는 그 위에 놓이거나 가장자리로 늘어진다. 안쪽 테이블은 바닥에 닿는 기둥과 받침이 있으며 가방은 선반 또는 바닥에 놓여 있다. 조리 용기는 상판 위에 있고 커튼은 상부에서 매달린 형태다. 지지 없이 떠 있는 물체나 불가능한 반사는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "문턱 밖에서 왼쪽 침대와 오른쪽 안쪽 생활 공간을 보는 배치는 맞지만, 침대가 더 크게 전경을 차지하고 금지된 가시 조명과 미지정 장식·소품을 추가했다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "침대의 화면 점유가 더 작고 오른쪽 생활 공간까지의 동선이 더 명료해 상대적으로 가깝지만, 미지정 사진·짐·주방 소품과 창문의 문자 표식을 만들어 엄격한 조건에는 실패했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물이나 시선, 겨냥하는 물체는 없다. 카메라는 열린 후면 출입구 바로 밖에서 실내를 비스듬히 내려다본다. 침대는 왼쪽 아래에서 안쪽으로 이어지고, 오른쪽 통로는 뒤편 테이블과 좌석을 향한다.",
        "built_space": "침대 한 개, 후면 출입구 한 곳, 왼쪽 측면 창 한 개와 안쪽 가로 창 한 개가 보인다. 안쪽에는 받침 기둥이 있는 테이블 한 개와 이를 둘러싼 벤치 좌석이 있고, 통로 양쪽에 수납장과 주방 설비가 있다. 문틀은 양쪽 가장자리에 있지만 침대는 중경보다 전경에 강하게 걸쳐 화면의 약 4할을 차지한다. 침대 옆 통로와 뒤편 휴식 공간의 연결은 가능하다. 상부에는 원형 조명 두 곳이 보인다. 창고 셔터는 화면 밖이라 상태를 확인할 수 없다.",
        "entities": "사람이나 얼굴은 없으며 침대와 생활 공간은 실제 캠퍼 내부로 읽힌다. 다만 침대 자체는 통상적인 두 사람용에 가까워 성인 네 명이 여유 있게 쉴 규모가 확실하지 않다. 침구 외에 벽 사진 여러 장, 수납 주머니, 컵과 용기, 러그, 주방 기기 등을 추가했다. 어두운 외부는 밤과 부합하지만 창고라는 장소를 입증하는 구조는 거의 보이지 않는다.",
        "hard_violations": [
         "새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 컵·용기와 주방 기기 등 미지정 물체를 다수 추가했다.",
         "가시적인 조명 기구를 만들지 말라는 명시적 조건과 달리 상부에 원형 조명 기구를 추가했다."
        ],
        "physics": "매트리스는 목재 침대 받침 위에 놓이고 이불은 매트리스 가장자리를 따라 처진다. 테이블은 바닥에 닿는 기둥으로 지지되며 컵과 용기는 상판에 놓여 있다. 좌석과 수납장은 바닥에 닿고 러그도 바닥 위에 있다. 떠 있거나 지지되지 않은 물체, 불가능한 신체나 반사는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "인물의 시선이나 무기, 이동 중인 물체는 없다. 문턱 바깥의 카메라가 왼쪽 침대의 윗면과 오른쪽 통로를 비스듬히 내려다본다. 통로는 화면 가운데 오른쪽에서 안쪽 테이블과 벤치 공간으로 이어진다.",
        "built_space": "침대 한 개와 열린 후면 출입구 한 곳, 왼쪽 측면 창 한 개 및 안쪽 가로 창 한 개가 보인다. 안쪽 테이블은 한 개이며 주변에 꺾여 이어지는 벤치 좌석이 있다. 왼쪽 침대 위 선반과 커튼, 통로 양쪽 수납·주방 설비가 배치되어 있다. 양쪽 문 가장자리는 주변부에 남고 침대는 화면의 약 3분의 1에서 4할 미만을 차지한다. 침대는 여전히 전경까지 내려오지만 A보다 오른쪽 생활 공간과 바닥 동선이 뚜렷하다. 창고 셔터는 보이지 않는다.",
        "entities": "사람과 얼굴은 없고 침대, 열린 후면 출입구, 안쪽 휴식 공간은 식별된다. 뒤편 좌석은 여러 명이 앉을 수 있어 보이지만 침대는 성인 네 명이 함께 여유 있게 쉬기에는 좁아 보인다. 벽 사진 여러 장, 상부 가방들, 통로의 여행 가방, 조리 용기와 주방 설비 등 미지정 물체가 있다. 왼쪽 창 하단에는 문구를 확정하기 어려운 문자성 표식도 보인다. 외부는 어둡지만 창고 위치는 확증하기 어렵다.",
        "hard_violations": [
         "새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 여러 가방과 조리 용기 등 미지정 물체를 다수 추가했다.",
         "허용된 원문이나 참조가 없는 문자성 표식을 왼쪽 창 하단에 추가했다."
        ],
        "physics": "침대와 매트리스는 목재 받침에 지지되고 침구는 그 위에 놓이거나 가장자리로 늘어진다. 안쪽 테이블은 바닥에 닿는 기둥과 받침이 있으며 가방은 선반 또는 바닥에 놓여 있다. 조리 용기는 상판 위에 있고 커튼은 상부에서 매달린 형태다. 지지 없이 떠 있는 물체나 불가능한 반사는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.464
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.214
   },
   "violations": {
    "B": [
     "[gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 컵·용기와 주방 기기 등 미지정 물체를 다수 추가했다.",
     "[gpt-high] 가시적인 조명 기구를 만들지 말라는 명시적 조건과 달리 상부에 원형 조명 기구를 추가했다."
    ],
    "A": [
     "[gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 여러 가방과 조리 용기 등 미지정 물체를 다수 추가했다.",
     "[gpt-high] 허용된 원문이나 참조가 없는 문자성 표식을 왼쪽 창 하단에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1214
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 카메라 구도와 조명 제한(눈에 띄는 조명 기구 배제)을 잘 준수하며, 요구된 캠핑카 내부의 깊이감과 공간 배치를 충실히 구현했습니다.  ★위반: [gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 여러 가방과 조리 용기 등 미지정 물체를 다수 추가했다. / [gpt-high] 허용된 원문이나 참조가 없는 문자성 표식을 왼쪽 창 하단에 추가했다."
   },
   {
    "label": "B",
    "score": 1214,
    "verdict_ko": "전반적인 공간 배치는 맞으나, 프레임 좌측에 차량 외부(후미등)가 과하게 노출되었고 프롬프트에서 금지한 켜진 조명 기구가 임의로 추가되었습니다.  ★위반: [gpt-high] 새로운 물체를 만들지 말라는 장소 제한에도 벽 사진, 컵·용기와 주방 기기 등 미지정 물체를 다수 추가했다. / [gpt-high] 가시적인 조명 기구를 만들지 말라는 명시적 조건과 달리 상부에 원형 조명 기구를 추가했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0afa-5080-77dd-93fd-2ca593bb698b",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh1::signage": {
  "fp": "acdd28391bff5a55",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh1": {
  "input_fingerprint": "7aadb73f2940dadb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 도로 위, 헤드라이트를 끈 낡은 캠핑카가 바퀴로 빗물을 튀기며 앞으로 내달리는 순간의 넓은 전경.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, under heavy rain. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the exterior track high behind and to the camper's left, looking diagonally down at its rear quarter while retaining a long stretch of road ahead. Keep the moving camper in the lower-left to central region at less than one-third of the image, traveling toward the upper-right depth, with rainwater displaced beside its wheels and no occupants discernible. Match its forward movement without changing the established viewing side, leaving the largely empty road to carry the sense of exposure before the explicit interior cut.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Moving camper truck in the lower-left of the frame, midground, moves toward Road extending into the upper-right distance; Road extending into the upper-right distance in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Moving along the road with headlights off and wheels displacing rainwater) — Rear and left side visible from an elevated rear-quarter angle, front directed into the upper-right depth; used as Moving focal subject within a broad environmental view; Outskirts road (Rain-covered and largely empty of other traffic) — Extends diagonally from the lower foreground toward the upper-right distance; used as Establishes forward direction and the exposed stretch ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, rain-muted nighttime illumination preserves the camper's outline and wheel spray while its headlights remain off.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper travels along a sparsely trafficked road at night in rain, with its headlights off and a single small light in the rear seating area. The supply bags are loaded aboard; the roof and window-frame leaks revealed shortly afterward are pre-existing defects, not newly acquired damage.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 도로 위, 헤드라이트를 끈 낡은 캠핑카가 바퀴로 빗물을 튀기며 앞으로 내달리는 순간의 넓은 전경.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, under heavy rain. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the exterior track high behind and to the camper's left, looking diagonally down at its rear quarter while retaining a long stretch of road ahead. Keep the moving camper in the lower-left to central region at less than one-third of the image, traveling toward the upper-right depth, with rainwater displaced beside its wheels and no occupants discernible. Match its forward movement without changing the established viewing side, leaving the largely empty road to carry the sense of exposure before the explicit interior cut.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Moving camper truck in the lower-left of the frame, midground, moves toward Road extending into the upper-right distance; Road extending into the upper-right distance in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Moving along the road with headlights off and wheels displacing rainwater) — Rear and left side visible from an elevated rear-quarter angle, front directed into the upper-right depth; used as Moving focal subject within a broad environmental view; Outskirts road (Rain-covered and largely empty of other traffic) — Extends diagonally from the lower foreground toward the upper-right distance; used as Establishes forward direction and the exposed stretch ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, rain-muted nighttime illumination preserves the camper's outline and wheel spray while its headlights remain off.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper travels along a sparsely trafficked road at night in rain, with its headlights off and a single small light in the rear seating area. The supply bags are loaded aboard; the roof and window-frame leaks revealed shortly afterward are pre-existing defects, not newly acquired damage.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 어두운 도로 위, 헤드라이트를 끈 낡은 캠핑카가 바퀴로 빗물을 튀기며 앞으로 내달리는 순간의 넓은 전경.\n\nLOCATION (lock): On a nearly deserted road at the city's outskirts at night, under heavy rain. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the exterior track high behind and to the camper's left, looking diagonally down at its rear quarter while retaining a long stretch of road ahead. Keep the moving camper in the lower-left to central region at less than one-third of the image, traveling toward the upper-right depth, with rainwater displaced beside its wheels and no occupants discernible. Match its forward movement without changing the established viewing side, leaving the largely empty road to carry the sense of exposure before the explicit interior cut.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Moving camper truck in the lower-left of the frame, midground, moves toward Road extending into the upper-right distance; Road extending into the upper-right distance in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Camper truck (Moving along the road with headlights off and wheels displacing rainwater) — Rear and left side visible from an elevated rear-quarter angle, front directed into the upper-right depth; used as Moving focal subject within a broad environmental view; Outskirts road (Rain-covered and largely empty of other traffic) — Extends diagonally from the lower foreground toward the upper-right distance; used as Establishes forward direction and the exposed stretch ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, rain-muted nighttime illumination preserves the camper's outline and wheel spray while its headlights remain off.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper travels along a sparsely trafficked road at night in rain, with its headlights off and a single small light in the rear seating area. The supply bags are loaded aboard; the roof and window-frame leaks revealed shortly afterward are pre-existing defects, not newly acquired damage.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 캠핑카의 좌측 후면 상단에서 피사체를 내려다보고 있으며, 캠핑카는 화면 우측 상단의 깊은 곳을 향해 주행하고 있습니다.",
    "built_space": "화면 좌측 하단에서 우측 상단으로 길게 뻗은 젖은 외곽 도로가 있으며, 우측으로 가로등이 줄지어 서 있고 멀리 도시의 윤곽이 보입니다.",
    "entities": "낡은 캠핑카가 묘사되어 있으며, 외부 조명은 꺼져 있고 후면 창문을 통해 단일 내부 조명이 보입니다. 식별 가능한 탑승자는 없습니다.",
    "hard_violations": [],
    "physics": "캠핑카의 바퀴가 지면에 닿아 있으며, 앞으로 나아가는 물리적 힘에 의해 바퀴 주변으로 빗물이 튀어오르는 모습이 자연스럽게 연출되었습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 캠핑카의 좌측 후방 높은 곳에서 하향으로 비스듬히 촬영하고 있으며, 캠핑카는 우측 상단의 원경을 향해 나아가고 있습니다.",
    "built_space": "다차선 도로가 우측 상단으로 뻗어 있으며, 우측에 늘어선 가로등과 멀리 떨어진 도시의 스카이라인이 보입니다.",
    "entities": "낡은 캠핑카가 주행 중이며, 후면 창문에 내부 조명이 보이지만 붉은색 테일램프가 밝게 켜져 있습니다. 식별 가능한 탑승자는 없습니다.",
    "hard_violations": [],
    "physics": "캠핑카가 젖은 노면 위를 주행하며 타이어를 통해 물보라를 일으키는 물리적 작용이 잘 표현되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 카메라 앵글, 프레이밍, 피사체의 위치와 방향을 완벽하게 구현했으며, 외부 조명이 모두 꺼진 상태에서 내부 후면 조명만 켜진 설정도 매우 정확하게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "구도와 카메라 앵글은 제시된 지침을 잘 따랐으나, 눈에 띄게 켜진 붉은색 테일램프가 '헤드라이트를 끈'이라는 어두운 조명 설정의 의도를 다소 약화시킵니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 캠핑카의 좌측 후면 상단에서 피사체를 내려다보고 있으며, 캠핑카는 화면 우측 상단의 깊은 곳을 향해 주행하고 있습니다.",
        "built_space": "화면 좌측 하단에서 우측 상단으로 길게 뻗은 젖은 외곽 도로가 있으며, 우측으로 가로등이 줄지어 서 있고 멀리 도시의 윤곽이 보입니다.",
        "entities": "낡은 캠핑카가 묘사되어 있으며, 외부 조명은 꺼져 있고 후면 창문을 통해 단일 내부 조명이 보입니다. 식별 가능한 탑승자는 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면에 닿아 있으며, 앞으로 나아가는 물리적 힘에 의해 바퀴 주변으로 빗물이 튀어오르는 모습이 자연스럽게 연출되었습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 캠핑카의 좌측 후방 높은 곳에서 하향으로 비스듬히 촬영하고 있으며, 캠핑카는 우측 상단의 원경을 향해 나아가고 있습니다.",
        "built_space": "다차선 도로가 우측 상단으로 뻗어 있으며, 우측에 늘어선 가로등과 멀리 떨어진 도시의 스카이라인이 보입니다.",
        "entities": "낡은 캠핑카가 주행 중이며, 후면 창문에 내부 조명이 보이지만 붉은색 테일램프가 밝게 켜져 있습니다. 식별 가능한 탑승자는 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카가 젖은 노면 위를 주행하며 타이어를 통해 물보라를 일으키는 물리적 작용이 잘 표현되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 카메라 앵글, 프레이밍, 피사체의 위치와 방향을 완벽하게 구현했으며, 외부 조명이 모두 꺼진 상태에서 내부 후면 조명만 켜진 설정도 매우 정확하게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "구도와 카메라 앵글은 제시된 지침을 잘 따랐으나, 눈에 띄게 켜진 붉은색 테일램프가 '헤드라이트를 끈'이라는 어두운 조명 설정의 의도를 다소 약화시킵니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 캠핑카의 좌측 후면 상단에서 피사체를 내려다보고 있으며, 캠핑카는 화면 우측 상단의 깊은 곳을 향해 주행하고 있습니다.",
        "built_space": "화면 좌측 하단에서 우측 상단으로 길게 뻗은 젖은 외곽 도로가 있으며, 우측으로 가로등이 줄지어 서 있고 멀리 도시의 윤곽이 보입니다.",
        "entities": "낡은 캠핑카가 묘사되어 있으며, 외부 조명은 꺼져 있고 후면 창문을 통해 단일 내부 조명이 보입니다. 식별 가능한 탑승자는 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면에 닿아 있으며, 앞으로 나아가는 물리적 힘에 의해 바퀴 주변으로 빗물이 튀어오르는 모습이 자연스럽게 연출되었습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 캠핑카의 좌측 후방 높은 곳에서 하향으로 비스듬히 촬영하고 있으며, 캠핑카는 우측 상단의 원경을 향해 나아가고 있습니다.",
        "built_space": "다차선 도로가 우측 상단으로 뻗어 있으며, 우측에 늘어선 가로등과 멀리 떨어진 도시의 스카이라인이 보입니다.",
        "entities": "낡은 캠핑카가 주행 중이며, 후면 창문에 내부 조명이 보이지만 붉은색 테일램프가 밝게 켜져 있습니다. 식별 가능한 탑승자는 없습니다.",
        "hard_violations": [],
        "physics": "캠핑카가 젖은 노면 위를 주행하며 타이어를 통해 물보라를 일으키는 물리적 작용이 잘 표현되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "작은 캠핑카를 좌하단 중경에 두고 긴 빗길을 남긴 넓은 전경은 더 충실하지만, 지정된 좌후방 대신 우후방에서 차량을 본다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "바퀴 물보라와 후방의 작은 실내등은 명확하지만, 차량이 더 크고 전경에 가까우며 좌후방 시점도 지키지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카 앞부분은 화면 우상단으로 이어지는 도로를 향하고, 뒤쪽은 좌하단을 향한다. 전진 방향과 도로의 소실 방향은 일치한다. 다만 후면 오른쪽으로 앞까지 이어지는 측면이 보여 차량의 오른쪽 측면을 보는 우후방 구도이며, 요구된 좌후방 구도와 반대다. 사람의 시선은 보이지 않고 전방 헤드라이트 광선도 없다.",
        "built_space": "젖은 도로가 아래쪽에서 우상단 먼 곳까지 이어지고 양옆에 가드레일이 있다. 여러 가로등과 전신주, 왼쪽의 별도 고가 구조물, 먼 고층 건물들이 보인다. 캠핑카는 좌하단 중경에서 화면 면적의 3분의 1보다 훨씬 작게 배치되어 있으며, 높은 외부 시점에서 지붕과 후면을 내려다본다. 후면에는 사다리 하나, 창 하나, 예비 타이어 하나가 있다. 노면의 길게 번진 조명 반사는 젖은 표면과 관찰 각도에 부합한다. 배경 시설의 구체적인 구성은 장소 설명보다 많이 구체화되어 있다.",
        "entities": "낡고 오염된 캠핑카 한 대, 빗물로 젖은 거의 빈 도로, 강한 야간 강우가 보인다. 다른 차량이나 식별 가능한 탑승자는 없다. 후면 창 안에 작은 따뜻한 불빛이 있으며 외부의 붉은 후미등과 구분된다. 헤드라이트 자체는 보이지 않지만 전방을 비추는 광선은 없다. 보급 가방과 누수 결함은 이 외부 원경에서 확인할 수 없으며, 등장인물의 신원 역시 판별 대상이 아니다.",
        "hard_violations": [],
        "physics": "차량은 타이어로 도로에 지지되어 있고, 바퀴 주변과 뒤쪽에 낮게 퍼지는 물보라가 보인다. 젖은 노면을 앞으로 달리며 물을 밀어내는 행동으로 설명 가능한 모습이다. 지붕 장비는 지붕에 얹혀 있으며 후면 사다리와 예비 타이어도 차체에 부착되어 있다. 지지 없이 뜬 물체나 불가능한 차체 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "캠핑카는 화면 우상단의 도로 깊이 방향으로 향하고 바퀴 물보라는 옆과 뒤로 퍼진다. 전진 목표와 도로 방향은 맞지만, 후면에서 화면 오른쪽으로 이어지는 차량 오른쪽 측면이 보여 지정된 좌후방이 아닌 우후방 시점이다. 탑승자의 시선은 보이지 않으며 차량 앞쪽에 헤드라이트 빛기둥도 없다.",
        "built_space": "도로는 아래쪽 전경에서 우상단으로 길게 이어진다. 양옆 가드레일, 오른쪽의 가로등과 전신주 열, 멀리 낮은 건물과 도시 건물들이 보인다. 높은 외부 시점이지만 캠핑카가 A보다 크게, 화면 아래쪽 전경에 더 가까이 놓여 중경의 작은 이동 대상이라는 요구에서는 덜 정확하다. 후면에 사다리 하나와 창 하나가 있고, 지붕에는 난간과 장비 및 어두운 적재물이 보인다. 가로등의 노면 반사는 가능한 방향으로 나타난다. 시설물의 상세 구성은 장소 설명에 없는 구체화다.",
        "entities": "오래되고 때가 탄 캠핑카 한 대가 비 내리는 밤의 빈 도로를 달린다. 뒤쪽 창 안에는 작은 따뜻한 등 하나와 커튼이 뚜렷하며 식별 가능한 사람은 없다. 후미등은 작고 어둡게 보이고 전방 조명 투사는 없다. 후면 번호판에 작은 문자 형태가 있으나 정확한 문구는 확정하기 어렵다. 지붕의 검은 적재물이 실내에 실린 보급 가방인지는 확인되지 않으며, 누수와 인물 외형도 이 프레임에서는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "보이는 타이어들이 노면에 닿아 차체를 지지하고, 앞뒤 바퀴 부근에서 물이 옆과 뒤로 튄다. 차량의 전진과 타이어의 수막 배제로 설명되는 물보라다. 지붕 적재물과 장비는 지붕 및 적재 난간에 받쳐져 있고 사다리도 차체에 붙어 있다. 떠 있는 차량이나 지지 없는 적재물은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "작은 캠핑카를 좌하단 중경에 두고 긴 빗길을 남긴 넓은 전경은 더 충실하지만, 지정된 좌후방 대신 우후방에서 차량을 본다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "바퀴 물보라와 후방의 작은 실내등은 명확하지만, 차량이 더 크고 전경에 가까우며 좌후방 시점도 지키지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 앞부분은 화면 우상단으로 이어지는 도로를 향하고, 뒤쪽은 좌하단을 향한다. 전진 방향과 도로의 소실 방향은 일치한다. 다만 후면 오른쪽으로 앞까지 이어지는 측면이 보여 차량의 오른쪽 측면을 보는 우후방 구도이며, 요구된 좌후방 구도와 반대다. 사람의 시선은 보이지 않고 전방 헤드라이트 광선도 없다.",
        "built_space": "젖은 도로가 아래쪽에서 우상단 먼 곳까지 이어지고 양옆에 가드레일이 있다. 여러 가로등과 전신주, 왼쪽의 별도 고가 구조물, 먼 고층 건물들이 보인다. 캠핑카는 좌하단 중경에서 화면 면적의 3분의 1보다 훨씬 작게 배치되어 있으며, 높은 외부 시점에서 지붕과 후면을 내려다본다. 후면에는 사다리 하나, 창 하나, 예비 타이어 하나가 있다. 노면의 길게 번진 조명 반사는 젖은 표면과 관찰 각도에 부합한다. 배경 시설의 구체적인 구성은 장소 설명보다 많이 구체화되어 있다.",
        "entities": "낡고 오염된 캠핑카 한 대, 빗물로 젖은 거의 빈 도로, 강한 야간 강우가 보인다. 다른 차량이나 식별 가능한 탑승자는 없다. 후면 창 안에 작은 따뜻한 불빛이 있으며 외부의 붉은 후미등과 구분된다. 헤드라이트 자체는 보이지 않지만 전방을 비추는 광선은 없다. 보급 가방과 누수 결함은 이 외부 원경에서 확인할 수 없으며, 등장인물의 신원 역시 판별 대상이 아니다.",
        "hard_violations": [],
        "physics": "차량은 타이어로 도로에 지지되어 있고, 바퀴 주변과 뒤쪽에 낮게 퍼지는 물보라가 보인다. 젖은 노면을 앞으로 달리며 물을 밀어내는 행동으로 설명 가능한 모습이다. 지붕 장비는 지붕에 얹혀 있으며 후면 사다리와 예비 타이어도 차체에 부착되어 있다. 지지 없이 뜬 물체나 불가능한 차체 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "캠핑카는 화면 우상단의 도로 깊이 방향으로 향하고 바퀴 물보라는 옆과 뒤로 퍼진다. 전진 목표와 도로 방향은 맞지만, 후면에서 화면 오른쪽으로 이어지는 차량 오른쪽 측면이 보여 지정된 좌후방이 아닌 우후방 시점이다. 탑승자의 시선은 보이지 않으며 차량 앞쪽에 헤드라이트 빛기둥도 없다.",
        "built_space": "도로는 아래쪽 전경에서 우상단으로 길게 이어진다. 양옆 가드레일, 오른쪽의 가로등과 전신주 열, 멀리 낮은 건물과 도시 건물들이 보인다. 높은 외부 시점이지만 캠핑카가 A보다 크게, 화면 아래쪽 전경에 더 가까이 놓여 중경의 작은 이동 대상이라는 요구에서는 덜 정확하다. 후면에 사다리 하나와 창 하나가 있고, 지붕에는 난간과 장비 및 어두운 적재물이 보인다. 가로등의 노면 반사는 가능한 방향으로 나타난다. 시설물의 상세 구성은 장소 설명에 없는 구체화다.",
        "entities": "오래되고 때가 탄 캠핑카 한 대가 비 내리는 밤의 빈 도로를 달린다. 뒤쪽 창 안에는 작은 따뜻한 등 하나와 커튼이 뚜렷하며 식별 가능한 사람은 없다. 후미등은 작고 어둡게 보이고 전방 조명 투사는 없다. 후면 번호판에 작은 문자 형태가 있으나 정확한 문구는 확정하기 어렵다. 지붕의 검은 적재물이 실내에 실린 보급 가방인지는 확인되지 않으며, 누수와 인물 외형도 이 프레임에서는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "보이는 타이어들이 노면에 닿아 차체를 지지하고, 앞뒤 바퀴 부근에서 물이 옆과 뒤로 튄다. 차량의 전진과 타이어의 수막 배제로 설명되는 물보라다. 지붕 적재물과 장비는 지붕 및 적재 난간에 받쳐져 있고 사다리도 차체에 붙어 있다. 떠 있는 차량이나 지지 없는 적재물은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.8
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1800
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "지정된 카메라 앵글, 프레이밍, 피사체의 위치와 방향을 완벽하게 구현했으며, 외부 조명이 모두 꺼진 상태에서 내부 후면 조명만 켜진 설정도 매우 정확하게 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 1800,
    "verdict_ko": "구도와 카메라 앵글은 제시된 지침을 잘 따랐으나, 눈에 띄게 켜진 붉은색 테일램프가 '헤드라이트를 끈'이라는 어두운 조명 설정의 의도를 다소 약화시킵니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0afe-f355-70b9-93aa-c1894d58b9f5",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh13::signage": {
  "fp": "e7ec6ddd9681120b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3666e4a4ce52ffbf": {
  "subjects": [],
  "subject_text": "경기도 화성 도로·오염된 들판, 도로 검문소, 산길 임도·캠핑카 정차 지점\n메마른 나무와 폐가가 흩어진 황량한 들판 사이 도로. 도로에 검문 구간이 있고, 옆으로 비포장 산길이 갈라져 산비탈을 오른다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L178",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::camper_living_space": {
  "input_fingerprint": "33286ce029b8c5b9",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camper_living_space",
    "tags": [
     "S44sh7",
     "S45sh13",
     "S55sh5"
    ]
   },
   "context_sig": "97ce65679ef31946"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n경기도 화성 도로·오염된 들판, 도로 검문소, 산길 임도·캠핑카 정차 지점: 황량한 중금속 오염지대와 경찰 통제선, 그리고 우회로인 비포장 산길이다. (특징: 메마른 나무와 회색빛 토양이 드러난 폐가 주변 들판; 경찰 제복과 방호 장비를 갖춘 인원들이 막아선 도로 검문소; 수풀이 우거진 비포장 흙길과 타이어가 펑크 나 주저앉은 캠핑카; 차체 아래 기어들어가 렌치로 나사를 조이는 현우) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비) / 폐공장 옆 캠핑카 보관 창고: 버려진 공장 지대 옆에 은닉된 낡은 창고로 어두운 셔터 문이 있다. (특징: 철제 셔터가 위로 말려 올라가는 입구; 먼지 덮인 낡고 작은 소형 캠핑카 외관)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 캠핑카의 뒷문을 열어보는 현우. 네 사람이 거뜬히 탈 만한, 침대도 있고...\n- 라울, 황급히 뒤에 실려 있던 가방에서 옷가지를 꺼내 천장에 대고, 앰버도 창틀에 가져다 대지만, 여기저기 새어 들어오는 빗물 막기엔 역부족...!\n- 55. 캠핑차 안 + 지방도로 -D\n\nTIME OF DAY (lock): night, rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n경기도 화성 도로·오염된 들판, 도로 검문소, 산길 임도·캠핑카 정차 지점: 황량한 중금속 오염지대와 경찰 통제선, 그리고 우회로인 비포장 산길이다. (특징: 메마른 나무와 회색빛 토양이 드러난 폐가 주변 들판; 경찰 제복과 방호 장비를 갖춘 인원들이 막아선 도로 검문소; 수풀이 우거진 비포장 흙길과 타이어가 펑크 나 주저앉은 캠핑카; 차체 아래 기어들어가 렌치로 나사를 조이는 현우) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비) / 폐공장 옆 캠핑카 보관 창고: 버려진 공장 지대 옆에 은닉된 낡은 창고로 어두운 셔터 문이 있다. (특징: 철제 셔터가 위로 말려 올라가는 입구; 먼지 덮인 낡고 작은 소형 캠핑카 외관)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 캠핑카의 뒷문을 열어보는 현우. 네 사람이 거뜬히 탈 만한, 침대도 있고...\n- 라울, 황급히 뒤에 실려 있던 가방에서 옷가지를 꺼내 천장에 대고, 앰버도 창틀에 가져다 대지만, 여기저기 새어 들어오는 빗물 막기엔 역부족...!\n- 55. 캠핑차 안 + 지방도로 -D\n\nTIME OF DAY (lock): night, rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camper_living_space_8fc641.png",
  "asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30",
  "input_asset_ids": [
   "c7e041f6-257d-4222-bf6a-78edf1c946f2"
  ],
  "origin_tag": "S45sh13",
  "place_text": "Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.",
  "origin_inputs": {
   "place_text": "Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.",
   "time_of_day_en": "night, rain",
   "conti_asset_id": "c7e041f6-257d-4222-bf6a-78edf1c946f2"
  }
 },
 "S45sh13::bgfirst_bg": {
  "input_fingerprint": "61efb7a10edde871",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 비가 새는 어두운 차 안에서 옷과 수건을 물구멍에 밀착시키고 힘껏 버티는 이현우, 앰버, 라울, 찰리의 복잡한 내부 구도.\n\nLOCATION (lock): Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.\n\nTIME OF DAY (lock): night, rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal into the compartment's left rear corner, holding a low, slightly upward diagonal toward the driver's seat; camera distance is the emphasized change as the interior opens into layered depth. Keep 찰리 seated along the left foreground edge, 앰버 braced at the right window, 라울 reaching toward the ceiling in the middle distance, and 이현우 remaining at the wheel beyond them, with gaps between their overlapping arms. 라울 looks up toward the leak, 앰버 concentrates on the window joint just outside the right edge, 찰리 watches the leaking ceiling, and 이현우 keeps his attention on the road beyond the windshield.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Compartment ceiling (Rain leaking through multiple points) — Its underside extends above the occupants toward the front seats; used as Connects the raised arms across the upper frame; Window frame (Leaking, with clothing pressed against it) — The interior joint is visible obliquely along the right edge; used as Establishes 앰버's separate effort without obscuring the driver; Front seats and steering wheel (이현우 remains in the driver's position) — Seat backs face the rear camera, with the steering wheel partially visible beyond; used as Anchors the deepest layer and preserves the distinction between driving and stopping leaks.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear lamp provides restrained visibility within the dark, rain-leaking interior, with subdued exposure and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 비가 새는 어두운 차 안에서 옷과 수건을 물구멍에 밀착시키고 힘껏 버티는 이현우, 앰버, 라울, 찰리의 복잡한 내부 구도.\n\nLOCATION (lock): Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior.\n\nTIME OF DAY (lock): night, rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal into the compartment's left rear corner, holding a low, slightly upward diagonal toward the driver's seat; camera distance is the emphasized change as the interior opens into layered depth. Keep 찰리 seated along the left foreground edge, 앰버 braced at the right window, 라울 reaching toward the ceiling in the middle distance, and 이현우 remaining at the wheel beyond them, with gaps between their overlapping arms. 라울 looks up toward the leak, 앰버 concentrates on the window joint just outside the right edge, 찰리 watches the leaking ceiling, and 이현우 keeps his attention on the road beyond the windshield.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Compartment ceiling (Rain leaking through multiple points) — Its underside extends above the occupants toward the front seats; used as Connects the raised arms across the upper frame; Window frame (Leaking, with clothing pressed against it) — The interior joint is visible obliquely along the right edge; used as Establishes 앰버's separate effort without obscuring the driver; Front seats and steering wheel (이현우 remains in the driver's position) — Seat backs face the rear camera, with the steering wheel partially visible beyond; used as Anchors the deepest layer and preserves the distinction between driving and stopping leaks.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear lamp provides restrained visibility within the dark, rain-leaking interior, with subdued exposure and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S45sh13__bgfirst_bg.png",
  "asset_id": "86a7bf1e-16a5-4e2c-98c0-5ae361a3dd53",
  "input_asset_ids": [
   "c7e041f6-257d-4222-bf6a-78edf1c946f2",
   "52016a0f-334c-4dbd-addb-6c0ad953ba30"
  ]
 },
 "S45sh13": {
  "input_fingerprint": "8bafd2b00bd5d1fb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비가 새는 어두운 차 안에서 옷과 수건을 물구멍에 밀착시키고 힘껏 버티는 이현우, 앰버, 라울, 찰리의 복잡한 내부 구도.\n\nLOCATION (lock): Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal into the compartment's left rear corner, holding a low, slightly upward diagonal toward the driver's seat; camera distance is the emphasized change as the interior opens into layered depth. Keep 찰리 seated along the left foreground edge, 앰버 braced at the right window, 라울 reaching toward the ceiling in the middle distance, and 이현우 remaining at the wheel beyond them, with gaps between their overlapping arms. 라울 looks up toward the leak, 앰버 concentrates on the window joint just outside the right edge, 찰리 watches the leaking ceiling, and 이현우 keeps his attention on the road beyond the windshield.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Compartment ceiling (Rain leaking through multiple points) — Its underside extends above the occupants toward the front seats; used as Connects the raised arms across the upper frame; Window frame (Leaking, with clothing pressed against it) — The interior joint is visible obliquely along the right edge; used as Establishes 앰버's separate effort without obscuring the driver; Front seats and steering wheel (이현우 remains in the driver's position) — Seat backs face the rear camera, with the steering wheel partially visible beyond; used as Anchors the deepest layer and preserves the distinction between driving and stopping leaks.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear lamp provides restrained visibility within the dark, rain-leaking interior, with subdued exposure and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain falls around the camper, whose headlights remain off while a single small light illuminates the rear compartment. Water leaks through the ceiling and window frames; the duffel bag and backpack contain the gathered supplies. 이현우: He remains at the wheel with an injured thigh and facial bruising. The contact card is concealed inside his shoe. 라울: He has pulled clothing from a bag and is pressing it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame in the rear compartment. 찰리: The old gorilla-shaped robot remains in the rear compartment with the blanket acquired at the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우, 앰버, 라울, 찰리 right now, so 이현우, 앰버, 라울, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우, 앰버, 라울, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비가 새는 어두운 차 안에서 옷과 수건을 물구멍에 밀착시키고 힘껏 버티는 이현우, 앰버, 라울, 찰리의 복잡한 내부 구도.\n\nLOCATION (lock): Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal into the compartment's left rear corner, holding a low, slightly upward diagonal toward the driver's seat; camera distance is the emphasized change as the interior opens into layered depth. Keep 찰리 seated along the left foreground edge, 앰버 braced at the right window, 라울 reaching toward the ceiling in the middle distance, and 이현우 remaining at the wheel beyond them, with gaps between their overlapping arms. 라울 looks up toward the leak, 앰버 concentrates on the window joint just outside the right edge, 찰리 watches the leaking ceiling, and 이현우 keeps his attention on the road beyond the windshield.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Compartment ceiling (Rain leaking through multiple points) — Its underside extends above the occupants toward the front seats; used as Connects the raised arms across the upper frame; Window frame (Leaking, with clothing pressed against it) — The interior joint is visible obliquely along the right edge; used as Establishes 앰버's separate effort without obscuring the driver; Front seats and steering wheel (이현우 remains in the driver's position) — Seat backs face the rear camera, with the steering wheel partially visible beyond; used as Anchors the deepest layer and preserves the distinction between driving and stopping leaks.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear lamp provides restrained visibility within the dark, rain-leaking interior, with subdued exposure and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain falls around the camper, whose headlights remain off while a single small light illuminates the rear compartment. Water leaks through the ceiling and window frames; the duffel bag and backpack contain the gathered supplies. 이현우: He remains at the wheel with an injured thigh and facial bruising. The contact card is concealed inside his shoe. 라울: He has pulled clothing from a bag and is pressing it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame in the rear compartment. 찰리: The old gorilla-shaped robot remains in the rear compartment with the blanket acquired at the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우, 앰버, 라울, 찰리 right now, so 이현우, 앰버, 라울, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우, 앰버, 라울, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 비가 새는 어두운 차 안에서 옷과 수건을 물구멍에 밀착시키고 힘껏 버티는 이현우, 앰버, 라울, 찰리의 복잡한 내부 구도.\n\nLOCATION (lock): Inside the camper's leaking passenger and living compartment on the rain-soaked road. A single small rear lamp lights the wet interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal into the compartment's left rear corner, holding a low, slightly upward diagonal toward the driver's seat; camera distance is the emphasized change as the interior opens into layered depth. Keep 찰리 seated along the left foreground edge, 앰버 braced at the right window, 라울 reaching toward the ceiling in the middle distance, and 이현우 remaining at the wheel beyond them, with gaps between their overlapping arms. 라울 looks up toward the leak, 앰버 concentrates on the window joint just outside the right edge, 찰리 watches the leaking ceiling, and 이현우 keeps his attention on the road beyond the windshield.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Compartment ceiling (Rain leaking through multiple points) — Its underside extends above the occupants toward the front seats; used as Connects the raised arms across the upper frame; Window frame (Leaking, with clothing pressed against it) — The interior joint is visible obliquely along the right edge; used as Establishes 앰버's separate effort without obscuring the driver; Front seats and steering wheel (이현우 remains in the driver's position) — Seat backs face the rear camera, with the steering wheel partially visible beyond; used as Anchors the deepest layer and preserves the distinction between driving and stopping leaks.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The small rear lamp provides restrained visibility within the dark, rain-leaking interior, with subdued exposure and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy rain falls around the camper, whose headlights remain off while a single small light illuminates the rear compartment. Water leaks through the ceiling and window frames; the duffel bag and backpack contain the gathered supplies. 이현우: He remains at the wheel with an injured thigh and facial bruising. The contact card is concealed inside his shoe. 라울: He has pulled clothing from a bag and is pressing it against the leaking ceiling. 앰버: She is pressing clothing against a leaking window frame in the rear compartment. 찰리: The old gorilla-shaped robot remains in the rear compartment with the blanket acquired at the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우, 앰버, 라울, 찰리 right now, so 이현우, 앰버, 라울, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우, 앰버, 라울, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S45sh13__bgfirst_bg.png",
     "asset_id": "86a7bf1e-16a5-4e2c-98c0-5ae361a3dd53",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S45sh13.png",
     "asset_id": "c7e041f6-257d-4222-bf6a-78edf1c946f2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camper_living_space_8fc641.png",
     "asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "라울은 양손으로 천장에 누른 천을 올려다보고, 앰버는 오른쪽 창틀에 밀착시킨 수건과 손 쪽을 본다. 찰리의 얼굴은 오른쪽 위를 향하지만 천장보다는 라울의 상체 쪽에 가까운 각도다. 이현우는 전방을 향해 앉아 있으나 고개가 오른쪽 아래로 숙여져 도로보다 계기판이나 조작부를 보는 듯하다. 무기나 이동 중인 인물은 없다.",
    "built_space": "양쪽 상부 수납장, 커튼 달린 측면 창열, 왼쪽 조리대, 양쪽 생활공간 좌석, 앞유리와 룸미러 하나, 작은 천장등 하나가 보인다. 앞좌석 등받이는 후방 카메라를 향하고 운전대 일부는 이현우 앞에 보인다. 찰리는 왼쪽 좌석, 앰버는 오른쪽 좌석, 라울은 가운데 통로, 이현우는 전방 운전 위치에 있다. 장소의 낡은 재료와 젖은 표면은 참조에 가깝다. 낮은 후방 와이드지만 시축은 비교적 중앙에 가깝고, 크게 잡힌 라울이 중간 공간을 많이 가린다.",
    "entities": "지정된 세 사람과 로봇 하나만 보인다. 찰리는 모래색 장갑판, 흰 기계 얼굴, 주황색 눈, 파란 가슴 원자로와 담요를 갖췄다. 이현우는 짧은 검은 머리와 어두운 셔츠의 젊은 남성이며, 후면 위주라 정확한 얼굴·멍·인이어는 확인하기 어렵다. 앰버는 어린 금발 소녀로 방진 마스크, 카키 작업복, 공구 벨트를 착용했다. 라울은 짙은 피부의 어린 소년으로 묶은 곱슬머리, 낡은 티셔츠와 반바지가 맞는다. 천장용 천과 창틀용 수건이 보이고 오른쪽 아래에 가방이 있으나 더플백과 배낭 두 종류 및 내용물은 구분하기 어렵다. 밤비와 여러 누수 지점이 표현됐다.",
    "hard_violations": [],
    "physics": "찰리는 왼쪽 좌석에 앉고 담요는 무릎과 기계 손에 받쳐져 있다. 라울은 아래쪽에 보이는 다리와 신발로 통로 바닥에 서서 양손으로 천을 천장에 누른다. 앰버의 접힌 다리와 무릎은 오른쪽 좌석에 지지되고 두 손은 수건을 창틀에 압착한다. 이현우의 몸은 운전석에 지지되며 팔은 운전대와 조작부 쪽으로 뻗어 있다. 천과 수건은 손으로 고정되고 물방울은 천장에서 아래로 떨어진다. 지지 없이 떠 있는 몸이나 물체는 없다."
   },
   {
    "label": "A",
    "direction": "라울은 손으로 누르는 천장 누수 지점을 올려다본다. 앰버는 오른쪽 가장자리의 창틀 이음매와 그곳에 댄 천을 바라본다. 찰리는 얼굴을 명확히 위로 기울여 새는 천장을 관찰한다. 이현우는 후두부를 카메라에 보이며 앞유리 너머를 향한다. 네 인물의 주의 대상이 각각 구분되고, 들어 올린 팔 사이에도 전방을 볼 틈이 남는다.",
    "built_space": "작은 천장등 하나, 룸미러 하나, 앞좌석 두 개, 양쪽 상부 수납장과 커튼 달린 측면 창열, 왼쪽 조리대, 좌우 생활공간 좌석이 보인다. 앞좌석 등받이는 카메라 쪽이고 운전대는 왼쪽 운전석 앞에 일부 가려져 있다. 찰리는 왼쪽 전경에 앉고 라울은 그 뒤 통로에서 천장으로 손을 뻗으며, 앰버는 오른쪽 좌석을 딛고 창틀을 막는다. 이현우는 가장 깊은 왼쪽 운전석에 있다. 낮은 후방 와이드에서 젖은 바닥과 앞좌석까지 깊이가 열려 A보다 후퇴한 카메라의 효과가 잘 드러난다. 다만 왼쪽 모서리에서의 사선성은 강하지 않다.",
    "entities": "추가 인물 없이 찰리, 이현우, 앰버, 라울이 보인다. 찰리의 육중한 기계 팔, 모래색 장갑, 흰 마스크형 얼굴과 파란 원자로는 참조와 일치하며 담요를 쥐고 있다. 이현우는 검은 머리와 낡은 어두운 옷을 입은 젊은 남성으로, 얼굴과 허벅지는 가려져 부상 여부를 판정할 수 없다. 앰버는 어린 금발 소녀이며 마스크·카키 작업복·공구 벨트가 맞지만 머리를 묶은 점은 참조와 다르다. 라울은 짙은 피부, 뒤로 모은 곱슬머리, 낡은 티셔츠와 반바지의 어린 소년이다. 두 사람이 누르는 천과 오른쪽 아래의 가방은 보이지만 가방 두 종류와 보급품 내용은 확인되지 않는다. 어두운 밤비, 작은 실내등, 다중 누수도 맞는다.",
    "hard_violations": [],
    "physics": "찰리의 골반과 다리는 왼쪽 좌석에 받쳐지고 담요는 두 기계 손과 무릎에 걸쳐 있다. 라울의 다리는 통로 바닥 방향으로 이어지며 발 접점은 하단 밖에 있지만, 몸이 떠 있는 모습은 아니다. 뻗은 두 손이 천을 천장에 눌러 고정한다. 앰버는 오른쪽 좌석에 굽힌 다리와 무릎을 대어 몸을 지지하고 양손으로 창틀의 천을 누른다. 이현우는 운전석에 앉아 전방으로 팔을 뻗고 있으나 손과 운전대의 접점은 대부분 가려져 있다. 떨어지는 물은 천장에서 바닥으로 이어지며, 명백하게 지지 없는 물체나 불가능한 자세는 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "인물 배치와 누수 차단 동작은 충실하지만, 라울이 전경에 더 크게 걸리고 찰리의 천장 주시와 이현우의 도로 주시가 B보다 불명확하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 전경의 찰리부터 중간의 라울, 깊숙한 운전석까지 공간이 잘 열리며, 천장·오른쪽 창틀·전방 도로를 향하는 시선과 동작이 지시를 더 정확히 따른다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울은 양손으로 천장에 누른 천을 올려다보고, 앰버는 오른쪽 창틀에 밀착시킨 수건과 손 쪽을 본다. 찰리의 얼굴은 오른쪽 위를 향하지만 천장보다는 라울의 상체 쪽에 가까운 각도다. 이현우는 전방을 향해 앉아 있으나 고개가 오른쪽 아래로 숙여져 도로보다 계기판이나 조작부를 보는 듯하다. 무기나 이동 중인 인물은 없다.",
        "built_space": "양쪽 상부 수납장, 커튼 달린 측면 창열, 왼쪽 조리대, 양쪽 생활공간 좌석, 앞유리와 룸미러 하나, 작은 천장등 하나가 보인다. 앞좌석 등받이는 후방 카메라를 향하고 운전대 일부는 이현우 앞에 보인다. 찰리는 왼쪽 좌석, 앰버는 오른쪽 좌석, 라울은 가운데 통로, 이현우는 전방 운전 위치에 있다. 장소의 낡은 재료와 젖은 표면은 참조에 가깝다. 낮은 후방 와이드지만 시축은 비교적 중앙에 가깝고, 크게 잡힌 라울이 중간 공간을 많이 가린다.",
        "entities": "지정된 세 사람과 로봇 하나만 보인다. 찰리는 모래색 장갑판, 흰 기계 얼굴, 주황색 눈, 파란 가슴 원자로와 담요를 갖췄다. 이현우는 짧은 검은 머리와 어두운 셔츠의 젊은 남성이며, 후면 위주라 정확한 얼굴·멍·인이어는 확인하기 어렵다. 앰버는 어린 금발 소녀로 방진 마스크, 카키 작업복, 공구 벨트를 착용했다. 라울은 짙은 피부의 어린 소년으로 묶은 곱슬머리, 낡은 티셔츠와 반바지가 맞는다. 천장용 천과 창틀용 수건이 보이고 오른쪽 아래에 가방이 있으나 더플백과 배낭 두 종류 및 내용물은 구분하기 어렵다. 밤비와 여러 누수 지점이 표현됐다.",
        "hard_violations": [],
        "physics": "찰리는 왼쪽 좌석에 앉고 담요는 무릎과 기계 손에 받쳐져 있다. 라울은 아래쪽에 보이는 다리와 신발로 통로 바닥에 서서 양손으로 천을 천장에 누른다. 앰버의 접힌 다리와 무릎은 오른쪽 좌석에 지지되고 두 손은 수건을 창틀에 압착한다. 이현우의 몸은 운전석에 지지되며 팔은 운전대와 조작부 쪽으로 뻗어 있다. 천과 수건은 손으로 고정되고 물방울은 천장에서 아래로 떨어진다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "라울은 손으로 누르는 천장 누수 지점을 올려다본다. 앰버는 오른쪽 가장자리의 창틀 이음매와 그곳에 댄 천을 바라본다. 찰리는 얼굴을 명확히 위로 기울여 새는 천장을 관찰한다. 이현우는 후두부를 카메라에 보이며 앞유리 너머를 향한다. 네 인물의 주의 대상이 각각 구분되고, 들어 올린 팔 사이에도 전방을 볼 틈이 남는다.",
        "built_space": "작은 천장등 하나, 룸미러 하나, 앞좌석 두 개, 양쪽 상부 수납장과 커튼 달린 측면 창열, 왼쪽 조리대, 좌우 생활공간 좌석이 보인다. 앞좌석 등받이는 카메라 쪽이고 운전대는 왼쪽 운전석 앞에 일부 가려져 있다. 찰리는 왼쪽 전경에 앉고 라울은 그 뒤 통로에서 천장으로 손을 뻗으며, 앰버는 오른쪽 좌석을 딛고 창틀을 막는다. 이현우는 가장 깊은 왼쪽 운전석에 있다. 낮은 후방 와이드에서 젖은 바닥과 앞좌석까지 깊이가 열려 A보다 후퇴한 카메라의 효과가 잘 드러난다. 다만 왼쪽 모서리에서의 사선성은 강하지 않다.",
        "entities": "추가 인물 없이 찰리, 이현우, 앰버, 라울이 보인다. 찰리의 육중한 기계 팔, 모래색 장갑, 흰 마스크형 얼굴과 파란 원자로는 참조와 일치하며 담요를 쥐고 있다. 이현우는 검은 머리와 낡은 어두운 옷을 입은 젊은 남성으로, 얼굴과 허벅지는 가려져 부상 여부를 판정할 수 없다. 앰버는 어린 금발 소녀이며 마스크·카키 작업복·공구 벨트가 맞지만 머리를 묶은 점은 참조와 다르다. 라울은 짙은 피부, 뒤로 모은 곱슬머리, 낡은 티셔츠와 반바지의 어린 소년이다. 두 사람이 누르는 천과 오른쪽 아래의 가방은 보이지만 가방 두 종류와 보급품 내용은 확인되지 않는다. 어두운 밤비, 작은 실내등, 다중 누수도 맞는다.",
        "hard_violations": [],
        "physics": "찰리의 골반과 다리는 왼쪽 좌석에 받쳐지고 담요는 두 기계 손과 무릎에 걸쳐 있다. 라울의 다리는 통로 바닥 방향으로 이어지며 발 접점은 하단 밖에 있지만, 몸이 떠 있는 모습은 아니다. 뻗은 두 손이 천을 천장에 눌러 고정한다. 앰버는 오른쪽 좌석에 굽힌 다리와 무릎을 대어 몸을 지지하고 양손으로 창틀의 천을 누른다. 이현우는 운전석에 앉아 전방으로 팔을 뻗고 있으나 손과 운전대의 접점은 대부분 가려져 있다. 떨어지는 물은 천장에서 바닥으로 이어지며, 명백하게 지지 없는 물체나 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물 배치와 누수 차단 동작은 충실하지만, 라울이 전경에 더 크게 걸리고 찰리의 천장 주시와 이현우의 도로 주시가 B보다 불명확하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 전경의 찰리부터 중간의 라울, 깊숙한 운전석까지 공간이 잘 열리며, 천장·오른쪽 창틀·전방 도로를 향하는 시선과 동작이 지시를 더 정확히 따른다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "라울은 양손으로 천장에 누른 천을 올려다보고, 앰버는 오른쪽 창틀에 밀착시킨 수건과 손 쪽을 본다. 찰리의 얼굴은 오른쪽 위를 향하지만 천장보다는 라울의 상체 쪽에 가까운 각도다. 이현우는 전방을 향해 앉아 있으나 고개가 오른쪽 아래로 숙여져 도로보다 계기판이나 조작부를 보는 듯하다. 무기나 이동 중인 인물은 없다.",
        "built_space": "양쪽 상부 수납장, 커튼 달린 측면 창열, 왼쪽 조리대, 양쪽 생활공간 좌석, 앞유리와 룸미러 하나, 작은 천장등 하나가 보인다. 앞좌석 등받이는 후방 카메라를 향하고 운전대 일부는 이현우 앞에 보인다. 찰리는 왼쪽 좌석, 앰버는 오른쪽 좌석, 라울은 가운데 통로, 이현우는 전방 운전 위치에 있다. 장소의 낡은 재료와 젖은 표면은 참조에 가깝다. 낮은 후방 와이드지만 시축은 비교적 중앙에 가깝고, 크게 잡힌 라울이 중간 공간을 많이 가린다.",
        "entities": "지정된 세 사람과 로봇 하나만 보인다. 찰리는 모래색 장갑판, 흰 기계 얼굴, 주황색 눈, 파란 가슴 원자로와 담요를 갖췄다. 이현우는 짧은 검은 머리와 어두운 셔츠의 젊은 남성이며, 후면 위주라 정확한 얼굴·멍·인이어는 확인하기 어렵다. 앰버는 어린 금발 소녀로 방진 마스크, 카키 작업복, 공구 벨트를 착용했다. 라울은 짙은 피부의 어린 소년으로 묶은 곱슬머리, 낡은 티셔츠와 반바지가 맞는다. 천장용 천과 창틀용 수건이 보이고 오른쪽 아래에 가방이 있으나 더플백과 배낭 두 종류 및 내용물은 구분하기 어렵다. 밤비와 여러 누수 지점이 표현됐다.",
        "hard_violations": [],
        "physics": "찰리는 왼쪽 좌석에 앉고 담요는 무릎과 기계 손에 받쳐져 있다. 라울은 아래쪽에 보이는 다리와 신발로 통로 바닥에 서서 양손으로 천을 천장에 누른다. 앰버의 접힌 다리와 무릎은 오른쪽 좌석에 지지되고 두 손은 수건을 창틀에 압착한다. 이현우의 몸은 운전석에 지지되며 팔은 운전대와 조작부 쪽으로 뻗어 있다. 천과 수건은 손으로 고정되고 물방울은 천장에서 아래로 떨어진다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "라울은 손으로 누르는 천장 누수 지점을 올려다본다. 앰버는 오른쪽 가장자리의 창틀 이음매와 그곳에 댄 천을 바라본다. 찰리는 얼굴을 명확히 위로 기울여 새는 천장을 관찰한다. 이현우는 후두부를 카메라에 보이며 앞유리 너머를 향한다. 네 인물의 주의 대상이 각각 구분되고, 들어 올린 팔 사이에도 전방을 볼 틈이 남는다.",
        "built_space": "작은 천장등 하나, 룸미러 하나, 앞좌석 두 개, 양쪽 상부 수납장과 커튼 달린 측면 창열, 왼쪽 조리대, 좌우 생활공간 좌석이 보인다. 앞좌석 등받이는 카메라 쪽이고 운전대는 왼쪽 운전석 앞에 일부 가려져 있다. 찰리는 왼쪽 전경에 앉고 라울은 그 뒤 통로에서 천장으로 손을 뻗으며, 앰버는 오른쪽 좌석을 딛고 창틀을 막는다. 이현우는 가장 깊은 왼쪽 운전석에 있다. 낮은 후방 와이드에서 젖은 바닥과 앞좌석까지 깊이가 열려 A보다 후퇴한 카메라의 효과가 잘 드러난다. 다만 왼쪽 모서리에서의 사선성은 강하지 않다.",
        "entities": "추가 인물 없이 찰리, 이현우, 앰버, 라울이 보인다. 찰리의 육중한 기계 팔, 모래색 장갑, 흰 마스크형 얼굴과 파란 원자로는 참조와 일치하며 담요를 쥐고 있다. 이현우는 검은 머리와 낡은 어두운 옷을 입은 젊은 남성으로, 얼굴과 허벅지는 가려져 부상 여부를 판정할 수 없다. 앰버는 어린 금발 소녀이며 마스크·카키 작업복·공구 벨트가 맞지만 머리를 묶은 점은 참조와 다르다. 라울은 짙은 피부, 뒤로 모은 곱슬머리, 낡은 티셔츠와 반바지의 어린 소년이다. 두 사람이 누르는 천과 오른쪽 아래의 가방은 보이지만 가방 두 종류와 보급품 내용은 확인되지 않는다. 어두운 밤비, 작은 실내등, 다중 누수도 맞는다.",
        "hard_violations": [],
        "physics": "찰리의 골반과 다리는 왼쪽 좌석에 받쳐지고 담요는 두 기계 손과 무릎에 걸쳐 있다. 라울의 다리는 통로 바닥 방향으로 이어지며 발 접점은 하단 밖에 있지만, 몸이 떠 있는 모습은 아니다. 뻗은 두 손이 천을 천장에 눌러 고정한다. 앰버는 오른쪽 좌석에 굽힌 다리와 무릎을 대어 몸을 지지하고 양손으로 창틀의 천을 누른다. 이현우는 운전석에 앉아 전방으로 팔을 뻗고 있으나 손과 운전대의 접점은 대부분 가려져 있다. 떨어지는 물은 천장에서 바닥으로 이어지며, 명백하게 지지 없는 물체나 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 8,
   "A": 9
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "인물 배치와 누수 차단 동작은 충실하지만, 라울이 전경에 더 크게 걸리고 찰리의 천장 주시와 이현우의 도로 주시가 B보다 불명확하다."
   },
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "왼쪽 전경의 찰리부터 중간의 라울, 깊숙한 운전석까지 공간이 잘 열리며, 천장·오른쪽 창틀·전방 도로를 향하는 시선과 동작이 지시를 더 정확히 따른다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camper_living_space_8fc641.png",
    "asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b04-451e-7f0f-9a5d-8be6ead925da",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S45sh13__bgfirst_bg.png",
   "bg_asset_id": "86a7bf1e-16a5-4e2c-98c0-5ae361a3dd53",
   "bg_record_key": "S45sh13::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "camper_living_space",
   "groupbg_asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S46sh1::signage": {
  "fp": "5e541cbc72644010",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S46sh1": {
  "input_fingerprint": "3d9fddf41a7b7496",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 마른 넝쿨이 덮이고 큼지막한 금이 간 도서관 건물 벽면 앞에 낡은 캠핑카가 멈춰 선 넓은 구도.\n\nLOCATION (lock): Outside an abandoned library at night in the rain, directly in front of its cracked, dead-vine-covered facade. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the elevated exterior position to one side of the camper's approach, looking diagonally downward across the stopped vehicle toward the library entrance. Hold the opening composition before the inward dolly begins, placing the camper in the lower-left third at less than a third of the image and distributing the cracked facade, clinging vines, and old sign across the deeper plane. Emphasize the expanded camera distance rather than introducing an additional visual flourish, and show no occupants through the vehicle.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old camper (Stopped in front of the library) — Seen obliquely from above, with its side and one end visible; used as Provides foreground scale and leads the eye toward the entrance; Library facade and entrance (Abandoned, with large cracks and dry vines clinging to the walls) — The entrance-facing facade recedes diagonally across the background; used as Establishes the destination and the route of the next camera movement; Old library sign (Aged) — The inscribed face bearing '미추홀도서관, 인천' is visible above the arrival area; used as Identifies the location within the architectural context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination appropriate to the rainy exterior preserves the facade's damage without introducing conspicuous light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rain continues at night around the parked, leaking camper. Dry vines cling to the library's deeply cracked walls beneath the old sign reading “미추홀도서관, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 마른 넝쿨이 덮이고 큼지막한 금이 간 도서관 건물 벽면 앞에 낡은 캠핑카가 멈춰 선 넓은 구도.\n\nLOCATION (lock): Outside an abandoned library at night in the rain, directly in front of its cracked, dead-vine-covered facade. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the elevated exterior position to one side of the camper's approach, looking diagonally downward across the stopped vehicle toward the library entrance. Hold the opening composition before the inward dolly begins, placing the camper in the lower-left third at less than a third of the image and distributing the cracked facade, clinging vines, and old sign across the deeper plane. Emphasize the expanded camera distance rather than introducing an additional visual flourish, and show no occupants through the vehicle.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old camper (Stopped in front of the library) — Seen obliquely from above, with its side and one end visible; used as Provides foreground scale and leads the eye toward the entrance; Library facade and entrance (Abandoned, with large cracks and dry vines clinging to the walls) — The entrance-facing facade recedes diagonally across the background; used as Establishes the destination and the route of the next camera movement; Old library sign (Aged) — The inscribed face bearing '미추홀도서관, 인천' is visible above the arrival area; used as Identifies the location within the architectural context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination appropriate to the rainy exterior preserves the facade's damage without introducing conspicuous light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rain continues at night around the parked, leaking camper. Dry vines cling to the library's deeply cracked walls beneath the old sign reading “미추홀도서관, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 마른 넝쿨이 덮이고 큼지막한 금이 간 도서관 건물 벽면 앞에 낡은 캠핑카가 멈춰 선 넓은 구도.\n\nLOCATION (lock): Outside an abandoned library at night in the rain, directly in front of its cracked, dead-vine-covered facade. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin from the elevated exterior position to one side of the camper's approach, looking diagonally downward across the stopped vehicle toward the library entrance. Hold the opening composition before the inward dolly begins, placing the camper in the lower-left third at less than a third of the image and distributing the cracked facade, clinging vines, and old sign across the deeper plane. Emphasize the expanded camera distance rather than introducing an additional visual flourish, and show no occupants through the vehicle.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old camper (Stopped in front of the library) — Seen obliquely from above, with its side and one end visible; used as Provides foreground scale and leads the eye toward the entrance; Library facade and entrance (Abandoned, with large cracks and dry vines clinging to the walls) — The entrance-facing facade recedes diagonally across the background; used as Establishes the destination and the route of the next camera movement; Old library sign (Aged) — The inscribed face bearing '미추홀도서관, 인천' is visible above the arrival area; used as Identifies the location within the architectural context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination appropriate to the rainy exterior preserves the facade's damage without introducing conspicuous light sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rain continues at night around the parked, leaking camper. Dry vines cling to the library's deeply cracked walls beneath the old sign reading “미추홀도서관, 인천.”\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
    "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 화분 등이 적절히 배치됨.",
    "entities": "좌측 하단의 낡은 캠핑카, 넝쿨이 덮인 도서관 벽면. 간판 텍스트가 '미추홀 도서관 인천'으로 표시되어 레퍼런스의 영문을 한글로 변형함. 야간, 비.",
    "hard_violations": [],
    "physics": "캠핑카가 지면에 정상적으로 닿아 있으며, 비가 자연스럽게 내리고 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
    "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 기둥, 화분 등의 요소가 정확한 위치에 있음.",
    "entities": "좌측 하단의 낡은 캠핑카, 금이 가고 넝쿨이 덮인 벽면. 간판은 '미추홀 도서관'과 작은 영문 'Michuhol Library Incheon'으로 레퍼런스와 정확히 일치. 야간, 비.",
    "hard_violations": [],
    "physics": "캠핑카가 젖은 바닥에 안정적으로 정차해 있으며, 빗방울이 중력에 맞게 떨어지고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 카메라 구도와 야간 비 내리는 분위기를 훌륭하게 구현했으며, 레퍼런스의 간판 텍스트(한영 혼용)를 임의로 변경하지 않고 정확히 재현하여 지침을 완벽히 준수함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 구도와 분위기는 잘 연출되었으나, 레퍼런스의 영문 간판 텍스트를 한글로 임의 번역 및 통합하여 스크립트 변경 금지 지침을 위반함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
        "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 기둥, 화분 등의 요소가 정확한 위치에 있음.",
        "entities": "좌측 하단의 낡은 캠핑카, 금이 가고 넝쿨이 덮인 벽면. 간판은 '미추홀 도서관'과 작은 영문 'Michuhol Library Incheon'으로 레퍼런스와 정확히 일치. 야간, 비.",
        "hard_violations": [],
        "physics": "캠핑카가 젖은 바닥에 안정적으로 정차해 있으며, 빗방울이 중력에 맞게 떨어지고 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
        "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 화분 등이 적절히 배치됨.",
        "entities": "좌측 하단의 낡은 캠핑카, 넝쿨이 덮인 도서관 벽면. 간판 텍스트가 '미추홀 도서관 인천'으로 표시되어 레퍼런스의 영문을 한글로 변형함. 야간, 비.",
        "hard_violations": [],
        "physics": "캠핑카가 지면에 정상적으로 닿아 있으며, 비가 자연스럽게 내리고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 카메라 구도와 야간 비 내리는 분위기를 훌륭하게 구현했으며, 레퍼런스의 간판 텍스트(한영 혼용)를 임의로 변경하지 않고 정확히 재현하여 지침을 완벽히 준수함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 구도와 분위기는 잘 연출되었으나, 레퍼런스의 영문 간판 텍스트를 한글로 임의 번역 및 통합하여 스크립트 변경 금지 지침을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
        "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 기둥, 화분 등의 요소가 정확한 위치에 있음.",
        "entities": "좌측 하단의 낡은 캠핑카, 금이 가고 넝쿨이 덮인 벽면. 간판은 '미추홀 도서관'과 작은 영문 'Michuhol Library Incheon'으로 레퍼런스와 정확히 일치. 야간, 비.",
        "hard_violations": [],
        "physics": "캠핑카가 젖은 바닥에 안정적으로 정차해 있으며, 빗방울이 중력에 맞게 떨어지고 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 캠핑카 위를 지나 도서관 입구를 향해 대각선 아래로 향함.",
        "built_space": "레퍼런스와 동일한 구조의 버려진 도서관. 입구 캐노피, 유리창, 바닥 타일, 화분 등이 적절히 배치됨.",
        "entities": "좌측 하단의 낡은 캠핑카, 넝쿨이 덮인 도서관 벽면. 간판 텍스트가 '미추홀 도서관 인천'으로 표시되어 레퍼런스의 영문을 한글로 변형함. 야간, 비.",
        "hard_violations": [],
        "physics": "캠핑카가 지면에 정상적으로 닿아 있으며, 비가 자연스럽게 내리고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캠핑카 지붕을 더 내려다보는 사선 시점과 좌하단 배치가 지정된 와이드 구도에 더 가깝고 원본 간판의 문자 체계도 유지하지만, 가로등과 출입구 조명은 요구보다 두드러진다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "비 오는 밤의 장소와 정차 상태는 맞지만, 낮은 카메라 시점이 지정된 하향 사선 구도에서 멀어지고 원본 간판의 영문 표기를 한글로 바꿨다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카는 뒤쪽이 화면 왼쪽 아래, 운전석 쪽이 오른쪽 위를 향하여 도서관 출입구 방향으로 놓여 있다. 카메라는 차량의 후면·측면·지붕을 함께 보며 출입구 쪽을 비스듬히 바라본다. 차량 지붕을 보는 하향 각도는 B보다 뚜렷하지만 높은 외부 시점이라는 인상은 다소 약하다. 사람이나 시선, 겨누는 물체는 없다.",
        "built_space": "청록색 수평 띠창이 모서리를 감싸는 석재 건물, 왼쪽의 낮은 건물 부분, 간판을 단 큰 입구 보와 왼쪽 사각 지지부가 보인다. 입구 아래 원형 기둥은 세 개, 유리 출입구는 한 곳이며 참고 사진의 주요 구조와 대응한다. 화분 두 개가 뚜렷하고 왼쪽 구역 일부는 차량에 가린다. 벽돌 포장과 출입구로 이어지는 노란 유도블록도 유지된다. 젖은 바닥의 조명 반사는 광원 위치와 모순되지 않는다.",
        "entities": "낡은 캠핑카 한 대가 화면 좌하단에서 전체 면적의 삼분의 일 미만을 차지하며 탑승자나 얼굴은 보이지 않는다. 깊게 갈라진 외벽과 마른 넝쿨, 오래된 도서관 간판, 야간의 빗줄기가 있다. 간판은 참고 사진처럼 한글 도서관명과 오른쪽 영문 도서관명·도시명을 유지한다. 차량 외부는 젖었지만 내부로 물이 새는지는 확인할 수 없다. 왼쪽 가로등과 입구 조명은 은은한 야간 조명이라는 요구보다 눈에 띈다.",
        "hard_violations": [],
        "physics": "캠핑카의 보이는 앞뒤 바퀴가 포장면에 닿아 차체를 지지하며 정차 상태로 읽힌다. 지붕 장비와 후면 사다리는 차량에 부착되어 있다. 입구 보는 사각 지지부와 건물에 연결되고, 안쪽 차양은 기둥으로 지지된다. 화분은 바닥에 놓이고 넝쿨은 벽에 붙어 있다. 비는 아래로 떨어지며 지면에 고인 물과 반사가 자연스럽다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "캠핑카 후면은 왼쪽 아래를 향하고 전면은 오른쪽 위의 출입구 방향을 향한다. 차량 측면과 후면은 보이지만 지붕 노출이 A보다 적고 건물 상부를 올려보는 느낌이 강하여, 지정된 높은 외부 위치에서 내려다보는 시점에 덜 맞는다. 사람의 시선이나 겨누는 물체는 없다.",
        "built_space": "모서리를 감싸는 청록색 띠창, 석재 외벽, 왼쪽 낮은 건물 부분, 큰 간판 보와 사각 지지부가 참고 사진에 가깝게 배치되어 있다. 출입구 안쪽에는 원형 기둥 세 개와 유리 출입구 한 곳이 보인다. 입구 앞 화분 두 개가 보이고 차량이 왼쪽 일부를 가린다. 벽돌 포장, 노란 유도블록, 오른쪽 벽 아래 배관도 유지된다. 바닥의 붉은 후미등 반사와 따뜻한 입구 조명 반사는 가능한 위치에 있다.",
        "entities": "낡고 녹슨 캠핑카 한 대가 좌하단에 있으며 후면 창에는 커튼과 실내 빛만 보이고 사람이나 얼굴은 없다. 도서관 벽의 균열, 마른 넝쿨, 빗줄기와 젖은 바닥이 표현되어 있다. 간판의 한글 도서관명은 맞지만 참고 사진 오른쪽의 영문 표기를 없애고 한글 '인천'으로 바꾸어 원본 문자 체계 유지 지시를 따르지 않았다. 차량 누수 여부는 확인되지 않는다. 가로등과 입구의 따뜻한 광원이 요구보다 두드러진다.",
        "hard_violations": [],
        "physics": "차량은 앞뒤 바퀴로 젖은 포장면에 지지되어 있으며 이동 중이라는 흔적 없이 멈춰 있다. 후면 사다리와 지붕 장비는 차체에 붙어 있다. 입구 보와 차양에는 구조적 지지부가 있고 화분은 바닥에 놓여 있다. 비와 웅덩이의 반사는 중력과 광원 방향에 어긋나지 않는다. 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "캠핑카 지붕을 더 내려다보는 사선 시점과 좌하단 배치가 지정된 와이드 구도에 더 가깝고 원본 간판의 문자 체계도 유지하지만, 가로등과 출입구 조명은 요구보다 두드러진다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "비 오는 밤의 장소와 정차 상태는 맞지만, 낮은 카메라 시점이 지정된 하향 사선 구도에서 멀어지고 원본 간판의 영문 표기를 한글로 바꿨다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카는 뒤쪽이 화면 왼쪽 아래, 운전석 쪽이 오른쪽 위를 향하여 도서관 출입구 방향으로 놓여 있다. 카메라는 차량의 후면·측면·지붕을 함께 보며 출입구 쪽을 비스듬히 바라본다. 차량 지붕을 보는 하향 각도는 B보다 뚜렷하지만 높은 외부 시점이라는 인상은 다소 약하다. 사람이나 시선, 겨누는 물체는 없다.",
        "built_space": "청록색 수평 띠창이 모서리를 감싸는 석재 건물, 왼쪽의 낮은 건물 부분, 간판을 단 큰 입구 보와 왼쪽 사각 지지부가 보인다. 입구 아래 원형 기둥은 세 개, 유리 출입구는 한 곳이며 참고 사진의 주요 구조와 대응한다. 화분 두 개가 뚜렷하고 왼쪽 구역 일부는 차량에 가린다. 벽돌 포장과 출입구로 이어지는 노란 유도블록도 유지된다. 젖은 바닥의 조명 반사는 광원 위치와 모순되지 않는다.",
        "entities": "낡은 캠핑카 한 대가 화면 좌하단에서 전체 면적의 삼분의 일 미만을 차지하며 탑승자나 얼굴은 보이지 않는다. 깊게 갈라진 외벽과 마른 넝쿨, 오래된 도서관 간판, 야간의 빗줄기가 있다. 간판은 참고 사진처럼 한글 도서관명과 오른쪽 영문 도서관명·도시명을 유지한다. 차량 외부는 젖었지만 내부로 물이 새는지는 확인할 수 없다. 왼쪽 가로등과 입구 조명은 은은한 야간 조명이라는 요구보다 눈에 띈다.",
        "hard_violations": [],
        "physics": "캠핑카의 보이는 앞뒤 바퀴가 포장면에 닿아 차체를 지지하며 정차 상태로 읽힌다. 지붕 장비와 후면 사다리는 차량에 부착되어 있다. 입구 보는 사각 지지부와 건물에 연결되고, 안쪽 차양은 기둥으로 지지된다. 화분은 바닥에 놓이고 넝쿨은 벽에 붙어 있다. 비는 아래로 떨어지며 지면에 고인 물과 반사가 자연스럽다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "캠핑카 후면은 왼쪽 아래를 향하고 전면은 오른쪽 위의 출입구 방향을 향한다. 차량 측면과 후면은 보이지만 지붕 노출이 A보다 적고 건물 상부를 올려보는 느낌이 강하여, 지정된 높은 외부 위치에서 내려다보는 시점에 덜 맞는다. 사람의 시선이나 겨누는 물체는 없다.",
        "built_space": "모서리를 감싸는 청록색 띠창, 석재 외벽, 왼쪽 낮은 건물 부분, 큰 간판 보와 사각 지지부가 참고 사진에 가깝게 배치되어 있다. 출입구 안쪽에는 원형 기둥 세 개와 유리 출입구 한 곳이 보인다. 입구 앞 화분 두 개가 보이고 차량이 왼쪽 일부를 가린다. 벽돌 포장, 노란 유도블록, 오른쪽 벽 아래 배관도 유지된다. 바닥의 붉은 후미등 반사와 따뜻한 입구 조명 반사는 가능한 위치에 있다.",
        "entities": "낡고 녹슨 캠핑카 한 대가 좌하단에 있으며 후면 창에는 커튼과 실내 빛만 보이고 사람이나 얼굴은 없다. 도서관 벽의 균열, 마른 넝쿨, 빗줄기와 젖은 바닥이 표현되어 있다. 간판의 한글 도서관명은 맞지만 참고 사진 오른쪽의 영문 표기를 없애고 한글 '인천'으로 바꾸어 원본 문자 체계 유지 지시를 따르지 않았다. 차량 누수 여부는 확인되지 않는다. 가로등과 입구의 따뜻한 광원이 요구보다 두드러진다.",
        "hard_violations": [],
        "physics": "차량은 앞뒤 바퀴로 젖은 포장면에 지지되어 있으며 이동 중이라는 흔적 없이 멈춰 있다. 후면 사다리와 지붕 장비는 차체에 붙어 있다. 입구 보와 차양에는 구조적 지지부가 있고 화분은 바닥에 놓여 있다. 비와 웅덩이의 반사는 중력과 광원 방향에 어긋나지 않는다. 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 카메라 구도와 야간 비 내리는 분위기를 훌륭하게 구현했으며, 레퍼런스의 간판 텍스트(한영 혼용)를 임의로 변경하지 않고 정확히 재현하여 지침을 완벽히 준수함."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "전반적인 구도와 분위기는 잘 연출되었으나, 레퍼런스의 영문 간판 텍스트를 한글로 임의 번역 및 통합하여 스크립트 변경 금지 지침을 위반함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_abandoned_library_exterior_sel.png",
    "asset_id": "c27d7844-3dc9-4630-8ad7-2fb4fcdd998d",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b0d-a969-78bb-bac5-2c8488268b0c",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S46sh19::signage": {
  "fp": "040c45a1ef131752",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::24ab141d8da50423": {
  "subjects": [],
  "subject_text": "버려진 도서관 외부\n오랫동안 방치된 도서관 건물. 외벽에 마른 덩굴이 달라붙어 있고, 벽 곳곳에 큰 균열이 나 있으며, 출입구에 낡은 간판이 걸려 있다.",
  "identity": "canonical",
  "scope_id": "L60",
  "scope_role": "location_exterior",
  "scope_sha": "70d3401b2497ef9e"
 },
 "groupbg::library_reading_room": {
  "input_fingerprint": "75f2ff4e9c9bd0d4",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "library_reading_room",
    "tags": [
     "S46sh19",
     "S46sh33"
    ]
   },
   "context_sig": "315f0fa715d881f3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 외부: 외벽에 금이 가고 넝쿨이 자란 폐건물 도서관이다. (특징: '미추홀도서관, 인천'이라는 낡고 기울어진 간판; 건물 벽면을 뒤덮은 마른 식물 넝쿨과 콘크리트 균열)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /열람실 안\n- 먼지 쌓여 있는 책장과 책들 사이로 열독중인 찰리가 보인다.\n\nTIME OF DAY (lock): night, rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 외부: 외벽에 금이 가고 넝쿨이 자란 폐건물 도서관이다. (특징: '미추홀도서관, 인천'이라는 낡고 기울어진 간판; 건물 벽면을 뒤덮은 마른 식물 넝쿨과 콘크리트 균열)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /열람실 안\n- 먼지 쌓여 있는 책장과 책들 사이로 열독중인 찰리가 보인다.\n\nTIME OF DAY (lock): night, rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_reading_room_e48422.png",
  "asset_id": "5d416f93-9628-40d3-9e35-f68f8dd976f8",
  "input_asset_ids": [
   "69886730-788e-4e34-bb48-b65d0ffb0eca"
  ],
  "origin_tag": "S46sh19",
  "place_text": "At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.",
  "origin_inputs": {
   "place_text": "At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.",
   "time_of_day_en": "night, rain",
   "conti_asset_id": "69886730-788e-4e34-bb48-b65d0ffb0eca"
  }
 },
 "S46sh19::bgfirst_bg": {
  "input_fingerprint": "aaff5dea5787c81e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 거친 손이 찰리가 들고 있던 책을 강하게 치는 mid-impact 순간, 책이 손에서 막 떨어져 허공에 뜬 역동적인 찰나.\n\nLOCATION (lock): At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.\n\nTIME OF DAY (lock): night, rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established lateral side of the exchange, settle the inward move just above book level with a shallow downward angle, emphasizing camera distance alone. Crop 이현우 and 찰리 to their opposing torso edges and forearms, with 이현우's striking hand entering from the left and 찰리's newly emptied hands on the right; suspend the released book between them at less than a third of the frame, leaving clear space beyond its leading edge. Their attention is directed downward toward the broken hand-to-book contact, while the visible torso fragments and reading-room floor keep the impact physically grounded.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open robotics book (Just released into the air after being struck) — The open next-generation robot chapter is tipped obliquely toward the camera, with no additional invented page content; used as Separates the striking hand from the hands that had offered it; Reading-room floor (Dusty from prolonged abandonment); used as Provides a subdued spatial field behind the airborne book.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast separate skin, the robot's precise hard surfaces, and the book without stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 거친 손이 찰리가 들고 있던 책을 강하게 치는 mid-impact 순간, 책이 손에서 막 떨어져 허공에 뜬 역동적인 찰나.\n\nLOCATION (lock): At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source.\n\nTIME OF DAY (lock): night, rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established lateral side of the exchange, settle the inward move just above book level with a shallow downward angle, emphasizing camera distance alone. Crop 이현우 and 찰리 to their opposing torso edges and forearms, with 이현우's striking hand entering from the left and 찰리's newly emptied hands on the right; suspend the released book between them at less than a third of the frame, leaving clear space beyond its leading edge. Their attention is directed downward toward the broken hand-to-book contact, while the visible torso fragments and reading-room floor keep the impact physically grounded.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open robotics book (Just released into the air after being struck) — The open next-generation robot chapter is tipped obliquely toward the camera, with no additional invented page content; used as Separates the striking hand from the hands that had offered it; Reading-room floor (Dusty from prolonged abandonment); used as Provides a subdued spatial field behind the airborne book.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast separate skin, the robot's precise hard surfaces, and the book without stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S46sh19__bgfirst_bg.png",
  "asset_id": "eee44424-140d-4ce4-9c15-1d79778d93e4",
  "input_asset_ids": [
   "69886730-788e-4e34-bb48-b65d0ffb0eca",
   "5d416f93-9628-40d3-9e35-f68f8dd976f8"
  ]
 },
 "S46sh19": {
  "input_fingerprint": "10557eb9125b1ee8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 찰리가 들고 있던 책을 강하게 치는 mid-impact 순간, 책이 손에서 막 떨어져 허공에 뜬 역동적인 찰나.\n\nLOCATION (lock): At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established lateral side of the exchange, settle the inward move just above book level with a shallow downward angle, emphasizing camera distance alone. Crop 이현우 and 찰리 to their opposing torso edges and forearms, with 이현우's striking hand entering from the left and 찰리's newly emptied hands on the right; suspend the released book between them at less than a third of the frame, leaving clear space beyond its leading edge. Their attention is directed downward toward the broken hand-to-book contact, while the visible torso fragments and reading-room floor keep the impact physically grounded.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open robotics book (Just released into the air after being struck) — The open next-generation robot chapter is tipped obliquely toward the camera, with no additional invented page content; used as Separates the striking hand from the hands that had offered it; Reading-room floor (Dusty from prolonged abandonment); used as Provides a subdued spatial field behind the airborne book.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast separate skin, the robot's precise hard surfaces, and the book without stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned reading room remains dusty, cobwebbed and lined with disordered shelves; the bags and canned food are at the group's resting place. The discarded Ubik book contains a chapter on next-generation robots. 이현우: His facial bruising and thigh wound remain untreated, and the contact card remains hidden inside his shoe. 찰리: The old gorilla-shaped robot still has the blanket from the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 찰리가 들고 있던 책을 강하게 치는 mid-impact 순간, 책이 손에서 막 떨어져 허공에 뜬 역동적인 찰나.\n\nLOCATION (lock): At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established lateral side of the exchange, settle the inward move just above book level with a shallow downward angle, emphasizing camera distance alone. Crop 이현우 and 찰리 to their opposing torso edges and forearms, with 이현우's striking hand entering from the left and 찰리's newly emptied hands on the right; suspend the released book between them at less than a third of the frame, leaving clear space beyond its leading edge. Their attention is directed downward toward the broken hand-to-book contact, while the visible torso fragments and reading-room floor keep the impact physically grounded.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open robotics book (Just released into the air after being struck) — The open next-generation robot chapter is tipped obliquely toward the camera, with no additional invented page content; used as Separates the striking hand from the hands that had offered it; Reading-room floor (Dusty from prolonged abandonment); used as Provides a subdued spatial field behind the airborne book.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast separate skin, the robot's precise hard surfaces, and the book without stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned reading room remains dusty, cobwebbed and lined with disordered shelves; the bags and canned food are at the group's resting place. The discarded Ubik book contains a chapter on next-generation robots. 이현우: His facial bruising and thigh wound remain untreated, and the contact card remains hidden inside his shoe. 찰리: The old gorilla-shaped robot still has the blanket from the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 이현우의 거친 손이 찰리가 들고 있던 책을 강하게 치는 mid-impact 순간, 책이 손에서 막 떨어져 허공에 뜬 역동적인 찰나.\n\nLOCATION (lock): At the makeshift meal spot inside the abandoned library's dusty reading room, beside disordered bookshelves. Nighttime illumination is not assigned a specific source. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established lateral side of the exchange, settle the inward move just above book level with a shallow downward angle, emphasizing camera distance alone. Crop 이현우 and 찰리 to their opposing torso edges and forearms, with 이현우's striking hand entering from the left and 찰리's newly emptied hands on the right; suspend the released book between them at less than a third of the frame, leaving clear space beyond its leading edge. Their attention is directed downward toward the broken hand-to-book contact, while the visible torso fragments and reading-room floor keep the impact physically grounded.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Open robotics book (Just released into the air after being struck) — The open next-generation robot chapter is tipped obliquely toward the camera, with no additional invented page content; used as Separates the striking hand from the hands that had offered it; Reading-room floor (Dusty from prolonged abandonment); used as Provides a subdued spatial field behind the airborne book.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast separate skin, the robot's precise hard surfaces, and the book without stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned reading room remains dusty, cobwebbed and lined with disordered shelves; the bags and canned food are at the group's resting place. The discarded Ubik book contains a chapter on next-generation robots. 이현우: His facial bruising and thigh wound remain untreated, and the contact card remains hidden inside his shoe. 찰리: The old gorilla-shaped robot still has the blanket from the unmanned shop.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S46sh19__bgfirst_bg.png",
     "asset_id": "eee44424-140d-4ce4-9c15-1d79778d93e4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S46sh19.png",
     "asset_id": "69886730-788e-4e34-bb48-b65d0ffb0eca",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 유빅사 로봇 도감: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1220897>",
     "asset_id": "253f339b-da28-4fdd-973e-2f7d0761f955",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_reading_room_e48422.png",
     "asset_id": "5d416f93-9628-40d3-9e35-f68f8dd976f8",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 유빅사 로봇 도감: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1220897>",
     "asset_id": "253f339b-da28-4fdd-973e-2f7d0761f955",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 찰리의 시선 및 몸의 방향이 중앙에 떠 있는 책을 향하고 있음.",
    "built_space": "물이 고인 폐도서관 바닥, 양옆으로 늘어선 책장, 비가 내리는 창밖 배경이 보임.",
    "entities": "피 묻은 이현우의 팔(왼쪽)과 담요를 두른 찰리의 기계 몸통(오른쪽)이 식별됨. 중앙의 책은 펼쳐져 있으나, 페이지 안의 그림이 참조된 고릴라 형태가 아닌 일반 인간형 로봇임.",
    "hard_violations": [
     "[gemini-pro] 소품 참조 이미지와 다른 임의의 로봇 그림을 책 페이지에 창작하여 삽입함 (no additional invented page content 지시 위반)",
     "[gpt-high] 추가 페이지 내용을 만들지 말라는 명시적 금지에도 도감 왼쪽 페이지에 참조에 없는 머리·다리 도해를 삽입했다."
    ],
    "physics": "책이 지지대 없이 허공에 떠 있으나, 프롬프트에서 요구한 타격 직후의 찰나 연출이므로 정당화됨."
   },
   {
    "label": "B",
    "direction": "이현우의 뻗은 손과 찰리의 빈손이 엇갈리며, 프레임 내의 주의가 공중에 뜬 책과 바닥 쪽으로 향함.",
    "built_space": "폐허가 된 도서관의 젖은 바닥, 어지러운 책장, 빗방울이 맺힌 창문이 정확히 구성됨.",
    "entities": "크롭 지시에 맞게 잘린 이현우의 팔과 셔츠(왼쪽), 찰리의 샌드 베이지색 기계 팔과 담요(오른쪽)가 등장하며, 허공의 책은 참조 이미지의 고릴라 로봇 도해와 한국어 텍스트를 그대로 담고 있음.",
    "hard_violations": [],
    "physics": "타격의 여파로 흩날리는 먼지와 함께 책이 허공에 자연스럽게 떠 있어 제시된 상황을 잘 뒷받침함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "요구된 토르소와 팔뚝 위주의 밀착된 크롭 구도를 완벽히 따랐으며, 특히 공중에 뜬 책 내부의 도해와 텍스트를 참조 이미지와 동일하게 구현해 내어 매우 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "타격 직후의 물리적 순간은 잘 연출했으나, 책 페이지에 참조에 없는 임의의 로봇 그림을 그려 넣어 엄격한 소품 일치 및 창작 금지 지시를 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 찰리의 시선 및 몸의 방향이 중앙에 떠 있는 책을 향하고 있음.",
        "built_space": "물이 고인 폐도서관 바닥, 양옆으로 늘어선 책장, 비가 내리는 창밖 배경이 보임.",
        "entities": "피 묻은 이현우의 팔(왼쪽)과 담요를 두른 찰리의 기계 몸통(오른쪽)이 식별됨. 중앙의 책은 펼쳐져 있으나, 페이지 안의 그림이 참조된 고릴라 형태가 아닌 일반 인간형 로봇임.",
        "hard_violations": [
         "소품 참조 이미지와 다른 임의의 로봇 그림을 책 페이지에 창작하여 삽입함 (no additional invented page content 지시 위반)"
        ],
        "physics": "책이 지지대 없이 허공에 떠 있으나, 프롬프트에서 요구한 타격 직후의 찰나 연출이므로 정당화됨."
       },
       {
        "label": "B",
        "direction": "이현우의 뻗은 손과 찰리의 빈손이 엇갈리며, 프레임 내의 주의가 공중에 뜬 책과 바닥 쪽으로 향함.",
        "built_space": "폐허가 된 도서관의 젖은 바닥, 어지러운 책장, 빗방울이 맺힌 창문이 정확히 구성됨.",
        "entities": "크롭 지시에 맞게 잘린 이현우의 팔과 셔츠(왼쪽), 찰리의 샌드 베이지색 기계 팔과 담요(오른쪽)가 등장하며, 허공의 책은 참조 이미지의 고릴라 로봇 도해와 한국어 텍스트를 그대로 담고 있음.",
        "hard_violations": [],
        "physics": "타격의 여파로 흩날리는 먼지와 함께 책이 허공에 자연스럽게 떠 있어 제시된 상황을 잘 뒷받침함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "요구된 토르소와 팔뚝 위주의 밀착된 크롭 구도를 완벽히 따랐으며, 특히 공중에 뜬 책 내부의 도해와 텍스트를 참조 이미지와 동일하게 구현해 내어 매우 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "타격 직후의 물리적 순간은 잘 연출했으나, 책 페이지에 참조에 없는 임의의 로봇 그림을 그려 넣어 엄격한 소품 일치 및 창작 금지 지시를 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 찰리의 시선 및 몸의 방향이 중앙에 떠 있는 책을 향하고 있음.",
        "built_space": "물이 고인 폐도서관 바닥, 양옆으로 늘어선 책장, 비가 내리는 창밖 배경이 보임.",
        "entities": "피 묻은 이현우의 팔(왼쪽)과 담요를 두른 찰리의 기계 몸통(오른쪽)이 식별됨. 중앙의 책은 펼쳐져 있으나, 페이지 안의 그림이 참조된 고릴라 형태가 아닌 일반 인간형 로봇임.",
        "hard_violations": [
         "소품 참조 이미지와 다른 임의의 로봇 그림을 책 페이지에 창작하여 삽입함 (no additional invented page content 지시 위반)"
        ],
        "physics": "책이 지지대 없이 허공에 떠 있으나, 프롬프트에서 요구한 타격 직후의 찰나 연출이므로 정당화됨."
       },
       {
        "label": "B",
        "direction": "이현우의 뻗은 손과 찰리의 빈손이 엇갈리며, 프레임 내의 주의가 공중에 뜬 책과 바닥 쪽으로 향함.",
        "built_space": "폐허가 된 도서관의 젖은 바닥, 어지러운 책장, 빗방울이 맺힌 창문이 정확히 구성됨.",
        "entities": "크롭 지시에 맞게 잘린 이현우의 팔과 셔츠(왼쪽), 찰리의 샌드 베이지색 기계 팔과 담요(오른쪽)가 등장하며, 허공의 책은 참조 이미지의 고릴라 로봇 도해와 한국어 텍스트를 그대로 담고 있음.",
        "hard_violations": [],
        "physics": "타격의 여파로 흩날리는 먼지와 함께 책이 허공에 자연스럽게 떠 있어 제시된 상황을 잘 뒷받침함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "양쪽 몸통 가장자리와 팔을 중심으로 충격 직후를 더 밀착해 포착하고 도감도 참조에 가깝지만, 책이 다소 크고 카메라의 하향각은 약하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "손에서 이탈한 책의 동작은 타당하지만 얼굴·허벅지·천장까지 드러내며 구도를 넓혔고, 금지된 추가 페이지 도해를 만들어 넣었다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 펼친 손이 왼쪽에서 중앙 책의 왼쪽 위 가장자리를 향하며, 찰리의 빈 두 손은 오른쪽에서 책 쪽으로 열려 있다. 책의 인쇄면은 카메라 쪽으로 비스듬히 노출된다. 이현우의 얼굴은 보이지 않고, 찰리의 잘린 얼굴은 책 쪽으로 숙여져 있으나 정확한 시선은 확인하기 어렵다.",
        "built_space": "왼쪽의 여러 목재 책장과 사다리 한 개, 뒤쪽의 기둥과 격자창, 책이 흩어진 젖고 더러운 바닥이 보인다. 참조 열람실의 재료와 야간 분위기에 부합하며 고정 설비의 명백한 중복은 없다. 양 인물은 화면 양끝을 차지하고 책 뒤로 바닥이 보이지만, 창과 책장도 상당 부분 보여 하향각은 약하다. 책은 화면 너비의 약 3분의 1을 조금 넘는다.",
        "entities": "왼쪽에는 상처와 오염이 있는 인간 손, 마른 팔과 낡은 어두운 셔츠가 보여 이현우의 보이는 신체·복장 조건에 부합한다. 얼굴이 제외되어 나이와 얼굴 정체성은 확인할 수 없다. 오른쪽 찰리는 샌드 베이지 기계 장갑, 관절 손, 흰 마스크 일부, 푸른 가슴 원자로와 담요를 갖췄다. 열린 책에는 왼쪽의 한국어 본문과 차세대 로봇 제목, 오른쪽의 청색 고릴라형 로봇 도해가 있어 참조 도감과 대체로 일치한다.",
        "hard_violations": [],
        "physics": "책에는 현재 닿아 있는 손이 없지만, 바로 옆의 타격 손과 벌어진 로봇 손 사이에서 기울어진 상태라 타격으로 막 이탈한 짧은 비행으로 설명된다. 책 주변의 먼지도 충격 직후라는 해석을 뒷받침한다. 펼친 페이지와 책등은 물리적인 책 형태를 유지한다. 팔은 각 몸통으로 연결되며, 발은 프레임 밖이므로 지면 접촉 자체는 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "이현우의 손바닥은 왼쪽에서 책 위쪽을 향하고, 책은 그 아래에서 기울어져 떨어지는 모습이다. 찰리의 두 손은 책을 향해 열린 채 떨어져 있다. 두 인물 모두 머리를 책 쪽으로 숙였으며, 찰리의 얼굴 방향은 책을 향한다. 책의 펼친 면은 지시대로 카메라에 비스듬히 보인다.",
        "built_space": "왼쪽 책장 열과 사다리 한 개, 식사 자리의 탁자 일부, 뒤쪽 격자창, 기둥, 파손된 천장과 젖은 바닥이 보인다. 장소의 주요 재료와 배치는 참조와 유사하고 불가능한 설비 중복은 없다. 다만 인물의 머리와 이현우의 허벅지, 찰리의 넓은 가슴까지 포함하여 몸통 가장자리와 전완만 남기는 지정 클로즈업보다 넓다. 책은 화면 너비의 약 3할이며 진행 방향 아래로 여백이 있다.",
        "entities": "이현우는 검은 머리, 젊은 얼굴 일부, 상처 난 팔과 피·먼지가 묻은 어두운 옷으로 묘사된다. 찰리는 흰 마스크, 주황색 눈, 베이지 장갑과 푸른 원자로, 어깨의 담요를 갖췄다. 그러나 도감 오른쪽의 로봇은 참조의 육중한 고릴라형보다 가늘고 긴 인간형이며, 왼쪽 페이지에는 참조에 없던 머리와 다리 도해가 추가되어 있다.",
        "hard_violations": [
         "추가 페이지 내용을 만들지 말라는 명시적 금지에도 도감 왼쪽 페이지에 참조에 없는 머리·다리 도해를 삽입했다."
        ],
        "physics": "책은 양쪽 손에서 분리되어 있지만 이현우의 뻗은 타격 팔 아래, 찰리의 벌어진 손 앞에 있어 타격 후 낙하하는 경로가 성립한다. 따라서 근거 없는 공중 정지로 볼 필요는 없다. 양쪽 팔과 기계 손은 몸통에 연결되어 있으며, 책의 휜 페이지와 회전 자세도 가능한 범위다. 인물의 발은 보이지 않아 바닥 지지는 직접 확인할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "양쪽 몸통 가장자리와 팔을 중심으로 충격 직후를 더 밀착해 포착하고 도감도 참조에 가깝지만, 책이 다소 크고 카메라의 하향각은 약하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "손에서 이탈한 책의 동작은 타당하지만 얼굴·허벅지·천장까지 드러내며 구도를 넓혔고, 금지된 추가 페이지 도해를 만들어 넣었다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 펼친 손이 왼쪽에서 중앙 책의 왼쪽 위 가장자리를 향하며, 찰리의 빈 두 손은 오른쪽에서 책 쪽으로 열려 있다. 책의 인쇄면은 카메라 쪽으로 비스듬히 노출된다. 이현우의 얼굴은 보이지 않고, 찰리의 잘린 얼굴은 책 쪽으로 숙여져 있으나 정확한 시선은 확인하기 어렵다.",
        "built_space": "왼쪽의 여러 목재 책장과 사다리 한 개, 뒤쪽의 기둥과 격자창, 책이 흩어진 젖고 더러운 바닥이 보인다. 참조 열람실의 재료와 야간 분위기에 부합하며 고정 설비의 명백한 중복은 없다. 양 인물은 화면 양끝을 차지하고 책 뒤로 바닥이 보이지만, 창과 책장도 상당 부분 보여 하향각은 약하다. 책은 화면 너비의 약 3분의 1을 조금 넘는다.",
        "entities": "왼쪽에는 상처와 오염이 있는 인간 손, 마른 팔과 낡은 어두운 셔츠가 보여 이현우의 보이는 신체·복장 조건에 부합한다. 얼굴이 제외되어 나이와 얼굴 정체성은 확인할 수 없다. 오른쪽 찰리는 샌드 베이지 기계 장갑, 관절 손, 흰 마스크 일부, 푸른 가슴 원자로와 담요를 갖췄다. 열린 책에는 왼쪽의 한국어 본문과 차세대 로봇 제목, 오른쪽의 청색 고릴라형 로봇 도해가 있어 참조 도감과 대체로 일치한다.",
        "hard_violations": [],
        "physics": "책에는 현재 닿아 있는 손이 없지만, 바로 옆의 타격 손과 벌어진 로봇 손 사이에서 기울어진 상태라 타격으로 막 이탈한 짧은 비행으로 설명된다. 책 주변의 먼지도 충격 직후라는 해석을 뒷받침한다. 펼친 페이지와 책등은 물리적인 책 형태를 유지한다. 팔은 각 몸통으로 연결되며, 발은 프레임 밖이므로 지면 접촉 자체는 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "이현우의 손바닥은 왼쪽에서 책 위쪽을 향하고, 책은 그 아래에서 기울어져 떨어지는 모습이다. 찰리의 두 손은 책을 향해 열린 채 떨어져 있다. 두 인물 모두 머리를 책 쪽으로 숙였으며, 찰리의 얼굴 방향은 책을 향한다. 책의 펼친 면은 지시대로 카메라에 비스듬히 보인다.",
        "built_space": "왼쪽 책장 열과 사다리 한 개, 식사 자리의 탁자 일부, 뒤쪽 격자창, 기둥, 파손된 천장과 젖은 바닥이 보인다. 장소의 주요 재료와 배치는 참조와 유사하고 불가능한 설비 중복은 없다. 다만 인물의 머리와 이현우의 허벅지, 찰리의 넓은 가슴까지 포함하여 몸통 가장자리와 전완만 남기는 지정 클로즈업보다 넓다. 책은 화면 너비의 약 3할이며 진행 방향 아래로 여백이 있다.",
        "entities": "이현우는 검은 머리, 젊은 얼굴 일부, 상처 난 팔과 피·먼지가 묻은 어두운 옷으로 묘사된다. 찰리는 흰 마스크, 주황색 눈, 베이지 장갑과 푸른 원자로, 어깨의 담요를 갖췄다. 그러나 도감 오른쪽의 로봇은 참조의 육중한 고릴라형보다 가늘고 긴 인간형이며, 왼쪽 페이지에는 참조에 없던 머리와 다리 도해가 추가되어 있다.",
        "hard_violations": [
         "추가 페이지 내용을 만들지 말라는 명시적 금지에도 도감 왼쪽 페이지에 참조에 없는 머리·다리 도해를 삽입했다."
        ],
        "physics": "책은 양쪽 손에서 분리되어 있지만 이현우의 뻗은 타격 팔 아래, 찰리의 벌어진 손 앞에 있어 타격 후 낙하하는 경로가 성립한다. 따라서 근거 없는 공중 정지로 볼 필요는 없다. 양쪽 팔과 기계 손은 몸통에 연결되어 있으며, 책의 휜 페이지와 회전 자세도 가능한 범위다. 인물의 발은 보이지 않아 바닥 지지는 직접 확인할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.944,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.694,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 소품 참조 이미지와 다른 임의의 로봇 그림을 책 페이지에 창작하여 삽입함 (no additional invented page content 지시 위반)",
     "[gpt-high] 추가 페이지 내용을 만들지 말라는 명시적 금지에도 도감 왼쪽 페이지에 참조에 없는 머리·다리 도해를 삽입했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 694
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 토르소와 팔뚝 위주의 밀착된 크롭 구도를 완벽히 따랐으며, 특히 공중에 뜬 책 내부의 도해와 텍스트를 참조 이미지와 동일하게 구현해 내어 매우 우수함."
   },
   {
    "label": "A",
    "score": 694,
    "verdict_ko": "타격 직후의 물리적 순간은 잘 연출했으나, 책 페이지에 참조에 없는 임의의 로봇 그림을 그려 넣어 엄격한 소품 일치 및 창작 금지 지시를 위반함.  ★위반: [gemini-pro] 소품 참조 이미지와 다른 임의의 로봇 그림을 책 페이지에 창작하여 삽입함 (no additional invented page content 지시 위반) / [gpt-high] 추가 페이지 내용을 만들지 말라는 명시적 금지에도 도감 왼쪽 페이지에 참조에 없는 머리·다리 도해를 삽입했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_reading_room_e48422.png",
    "asset_id": "5d416f93-9628-40d3-9e35-f68f8dd976f8",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 유빅사 로봇 도감: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1220897>",
    "asset_id": "253f339b-da28-4fdd-973e-2f7d0761f955",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b12-fb36-731b-8dac-5050862edb5f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S46sh19__bgfirst_bg.png",
   "bg_asset_id": "eee44424-140d-4ce4-9c15-1d79778d93e4",
   "bg_record_key": "S46sh19::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "library_reading_room",
   "groupbg_asset_id": "5d416f93-9628-40d3-9e35-f68f8dd976f8"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S46sh33::signage": {
  "fp": "63d0e8ed830a3d51",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S46sh33": {
  "input_fingerprint": "1437ffbcae17ca2b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 금속 눈매를 둥글게 휘며 환하게 웃는 찰리의 상체.\n\nLOCATION (lock): At the old computer among the abandoned library's dusty bookshelves. The active computer provides a local screen glow in the nighttime interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the tracking move settle beside the computer at its established side three-quarter endpoint, with the lens slightly below 찰리's eyes and outside the screen-to-face axis. Place his upper body in the left-center of the image and retain open space to the right toward the computer just outside frame; his rounded metal eye shapes and broad smile remain directed toward that screen, never toward the lens. Hold camera distance and body placement steady so the emphasis is his attentive gaze resolving into satisfaction, before any downward look toward the crayons.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bookshelves and books (Disordered and coated with accumulated dust) — Shelf openings remain oblique and softly legible behind 찰리; used as Keeps his private discovery rooted in the abandoned reading room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves readable facial mechanisms and gentle tonal separation, conveying warmth through expression rather than an invented colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dusty books and shelves surround the old computer in the abandoned library. Boxes and books serve as bedding in one corner, and crayons lie scattered on the floor. 찰리: He is awake by the old computer after reading among the dusty shelves, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 금속 눈매를 둥글게 휘며 환하게 웃는 찰리의 상체.\n\nLOCATION (lock): At the old computer among the abandoned library's dusty bookshelves. The active computer provides a local screen glow in the nighttime interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the tracking move settle beside the computer at its established side three-quarter endpoint, with the lens slightly below 찰리's eyes and outside the screen-to-face axis. Place his upper body in the left-center of the image and retain open space to the right toward the computer just outside frame; his rounded metal eye shapes and broad smile remain directed toward that screen, never toward the lens. Hold camera distance and body placement steady so the emphasis is his attentive gaze resolving into satisfaction, before any downward look toward the crayons.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bookshelves and books (Disordered and coated with accumulated dust) — Shelf openings remain oblique and softly legible behind 찰리; used as Keeps his private discovery rooted in the abandoned reading room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves readable facial mechanisms and gentle tonal separation, conveying warmth through expression rather than an invented colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dusty books and shelves surround the old computer in the abandoned library. Boxes and books serve as bedding in one corner, and crayons lie scattered on the floor. 찰리: He is awake by the old computer after reading among the dusty shelves, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, rain.\n\nSHOT TEXT (authoritative, Korean): 금속 눈매를 둥글게 휘며 환하게 웃는 찰리의 상체.\n\nLOCATION (lock): At the old computer among the abandoned library's dusty bookshelves. The active computer provides a local screen glow in the nighttime interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the tracking move settle beside the computer at its established side three-quarter endpoint, with the lens slightly below 찰리's eyes and outside the screen-to-face axis. Place his upper body in the left-center of the image and retain open space to the right toward the computer just outside frame; his rounded metal eye shapes and broad smile remain directed toward that screen, never toward the lens. Hold camera distance and body placement steady so the emphasis is his attentive gaze resolving into satisfaction, before any downward look toward the crayons.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bookshelves and books (Disordered and coated with accumulated dust) — Shelf openings remain oblique and softly legible behind 찰리; used as Keeps his private discovery rooted in the abandoned reading room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves readable facial mechanisms and gentle tonal separation, conveying warmth through expression rather than an invented colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dusty books and shelves surround the old computer in the abandoned library. Boxes and books serve as bedding in one corner, and crayons lie scattered on the floor. 찰리: He is awake by the old computer after reading among the dusty shelves, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선과 고개가 화면 우측 전경에 있는 컴퓨터 모니터를 향하고 있음.",
    "built_space": "비 내리는 창문과 낡은 책장이 있는 도서관 배경이며, 우측에 모니터가 배치됨.",
    "entities": "장갑판과 덮고 있는 담요는 참조와 일치하나, 흰색 마스크에 유기적인 뺨 굴곡이 생기고 입 선이 주황색으로 발광함.",
    "hard_violations": [],
    "physics": "찰리의 상체는 양손을 책상에 얹어 안정적으로 지지된 상태임."
   },
   {
    "label": "B",
    "direction": "찰리의 시선과 얼굴 정면이 우측에 배치된 구형 모니터를 정확히 향함.",
    "built_space": "어두운 도서관 내부로 뒤쪽에 책장과 창문이 보이며, 우측 앞쪽에 큰 모니터가 있음.",
    "entities": "기계 몸체의 디테일과 이전 샷의 담요가 정확히 일치하며, 마스크의 입 선이 발광하지 않는 상태로 둥글게 휘어 미소를 표현함.",
    "hard_violations": [],
    "physics": "책상과 키보드 위에 양손을 올려 몸의 무게를 자연스럽게 지탱하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "기계 마스크의 재질과 비발광 입 선 등 참조 이미지의 설정을 잘 유지하며 미소를 표현했으나, 컴퓨터가 프레임 바깥이 아닌 안쪽에 위치해 여백이 부족한 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지정된 샷 크기와 조명 분위기는 좋으나, 딱딱한 금속 마스크에 인간의 뺨 근육 같은 굴곡을 임의로 더하고 입 선이 눈처럼 발광하도록 훼손한 점이 큰 감점 요인입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 고개가 화면 우측 전경에 있는 컴퓨터 모니터를 향하고 있음.",
        "built_space": "비 내리는 창문과 낡은 책장이 있는 도서관 배경이며, 우측에 모니터가 배치됨.",
        "entities": "장갑판과 덮고 있는 담요는 참조와 일치하나, 흰색 마스크에 유기적인 뺨 굴곡이 생기고 입 선이 주황색으로 발광함.",
        "hard_violations": [],
        "physics": "찰리의 상체는 양손을 책상에 얹어 안정적으로 지지된 상태임."
       },
       {
        "label": "B",
        "direction": "찰리의 시선과 얼굴 정면이 우측에 배치된 구형 모니터를 정확히 향함.",
        "built_space": "어두운 도서관 내부로 뒤쪽에 책장과 창문이 보이며, 우측 앞쪽에 큰 모니터가 있음.",
        "entities": "기계 몸체의 디테일과 이전 샷의 담요가 정확히 일치하며, 마스크의 입 선이 발광하지 않는 상태로 둥글게 휘어 미소를 표현함.",
        "hard_violations": [],
        "physics": "책상과 키보드 위에 양손을 올려 몸의 무게를 자연스럽게 지탱하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "기계 마스크의 재질과 비발광 입 선 등 참조 이미지의 설정을 잘 유지하며 미소를 표현했으나, 컴퓨터가 프레임 바깥이 아닌 안쪽에 위치해 여백이 부족한 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지정된 샷 크기와 조명 분위기는 좋으나, 딱딱한 금속 마스크에 인간의 뺨 근육 같은 굴곡을 임의로 더하고 입 선이 눈처럼 발광하도록 훼손한 점이 큰 감점 요인입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 고개가 화면 우측 전경에 있는 컴퓨터 모니터를 향하고 있음.",
        "built_space": "비 내리는 창문과 낡은 책장이 있는 도서관 배경이며, 우측에 모니터가 배치됨.",
        "entities": "장갑판과 덮고 있는 담요는 참조와 일치하나, 흰색 마스크에 유기적인 뺨 굴곡이 생기고 입 선이 주황색으로 발광함.",
        "hard_violations": [],
        "physics": "찰리의 상체는 양손을 책상에 얹어 안정적으로 지지된 상태임."
       },
       {
        "label": "B",
        "direction": "찰리의 시선과 얼굴 정면이 우측에 배치된 구형 모니터를 정확히 향함.",
        "built_space": "어두운 도서관 내부로 뒤쪽에 책장과 창문이 보이며, 우측 앞쪽에 큰 모니터가 있음.",
        "entities": "기계 몸체의 디테일과 이전 샷의 담요가 정확히 일치하며, 마스크의 입 선이 발광하지 않는 상태로 둥글게 휘어 미소를 표현함.",
        "hard_violations": [],
        "physics": "책상과 키보드 위에 양손을 올려 몸의 무게를 자연스럽게 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "허리까지 담은 미디엄 숏과 좌측 중심 배치, 화면을 향한 미소가 더 충실하지만, 둥글게 휘는 눈매가 없고 프레임 밖이어야 할 모니터가 오른쪽 여백을 크게 차지한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "컴퓨터를 보는 방향과 담요는 맞지만, 더 조인 상체 구도와 약한 미소, 추가된 입의 주황 발광이 지시에서 멀어지며 눈매와 모니터의 프레임 밖 배치도 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 두 눈은 오른쪽 전경의 모니터를 향하며 렌즈나 바닥을 보지 않는다. 카메라는 얼굴과 화면을 잇는 축에서 벗어나 얼굴의 측면과 정면을 함께 보여준다. 모니터는 뒷면과 측면이 카메라에 보이고 표시 면은 찰리를 향해 있어 사용 방향이 맞다.",
        "built_space": "찰리는 왼쪽 중심에 있고 머리부터 허리 부근까지 보인다. 오른쪽 전경에 모니터 한 대와 하단 책상 면이 있으며, 뒤에는 오른쪽으로 이어지는 다단 책장과 책 무더기, 사다리 한 개, 왼쪽의 빗물 맺힌 격자 창이 보인다. 책장 칸은 비스듬히 이어져 요구된 배경 깊이에 가깝다. 다만 모니터가 프레임 안을 크게 점유하여 컴퓨터 쪽의 열린 여백을 줄인다. 고정 설비의 명백한 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "등장 인물은 기계 찰리 하나뿐이다. 샌드 베이지 장갑판, 육중한 팔, 흰 마스크, 주황색 눈 두 개, 검은 입 선, 푸른 원형 가슴 원자로와 낡은 적갈색 계열 담요가 참조와 대체로 일치한다. 인간 피부나 치아는 없다. 입은 웃는 곡선이지만 두 눈은 참조처럼 둥근 구멍으로 남아, 웃으며 휘어지는 금속 눈매는 구현하지 않았다. 먼지 낀 책과 책장, 오래된 컴퓨터가 보이고 비 오는 밤의 창도 유지된다. 크레용과 침구용 상자는 해당 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "머리와 팔은 목 및 어깨의 기계 관절로 몸통에 연결되어 있고, 담요는 머리 뒤와 어깨에 걸쳐 중력 방향으로 늘어진다. 하체와 좌석 접점은 화면 밖이므로 지지 방식을 확정할 수 없지만 공중에 뜬 몸으로 보이지 않는다. 모니터와 책 더미는 책상 및 선반 위에 놓여 있으며 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 오른쪽 모니터를 향하고 있으며 카메라를 응시하거나 손 쪽으로 내려다보지 않는다. 모니터의 어두운 뒷면이 카메라를 향하고 표시 면은 찰리 쪽이므로 관찰 대상과 기기의 방향은 맞는다.",
        "built_space": "왼쪽 중심의 찰리를 머리부터 복부와 모은 손까지 담았으며 A보다 몸통 중심으로 조인 구도다. 오른쪽 전경에는 모니터 한 대, 하단에는 책상 면과 책 더미가 보인다. 배경 왼쪽에 다단 책장과 사다리 한 개, 중앙 뒤에 빗물 맺힌 큰 격자 창이 있다. 낡은 도서관의 재료와 야간 분위기는 유지되지만, 찰리 뒤를 채우는 것은 주로 창이고 비스듬한 책장 칸의 존재감은 약하다. 모니터가 요구된 프레임 밖 위치 대신 오른쪽 여백을 크게 막는다.",
        "entities": "기계 찰리 하나만 등장하며 베이지 장갑판, 흰 마스크, 주황색 눈 두 개, 푸른 가슴 원자로와 낡은 담요는 일치한다. 팔과 손도 육중한 기계 구조이고 인간의 피부나 치아는 없다. 눈은 휘어진 웃음 눈매가 아니라 원형이며 입도 작은 미소에 가깝다. 입 선에 주황 발광이 추가되어 참조의 검은 입과 다르고, 표정 대신 새 발광으로 따뜻함을 보탠다. 책장과 책, 오래된 모니터는 확인되며 크레용과 침구 구역은 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 팔은 어깨와 팔꿈치 관절로 연결되고, 하단의 기계 손들은 서로 겹쳐 접촉한 자세다. 팔과 손이 분리되어 떠 있는 모습은 아니다. 담요는 어깨에 걸쳐 아래로 처지고, 모니터와 책 더미는 책상에 놓여 있다. 하체의 바닥 또는 좌석 접점은 가려져 있지만, 명백한 무지지 부유나 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "허리까지 담은 미디엄 숏과 좌측 중심 배치, 화면을 향한 미소가 더 충실하지만, 둥글게 휘는 눈매가 없고 프레임 밖이어야 할 모니터가 오른쪽 여백을 크게 차지한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "컴퓨터를 보는 방향과 담요는 맞지만, 더 조인 상체 구도와 약한 미소, 추가된 입의 주황 발광이 지시에서 멀어지며 눈매와 모니터의 프레임 밖 배치도 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 두 눈은 오른쪽 전경의 모니터를 향하며 렌즈나 바닥을 보지 않는다. 카메라는 얼굴과 화면을 잇는 축에서 벗어나 얼굴의 측면과 정면을 함께 보여준다. 모니터는 뒷면과 측면이 카메라에 보이고 표시 면은 찰리를 향해 있어 사용 방향이 맞다.",
        "built_space": "찰리는 왼쪽 중심에 있고 머리부터 허리 부근까지 보인다. 오른쪽 전경에 모니터 한 대와 하단 책상 면이 있으며, 뒤에는 오른쪽으로 이어지는 다단 책장과 책 무더기, 사다리 한 개, 왼쪽의 빗물 맺힌 격자 창이 보인다. 책장 칸은 비스듬히 이어져 요구된 배경 깊이에 가깝다. 다만 모니터가 프레임 안을 크게 점유하여 컴퓨터 쪽의 열린 여백을 줄인다. 고정 설비의 명백한 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "등장 인물은 기계 찰리 하나뿐이다. 샌드 베이지 장갑판, 육중한 팔, 흰 마스크, 주황색 눈 두 개, 검은 입 선, 푸른 원형 가슴 원자로와 낡은 적갈색 계열 담요가 참조와 대체로 일치한다. 인간 피부나 치아는 없다. 입은 웃는 곡선이지만 두 눈은 참조처럼 둥근 구멍으로 남아, 웃으며 휘어지는 금속 눈매는 구현하지 않았다. 먼지 낀 책과 책장, 오래된 컴퓨터가 보이고 비 오는 밤의 창도 유지된다. 크레용과 침구용 상자는 해당 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "머리와 팔은 목 및 어깨의 기계 관절로 몸통에 연결되어 있고, 담요는 머리 뒤와 어깨에 걸쳐 중력 방향으로 늘어진다. 하체와 좌석 접점은 화면 밖이므로 지지 방식을 확정할 수 없지만 공중에 뜬 몸으로 보이지 않는다. 모니터와 책 더미는 책상 및 선반 위에 놓여 있으며 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 오른쪽 모니터를 향하고 있으며 카메라를 응시하거나 손 쪽으로 내려다보지 않는다. 모니터의 어두운 뒷면이 카메라를 향하고 표시 면은 찰리 쪽이므로 관찰 대상과 기기의 방향은 맞는다.",
        "built_space": "왼쪽 중심의 찰리를 머리부터 복부와 모은 손까지 담았으며 A보다 몸통 중심으로 조인 구도다. 오른쪽 전경에는 모니터 한 대, 하단에는 책상 면과 책 더미가 보인다. 배경 왼쪽에 다단 책장과 사다리 한 개, 중앙 뒤에 빗물 맺힌 큰 격자 창이 있다. 낡은 도서관의 재료와 야간 분위기는 유지되지만, 찰리 뒤를 채우는 것은 주로 창이고 비스듬한 책장 칸의 존재감은 약하다. 모니터가 요구된 프레임 밖 위치 대신 오른쪽 여백을 크게 막는다.",
        "entities": "기계 찰리 하나만 등장하며 베이지 장갑판, 흰 마스크, 주황색 눈 두 개, 푸른 가슴 원자로와 낡은 담요는 일치한다. 팔과 손도 육중한 기계 구조이고 인간의 피부나 치아는 없다. 눈은 휘어진 웃음 눈매가 아니라 원형이며 입도 작은 미소에 가깝다. 입 선에 주황 발광이 추가되어 참조의 검은 입과 다르고, 표정 대신 새 발광으로 따뜻함을 보탠다. 책장과 책, 오래된 모니터는 확인되며 크레용과 침구 구역은 화면에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 팔은 어깨와 팔꿈치 관절로 연결되고, 하단의 기계 손들은 서로 겹쳐 접촉한 자세다. 팔과 손이 분리되어 떠 있는 모습은 아니다. 담요는 어깨에 걸쳐 아래로 처지고, 모니터와 책 더미는 책상에 놓여 있다. 하체의 바닥 또는 좌석 접점은 가려져 있지만, 명백한 무지지 부유나 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.286,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1286
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "기계 마스크의 재질과 비발광 입 선 등 참조 이미지의 설정을 잘 유지하며 미소를 표현했으나, 컴퓨터가 프레임 바깥이 아닌 안쪽에 위치해 여백이 부족한 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 1286,
    "verdict_ko": "지정된 샷 크기와 조명 분위기는 좋으나, 딱딱한 금속 마스크에 인간의 뺨 근육 같은 굴곡을 임의로 더하고 입 선이 눈처럼 발광하도록 훼손한 점이 큰 감점 요인입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S46sh19_sel.png",
    "asset_id": "d8b67c9a-1556-47dd-bf73-2a7c286b4a5b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b1d-2f50-7614-bfd0-f5a75dba4210",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh19"
  }
 },
 "S47sh9::signage": {
  "fp": "cd8ab20bf2b89186",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::076de62aef16348c": {
  "subjects": [],
  "subject_text": "버려진 도서관 열람실·서가, 복도·외진 구석\n먼지와 거미줄이 쌓인 열람실과 서가. 부서진 책장과 낡은 책, 고물 컴퓨터가 남아 있으며, 깨진 천창에서 잿빛 빛이 들어온다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L171",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::library_mural_recess": {
  "input_fingerprint": "0b1fde5884d70c87",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "library_mural_recess",
    "tags": [
     "S47sh9"
    ]
   },
   "context_sig": "31c0a455d653214a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 열람실·서가, 복도·외진 구석: 거미줄과 먼지가 가득한 구형 도서관 내부로 버려진 물품들이 방치되어 있다. (특징: 거미줄이 쳐진 목재 책꽂이와 바닥에 뒹구는 곰팡이 핀 책들; 바닥에 깔린 종이 박스와 웅크린 아이들; 작동 중인 구형 모니터 컴퓨터와 파란색 크레파스; 벽 한 면을 가득 채운 파란색 바다와 로봇, 야자수 크레파스 벽화)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 바다가 그려진 벽을 따라 천천히 걸어가는 앰버. 안쪽으로, 더욱 안쪽으로 들어가자, 벽 한 면이 온통 그림으로 뒤덮인.\n\nTIME OF DAY (lock): early morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 열람실·서가, 복도·외진 구석: 거미줄과 먼지가 가득한 구형 도서관 내부로 버려진 물품들이 방치되어 있다. (특징: 거미줄이 쳐진 목재 책꽂이와 바닥에 뒹구는 곰팡이 핀 책들; 바닥에 깔린 종이 박스와 웅크린 아이들; 작동 중인 구형 모니터 컴퓨터와 파란색 크레파스; 벽 한 면을 가득 채운 파란색 바다와 로봇, 야자수 크레파스 벽화)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 바다가 그려진 벽을 따라 천천히 걸어가는 앰버. 안쪽으로, 더욱 안쪽으로 들어가자, 벽 한 면이 온통 그림으로 뒤덮인.\n\nTIME OF DAY (lock): early morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_mural_recess_76b617.png",
  "asset_id": "a4fd405f-d731-4df5-b17a-6ffbfd317ce2",
  "input_asset_ids": [
   "d32c08c6-b25d-4171-ae56-70ae7eda22cc"
  ],
  "origin_tag": "S47sh9",
  "place_text": "Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.",
  "origin_inputs": {
   "place_text": "Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.",
   "time_of_day_en": "early morning",
   "conti_asset_id": "d32c08c6-b25d-4171-ae56-70ae7eda22cc"
  }
 },
 "S47sh9::bgfirst_bg": {
  "input_fingerprint": "275d98cfbd613155",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 벽에 크레파스를 댄 찰리의 낡은 뒷모습을 경이로운 눈빛으로 올려다보는 앰버의 상체.\n\nLOCATION (lock): Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage behind 찰리, laterally clear of his shoulder and at 앰버's upper-chest height, looking slightly upward into her three-quarter view. Keep his worn back as a narrow left foreground edge and his crayon touching the wall beside it, while 앰버's upper body occupies the right middle distance with enough separation for her gently raised chin to remain natural. Emphasize her gaze as she pauses to look up in wonder at 찰리, while he continues drawing and attends to the continuation of the mural beyond the left crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crayon mural on the library wall (Extensive drawing still being worked on) — The drawn face is seen obliquely beyond 찰리, showing a partial blue sea and the paradise imagery rather than a complete reproduction; used as Links the drawing hand to 앰버's wonder while keeping her reaction primary; Crayon (Held against the wall by 찰리) — Its contacting tip remains visible beside the foreground arm; used as Provides a small, concrete sign of the ongoing act of drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained early-morning illumination keeps the room subdued while allowing the mural's explicitly colorful sea imagery to carry the scene's tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 벽에 크레파스를 댄 찰리의 낡은 뒷모습을 경이로운 눈빛으로 올려다보는 앰버의 상체.\n\nLOCATION (lock): Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage behind 찰리, laterally clear of his shoulder and at 앰버's upper-chest height, looking slightly upward into her three-quarter view. Keep his worn back as a narrow left foreground edge and his crayon touching the wall beside it, while 앰버's upper body occupies the right middle distance with enough separation for her gently raised chin to remain natural. Emphasize her gaze as she pauses to look up in wonder at 찰리, while he continues drawing and attends to the continuation of the mural beyond the left crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crayon mural on the library wall (Extensive drawing still being worked on) — The drawn face is seen obliquely beyond 찰리, showing a partial blue sea and the paradise imagery rather than a complete reproduction; used as Links the drawing hand to 앰버's wonder while keeping her reaction primary; Crayon (Held against the wall by 찰리) — Its contacting tip remains visible beside the foreground arm; used as Provides a small, concrete sign of the ongoing act of drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained early-morning illumination keeps the room subdued while allowing the mural's explicitly colorful sea imagery to carry the scene's tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh9__bgfirst_bg.png",
  "asset_id": "41d48ac3-ec13-4552-8ceb-43554d15913a",
  "input_asset_ids": [
   "d32c08c6-b25d-4171-ae56-70ae7eda22cc",
   "a4fd405f-d731-4df5-b17a-6ffbfd317ce2"
  ]
 },
 "S47sh9": {
  "input_fingerprint": "49533baddc6682bb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 벽에 크레파스를 댄 찰리의 낡은 뒷모습을 경이로운 눈빛으로 올려다보는 앰버의 상체.\n\nLOCATION (lock): Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage behind 찰리, laterally clear of his shoulder and at 앰버's upper-chest height, looking slightly upward into her three-quarter view. Keep his worn back as a narrow left foreground edge and his crayon touching the wall beside it, while 앰버's upper body occupies the right middle distance with enough separation for her gently raised chin to remain natural. Emphasize her gaze as she pauses to look up in wonder at 찰리, while he continues drawing and attends to the continuation of the mural beyond the left crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crayon mural on the library wall (Extensive drawing still being worked on) — The drawn face is seen obliquely beyond 찰리, showing a partial blue sea and the paradise imagery rather than a complete reproduction; used as Links the drawing hand to 앰버's wonder while keeping her reaction primary; Crayon (Held against the wall by 찰리) — Its contacting tip remains visible beside the foreground arm; used as Provides a small, concrete sign of the ongoing act of drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained early-morning illumination keeps the room subdued while allowing the mural's explicitly colorful sea imagery to carry the scene's tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the early-morning library, a blue crayon sea leads into a wall-sized mural of Charlie, Hyunwoo, Raul and Amber arriving in Haenam by camper, with Pedro, Miyeon and the priest amid marine life, palms and vivid scenery. Crayon writing describes Haenam and Soyoung; boxes and books remain as makeshift bedding. 찰리: He is still drawing on the wall with crayons, with the blanket from the shop retained. 앰버: She is awake and has followed the mural farther into the library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 벽에 크레파스를 댄 찰리의 낡은 뒷모습을 경이로운 눈빛으로 올려다보는 앰버의 상체.\n\nLOCATION (lock): Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage behind 찰리, laterally clear of his shoulder and at 앰버's upper-chest height, looking slightly upward into her three-quarter view. Keep his worn back as a narrow left foreground edge and his crayon touching the wall beside it, while 앰버's upper body occupies the right middle distance with enough separation for her gently raised chin to remain natural. Emphasize her gaze as she pauses to look up in wonder at 찰리, while he continues drawing and attends to the continuation of the mural beyond the left crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crayon mural on the library wall (Extensive drawing still being worked on) — The drawn face is seen obliquely beyond 찰리, showing a partial blue sea and the paradise imagery rather than a complete reproduction; used as Links the drawing hand to 앰버's wonder while keeping her reaction primary; Crayon (Held against the wall by 찰리) — Its contacting tip remains visible beside the foreground arm; used as Provides a small, concrete sign of the ongoing act of drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained early-morning illumination keeps the room subdued while allowing the mural's explicitly colorful sea imagery to carry the scene's tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the early-morning library, a blue crayon sea leads into a wall-sized mural of Charlie, Hyunwoo, Raul and Amber arriving in Haenam by camper, with Pedro, Miyeon and the priest amid marine life, palms and vivid scenery. Crayon writing describes Haenam and Soyoung; boxes and books remain as makeshift bedding. 찰리: He is still drawing on the wall with crayons, with the blanket from the shop retained. 앰버: She is awake and has followed the mural farther into the library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 벽에 크레파스를 댄 찰리의 낡은 뒷모습을 경이로운 눈빛으로 올려다보는 앰버의 상체.\n\nLOCATION (lock): Along the mural-covered inner wall of the abandoned library, farther inside the reading-room area. Early-morning daylight reaches the interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the close-observation stage behind 찰리, laterally clear of his shoulder and at 앰버's upper-chest height, looking slightly upward into her three-quarter view. Keep his worn back as a narrow left foreground edge and his crayon touching the wall beside it, while 앰버's upper body occupies the right middle distance with enough separation for her gently raised chin to remain natural. Emphasize her gaze as she pauses to look up in wonder at 찰리, while he continues drawing and attends to the continuation of the mural beyond the left crop.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Crayon mural on the library wall (Extensive drawing still being worked on) — The drawn face is seen obliquely beyond 찰리, showing a partial blue sea and the paradise imagery rather than a complete reproduction; used as Links the drawing hand to 앰버's wonder while keeping her reaction primary; Crayon (Held against the wall by 찰리) — Its contacting tip remains visible beside the foreground arm; used as Provides a small, concrete sign of the ongoing act of drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained early-morning illumination keeps the room subdued while allowing the mural's explicitly colorful sea imagery to carry the scene's tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the early-morning library, a blue crayon sea leads into a wall-sized mural of Charlie, Hyunwoo, Raul and Amber arriving in Haenam by camper, with Pedro, Miyeon and the priest amid marine life, palms and vivid scenery. Crayon writing describes Haenam and Soyoung; boxes and books remain as makeshift bedding. 찰리: He is still drawing on the wall with crayons, with the blanket from the shop retained. 앰버: She is awake and has followed the mural farther into the library.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh9__bgfirst_bg.png",
     "asset_id": "41d48ac3-ec13-4552-8ceb-43554d15913a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S47sh9.png",
     "asset_id": "d32c08c6-b25d-4171-ae56-70ae7eda22cc",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_mural_recess_76b617.png",
     "asset_id": "a4fd405f-d731-4df5-b17a-6ffbfd317ce2",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버의 시선은 찰리의 머리를 향해 위로 향하고 있으며, 찰리의 왼팔과 크레파스는 왼쪽 벽면을 향하고 있습니다.",
    "built_space": "폐도서관 내부로, 왼쪽에 벽화가 있는 벽면과 오른쪽에 책장들이 위치해 있으며 뒷편 창문으로 아침 햇살이 들어옵니다.",
    "entities": "앰버는 금발, 작업복, 방진 마스크를 착용한 소녀로 일치하며, 찰리 역시 둥근 헬멧 형태의 베이지색 기계 몸체와 담요를 두른 모습이 레퍼런스와 부합합니다.",
    "hard_violations": [],
    "physics": "찰리와 앰버 모두 바닥에 안정적으로 서 있으며, 찰리의 손이 크레파스를 쥐고 벽을 누르는 동작이 물리적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "앰버가 찰리를 올려다보고 있으며, 찰리는 왼쪽 벽면을 향해 팔을 뻗고 있습니다.",
    "built_space": "동일한 폐도서관 구조로, 벽화와 책장, 창문의 배치가 레퍼런스와 일치합니다.",
    "entities": "앰버의 외형은 레퍼런스와 일치하지만, 찰리의 머리 외형이 둥근 형태가 아닌 각진 다각형 형태로 생성되어 레퍼런스와 차이가 있습니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 서 있으나, 찰리가 크레파스를 쥔 손가락 부분이 뭉개져 벽과 맞닿은 부분의 구조가 불명확합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 외형(둥근 머리 형태와 관절부)을 레퍼런스에 가깝게 재현하였으며, 크레파스를 쥔 손의 형태와 질감이 자연스럽게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리의 머리가 레퍼런스의 둥근 형태와 달리 각진 다각형 형태로 왜곡되었고, 크레파스를 쥔 손가락의 묘사가 다소 불분명합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 찰리의 머리를 향해 위로 향하고 있으며, 찰리의 왼팔과 크레파스는 왼쪽 벽면을 향하고 있습니다.",
        "built_space": "폐도서관 내부로, 왼쪽에 벽화가 있는 벽면과 오른쪽에 책장들이 위치해 있으며 뒷편 창문으로 아침 햇살이 들어옵니다.",
        "entities": "앰버는 금발, 작업복, 방진 마스크를 착용한 소녀로 일치하며, 찰리 역시 둥근 헬멧 형태의 베이지색 기계 몸체와 담요를 두른 모습이 레퍼런스와 부합합니다.",
        "hard_violations": [],
        "physics": "찰리와 앰버 모두 바닥에 안정적으로 서 있으며, 찰리의 손이 크레파스를 쥐고 벽을 누르는 동작이 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "앰버가 찰리를 올려다보고 있으며, 찰리는 왼쪽 벽면을 향해 팔을 뻗고 있습니다.",
        "built_space": "동일한 폐도서관 구조로, 벽화와 책장, 창문의 배치가 레퍼런스와 일치합니다.",
        "entities": "앰버의 외형은 레퍼런스와 일치하지만, 찰리의 머리 외형이 둥근 형태가 아닌 각진 다각형 형태로 생성되어 레퍼런스와 차이가 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서 있으나, 찰리가 크레파스를 쥔 손가락 부분이 뭉개져 벽과 맞닿은 부분의 구조가 불명확합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 외형(둥근 머리 형태와 관절부)을 레퍼런스에 가깝게 재현하였으며, 크레파스를 쥔 손의 형태와 질감이 자연스럽게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리의 머리가 레퍼런스의 둥근 형태와 달리 각진 다각형 형태로 왜곡되었고, 크레파스를 쥔 손가락의 묘사가 다소 불분명합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 시선은 찰리의 머리를 향해 위로 향하고 있으며, 찰리의 왼팔과 크레파스는 왼쪽 벽면을 향하고 있습니다.",
        "built_space": "폐도서관 내부로, 왼쪽에 벽화가 있는 벽면과 오른쪽에 책장들이 위치해 있으며 뒷편 창문으로 아침 햇살이 들어옵니다.",
        "entities": "앰버는 금발, 작업복, 방진 마스크를 착용한 소녀로 일치하며, 찰리 역시 둥근 헬멧 형태의 베이지색 기계 몸체와 담요를 두른 모습이 레퍼런스와 부합합니다.",
        "hard_violations": [],
        "physics": "찰리와 앰버 모두 바닥에 안정적으로 서 있으며, 찰리의 손이 크레파스를 쥐고 벽을 누르는 동작이 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "앰버가 찰리를 올려다보고 있으며, 찰리는 왼쪽 벽면을 향해 팔을 뻗고 있습니다.",
        "built_space": "동일한 폐도서관 구조로, 벽화와 책장, 창문의 배치가 레퍼런스와 일치합니다.",
        "entities": "앰버의 외형은 레퍼런스와 일치하지만, 찰리의 머리 외형이 둥근 형태가 아닌 각진 다각형 형태로 생성되어 레퍼런스와 차이가 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서 있으나, 찰리가 크레파스를 쥔 손가락 부분이 뭉개져 벽과 맞닿은 부분의 구조가 불명확합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 상체와 찰리를 올려다보는 경이로운 눈빛을 더 가까이 담아 우세하지만, 찰리의 등이 좁은 왼쪽 가장자리를 넘어 화면 절반가량을 차지하고 카메라도 지정된 낮은 위치보다 높다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "벽에 닿은 크레파스와 앰버의 올려다보는 시선은 맞지만, 찰리의 몸통과 도서관을 넓게 보여 주느라 앰버의 상체 반응을 중심으로 한 지정 구도에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 눈은 왼쪽 위 찰리의 머리와 윗등을 향하며 턱도 자연스럽게 올라가 있다. 찰리는 카메라에 등을 보이고 벽 쪽으로 머리를 돌렸으며 눈 자체는 보이지 않는다. 기계 손이 쥔 파란 크레파스는 왼쪽 위 벽면의 푸른 바다 그림에 끝을 대고 있다.",
        "built_space": "왼쪽 벽화 벽 하나, 오른쪽으로 이어지는 여러 책장, 뒤쪽 격자창과 열람용 탁자·의자가 보인다. 천장에는 길이 방향으로 형광등 기구가 적어도 두 개 식별되며, 바닥에는 책 더미·상자·담요가 놓여 있다. 재질과 공간의 배치는 장소 참조에 부합한다. 찰리는 벽 바로 앞 왼쪽 전경, 앰버는 그 뒤 오른쪽 통로에 있지만 찰리의 등이 지나치게 넓다. 카메라는 앰버의 윗가슴 높이에서 올려다보기보다 얼굴보다 높은 위치에서 내려다보는 쪽으로 읽힌다.",
        "entities": "실제 등장 개체는 찰리와 앰버뿐이다. 찰리의 보이는 머리·어깨·손은 마모된 샌드 베이지 장갑판과 기계 관절이며 인간 피부가 없다. 어깨와 등에 낡은 담요가 걸쳐져 있다. 얼굴과 가슴 원자로는 뒷모습 구도에 가려진다. 앰버는 참조와 유사한 금발의 어린 여자아이로, 밝은 피부와 큰 눈, 때 묻은 카키 작업복, 목에 걸린 정교한 방진 마스크가 보인다. 혼혈 배경 자체는 외형만으로 확정할 수 없다. 공구 벨트는 프레임 밖이다. 벽에는 파란 바다·물고기·야자수·풍경 일부가 표면의 마모를 따라 보이며, 추가 문구나 사진 위 자막은 없다.",
        "hard_violations": [],
        "physics": "크레파스는 찰리의 기계 손가락에 잡혀 있고 끝이 벽에 접촉한다. 그리는 손으로 이어지는 팔 일부는 전경 어깨에 가려지지만 불가능한 연결은 보이지 않는다. 담요는 어깨에 지지되어 등에 늘어진다. 두 인물은 서 있는 상체 자세이며 발은 프레임 밖이므로 직접적인 바닥 접촉은 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 책과 상자는 바닥 또는 다른 책 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "앰버는 왼쪽 위 찰리의 머리 쪽을 바라보고 입을 조금 벌려 놀라움을 표현한다. 찰리의 머리와 몸은 벽을 향하고 얼굴은 숨겨져 있다. 왼쪽에 들어 올린 기계 손의 파란 크레파스 끝은 벽화의 고래 꼬리 주변 푸른 면을 향해 실제로 닿아 있다.",
        "built_space": "왼쪽 벽화 벽 하나와 그 위 창 구획, 오른쪽 책장 열, 끝의 큰 격자창, 뒤쪽 열람용 탁자·의자가 보인다. 천장 형광등은 앞에서 뒤로 네 개가 식별된다. 바닥의 책·상자·담요 배치와 콘크리트 보 구조는 장소 참조와 잘 맞는다. 찰리는 벽 가까이에, 앰버는 오른쪽 통로에 서 있어 공간 사용은 가능하다. 다만 찰리의 등과 팔 전체가 화면 왼쪽 절반을 채우며 앰버는 허리까지 작게 들어간다. 높은 카메라와 넓은 공간 제시 때문에 지정된 낮은 관찰 시점의 상체 중심 구도와 다르다.",
        "entities": "추가 인물 없이 찰리와 앰버가 보인다. 찰리는 참조에 가까운 각진 베이지 기계 장갑과 육중한 팔을 갖고 등에 담요를 걸쳤으며 인간 얼굴이나 피부가 추가되지 않았다. 얼굴과 가슴 표식은 뒤로 돌아선 몸에 가려져 있다. 앰버는 금발의 어린 여자아이로 참조와 유사한 얼굴, 카키 작업복, 목에 걸린 기계식 방진 마스크와 허리의 가죽 공구 벨트를 갖춘다. 벽화에는 바다·고래·야자수·캠핑카·등대가 보이고 그림은 낡은 벽 표면에 붙어 있다. 해남과 소영에 관한 읽을 수 있는 문구는 이 범위에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 굽힌 팔은 어깨·팔꿈치·손목으로 연속되며 손가락이 크레파스를 잡고 벽에 끝을 댄다. 담요는 목과 어깨에 걸쳐 중력 방향으로 늘어진다. 앰버의 마스크는 목의 끈에, 공구 벨트는 허리에 지지된다. 두 인물의 발은 잘렸지만 몸통 자세에 부유나 불가능한 지지 관계는 없다. 바닥의 상자와 책 더미도 접촉면을 갖는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앰버의 상체와 찰리를 올려다보는 경이로운 눈빛을 더 가까이 담아 우세하지만, 찰리의 등이 좁은 왼쪽 가장자리를 넘어 화면 절반가량을 차지하고 카메라도 지정된 낮은 위치보다 높다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "벽에 닿은 크레파스와 앰버의 올려다보는 시선은 맞지만, 찰리의 몸통과 도서관을 넓게 보여 주느라 앰버의 상체 반응을 중심으로 한 지정 구도에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 눈은 왼쪽 위 찰리의 머리와 윗등을 향하며 턱도 자연스럽게 올라가 있다. 찰리는 카메라에 등을 보이고 벽 쪽으로 머리를 돌렸으며 눈 자체는 보이지 않는다. 기계 손이 쥔 파란 크레파스는 왼쪽 위 벽면의 푸른 바다 그림에 끝을 대고 있다.",
        "built_space": "왼쪽 벽화 벽 하나, 오른쪽으로 이어지는 여러 책장, 뒤쪽 격자창과 열람용 탁자·의자가 보인다. 천장에는 길이 방향으로 형광등 기구가 적어도 두 개 식별되며, 바닥에는 책 더미·상자·담요가 놓여 있다. 재질과 공간의 배치는 장소 참조에 부합한다. 찰리는 벽 바로 앞 왼쪽 전경, 앰버는 그 뒤 오른쪽 통로에 있지만 찰리의 등이 지나치게 넓다. 카메라는 앰버의 윗가슴 높이에서 올려다보기보다 얼굴보다 높은 위치에서 내려다보는 쪽으로 읽힌다.",
        "entities": "실제 등장 개체는 찰리와 앰버뿐이다. 찰리의 보이는 머리·어깨·손은 마모된 샌드 베이지 장갑판과 기계 관절이며 인간 피부가 없다. 어깨와 등에 낡은 담요가 걸쳐져 있다. 얼굴과 가슴 원자로는 뒷모습 구도에 가려진다. 앰버는 참조와 유사한 금발의 어린 여자아이로, 밝은 피부와 큰 눈, 때 묻은 카키 작업복, 목에 걸린 정교한 방진 마스크가 보인다. 혼혈 배경 자체는 외형만으로 확정할 수 없다. 공구 벨트는 프레임 밖이다. 벽에는 파란 바다·물고기·야자수·풍경 일부가 표면의 마모를 따라 보이며, 추가 문구나 사진 위 자막은 없다.",
        "hard_violations": [],
        "physics": "크레파스는 찰리의 기계 손가락에 잡혀 있고 끝이 벽에 접촉한다. 그리는 손으로 이어지는 팔 일부는 전경 어깨에 가려지지만 불가능한 연결은 보이지 않는다. 담요는 어깨에 지지되어 등에 늘어진다. 두 인물은 서 있는 상체 자세이며 발은 프레임 밖이므로 직접적인 바닥 접촉은 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 책과 상자는 바닥 또는 다른 책 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "앰버는 왼쪽 위 찰리의 머리 쪽을 바라보고 입을 조금 벌려 놀라움을 표현한다. 찰리의 머리와 몸은 벽을 향하고 얼굴은 숨겨져 있다. 왼쪽에 들어 올린 기계 손의 파란 크레파스 끝은 벽화의 고래 꼬리 주변 푸른 면을 향해 실제로 닿아 있다.",
        "built_space": "왼쪽 벽화 벽 하나와 그 위 창 구획, 오른쪽 책장 열, 끝의 큰 격자창, 뒤쪽 열람용 탁자·의자가 보인다. 천장 형광등은 앞에서 뒤로 네 개가 식별된다. 바닥의 책·상자·담요 배치와 콘크리트 보 구조는 장소 참조와 잘 맞는다. 찰리는 벽 가까이에, 앰버는 오른쪽 통로에 서 있어 공간 사용은 가능하다. 다만 찰리의 등과 팔 전체가 화면 왼쪽 절반을 채우며 앰버는 허리까지 작게 들어간다. 높은 카메라와 넓은 공간 제시 때문에 지정된 낮은 관찰 시점의 상체 중심 구도와 다르다.",
        "entities": "추가 인물 없이 찰리와 앰버가 보인다. 찰리는 참조에 가까운 각진 베이지 기계 장갑과 육중한 팔을 갖고 등에 담요를 걸쳤으며 인간 얼굴이나 피부가 추가되지 않았다. 얼굴과 가슴 표식은 뒤로 돌아선 몸에 가려져 있다. 앰버는 금발의 어린 여자아이로 참조와 유사한 얼굴, 카키 작업복, 목에 걸린 기계식 방진 마스크와 허리의 가죽 공구 벨트를 갖춘다. 벽화에는 바다·고래·야자수·캠핑카·등대가 보이고 그림은 낡은 벽 표면에 붙어 있다. 해남과 소영에 관한 읽을 수 있는 문구는 이 범위에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 굽힌 팔은 어깨·팔꿈치·손목으로 연속되며 손가락이 크레파스를 잡고 벽에 끝을 댄다. 담요는 목과 어깨에 걸쳐 중력 방향으로 늘어진다. 앰버의 마스크는 목의 끈에, 공구 벨트는 허리에 지지된다. 두 인물의 발은 잘렸지만 몸통 자세에 부유나 불가능한 지지 관계는 없다. 바닥의 상자와 책 더미도 접촉면을 갖는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1714,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "찰리의 외형(둥근 머리 형태와 관절부)을 레퍼런스에 가깝게 재현하였으며, 크레파스를 쥔 손의 형태와 질감이 자연스럽게 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "찰리의 머리가 레퍼런스의 둥근 형태와 달리 각진 다각형 형태로 왜곡되었고, 크레파스를 쥔 손가락의 묘사가 다소 불분명합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_mural_recess_76b617.png",
    "asset_id": "a4fd405f-d731-4df5-b17a-6ffbfd317ce2",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b22-37a7-7d13-a2ff-7986991324fb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh9__bgfirst_bg.png",
   "bg_asset_id": "41d48ac3-ec13-4552-8ceb-43554d15913a",
   "bg_record_key": "S47sh9::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "library_mural_recess",
   "groupbg_asset_id": "a4fd405f-d731-4df5-b17a-6ffbfd317ce2"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh17::signage": {
  "fp": "08fa191951bd63ae",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::library_secluded_exterior": {
  "input_fingerprint": "d08ed395db15577c",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "library_secluded_exterior",
    "tags": [
     "S47sh17",
     "S47sh21"
    ]
   },
   "context_sig": "f11b103199dd2b1b"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 열람실·서가, 복도·외진 구석: 거미줄과 먼지가 가득한 구형 도서관 내부로 버려진 물품들이 방치되어 있다. (특징: 거미줄이 쳐진 목재 책꽂이와 바닥에 뒹구는 곰팡이 핀 책들; 바닥에 깔린 종이 박스와 웅크린 아이들; 작동 중인 구형 모니터 컴퓨터와 파란색 크레파스; 벽 한 면을 가득 채운 파란색 바다와 로봇, 야자수 크레파스 벽화)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 주머니 속에 일회용 핸드폰 (무인점포에서 집어온)을 들고 슬쩍 밖으로 나간다.\n- / 도서관 외진 구석\n- 한숨 쉬며 돌아서는 현우, 아뿔싸! 찰리가 곁에 와 있다.\n\nTIME OF DAY (lock): early morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n버려진 도서관 열람실·서가, 복도·외진 구석: 거미줄과 먼지가 가득한 구형 도서관 내부로 버려진 물품들이 방치되어 있다. (특징: 거미줄이 쳐진 목재 책꽂이와 바닥에 뒹구는 곰팡이 핀 책들; 바닥에 깔린 종이 박스와 웅크린 아이들; 작동 중인 구형 모니터 컴퓨터와 파란색 크레파스; 벽 한 면을 가득 채운 파란색 바다와 로봇, 야자수 크레파스 벽화)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 주머니 속에 일회용 핸드폰 (무인점포에서 집어온)을 들고 슬쩍 밖으로 나간다.\n- / 도서관 외진 구석\n- 한숨 쉬며 돌아서는 현우, 아뿔싸! 찰리가 곁에 와 있다.\n\nTIME OF DAY (lock): early morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_secluded_exterior_66c1ef.png",
  "asset_id": "2a733389-3bb0-4279-b042-4b4a61eccf49",
  "input_asset_ids": [
   "8fe4197f-d3b8-4326-baa1-d7b771f793ef"
  ],
  "origin_tag": "S47sh17",
  "place_text": "In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.",
  "origin_inputs": {
   "place_text": "In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.",
   "time_of_day_en": "early morning",
   "conti_asset_id": "8fe4197f-d3b8-4326-baa1-d7b771f793ef"
  }
 },
 "S47sh17::bgfirst_bg": {
  "input_fingerprint": "77abbc85f11820f5",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 일회용 핸드폰을 귀에 바짝 댄 이현우가 어두운 복도 구석에서 긴장된 표정으로 서 있는 상체.\n\nLOCATION (lock): In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the private-call stage from 이현우's front three-quarter side, slightly above his eye line with a gentle downward tilt, before advancing closer. Place his upper body in the right third, elbow folded tightly as he presses the disposable phone to his ear, leaving negative space to the left along his averted gaze toward the secluded corridor outside frame. His lifted shoulder and arrested breath express concentration on the caller; exclude the route behind him and use camera distance as the approach's sole emphasized change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Secluded corridor corner (Separated from the group's gathering area) — The intersecting walls sit obliquely behind 이현우; the approach behind him is excluded; used as Makes the call feel private without revealing 찰리 prematurely; Disposable phone (In use, held tightly to 이현우's ear) — Its outer side is visible; the display is not presented to camera; used as A small focal accent beside the tense expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the corridor corner dim within the early-morning setting, preserving the tense face with restrained ambient visibility and no added practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 일회용 핸드폰을 귀에 바짝 댄 이현우가 어두운 복도 구석에서 긴장된 표정으로 서 있는 상체.\n\nLOCATION (lock): In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the private-call stage from 이현우's front three-quarter side, slightly above his eye line with a gentle downward tilt, before advancing closer. Place his upper body in the right third, elbow folded tightly as he presses the disposable phone to his ear, leaving negative space to the left along his averted gaze toward the secluded corridor outside frame. His lifted shoulder and arrested breath express concentration on the caller; exclude the route behind him and use camera distance as the approach's sole emphasized change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Secluded corridor corner (Separated from the group's gathering area) — The intersecting walls sit obliquely behind 이현우; the approach behind him is excluded; used as Makes the call feel private without revealing 찰리 prematurely; Disposable phone (In use, held tightly to 이현우's ear) — Its outer side is visible; the display is not presented to camera; used as A small focal accent beside the tense expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the corridor corner dim within the early-morning setting, preserving the tense face with restrained ambient visibility and no added practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh17__bgfirst_bg.png",
  "asset_id": "b893036f-70de-45d9-b267-915169e25586",
  "input_asset_ids": [
   "8fe4197f-d3b8-4326-baa1-d7b771f793ef",
   "2a733389-3bb0-4279-b042-4b4a61eccf49"
  ]
 },
 "S47sh17": {
  "input_fingerprint": "da71e5796648d788",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 일회용 핸드폰을 귀에 바짝 댄 이현우가 어두운 복도 구석에서 긴장된 표정으로 서 있는 상체.\n\nLOCATION (lock): In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the private-call stage from 이현우's front three-quarter side, slightly above his eye line with a gentle downward tilt, before advancing closer. Place his upper body in the right third, elbow folded tightly as he presses the disposable phone to his ear, leaving negative space to the left along his averted gaze toward the secluded corridor outside frame. His lifted shoulder and arrested breath express concentration on the caller; exclude the route behind him and use camera distance as the approach's sole emphasized change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Secluded corridor corner (Separated from the group's gathering area) — The intersecting walls sit obliquely behind 이현우; the approach behind him is excluded; used as Makes the call feel private without revealing 찰리 prematurely; Disposable phone (In use, held tightly to 이현우's ear) — Its outer side is visible; the display is not presented to camera; used as A small focal accent beside the tense expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the corridor corner dim within the early-morning setting, preserving the tense face with restrained ambient visibility and no added practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sea mural, the Haenam arrival scene and their crayon writing remain on the library walls. The dusty shelves and makeshift bedding remain in place. 이현우: He is in a secluded corner using the disposable phone, having removed a shoe to retrieve the intact contact card. His facial bruising and thigh wound remain untreated.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 일회용 핸드폰을 귀에 바짝 댄 이현우가 어두운 복도 구석에서 긴장된 표정으로 서 있는 상체.\n\nLOCATION (lock): In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the private-call stage from 이현우's front three-quarter side, slightly above his eye line with a gentle downward tilt, before advancing closer. Place his upper body in the right third, elbow folded tightly as he presses the disposable phone to his ear, leaving negative space to the left along his averted gaze toward the secluded corridor outside frame. His lifted shoulder and arrested breath express concentration on the caller; exclude the route behind him and use camera distance as the approach's sole emphasized change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Secluded corridor corner (Separated from the group's gathering area) — The intersecting walls sit obliquely behind 이현우; the approach behind him is excluded; used as Makes the call feel private without revealing 찰리 prematurely; Disposable phone (In use, held tightly to 이현우's ear) — Its outer side is visible; the display is not presented to camera; used as A small focal accent beside the tense expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the corridor corner dim within the early-morning setting, preserving the tense face with restrained ambient visibility and no added practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sea mural, the Haenam arrival scene and their crayon writing remain on the library walls. The dusty shelves and makeshift bedding remain in place. 이현우: He is in a secluded corner using the disposable phone, having removed a shoe to retrieve the intact contact card. His facial bruising and thigh wound remain untreated.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 일회용 핸드폰을 귀에 바짝 댄 이현우가 어두운 복도 구석에서 긴장된 표정으로 서 있는 상체.\n\nLOCATION (lock): In a secluded, dim corridor corner of the abandoned library, away from the others. Morning daylight is weaker here than in the main reading area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the private-call stage from 이현우's front three-quarter side, slightly above his eye line with a gentle downward tilt, before advancing closer. Place his upper body in the right third, elbow folded tightly as he presses the disposable phone to his ear, leaving negative space to the left along his averted gaze toward the secluded corridor outside frame. His lifted shoulder and arrested breath express concentration on the caller; exclude the route behind him and use camera distance as the approach's sole emphasized change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Secluded corridor corner (Separated from the group's gathering area) — The intersecting walls sit obliquely behind 이현우; the approach behind him is excluded; used as Makes the call feel private without revealing 찰리 prematurely; Disposable phone (In use, held tightly to 이현우's ear) — Its outer side is visible; the display is not presented to camera; used as A small focal accent beside the tense expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the corridor corner dim within the early-morning setting, preserving the tense face with restrained ambient visibility and no added practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sea mural, the Haenam arrival scene and their crayon writing remain on the library walls. The dusty shelves and makeshift bedding remain in place. 이현우: He is in a secluded corner using the disposable phone, having removed a shoe to retrieve the intact contact card. His facial bruising and thigh wound remain untreated.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh17__bgfirst_bg.png",
     "asset_id": "b893036f-70de-45d9-b267-915169e25586",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S47sh17.png",
     "asset_id": "8fe4197f-d3b8-4326-baa1-d7b771f793ef",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_secluded_exterior_66c1ef.png",
     "asset_id": "2a733389-3bb0-4279-b042-4b4a61eccf49",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 프롬프트의 지시대로 화면 왼쪽의 여백을 향해 시선을 두고 있으며, 왼손으로 귀에 휴대폰을 바짝 대고 있음.",
    "built_space": "기준 사진의 도서관 복도가 나타나지만, 기둥 정면에 있어야 할 벽화가 우측 복도 벽면에 맞춰 깊이를 가지도록 투시가 변형됨. 지시문에서 제외하라고 명시된 인물 뒤편의 긴 복도 통로가 화면 왼쪽에 그대로 노출되어 있음.",
    "entities": "이현우의 외모, 상처, 옷차림은 캐릭터 기준 이미지와 일치함. 일회용 휴대폰의 형태와 벽화의 주요 구성 요소(등대, 배, 로봇, 야자수 등)도 명확하게 묘사됨.",
    "hard_violations": [],
    "physics": "인물의 자세는 안정적이며, 왼손이 휴대폰을 쥐고 지탱하는 모습과 관절 형태가 자연스럽고 물리적인 오류가 없음."
   },
   {
    "label": "B",
    "direction": "이현우는 프롬프트의 의도와 반대로 왼쪽 여백이 아닌 화면 오른쪽 테두리 밖으로 시선을 돌리고 있음. 오른손으로 휴대폰을 귀에 대고 있음.",
    "built_space": "후보 A와 마찬가지로 벽화가 우측 벽면에 덧입혀져 원본의 공간 구조(기둥 정면)를 왜곡했으며, 프롬프트에서 배제하라 지시한 긴 복도 통로가 화면 왼쪽에 깊게 노출됨.",
    "entities": "이현우의 인물 특징과 의상은 기준과 잘 맞음. 휴대폰과 벽화의 주요 디테일도 포함되어 있음.",
    "hard_violations": [],
    "physics": "오른손으로 휴대폰을 쥔 형태가 다소 뭉툭하지만 지지점과 관절 구조 자체는 물리적으로 불가능한 수준이 아니며, 인물의 서 있는 자세도 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "시선을 왼쪽 여백으로 향해 프레이밍 지시를 잘 따랐고 통화에 집중하는 긴장된 표정이 우수하나, 제외해야 할 배경의 복도 통로가 노출되고 벽화의 공간 구조가 변경된 점은 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "제외되어야 할 긴 복도가 배경에 나타난 데다, 인물이 왼쪽 여백이 아닌 오른쪽 프레임 밖을 응시하고 있어 지시된 시선 및 프레이밍 조건을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프롬프트의 지시대로 화면 왼쪽의 여백을 향해 시선을 두고 있으며, 왼손으로 귀에 휴대폰을 바짝 대고 있음.",
        "built_space": "기준 사진의 도서관 복도가 나타나지만, 기둥 정면에 있어야 할 벽화가 우측 복도 벽면에 맞춰 깊이를 가지도록 투시가 변형됨. 지시문에서 제외하라고 명시된 인물 뒤편의 긴 복도 통로가 화면 왼쪽에 그대로 노출되어 있음.",
        "entities": "이현우의 외모, 상처, 옷차림은 캐릭터 기준 이미지와 일치함. 일회용 휴대폰의 형태와 벽화의 주요 구성 요소(등대, 배, 로봇, 야자수 등)도 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "인물의 자세는 안정적이며, 왼손이 휴대폰을 쥐고 지탱하는 모습과 관절 형태가 자연스럽고 물리적인 오류가 없음."
       },
       {
        "label": "B",
        "direction": "이현우는 프롬프트의 의도와 반대로 왼쪽 여백이 아닌 화면 오른쪽 테두리 밖으로 시선을 돌리고 있음. 오른손으로 휴대폰을 귀에 대고 있음.",
        "built_space": "후보 A와 마찬가지로 벽화가 우측 벽면에 덧입혀져 원본의 공간 구조(기둥 정면)를 왜곡했으며, 프롬프트에서 배제하라 지시한 긴 복도 통로가 화면 왼쪽에 깊게 노출됨.",
        "entities": "이현우의 인물 특징과 의상은 기준과 잘 맞음. 휴대폰과 벽화의 주요 디테일도 포함되어 있음.",
        "hard_violations": [],
        "physics": "오른손으로 휴대폰을 쥔 형태가 다소 뭉툭하지만 지지점과 관절 구조 자체는 물리적으로 불가능한 수준이 아니며, 인물의 서 있는 자세도 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "시선을 왼쪽 여백으로 향해 프레이밍 지시를 잘 따랐고 통화에 집중하는 긴장된 표정이 우수하나, 제외해야 할 배경의 복도 통로가 노출되고 벽화의 공간 구조가 변경된 점은 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "제외되어야 할 긴 복도가 배경에 나타난 데다, 인물이 왼쪽 여백이 아닌 오른쪽 프레임 밖을 응시하고 있어 지시된 시선 및 프레이밍 조건을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 프롬프트의 지시대로 화면 왼쪽의 여백을 향해 시선을 두고 있으며, 왼손으로 귀에 휴대폰을 바짝 대고 있음.",
        "built_space": "기준 사진의 도서관 복도가 나타나지만, 기둥 정면에 있어야 할 벽화가 우측 복도 벽면에 맞춰 깊이를 가지도록 투시가 변형됨. 지시문에서 제외하라고 명시된 인물 뒤편의 긴 복도 통로가 화면 왼쪽에 그대로 노출되어 있음.",
        "entities": "이현우의 외모, 상처, 옷차림은 캐릭터 기준 이미지와 일치함. 일회용 휴대폰의 형태와 벽화의 주요 구성 요소(등대, 배, 로봇, 야자수 등)도 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "인물의 자세는 안정적이며, 왼손이 휴대폰을 쥐고 지탱하는 모습과 관절 형태가 자연스럽고 물리적인 오류가 없음."
       },
       {
        "label": "B",
        "direction": "이현우는 프롬프트의 의도와 반대로 왼쪽 여백이 아닌 화면 오른쪽 테두리 밖으로 시선을 돌리고 있음. 오른손으로 휴대폰을 귀에 대고 있음.",
        "built_space": "후보 A와 마찬가지로 벽화가 우측 벽면에 덧입혀져 원본의 공간 구조(기둥 정면)를 왜곡했으며, 프롬프트에서 배제하라 지시한 긴 복도 통로가 화면 왼쪽에 깊게 노출됨.",
        "entities": "이현우의 인물 특징과 의상은 기준과 잘 맞음. 휴대폰과 벽화의 주요 디테일도 포함되어 있음.",
        "hard_violations": [],
        "physics": "오른손으로 휴대폰을 쥔 형태가 다소 뭉툭하지만 지지점과 관절 구조 자체는 물리적으로 불가능한 수준이 아니며, 인물의 서 있는 자세도 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "오른쪽 상체 배치, 왼쪽으로 피한 시선, 귀에 밀착한 전화와 접힌 팔은 맞지만, 제외해야 할 뒤쪽 접근 복도를 넓게 드러낸다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "통화 동작과 인물·장소는 부합하지만, 시선이 왼쪽 여백과 반대인 오른쪽을 향하고 뒤쪽 접근 복도까지 노출한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈이 화면 오른쪽 바깥을 향한다. 지시된 왼쪽 여백 너머의 외진 복도를 보는 방향과 반대다. 전화는 귀에 붙어 있고, 카메라에는 화면이 아닌 외장이 보인다.",
        "built_space": "인물은 화면 오른쪽의 벽 모서리 옆에 서 있다. 바다 벽화 한 면, 양쪽으로 이어지는 책장 열, 끝 창 한 개, 왼쪽 측면 창들과 천장 등기구 여러 개가 보인다. 낡은 회벽과 서가, 바닥의 책·상자는 장소 참조에 부합한다. 그러나 왼쪽 화면 대부분에 뒤로 이어지는 접근 복도가 드러나, 교차하는 벽만으로 통화 공간을 닫고 접근로를 제외하라는 구도와 다르다. 상체 미디엄 숏이지만 카메라의 하향 각도는 뚜렷하지 않다.",
        "entities": "짧고 헝클어진 검은 머리의 동아시아계 젊은 남성 한 명이며, 참조 인물과 얼굴·마른 체격이 대체로 부합한다. 어두운 셔츠에 먼지와 핏자국이 있고 얼굴의 멍도 보인다. 작은 검은 전화 한 개를 들고 있지만 일회용 여부는 외관만으로 확정할 수 없다. 인이어 무전기는 확인되지 않는다. 신발, 연락처 카드, 허벅지 상처는 프레임 밖이므로 판단하지 않는다. 추가 인물이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "전화는 인물의 손가락과 손바닥에 잡혀 귀에 눌려 있다. 손목과 전완은 셔츠 소매로 자연스럽게 이어지고, 팔꿈치가 굽혀져 통화 자세를 지탱한다. 몸통은 세워져 있으며 발은 프레임 밖이다. 공중에 뜬 자세나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴과 눈이 화면 왼쪽의 복도 쪽을 향해, 오른쪽에 배치된 인물의 시선 앞에 왼쪽 여백이 생긴다. 다만 시선이 향하는 복도 자체도 화면 안에 크게 보인다. 전화는 귀에 밀착하고 외장이 카메라를 향해 화면 노출 금지를 지킨다.",
        "built_space": "인물의 상체는 오른쪽 삼분할 부근에 있고, 벽 모서리에 바짝 붙어 있다. 바다 벽화 한 면, 양쪽 서가 열, 끝 창 한 개, 왼쪽 측면 창들과 천장 등기구 여러 개가 보인다. 책장·회벽·짙은 걸레받이·바닥의 책과 상자는 장소 참조와 잘 연결된다. 그러나 뒤쪽 접근로가 화면 왼쪽에 길게 열려 있어 통화 공간을 숨기는 배경 지시를 충족하지 못한다. 미디엄 숏과 약간 높은 전방 사선 시점은 대체로 맞는다.",
        "entities": "동아시아계의 10대 후반으로 보이는 남성 한 명으로, 검은 헝클어진 머리와 얼굴·체격이 참조에 대체로 부합한다. 낡은 어두운 셔츠의 흙먼지와 핏자국, 치료되지 않은 얼굴 멍이 보인다. 작은 검은 전화 한 개를 사용하며, 인이어 무전기는 전화와 머리카락에 가려 확인하기 어렵다. 신발 제거 상태, 카드, 허벅지 상처는 이 상체 프레이밍에서 확인 대상이 아니다. 다른 인물이나 추가 문자는 없다.",
        "hard_violations": [],
        "physics": "인물 자신의 손이 전화를 감싸 잡아 귀에 고정하며, 손목과 전완이 자연스럽게 연결된다. 팔꿈치를 몸 가까이 접고 어깨를 조금 올린 통화 자세가 물리적으로 가능하다. 몸통은 벽 곁에서 세워져 있고 발은 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "오른쪽 상체 배치, 왼쪽으로 피한 시선, 귀에 밀착한 전화와 접힌 팔은 맞지만, 제외해야 할 뒤쪽 접근 복도를 넓게 드러낸다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "통화 동작과 인물·장소는 부합하지만, 시선이 왼쪽 여백과 반대인 오른쪽을 향하고 뒤쪽 접근 복도까지 노출한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈이 화면 오른쪽 바깥을 향한다. 지시된 왼쪽 여백 너머의 외진 복도를 보는 방향과 반대다. 전화는 귀에 붙어 있고, 카메라에는 화면이 아닌 외장이 보인다.",
        "built_space": "인물은 화면 오른쪽의 벽 모서리 옆에 서 있다. 바다 벽화 한 면, 양쪽으로 이어지는 책장 열, 끝 창 한 개, 왼쪽 측면 창들과 천장 등기구 여러 개가 보인다. 낡은 회벽과 서가, 바닥의 책·상자는 장소 참조에 부합한다. 그러나 왼쪽 화면 대부분에 뒤로 이어지는 접근 복도가 드러나, 교차하는 벽만으로 통화 공간을 닫고 접근로를 제외하라는 구도와 다르다. 상체 미디엄 숏이지만 카메라의 하향 각도는 뚜렷하지 않다.",
        "entities": "짧고 헝클어진 검은 머리의 동아시아계 젊은 남성 한 명이며, 참조 인물과 얼굴·마른 체격이 대체로 부합한다. 어두운 셔츠에 먼지와 핏자국이 있고 얼굴의 멍도 보인다. 작은 검은 전화 한 개를 들고 있지만 일회용 여부는 외관만으로 확정할 수 없다. 인이어 무전기는 확인되지 않는다. 신발, 연락처 카드, 허벅지 상처는 프레임 밖이므로 판단하지 않는다. 추가 인물이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "전화는 인물의 손가락과 손바닥에 잡혀 귀에 눌려 있다. 손목과 전완은 셔츠 소매로 자연스럽게 이어지고, 팔꿈치가 굽혀져 통화 자세를 지탱한다. 몸통은 세워져 있으며 발은 프레임 밖이다. 공중에 뜬 자세나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴과 눈이 화면 왼쪽의 복도 쪽을 향해, 오른쪽에 배치된 인물의 시선 앞에 왼쪽 여백이 생긴다. 다만 시선이 향하는 복도 자체도 화면 안에 크게 보인다. 전화는 귀에 밀착하고 외장이 카메라를 향해 화면 노출 금지를 지킨다.",
        "built_space": "인물의 상체는 오른쪽 삼분할 부근에 있고, 벽 모서리에 바짝 붙어 있다. 바다 벽화 한 면, 양쪽 서가 열, 끝 창 한 개, 왼쪽 측면 창들과 천장 등기구 여러 개가 보인다. 책장·회벽·짙은 걸레받이·바닥의 책과 상자는 장소 참조와 잘 연결된다. 그러나 뒤쪽 접근로가 화면 왼쪽에 길게 열려 있어 통화 공간을 숨기는 배경 지시를 충족하지 못한다. 미디엄 숏과 약간 높은 전방 사선 시점은 대체로 맞는다.",
        "entities": "동아시아계의 10대 후반으로 보이는 남성 한 명으로, 검은 헝클어진 머리와 얼굴·체격이 참조에 대체로 부합한다. 낡은 어두운 셔츠의 흙먼지와 핏자국, 치료되지 않은 얼굴 멍이 보인다. 작은 검은 전화 한 개를 사용하며, 인이어 무전기는 전화와 머리카락에 가려 확인하기 어렵다. 신발 제거 상태, 카드, 허벅지 상처는 이 상체 프레이밍에서 확인 대상이 아니다. 다른 인물이나 추가 문자는 없다.",
        "hard_violations": [],
        "physics": "인물 자신의 손이 전화를 감싸 잡아 귀에 고정하며, 손목과 전완이 자연스럽게 연결된다. 팔꿈치를 몸 가까이 접고 어깨를 조금 올린 통화 자세가 물리적으로 가능하다. 몸통은 벽 곁에서 세워져 있고 발은 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.238
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.238
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1238
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "시선을 왼쪽 여백으로 향해 프레이밍 지시를 잘 따랐고 통화에 집중하는 긴장된 표정이 우수하나, 제외해야 할 배경의 복도 통로가 노출되고 벽화의 공간 구조가 변경된 점은 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1238,
    "verdict_ko": "제외되어야 할 긴 복도가 배경에 나타난 데다, 인물이 왼쪽 여백이 아닌 오른쪽 프레임 밖을 응시하고 있어 지시된 시선 및 프레이밍 조건을 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_library_secluded_exterior_66c1ef.png",
    "asset_id": "2a733389-3bb0-4279-b042-4b4a61eccf49",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b2b-b907-7c30-87f4-a113117ddd9d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh17__bgfirst_bg.png",
   "bg_asset_id": "b893036f-70de-45d9-b267-915169e25586",
   "bg_record_key": "S47sh17::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "library_secluded_exterior",
   "groupbg_asset_id": "2a733389-3bb0-4279-b042-4b4a61eccf49"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh21::signage": {
  "fp": "c9845850d9884c57",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh21": {
  "input_fingerprint": "aef698b3f8a6280c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 몸을 돌린 이현우의 바로 눈앞에 거대한 찰리가 조용히 서 있는 구도.\n\nLOCATION (lock): In the same secluded corridor corner inside the abandoned library, under subdued morning light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the pan at 이현우's shoulder height, looking past his rear three-quarter shoulder with a slight upward angle toward 찰리, without crossing their established axis. Keep 이현우's interrupted turn in the left foreground quarter and reveal 찰리's upper body just beyond him on the right at close conversational separation, with the corridor corner connecting both depth planes. Emphasize their changed facing relationship: 이현우 freezes while looking at 찰리, and 찰리 quietly inclines his head toward 이현우 rather than addressing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우's turned shoulder in the middle-left of the frame, foreground, looks toward 찰리 immediately beyond his shoulder; 찰리 at close conversational separation in the middle-right of the frame, midground, looks toward 이현우 in the same corridor corner.\n- KEY BACKGROUND ELEMENTS: Library corridor corner (The same secluded area used for the call) — The wall junction continues behind both figures on the established side of their axis; used as Demonstrates that 찰리 is physically beside 이현우 in the same space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corner's subdued morning ambience and unchanged contrast so the revelation comes from spatial disclosure, not a lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the secluded library corner's worn surfaces, dust, and subdued daytime illumination from the reference. Exclude the mural wall and furnishings from the separate sleeping and reading area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon mural and writing remain intact in the abandoned library, with dusty shelves and the makeshift bedding unchanged. 이현우: He remains in the secluded library corner with the disposable phone and the retrieved contact card. His facial bruising and thigh wound persist. 찰리: He has come to the secluded corner and stands quietly there, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 몸을 돌린 이현우의 바로 눈앞에 거대한 찰리가 조용히 서 있는 구도.\n\nLOCATION (lock): In the same secluded corridor corner inside the abandoned library, under subdued morning light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the pan at 이현우's shoulder height, looking past his rear three-quarter shoulder with a slight upward angle toward 찰리, without crossing their established axis. Keep 이현우's interrupted turn in the left foreground quarter and reveal 찰리's upper body just beyond him on the right at close conversational separation, with the corridor corner connecting both depth planes. Emphasize their changed facing relationship: 이현우 freezes while looking at 찰리, and 찰리 quietly inclines his head toward 이현우 rather than addressing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우's turned shoulder in the middle-left of the frame, foreground, looks toward 찰리 immediately beyond his shoulder; 찰리 at close conversational separation in the middle-right of the frame, midground, looks toward 이현우 in the same corridor corner.\n- KEY BACKGROUND ELEMENTS: Library corridor corner (The same secluded area used for the call) — The wall junction continues behind both figures on the established side of their axis; used as Demonstrates that 찰리 is physically beside 이현우 in the same space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corner's subdued morning ambience and unchanged contrast so the revelation comes from spatial disclosure, not a lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the secluded library corner's worn surfaces, dust, and subdued daytime illumination from the reference. Exclude the mural wall and furnishings from the separate sleeping and reading area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon mural and writing remain intact in the abandoned library, with dusty shelves and the makeshift bedding unchanged. 이현우: He remains in the secluded library corner with the disposable phone and the retrieved contact card. His facial bruising and thigh wound persist. 찰리: He has come to the secluded corner and stands quietly there, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 몸을 돌린 이현우의 바로 눈앞에 거대한 찰리가 조용히 서 있는 구도.\n\nLOCATION (lock): In the same secluded corridor corner inside the abandoned library, under subdued morning light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the pan at 이현우's shoulder height, looking past his rear three-quarter shoulder with a slight upward angle toward 찰리, without crossing their established axis. Keep 이현우's interrupted turn in the left foreground quarter and reveal 찰리's upper body just beyond him on the right at close conversational separation, with the corridor corner connecting both depth planes. Emphasize their changed facing relationship: 이현우 freezes while looking at 찰리, and 찰리 quietly inclines his head toward 이현우 rather than addressing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이현우's turned shoulder in the middle-left of the frame, foreground, looks toward 찰리 immediately beyond his shoulder; 찰리 at close conversational separation in the middle-right of the frame, midground, looks toward 이현우 in the same corridor corner.\n- KEY BACKGROUND ELEMENTS: Library corridor corner (The same secluded area used for the call) — The wall junction continues behind both figures on the established side of their axis; used as Demonstrates that 찰리 is physically beside 이현우 in the same space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the corner's subdued morning ambience and unchanged contrast so the revelation comes from spatial disclosure, not a lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the secluded library corner's worn surfaces, dust, and subdued daytime illumination from the reference. Exclude the mural wall and furnishings from the separate sleeping and reading area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The crayon mural and writing remain intact in the abandoned library, with dusty shelves and the makeshift bedding unchanged. 이현우: He remains in the secluded library corner with the disposable phone and the retrieved contact card. His facial bruising and thigh wound persist. 찰리: He has come to the secluded corner and stands quietly there, still with the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 위쪽의 찰리를 향하고, 찰리는 이현우를 향해 고개를 약간 숙이고 있음.",
    "built_space": "왼쪽의 책장과 오른쪽의 낡은 벽면이 있는 도서관 복도. 지시대로 벽화가 제외됨.",
    "entities": "이현우(뒷모습, 낡은 셔츠, 상처)와 찰리(기계 외형, 어깨의 담요) 모두 참조와 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥을 딛고 자연스럽게 서 있음."
   },
   {
    "label": "B",
    "direction": "이현우는 찰리를 바라보고, 찰리 역시 이현우를 마주 보고 있음.",
    "built_space": "폐도서관 복도 구조가 맞으며, 오른쪽 벽면의 벽화가 올바르게 제외됨.",
    "entities": "이현우와 찰리의 외형, 복장, 담요 소품 모두 정확히 반영됨.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 오버 더 숄더 카메라 구도와 찰리가 고개를 숙이는 동작, 벽화 제외 지시를 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구사항을 전반적으로 잘 따랐으나, 찰리의 고개 기울임이 A에 비해 덜 명확함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 위쪽의 찰리를 향하고, 찰리는 이현우를 향해 고개를 약간 숙이고 있음.",
        "built_space": "왼쪽의 책장과 오른쪽의 낡은 벽면이 있는 도서관 복도. 지시대로 벽화가 제외됨.",
        "entities": "이현우(뒷모습, 낡은 셔츠, 상처)와 찰리(기계 외형, 어깨의 담요) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 찰리를 바라보고, 찰리 역시 이현우를 마주 보고 있음.",
        "built_space": "폐도서관 복도 구조가 맞으며, 오른쪽 벽면의 벽화가 올바르게 제외됨.",
        "entities": "이현우와 찰리의 외형, 복장, 담요 소품 모두 정확히 반영됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 오버 더 숄더 카메라 구도와 찰리가 고개를 숙이는 동작, 벽화 제외 지시를 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구사항을 전반적으로 잘 따랐으나, 찰리의 고개 기울임이 A에 비해 덜 명확함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 위쪽의 찰리를 향하고, 찰리는 이현우를 향해 고개를 약간 숙이고 있음.",
        "built_space": "왼쪽의 책장과 오른쪽의 낡은 벽면이 있는 도서관 복도. 지시대로 벽화가 제외됨.",
        "entities": "이현우(뒷모습, 낡은 셔츠, 상처)와 찰리(기계 외형, 어깨의 담요) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 찰리를 바라보고, 찰리 역시 이현우를 마주 보고 있음.",
        "built_space": "폐도서관 복도 구조가 맞으며, 오른쪽 벽면의 벽화가 올바르게 제외됨.",
        "entities": "이현우와 찰리의 외형, 복장, 담요 소품 모두 정확히 반영됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 전경의 뒤쪽 사선 어깨 너머로 가까운 찰리를 보는 미디엄 구도와 상호 대면 방향이 더 정확하지만, 찰리의 작은 체구 설정과 보이는 귀의 인이어는 충분히 반영되지 않았다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "돌아보다 멈춘 동작과 찰리의 고개 기울임은 좋지만, 이현우의 앞가슴을 드러내 지정된 뒤쪽 어깨 시점에서 벗어나고 찰리의 얼굴도 상대보다 카메라 쪽에 더 열려 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 오른쪽의 찰리를 향해 얼굴을 돌리고 있다. 찰리의 흰 얼굴은 왼쪽 아래의 이현우 얼굴 쪽을 향해 있어 가까운 대면 관계가 읽힌다. 발광하는 눈에는 동공이 없어 세밀한 시선은 판별하기 어렵지만 머리 방향은 렌즈보다 이현우를 가리킨다. 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽 벽을 따라 이어지는 책장 열 하나와 인물 사이 뒤쪽 책장 열 하나, 복도 끝 창 하나, 천장에 연속된 선형 조명 약 네 개가 보인다. 오른쪽의 벗겨진 벽과 중앙의 벽 모서리가 두 인물의 공간을 연결한다. 이현우는 왼쪽 전경, 찰리는 바로 오른쪽 중경에 있으며 책장이나 벽과 부딪히지 않는다. 이전 장면의 낡은 도서관 재질과 어두운 아침 자연광을 유지하고 벽화는 노출하지 않는다. 반사상이나 명백히 중복된 고정 설비는 없다.",
        "entities": "등장 개체는 젊은 동아시아계 남성 이현우 한 명과 기계 찰리 한 대뿐이다. 이현우의 헝클어진 검은 머리, 마른 체격, 먼지와 핏자국이 묻은 어두운 셔츠, 옆얼굴의 멍은 참고와 부합한다. 보이는 귀에서는 인이어 무전기를 확인할 수 없다. 찰리는 샌드 베이지 장갑, 육중한 팔, 흰 마스크, 주황색 원형 눈 두 개, 검은 입 선 하나, 파란 가슴 원자로를 갖췄으며 인간 피부나 눈이 추가되지 않았다. 다만 이현우보다 머리가 높고 몸통도 커서 키 작은 성인 남성 정도라는 크기 설정보다 크게 읽힌다. 담요 한 장이 화면 오른쪽 어깨에 걸려 있다. 전화기와 연락처 카드는 손이 잘린 구도에서 확인할 수 없고 허벅지 상처도 프레임 밖이다.",
        "hard_violations": [],
        "physics": "두 인물의 하체와 발은 프레임 밖이므로 바닥 접촉을 직접 확인할 수 없지만, 몸통은 자연스러운 직립 자세로 아래 프레임에 이어지며 공중에 떠 있다는 징후는 없다. 찰리의 팔은 어깨와 팔꿈치 관절에 연결되어 아래로 내려간다. 담요는 어깨 장갑 위에 얹혀 지지되고 옆으로 처진다. 이현우의 목과 어깨 회전도 가능한 범위다."
       },
       {
        "label": "B",
        "direction": "이현우는 몸통을 비튼 채 오른쪽 뒤의 찰리를 바라본다. 찰리는 머리를 화면 왼쪽으로 기울였지만 마스크 전면이 카메라에 비교적 정면으로 드러나, 이현우의 얼굴을 정확히 향하는 관계는 A보다 약하다. 발광 눈에는 동공이 없어 정확한 응시점은 확정할 수 없다. 무기나 이동하는 물체는 없다.",
        "built_space": "왼쪽의 긴 책장 열 하나, 인물 뒤 중앙의 책장 열 하나, 복도 끝 창 하나와 천장의 선형 조명 약 네 개가 보인다. 오른쪽 벗겨진 벽과 그 모서리가 두 인물 뒤로 이어지고, 인물은 가까운 거리에서 같은 복도에 서 있다. 낡은 목재와 회벽, 억제된 아침빛은 참고와 일치하며 벽화는 보이지 않는다. 다만 이현우의 앞가슴과 셔츠 앞여밈이 크게 보여 카메라는 요구된 뒤쪽 사선 어깨 시점보다 몸의 앞쪽에 놓인 인상이다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "이현우 한 명과 찰리 한 대만 보인다. 이현우의 젊은 동아시아계 외형, 헝클어진 검은 머리, 마른 체격과 오염된 어두운 셔츠는 참고에 부합한다. 얼굴이 옆뒤로 가려져 멍의 세부는 확인하기 어렵고 보이는 귀의 인이어도 뚜렷하지 않다. 찰리는 베이지 장갑과 육중한 팔, 흰 마스크, 주황색 눈 두 개와 입 선, 파란 원자로를 유지한다. 머리와 몸통의 상대 크기는 키 작은 성인 남성 정도라는 설정보다 크게 읽힌다. 화면 오른쪽 어깨에는 낡은 담요 한 장이 걸려 있다. 손의 전화기와 연락처 카드, 이현우의 허벅지 상처는 구도상 확인할 수 없다.",
        "hard_violations": [],
        "physics": "이현우는 몸통과 목을 서로 다른 각도로 돌려 동작이 중단된 순간을 자연스럽게 표현한다. 찰리의 기울어진 머리는 목 기계부에 연결되고 팔도 관절에 정상적으로 연결되어 있다. 담요는 어깨 장갑이 받치며 아래로 늘어진다. 두 인물의 발은 잘려 있지만 몸이 아래로 이어져 있으며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 전경의 뒤쪽 사선 어깨 너머로 가까운 찰리를 보는 미디엄 구도와 상호 대면 방향이 더 정확하지만, 찰리의 작은 체구 설정과 보이는 귀의 인이어는 충분히 반영되지 않았다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "돌아보다 멈춘 동작과 찰리의 고개 기울임은 좋지만, 이현우의 앞가슴을 드러내 지정된 뒤쪽 어깨 시점에서 벗어나고 찰리의 얼굴도 상대보다 카메라 쪽에 더 열려 있다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 화면 오른쪽의 찰리를 향해 얼굴을 돌리고 있다. 찰리의 흰 얼굴은 왼쪽 아래의 이현우 얼굴 쪽을 향해 있어 가까운 대면 관계가 읽힌다. 발광하는 눈에는 동공이 없어 세밀한 시선은 판별하기 어렵지만 머리 방향은 렌즈보다 이현우를 가리킨다. 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽 벽을 따라 이어지는 책장 열 하나와 인물 사이 뒤쪽 책장 열 하나, 복도 끝 창 하나, 천장에 연속된 선형 조명 약 네 개가 보인다. 오른쪽의 벗겨진 벽과 중앙의 벽 모서리가 두 인물의 공간을 연결한다. 이현우는 왼쪽 전경, 찰리는 바로 오른쪽 중경에 있으며 책장이나 벽과 부딪히지 않는다. 이전 장면의 낡은 도서관 재질과 어두운 아침 자연광을 유지하고 벽화는 노출하지 않는다. 반사상이나 명백히 중복된 고정 설비는 없다.",
        "entities": "등장 개체는 젊은 동아시아계 남성 이현우 한 명과 기계 찰리 한 대뿐이다. 이현우의 헝클어진 검은 머리, 마른 체격, 먼지와 핏자국이 묻은 어두운 셔츠, 옆얼굴의 멍은 참고와 부합한다. 보이는 귀에서는 인이어 무전기를 확인할 수 없다. 찰리는 샌드 베이지 장갑, 육중한 팔, 흰 마스크, 주황색 원형 눈 두 개, 검은 입 선 하나, 파란 가슴 원자로를 갖췄으며 인간 피부나 눈이 추가되지 않았다. 다만 이현우보다 머리가 높고 몸통도 커서 키 작은 성인 남성 정도라는 크기 설정보다 크게 읽힌다. 담요 한 장이 화면 오른쪽 어깨에 걸려 있다. 전화기와 연락처 카드는 손이 잘린 구도에서 확인할 수 없고 허벅지 상처도 프레임 밖이다.",
        "hard_violations": [],
        "physics": "두 인물의 하체와 발은 프레임 밖이므로 바닥 접촉을 직접 확인할 수 없지만, 몸통은 자연스러운 직립 자세로 아래 프레임에 이어지며 공중에 떠 있다는 징후는 없다. 찰리의 팔은 어깨와 팔꿈치 관절에 연결되어 아래로 내려간다. 담요는 어깨 장갑 위에 얹혀 지지되고 옆으로 처진다. 이현우의 목과 어깨 회전도 가능한 범위다."
       },
       {
        "label": "A",
        "direction": "이현우는 몸통을 비튼 채 오른쪽 뒤의 찰리를 바라본다. 찰리는 머리를 화면 왼쪽으로 기울였지만 마스크 전면이 카메라에 비교적 정면으로 드러나, 이현우의 얼굴을 정확히 향하는 관계는 A보다 약하다. 발광 눈에는 동공이 없어 정확한 응시점은 확정할 수 없다. 무기나 이동하는 물체는 없다.",
        "built_space": "왼쪽의 긴 책장 열 하나, 인물 뒤 중앙의 책장 열 하나, 복도 끝 창 하나와 천장의 선형 조명 약 네 개가 보인다. 오른쪽 벗겨진 벽과 그 모서리가 두 인물 뒤로 이어지고, 인물은 가까운 거리에서 같은 복도에 서 있다. 낡은 목재와 회벽, 억제된 아침빛은 참고와 일치하며 벽화는 보이지 않는다. 다만 이현우의 앞가슴과 셔츠 앞여밈이 크게 보여 카메라는 요구된 뒤쪽 사선 어깨 시점보다 몸의 앞쪽에 놓인 인상이다. 불가능한 반사나 설비 중복은 보이지 않는다.",
        "entities": "이현우 한 명과 찰리 한 대만 보인다. 이현우의 젊은 동아시아계 외형, 헝클어진 검은 머리, 마른 체격과 오염된 어두운 셔츠는 참고에 부합한다. 얼굴이 옆뒤로 가려져 멍의 세부는 확인하기 어렵고 보이는 귀의 인이어도 뚜렷하지 않다. 찰리는 베이지 장갑과 육중한 팔, 흰 마스크, 주황색 눈 두 개와 입 선, 파란 원자로를 유지한다. 머리와 몸통의 상대 크기는 키 작은 성인 남성 정도라는 설정보다 크게 읽힌다. 화면 오른쪽 어깨에는 낡은 담요 한 장이 걸려 있다. 손의 전화기와 연락처 카드, 이현우의 허벅지 상처는 구도상 확인할 수 없다.",
        "hard_violations": [],
        "physics": "이현우는 몸통과 목을 서로 다른 각도로 돌려 동작이 중단된 순간을 자연스럽게 표현한다. 찰리의 기울어진 머리는 목 기계부에 연결되고 팔도 관절에 정상적으로 연결되어 있다. 담요는 어깨 장갑이 받치며 아래로 늘어진다. 두 인물의 발은 잘려 있지만 몸이 아래로 이어져 있으며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지시된 오버 더 숄더 카메라 구도와 찰리가 고개를 숙이는 동작, 벽화 제외 지시를 완벽히 구현함."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "요구사항을 전반적으로 잘 따랐으나, 찰리의 고개 기울임이 A에 비해 덜 명확함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S47sh17_sel.png",
    "asset_id": "00603913-b7c0-4933-a315-e0b5902bfbd2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b34-d01c-71d6-a70a-3e373785d053",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S47sh17"
  }
 },
 "S48sh9::signage": {
  "fp": "c89f6ca0bb10eb07",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh9": {
  "input_fingerprint": "5b901c94f13b936a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지를 일으키며 좁고 가파른 산길 임도 쪽으로 차체가 크게 기울어진 채 방향을 틀고 있는 캠핑카의 외관 전경.\n\nLOCATION (lock): At the turn from the main road onto a steep, narrow dirt forest track, away from the road checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the high endpoint of the crane outside the junction, looking diagonally downward while retaining the camper's rear three-quarter view and full body. Place the sharply leaning vehicle near center at less than a third of the image, turning from the lower-left road toward the upper-right mountain track, with its dust trail remaining behind the turn. Emphasize the vehicle's changing position within the two-route layout, preserve the uphill screen direction for the following cut, and keep occupants out of view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uphill mountain track in the upper-right of the frame, background; Road being left behind in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camper (Turning sharply uphill with its body visibly leaning) — Rear three-quarter view, nose directed toward the upper-right track; used as Makes the abrupt evasive turn readable within the wider terrain; Road and mountain-track junction (The vehicle is leaving the road for the narrow, steep track) — The road crosses the lower-left area while the track climbs toward the upper right; used as Clarifies the choice of route and establishes the next shot's uphill direction; Raised earth dust (Thrown up by the turning vehicle); used as Marks the path already traveled without concealing the route ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast retain the vehicle's shape through the raised dust without sensational color or glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper, still carrying the gathered supplies, turns away from the distant checkpoint toward a mountain dirt road. Barren fields, dead trees and abandoned houses surround the route.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지를 일으키며 좁고 가파른 산길 임도 쪽으로 차체가 크게 기울어진 채 방향을 틀고 있는 캠핑카의 외관 전경.\n\nLOCATION (lock): At the turn from the main road onto a steep, narrow dirt forest track, away from the road checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the high endpoint of the crane outside the junction, looking diagonally downward while retaining the camper's rear three-quarter view and full body. Place the sharply leaning vehicle near center at less than a third of the image, turning from the lower-left road toward the upper-right mountain track, with its dust trail remaining behind the turn. Emphasize the vehicle's changing position within the two-route layout, preserve the uphill screen direction for the following cut, and keep occupants out of view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uphill mountain track in the upper-right of the frame, background; Road being left behind in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camper (Turning sharply uphill with its body visibly leaning) — Rear three-quarter view, nose directed toward the upper-right track; used as Makes the abrupt evasive turn readable within the wider terrain; Road and mountain-track junction (The vehicle is leaving the road for the narrow, steep track) — The road crosses the lower-left area while the track climbs toward the upper right; used as Clarifies the choice of route and establishes the next shot's uphill direction; Raised earth dust (Thrown up by the turning vehicle); used as Marks the path already traveled without concealing the route ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast retain the vehicle's shape through the raised dust without sensational color or glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper, still carrying the gathered supplies, turns away from the distant checkpoint toward a mountain dirt road. Barren fields, dead trees and abandoned houses surround the route.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙먼지를 일으키며 좁고 가파른 산길 임도 쪽으로 차체가 크게 기울어진 채 방향을 틀고 있는 캠핑카의 외관 전경.\n\nLOCATION (lock): At the turn from the main road onto a steep, narrow dirt forest track, away from the road checkpoint. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the high endpoint of the crane outside the junction, looking diagonally downward while retaining the camper's rear three-quarter view and full body. Place the sharply leaning vehicle near center at less than a third of the image, turning from the lower-left road toward the upper-right mountain track, with its dust trail remaining behind the turn. Emphasize the vehicle's changing position within the two-route layout, preserve the uphill screen direction for the following cut, and keep occupants out of view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uphill mountain track in the upper-right of the frame, background; Road being left behind in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Camper (Turning sharply uphill with its body visibly leaning) — Rear three-quarter view, nose directed toward the upper-right track; used as Makes the abrupt evasive turn readable within the wider terrain; Road and mountain-track junction (The vehicle is leaving the road for the narrow, steep track) — The road crosses the lower-left area while the track climbs toward the upper right; used as Clarifies the choice of route and establishes the next shot's uphill direction; Raised earth dust (Thrown up by the turning vehicle); used as Marks the path already traveled without concealing the route ahead.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast retain the vehicle's shape through the raised dust without sensational color or glare.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old camper, still carrying the gathered supplies, turns away from the distant checkpoint toward a mountain dirt road. Barren fields, dead trees and abandoned houses surround the route.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카 기수가 우측 상단 산길을 향하고 있으며, 카메라는 대각선 아래로 향함.",
    "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 있으며, 흙길 바로 옆에 폐가와 마른 나무들이 있음.",
    "entities": "지붕에 짐을 실은 캠핑카, 폐가 2채가 뚜렷이 보이며 인물은 없음.",
    "hard_violations": [],
    "physics": "바퀴가 지면에 닿아 흙먼지를 일으키고 있으나, 차체가 지형에 맞춰 평행하게 이동 중이며 회전 시의 기울어짐은 보이지 않음."
   },
   {
    "label": "B",
    "direction": "캠핑카 기수가 우측 상단의 가파른 산길을 향해 방향을 틀고 있으며, 카메라는 대각선 아래를 향함.",
    "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 만나는 교차로가 묘사되며, 원경에 폐가와 굽이치는 산길이 배치됨.",
    "entities": "지붕에 짐을 실은 낡은 캠핑카, 황량한 들판, 마른 나무와 폐가들이 보이며 인물은 없음.",
    "hard_violations": [
     "[gpt-high] 장소 설명 밖의 요소를 발명하지 말라는 제한에도 교차로에 도로 표지판과 금속 가드레일을 추가했다."
    ],
    "physics": "캠핑카 바퀴가 지면에 닿은 채 산길 진입로의 단차로 인해 차체가 기울어지며 거친 흙먼지를 일으킴."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "포장도로에서 산길로 진입하는 교차로 시점과 흙먼지를 일으키며 방향을 트는 순간을 잘 포착했으며, 요구된 구도와 영화적 사실감을 훌륭하게 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이미 방향을 다 틀고 흙길을 직진 중인 모습으로 렌더링되어, 샷 텍스트가 요구한 '방향을 틀며 차체가 크게 기울어지는' 교차점에서의 역동성이 부족함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 기수가 우측 상단의 가파른 산길을 향해 방향을 틀고 있으며, 카메라는 대각선 아래를 향함.",
        "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 만나는 교차로가 묘사되며, 원경에 폐가와 굽이치는 산길이 배치됨.",
        "entities": "지붕에 짐을 실은 낡은 캠핑카, 황량한 들판, 마른 나무와 폐가들이 보이며 인물은 없음.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 지면에 닿은 채 산길 진입로의 단차로 인해 차체가 기울어지며 거친 흙먼지를 일으킴."
       },
       {
        "label": "A",
        "direction": "캠핑카 기수가 우측 상단 산길을 향하고 있으며, 카메라는 대각선 아래로 향함.",
        "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 있으며, 흙길 바로 옆에 폐가와 마른 나무들이 있음.",
        "entities": "지붕에 짐을 실은 캠핑카, 폐가 2채가 뚜렷이 보이며 인물은 없음.",
        "hard_violations": [],
        "physics": "바퀴가 지면에 닿아 흙먼지를 일으키고 있으나, 차체가 지형에 맞춰 평행하게 이동 중이며 회전 시의 기울어짐은 보이지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "포장도로에서 산길로 진입하는 교차로 시점과 흙먼지를 일으키며 방향을 트는 순간을 잘 포착했으며, 요구된 구도와 영화적 사실감을 훌륭하게 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "이미 방향을 다 틀고 흙길을 직진 중인 모습으로 렌더링되어, 샷 텍스트가 요구한 '방향을 틀며 차체가 크게 기울어지는' 교차점에서의 역동성이 부족함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카 기수가 우측 상단의 가파른 산길을 향해 방향을 틀고 있으며, 카메라는 대각선 아래를 향함.",
        "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 만나는 교차로가 묘사되며, 원경에 폐가와 굽이치는 산길이 배치됨.",
        "entities": "지붕에 짐을 실은 낡은 캠핑카, 황량한 들판, 마른 나무와 폐가들이 보이며 인물은 없음.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 지면에 닿은 채 산길 진입로의 단차로 인해 차체가 기울어지며 거친 흙먼지를 일으킴."
       },
       {
        "label": "A",
        "direction": "캠핑카 기수가 우측 상단 산길을 향하고 있으며, 카메라는 대각선 아래로 향함.",
        "built_space": "좌측 하단의 포장도로와 우측 상단 흙길이 있으며, 흙길 바로 옆에 폐가와 마른 나무들이 있음.",
        "entities": "지붕에 짐을 실은 캠핑카, 폐가 2채가 뚜렷이 보이며 인물은 없음.",
        "hard_violations": [],
        "physics": "바퀴가 지면에 닿아 흙먼지를 일으키고 있으나, 차체가 지형에 맞춰 평행하게 이동 중이며 회전 시의 기울어짐은 보이지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "고각 와이드와 우상향 진입은 맞지만, 지정되지 않은 표지판·가드레일을 추가했고 급회전의 기울기와 흙먼지도 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "작게 배치한 캠핑카의 후면 사선 전경, 크게 기운 차체, 우상향 임도와 회전 뒤에 남은 흙먼지가 요구된 순간을 가장 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 앞부분은 화면 우상단의 오르막 흙길을 향하고, 후면은 좌하단 도로 쪽을 향한다. 먼지는 차량 뒤에서 좌하단으로 이어져 지나온 경로를 표시한다. 사람이나 시선은 보이지 않는다.",
        "built_space": "좌하단의 포장도로 한 줄기와 우상단으로 올라가는 좁은 임도 한 줄기가 연결된다. 교차로 바깥의 높은 위치에서 내려다보며 캠핑카 전체와 후면·옆면을 보여준다. 차량은 중앙 부근에서 화면의 3분의 1보다 작다. 왼쪽에는 폐가 여러 채가 있고, 도로 가장자리에는 금속 가드레일 한 구간과 뒷면이 보이는 표지판 한 기가 추가되어 있다.",
        "entities": "낡고 오염된 캠핑카 한 대, 지붕 적재물, 흙먼지, 황폐한 밭, 잎 없는 나무와 폐가가 보인다. 탑승자나 다른 사람은 없다. 낮의 절제된 색과 실제 금속·흙·암석 질감은 요구에 맞는다. 차량 외형을 대조할 별도 참조 이미지는 없다.",
        "hard_violations": [
         "장소 설명 밖의 요소를 발명하지 말라는 제한에도 교차로에 도로 표지판과 금속 가드레일을 추가했다."
        ],
        "physics": "보이는 앞뒤 바퀴가 흙길에 닿아 차체를 지지하고, 차체는 화면 왼쪽으로 기울어 있다. 지붕 짐은 적재대 위에 놓여 있다. 뒤쪽 바퀴 부근에서 일어난 먼지가 후방으로 퍼지는 것은 주행으로 설명된다. 지지 없이 떠 있는 물체는 없지만, 급격한 회전보다는 이미 임도로 진입한 상태에 가깝게 읽힌다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 코는 우상단으로 이어지는 산길을 향하며 후면은 좌하단 포장도로를 향한다. 흙먼지가 후륜 뒤에서 좌하단 접속부로 퍼지고, 진행할 우상단 길은 가리지 않는다. 보이는 사람이나 시선은 없다.",
        "built_space": "좌하단 전경을 가로지르는 포장도로 한 줄기에서 좁은 흙길 한 줄기가 갈라져 우상단 배경으로 올라간다. 교차로 바깥 고각 시점이며, 중앙보다 약간 오른쪽의 캠핑카 전체를 후면 사선으로 담았다. 차량 크기는 화면의 3분의 1 미만이다. 왼쪽 배경에는 폐가들이 있고 주변에는 황폐한 밭과 죽은 나무가 배치되어 있다.",
        "entities": "낡은 캠핑카 한 대에 지붕 보급품과 후면 적재물이 실려 있다. 거친 흙먼지, 좁은 산길, 포장도로, 폐가, 황폐한 밭과 잎 없는 나무가 모두 보인다. 얼굴이나 탑승자는 드러나지 않는다. 낮의 조명과 억제된 색감, 차량과 지면의 물성도 요청에 부합한다.",
        "hard_violations": [],
        "physics": "보이는 바퀴가 경사진 흙길에 접촉하여 차량을 받치며, 차체 상부가 화면 왼쪽으로 뚜렷하게 기운다. 급회전과 노면 경사로 가능한 자세다. 지붕 짐은 적재대에 놓이고 결속되어 있으며 후면 물품도 거치부에 붙어 있다. 바퀴 뒤의 먼지는 회전하며 흙을 밀어낸 결과로 자연스럽게 이어지고, 차량이나 짐이 공중에 떠 있지 않다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "고각 와이드와 우상향 진입은 맞지만, 지정되지 않은 표지판·가드레일을 추가했고 급회전의 기울기와 흙먼지도 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "작게 배치한 캠핑카의 후면 사선 전경, 크게 기운 차체, 우상향 임도와 회전 뒤에 남은 흙먼지가 요구된 순간을 가장 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카의 앞부분은 화면 우상단의 오르막 흙길을 향하고, 후면은 좌하단 도로 쪽을 향한다. 먼지는 차량 뒤에서 좌하단으로 이어져 지나온 경로를 표시한다. 사람이나 시선은 보이지 않는다.",
        "built_space": "좌하단의 포장도로 한 줄기와 우상단으로 올라가는 좁은 임도 한 줄기가 연결된다. 교차로 바깥의 높은 위치에서 내려다보며 캠핑카 전체와 후면·옆면을 보여준다. 차량은 중앙 부근에서 화면의 3분의 1보다 작다. 왼쪽에는 폐가 여러 채가 있고, 도로 가장자리에는 금속 가드레일 한 구간과 뒷면이 보이는 표지판 한 기가 추가되어 있다.",
        "entities": "낡고 오염된 캠핑카 한 대, 지붕 적재물, 흙먼지, 황폐한 밭, 잎 없는 나무와 폐가가 보인다. 탑승자나 다른 사람은 없다. 낮의 절제된 색과 실제 금속·흙·암석 질감은 요구에 맞는다. 차량 외형을 대조할 별도 참조 이미지는 없다.",
        "hard_violations": [
         "장소 설명 밖의 요소를 발명하지 말라는 제한에도 교차로에 도로 표지판과 금속 가드레일을 추가했다."
        ],
        "physics": "보이는 앞뒤 바퀴가 흙길에 닿아 차체를 지지하고, 차체는 화면 왼쪽으로 기울어 있다. 지붕 짐은 적재대 위에 놓여 있다. 뒤쪽 바퀴 부근에서 일어난 먼지가 후방으로 퍼지는 것은 주행으로 설명된다. 지지 없이 떠 있는 물체는 없지만, 급격한 회전보다는 이미 임도로 진입한 상태에 가깝게 읽힌다."
       },
       {
        "label": "A",
        "direction": "캠핑카의 코는 우상단으로 이어지는 산길을 향하며 후면은 좌하단 포장도로를 향한다. 흙먼지가 후륜 뒤에서 좌하단 접속부로 퍼지고, 진행할 우상단 길은 가리지 않는다. 보이는 사람이나 시선은 없다.",
        "built_space": "좌하단 전경을 가로지르는 포장도로 한 줄기에서 좁은 흙길 한 줄기가 갈라져 우상단 배경으로 올라간다. 교차로 바깥 고각 시점이며, 중앙보다 약간 오른쪽의 캠핑카 전체를 후면 사선으로 담았다. 차량 크기는 화면의 3분의 1 미만이다. 왼쪽 배경에는 폐가들이 있고 주변에는 황폐한 밭과 죽은 나무가 배치되어 있다.",
        "entities": "낡은 캠핑카 한 대에 지붕 보급품과 후면 적재물이 실려 있다. 거친 흙먼지, 좁은 산길, 포장도로, 폐가, 황폐한 밭과 잎 없는 나무가 모두 보인다. 얼굴이나 탑승자는 드러나지 않는다. 낮의 조명과 억제된 색감, 차량과 지면의 물성도 요청에 부합한다.",
        "hard_violations": [],
        "physics": "보이는 바퀴가 경사진 흙길에 접촉하여 차량을 받치며, 차체 상부가 화면 왼쪽으로 뚜렷하게 기운다. 급회전과 노면 경사로 가능한 자세다. 지붕 짐은 적재대에 놓이고 결속되어 있으며 후면 물품도 거치부에 붙어 있다. 바퀴 뒤의 먼지는 회전하며 흙을 밀어낸 결과로 자연스럽게 이어지고, 차량이나 짐이 공중에 떠 있지 않다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.556
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.306
   },
   "violations": {
    "B": [
     "[gpt-high] 장소 설명 밖의 요소를 발명하지 말라는 제한에도 교차로에 도로 표지판과 금속 가드레일을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1306,
   "A": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1306,
    "verdict_ko": "포장도로에서 산길로 진입하는 교차로 시점과 흙먼지를 일으키며 방향을 트는 순간을 잘 포착했으며, 요구된 구도와 영화적 사실감을 훌륭하게 구현함.  ★위반: [gpt-high] 장소 설명 밖의 요소를 발명하지 말라는 제한에도 교차로에 도로 표지판과 금속 가드레일을 추가했다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "이미 방향을 다 틀고 흙길을 직진 중인 모습으로 렌더링되어, 샷 텍스트가 요구한 '방향을 틀며 차체가 크게 기울어지는' 교차점에서의 역동성이 부족함."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b39-e83e-72a3-ae97-344cce4dcf8d",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S48sh24::signage": {
  "fp": "dd25aaa57d8ef8a6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::mountain_track_descent": {
  "input_fingerprint": "db9e913e2394a3d6",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "mountain_track_descent",
    "tags": [
     "S48sh24",
     "S48sh29"
    ]
   },
   "context_sig": "5ae06324aa6e0c29"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n경기도 화성 도로·오염된 들판, 도로 검문소, 산길 임도·캠핑카 정차 지점: 황량한 중금속 오염지대와 경찰 통제선, 그리고 우회로인 비포장 산길이다. (특징: 메마른 나무와 회색빛 토양이 드러난 폐가 주변 들판; 경찰 제복과 방호 장비를 갖춘 인원들이 막아선 도로 검문소; 수풀이 우거진 비포장 흙길과 타이어가 펑크 나 주저앉은 캠핑카; 차체 아래 기어들어가 렌치로 나사를 조이는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산길을 내려가는 앰버와 라울, 찰리와 함께 걸으며\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n경기도 화성 도로·오염된 들판, 도로 검문소, 산길 임도·캠핑카 정차 지점: 황량한 중금속 오염지대와 경찰 통제선, 그리고 우회로인 비포장 산길이다. (특징: 메마른 나무와 회색빛 토양이 드러난 폐가 주변 들판; 경찰 제복과 방호 장비를 갖춘 인원들이 막아선 도로 검문소; 수풀이 우거진 비포장 흙길과 타이어가 펑크 나 주저앉은 캠핑카; 차체 아래 기어들어가 렌치로 나사를 조이는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산길을 내려가는 앰버와 라울, 찰리와 함께 걸으며\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_mountain_track_descent_99532e.png",
  "asset_id": "4f4c9a92-a4c5-4e97-93ba-d9382ad6b1cf",
  "input_asset_ids": [
   "9da1ce80-366e-4825-bdc9-476fd1c169f9"
  ],
  "origin_tag": "S48sh24",
  "place_text": "On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.",
  "origin_inputs": {
   "place_text": "On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.",
   "time_of_day_en": "day",
   "conti_asset_id": "9da1ce80-366e-4825-bdc9-476fd1c169f9"
  }
 },
 "S48sh24::bgfirst_bg": {
  "input_fingerprint": "92c320957fecc306",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차 밖 산길에 각자 짐을 멘 채 서 있는 앰버와 찰리의 전신 구도.\n\nLOCATION (lock): On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the exterior regroup's opening position on the downhill side of the vehicle, with a low, gently upward three-quarter view that includes 앰버 and 찰리 from head to foot. Place 앰버 left of center settling her carried load and 찰리 to the right with his weight braced beneath his belongings, keeping their pauses asymmetrical and a portion of the vehicle behind them for scale. Both attend to the downhill route beyond the left edge without yet stepping off; emphasize their new exterior placement and defer the parallel tracking move until they begin walking.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain track (The group has regrouped beside the vehicle before descending) — Its downhill continuation passes toward the left edge; used as Supports the full figures and prepares the lateral walking axis; Camper exterior (Beside the pair after the vehicle trouble) — A partial side view remains behind the figures without obscuring their silhouettes; used as Establishes that they are now outside and supplies a restrained scale reference; Carried belongings (Borne by 앰버 and 찰리) — Visible against their bodies rather than enlarged in the foreground; used as Makes their preparation to descend legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight provides clear full-body separation and tactile detail without an added change in lighting emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차 밖 산길에 각자 짐을 멘 채 서 있는 앰버와 찰리의 전신 구도.\n\nLOCATION (lock): On the dirt mountain track beside the stopped camper, before the group descends toward the roadside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the exterior regroup's opening position on the downhill side of the vehicle, with a low, gently upward three-quarter view that includes 앰버 and 찰리 from head to foot. Place 앰버 left of center settling her carried load and 찰리 to the right with his weight braced beneath his belongings, keeping their pauses asymmetrical and a portion of the vehicle behind them for scale. Both attend to the downhill route beyond the left edge without yet stepping off; emphasize their new exterior placement and defer the parallel tracking move until they begin walking.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain track (The group has regrouped beside the vehicle before descending) — Its downhill continuation passes toward the left edge; used as Supports the full figures and prepares the lateral walking axis; Camper exterior (Beside the pair after the vehicle trouble) — A partial side view remains behind the figures without obscuring their silhouettes; used as Establishes that they are now outside and supplies a restrained scale reference; Carried belongings (Borne by 앰버 and 찰리) — Visible against their bodies rather than enlarged in the foreground; used as Makes their preparation to descend legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight provides clear full-body separation and tactile detail without an added change in lighting emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S48sh24__bgfirst_bg.png",
  "asset_id": "313b8dce-39a5-4733-84a6-c5f14c80820e",
  "input_asset_ids": [
   "9da1ce80-366e-4825-bdc9-476fd1c169f9",
   "4f4c9a92-a4c5-4e97-93ba-d9382ad6b1cf"
  ]
 },
 "S48sh24": {
  "input_fingerprint": "ceaf675aa068d31d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차 밖 산길에 각자 짐을 멘 채 서 있는 앰버와 찰리의 전신 구도.\n\nLOCATION (lock): On the dirt mountain track beside the stopped camper, before the group descends toward the roadside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the exterior regroup's opening position on the downhill side of the vehicle, with a low, gently upward three-quarter view that includes 앰버 and 찰리 from head to foot. Place 앰버 left of center settling her carried load and 찰리 to the right with his weight braced beneath his belongings, keeping their pauses asymmetrical and a portion of the vehicle behind them for scale. Both attend to the downhill route beyond the left edge without yet stepping off; emphasize their new exterior placement and defer the parallel tracking move until they begin walking.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain track (The group has regrouped beside the vehicle before descending) — Its downhill continuation passes toward the left edge; used as Supports the full figures and prepares the lateral walking axis; Camper exterior (Beside the pair after the vehicle trouble) — A partial side view remains behind the figures without obscuring their silhouettes; used as Establishes that they are now outside and supplies a restrained scale reference; Carried belongings (Borne by 앰버 and 찰리) — Visible against their bodies rather than enlarged in the foreground; used as Makes their preparation to descend legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight provides clear full-body separation and tactile detail without an added change in lighting emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has very little fuel and a punctured tire; its wheel bolts have just been tightened. The route descends from the mountain track toward the roadside. 앰버: She is outside the camper, descending the mountain path on foot. 찰리: He is walking down the mountain path, still with the shop blanket.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버, 찰리 right now, so 앰버, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차 밖 산길에 각자 짐을 멘 채 서 있는 앰버와 찰리의 전신 구도.\n\nLOCATION (lock): On the dirt mountain track beside the stopped camper, before the group descends toward the roadside. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the exterior regroup's opening position on the downhill side of the vehicle, with a low, gently upward three-quarter view that includes 앰버 and 찰리 from head to foot. Place 앰버 left of center settling her carried load and 찰리 to the right with his weight braced beneath his belongings, keeping their pauses asymmetrical and a portion of the vehicle behind them for scale. Both attend to the downhill route beyond the left edge without yet stepping off; emphasize their new exterior placement and defer the parallel tracking move until they begin walking.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain track (The group has regrouped beside the vehicle before descending) — Its downhill continuation passes toward the left edge; used as Supports the full figures and prepares the lateral walking axis; Camper exterior (Beside the pair after the vehicle trouble) — A partial side view remains behind the figures without obscuring their silhouettes; used as Establishes that they are now outside and supplies a restrained scale reference; Carried belongings (Borne by 앰버 and 찰리) — Visible against their bodies rather than enlarged in the foreground; used as Makes their preparation to descend legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight provides clear full-body separation and tactile detail without an added change in lighting emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has very little fuel and a punctured tire; its wheel bolts have just been tightened. The route descends from the mountain track toward the roadside. 앰버: She is outside the camper, descending the mountain path on foot. 찰리: He is walking down the mountain path, still with the shop blanket.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버, 찰리 right now, so 앰버, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차 밖 산길에 각자 짐을 멘 채 서 있는 앰버와 찰리의 전신 구도.\n\nLOCATION (lock): On the dirt mountain track beside the stopped camper, before the group descends toward the roadside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the exterior regroup's opening position on the downhill side of the vehicle, with a low, gently upward three-quarter view that includes 앰버 and 찰리 from head to foot. Place 앰버 left of center settling her carried load and 찰리 to the right with his weight braced beneath his belongings, keeping their pauses asymmetrical and a portion of the vehicle behind them for scale. Both attend to the downhill route beyond the left edge without yet stepping off; emphasize their new exterior placement and defer the parallel tracking move until they begin walking.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain track (The group has regrouped beside the vehicle before descending) — Its downhill continuation passes toward the left edge; used as Supports the full figures and prepares the lateral walking axis; Camper exterior (Beside the pair after the vehicle trouble) — A partial side view remains behind the figures without obscuring their silhouettes; used as Establishes that they are now outside and supplies a restrained scale reference; Carried belongings (Borne by 앰버 and 찰리) — Visible against their bodies rather than enlarged in the foreground; used as Makes their preparation to descend legible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight provides clear full-body separation and tactile detail without an added change in lighting emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper has very little fuel and a punctured tire; its wheel bolts have just been tightened. The route descends from the mountain track toward the roadside. 앰버: She is outside the camper, descending the mountain path on foot. 찰리: He is walking down the mountain path, still with the shop blanket.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버, 찰리 right now, so 앰버, 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버, 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S48sh24__bgfirst_bg.png",
     "asset_id": "313b8dce-39a5-4733-84a6-c5f14c80820e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S48sh24.png",
     "asset_id": "9da1ce80-366e-4825-bdc9-476fd1c169f9",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_mountain_track_descent_99532e.png",
     "asset_id": "4f4c9a92-a4c5-4e97-93ba-d9382ad6b1cf",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "앰버와 찰리 모두 화면 왼쪽 내리막길 방향으로 시선과 자세를 향하고 있습니다.",
    "built_space": "우측에 캠핑카 측면이 위치하고, 뒤쪽으로 산길 로케이션이 정확하게 배치되어 있습니다.",
    "entities": "앰버(목에 걸친 마스크, 작업복)와 찰리(고릴라형 기계, 흰 마스크, 담요) 모두 레퍼런스와 완벽히 일치합니다.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 흙길 위에 안정적으로 서 있으며, 앰버는 가방끈을 쥐고 찰리는 담요를 팔에 끼고 있습니다."
   },
   {
    "label": "B",
    "direction": "두 캐릭터 모두 왼쪽 앞 방향을 바라보며 이동하려는 자세를 취하고 있습니다.",
    "built_space": "우측 캠핑카와 배경의 내리막길 풍경이 지정된 로케이션과 잘 들어맞습니다.",
    "entities": "찰리는 외형이 일치하나, 앰버는 마스크를 얼굴에 착용하고 있어 레퍼런스 얼굴이 보이지 않습니다.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 지면에 발을 딛고 있으며 자세에 무리가 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 마스크 위치가 레퍼런스와 일치하여 얼굴이 잘 드러나며, 두 캐릭터가 짐을 멘 채 화면 왼쪽을 주시하는 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "앰버가 마스크로 얼굴을 가려 레퍼런스의 신원을 확인하기 어렵고, 찰리의 담요 묘사가 다소 부자연스럽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 찰리 모두 화면 왼쪽 내리막길 방향으로 시선과 자세를 향하고 있습니다.",
        "built_space": "우측에 캠핑카 측면이 위치하고, 뒤쪽으로 산길 로케이션이 정확하게 배치되어 있습니다.",
        "entities": "앰버(목에 걸친 마스크, 작업복)와 찰리(고릴라형 기계, 흰 마스크, 담요) 모두 레퍼런스와 완벽히 일치합니다.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 흙길 위에 안정적으로 서 있으며, 앰버는 가방끈을 쥐고 찰리는 담요를 팔에 끼고 있습니다."
       },
       {
        "label": "B",
        "direction": "두 캐릭터 모두 왼쪽 앞 방향을 바라보며 이동하려는 자세를 취하고 있습니다.",
        "built_space": "우측 캠핑카와 배경의 내리막길 풍경이 지정된 로케이션과 잘 들어맞습니다.",
        "entities": "찰리는 외형이 일치하나, 앰버는 마스크를 얼굴에 착용하고 있어 레퍼런스 얼굴이 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 지면에 발을 딛고 있으며 자세에 무리가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 마스크 위치가 레퍼런스와 일치하여 얼굴이 잘 드러나며, 두 캐릭터가 짐을 멘 채 화면 왼쪽을 주시하는 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "앰버가 마스크로 얼굴을 가려 레퍼런스의 신원을 확인하기 어렵고, 찰리의 담요 묘사가 다소 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 찰리 모두 화면 왼쪽 내리막길 방향으로 시선과 자세를 향하고 있습니다.",
        "built_space": "우측에 캠핑카 측면이 위치하고, 뒤쪽으로 산길 로케이션이 정확하게 배치되어 있습니다.",
        "entities": "앰버(목에 걸친 마스크, 작업복)와 찰리(고릴라형 기계, 흰 마스크, 담요) 모두 레퍼런스와 완벽히 일치합니다.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 흙길 위에 안정적으로 서 있으며, 앰버는 가방끈을 쥐고 찰리는 담요를 팔에 끼고 있습니다."
       },
       {
        "label": "B",
        "direction": "두 캐릭터 모두 왼쪽 앞 방향을 바라보며 이동하려는 자세를 취하고 있습니다.",
        "built_space": "우측 캠핑카와 배경의 내리막길 풍경이 지정된 로케이션과 잘 들어맞습니다.",
        "entities": "찰리는 외형이 일치하나, 앰버는 마스크를 얼굴에 착용하고 있어 레퍼런스 얼굴이 보이지 않습니다.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 지면에 발을 딛고 있으며 자세에 무리가 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 사선 앙각의 전신 구도와 왼쪽 앰버·오른쪽 찰리의 비대칭 정지가 더 정확하지만, 찰리의 체격이 지정된 짧고 육중한 비율보다 길쭉하고 햇빛이 강하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장소와 왼쪽을 향한 주의, 담요를 받친 손은 충실하지만, 앙각이 약하고 앰버의 보폭이 이미 내려가기 시작한 순간처럼 보여 출발 전 정지 지시에서 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 눈과 얼굴은 화면 왼쪽 바깥의 먼 방향을 향하며 시선이 약간 높다. 찰리의 마스크도 왼쪽 사선으로 돌아가 있다. 둘 다 카메라를 응시하지 않고 왼쪽 경로에 주의를 두지만, 바로 아래의 노면을 내려다보는 모습은 아니다. 무기나 별도로 겨냥하는 물건은 없다.",
        "built_space": "돌과 흙으로 된 산길이 왼쪽 아래로 이어지고, 왼쪽 계곡에 도로와 작은 시설들이 보인다. 캠퍼 한 대가 두 인물 뒤 오른쪽에 부분적으로 남는다. 펼치지 않은 차양 한 개, 출입문 한 개와 그 창, 문 오른쪽의 좁은 창 한 개, 운전석 창 한 개가 보인다. 뒤쪽 측면과 바퀴는 찰리에 상당 부분 가려지고 앞바퀴는 오른쪽 끝에 걸린다. 앰버는 중앙 왼쪽, 찰리는 오른쪽에서 모두 머리부터 발까지 보이며, 낮은 카메라가 두 인물을 완만하게 올려다본다. 장소의 재료와 계곡 배치는 참조에 부합하고 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "등장 주체는 앰버와 찰리뿐이다. 앰버는 금발과 밝은 피부의 어린 여자아이로, 참조의 혼혈 인상과 대체로 부합하나 얼굴 아래쪽은 방진 마스크에 가려져 있다. 카키 작업복, 가죽 공구 벨트, 부츠, 배낭이 보인다. 찰리는 샌드 베이지 장갑판, 흰 마스크, 주황색 눈 두 개와 입 선, 푸른 원형 흉부 장치를 갖춘 기계이며 인간 피부나 치아가 없다. 다만 짧고 뚱뚱한 성인 크기의 고릴라형보다는 키와 다리가 길게 읽힌다. 배낭과 줄무늬 담요가 몸에 붙어 있다. 캠퍼의 연료량과 타이어 펑크, 볼트 조임 상태는 이 화면으로 확정할 수 없다. 낮은 맞지만 직사광과 그림자가 요구된 차분한 광질보다 강하다.",
        "hard_violations": [],
        "physics": "앰버는 두 부츠를 흙길에 딛고 양손으로 배낭 끈을 잡는다. 배낭은 어깨끈으로 지지된다. 찰리는 벌린 두 발로 지면을 딛고 한 손으로 어깨끈을 잡아 짐 아래에서 버틴다. 담요는 배낭 옆에 걸려 고정된 상태로 읽힌다. 양쪽의 팔 위치와 체중 배분이 달라 출발 전 잠깐 멈춘 자세가 성립한다. 캠퍼는 바퀴로 지면에 지지되며, 근거 없이 공중에 떠 있는 몸이나 짐은 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 얼굴과 눈을 분명하게 화면 왼쪽 바깥의 경로 쪽으로 돌린다. 찰리도 왼쪽 사선을 향한다. 앰버의 앞쪽 발 역시 왼쪽 진행 방향으로 나가 있어 이동 축은 정확하지만, 아직 출발하지 않았다는 순간은 덜 명확하다. 겨냥하는 무기나 도구는 없다.",
        "built_space": "참조와 같은 돌 많은 흙길, 왼쪽 아래 도로와 계곡, 오른쪽 수풀 사면과 캠퍼 한 대가 보인다. 캠퍼에는 접힌 차양 한 개, 출입문 한 개와 그 창, 문 오른쪽의 좁은 창 한 개, 운전석 창 한 개가 보이며, 뒤 바퀴 한 개와 오른쪽 끝의 앞바퀴 일부가 드러난다. 나머지 측면 설비는 인물에 가려져 중복 여부를 문제 삼을 근거가 없다. 앰버는 중앙 왼쪽, 찰리는 오른쪽에 전신으로 배치되고 차량은 뒤에 남는다. 다만 카메라는 A보다 높고 수평에 가까워, 지정된 낮은 앙각이 약하다. 불가능한 반사는 보이지 않는다.",
        "entities": "앰버와 찰리 외에 추가 인물은 없다. 앰버는 참조와 유사한 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 벨트, 부츠와 배낭을 갖췄다. 방진 마스크는 참조처럼 목에 걸려 얼굴이 드러난다. 찰리는 베이지 장갑판과 긴 팔, 흰 얼굴판의 주황색 눈 두 개와 입 선, 푸른 흉부 장치를 유지한다. 다만 지정된 짧고 육중한 체형보다 다리가 길다. 찰리의 배낭과 팔에 안은 줄무늬 담요가 확인된다. 캠퍼의 연료 부족이나 타이어 수리 상태는 확정할 수 없다. 낮 풍경은 맞지만 햇빛과 그림자가 차분한 확산광보다는 강하다.",
        "hard_violations": [],
        "physics": "앰버는 양손으로 어깨끈을 잡고 배낭을 지지하며, 두 부츠 모두 노면에 닿아 있다. 다만 왼쪽으로 길게 내민 발과 넓은 보폭이 정지 중 짐을 추스르기보다 첫걸음에 가깝게 읽힌다. 찰리는 두 발을 벌리고 무릎을 굽혀 체중을 받으며, 말린 담요를 손과 굽힌 팔로 몸에 받친다. 배낭은 어깨끈으로 지지되고 캠퍼는 바퀴로 지면에 서 있다. 지지 없는 물체나 불가능한 신체 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 사선 앙각의 전신 구도와 왼쪽 앰버·오른쪽 찰리의 비대칭 정지가 더 정확하지만, 찰리의 체격이 지정된 짧고 육중한 비율보다 길쭉하고 햇빛이 강하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장소와 왼쪽을 향한 주의, 담요를 받친 손은 충실하지만, 앙각이 약하고 앰버의 보폭이 이미 내려가기 시작한 순간처럼 보여 출발 전 정지 지시에서 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 눈과 얼굴은 화면 왼쪽 바깥의 먼 방향을 향하며 시선이 약간 높다. 찰리의 마스크도 왼쪽 사선으로 돌아가 있다. 둘 다 카메라를 응시하지 않고 왼쪽 경로에 주의를 두지만, 바로 아래의 노면을 내려다보는 모습은 아니다. 무기나 별도로 겨냥하는 물건은 없다.",
        "built_space": "돌과 흙으로 된 산길이 왼쪽 아래로 이어지고, 왼쪽 계곡에 도로와 작은 시설들이 보인다. 캠퍼 한 대가 두 인물 뒤 오른쪽에 부분적으로 남는다. 펼치지 않은 차양 한 개, 출입문 한 개와 그 창, 문 오른쪽의 좁은 창 한 개, 운전석 창 한 개가 보인다. 뒤쪽 측면과 바퀴는 찰리에 상당 부분 가려지고 앞바퀴는 오른쪽 끝에 걸린다. 앰버는 중앙 왼쪽, 찰리는 오른쪽에서 모두 머리부터 발까지 보이며, 낮은 카메라가 두 인물을 완만하게 올려다본다. 장소의 재료와 계곡 배치는 참조에 부합하고 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "등장 주체는 앰버와 찰리뿐이다. 앰버는 금발과 밝은 피부의 어린 여자아이로, 참조의 혼혈 인상과 대체로 부합하나 얼굴 아래쪽은 방진 마스크에 가려져 있다. 카키 작업복, 가죽 공구 벨트, 부츠, 배낭이 보인다. 찰리는 샌드 베이지 장갑판, 흰 마스크, 주황색 눈 두 개와 입 선, 푸른 원형 흉부 장치를 갖춘 기계이며 인간 피부나 치아가 없다. 다만 짧고 뚱뚱한 성인 크기의 고릴라형보다는 키와 다리가 길게 읽힌다. 배낭과 줄무늬 담요가 몸에 붙어 있다. 캠퍼의 연료량과 타이어 펑크, 볼트 조임 상태는 이 화면으로 확정할 수 없다. 낮은 맞지만 직사광과 그림자가 요구된 차분한 광질보다 강하다.",
        "hard_violations": [],
        "physics": "앰버는 두 부츠를 흙길에 딛고 양손으로 배낭 끈을 잡는다. 배낭은 어깨끈으로 지지된다. 찰리는 벌린 두 발로 지면을 딛고 한 손으로 어깨끈을 잡아 짐 아래에서 버틴다. 담요는 배낭 옆에 걸려 고정된 상태로 읽힌다. 양쪽의 팔 위치와 체중 배분이 달라 출발 전 잠깐 멈춘 자세가 성립한다. 캠퍼는 바퀴로 지면에 지지되며, 근거 없이 공중에 떠 있는 몸이나 짐은 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 얼굴과 눈을 분명하게 화면 왼쪽 바깥의 경로 쪽으로 돌린다. 찰리도 왼쪽 사선을 향한다. 앰버의 앞쪽 발 역시 왼쪽 진행 방향으로 나가 있어 이동 축은 정확하지만, 아직 출발하지 않았다는 순간은 덜 명확하다. 겨냥하는 무기나 도구는 없다.",
        "built_space": "참조와 같은 돌 많은 흙길, 왼쪽 아래 도로와 계곡, 오른쪽 수풀 사면과 캠퍼 한 대가 보인다. 캠퍼에는 접힌 차양 한 개, 출입문 한 개와 그 창, 문 오른쪽의 좁은 창 한 개, 운전석 창 한 개가 보이며, 뒤 바퀴 한 개와 오른쪽 끝의 앞바퀴 일부가 드러난다. 나머지 측면 설비는 인물에 가려져 중복 여부를 문제 삼을 근거가 없다. 앰버는 중앙 왼쪽, 찰리는 오른쪽에 전신으로 배치되고 차량은 뒤에 남는다. 다만 카메라는 A보다 높고 수평에 가까워, 지정된 낮은 앙각이 약하다. 불가능한 반사는 보이지 않는다.",
        "entities": "앰버와 찰리 외에 추가 인물은 없다. 앰버는 참조와 유사한 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 벨트, 부츠와 배낭을 갖췄다. 방진 마스크는 참조처럼 목에 걸려 얼굴이 드러난다. 찰리는 베이지 장갑판과 긴 팔, 흰 얼굴판의 주황색 눈 두 개와 입 선, 푸른 흉부 장치를 유지한다. 다만 지정된 짧고 육중한 체형보다 다리가 길다. 찰리의 배낭과 팔에 안은 줄무늬 담요가 확인된다. 캠퍼의 연료 부족이나 타이어 수리 상태는 확정할 수 없다. 낮 풍경은 맞지만 햇빛과 그림자가 차분한 확산광보다는 강하다.",
        "hard_violations": [],
        "physics": "앰버는 양손으로 어깨끈을 잡고 배낭을 지지하며, 두 부츠 모두 노면에 닿아 있다. 다만 왼쪽으로 길게 내민 발과 넓은 보폭이 정지 중 짐을 추스르기보다 첫걸음에 가깝게 읽힌다. 찰리는 두 발을 벌리고 무릎을 굽혀 체중을 받으며, 말린 담요를 손과 굽힌 팔로 몸에 받친다. 배낭은 어깨끈으로 지지되고 캠퍼는 바퀴로 지면에 서 있다. 지지 없는 물체나 불가능한 신체 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "앰버의 마스크 위치가 레퍼런스와 일치하여 얼굴이 잘 드러나며, 두 캐릭터가 짐을 멘 채 화면 왼쪽을 주시하는 구도를 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "앰버가 마스크로 얼굴을 가려 레퍼런스의 신원을 확인하기 어렵고, 찰리의 담요 묘사가 다소 부자연스럽습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_mountain_track_descent_99532e.png",
    "asset_id": "4f4c9a92-a4c5-4e97-93ba-d9382ad6b1cf",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b3e-f040-780f-ac6c-b3b51285a087",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S48sh24__bgfirst_bg.png",
   "bg_asset_id": "313b8dce-39a5-4733-84a6-c5f14c80820e",
   "bg_record_key": "S48sh24::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "mountain_track_descent",
   "groupbg_asset_id": "4f4c9a92-a4c5-4e97-93ba-d9382ad6b1cf"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S48sh29::signage": {
  "fp": "22a3444c1c28a05b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh29": {
  "input_fingerprint": "96cf0fff3d08032c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손을 활짝 펴 찰리의 거대한 금속 손 전체를 따뜻하게 감싸 쥐고 있는 앰버의 찰나.\n\nLOCATION (lock): On the mountain path descending toward the road, beside the route taken by the damaged camper. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move on the established lateral side, positioning the lens just above the joined hands and looking down obliquely; camera distance alone carries the emphasis into the promise. 앰버's two hands enter from the left to enclose 찰리's larger metal hand extending from the right, with the clasp centered and their forearms and small torso margins preserving believable scale. Keep their faces outside the crop and their attention directed down toward the shared grip, holding the moment after contact without adding a squeeze, shake, or other new action.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain track beneath their hands (The same descent route); used as A softly resolved contextual strip around the forearms prevents the clasp from becoming a disconnected object study.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight and gentle contrast, letting the contact between human hands and precisely rendered metal communicate warmth without a color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt path, surrounding hillside, and daylight colors from the reference. Exclude checkpoint barriers and road fixtures from the separate main-road location.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains low on fuel with a punctured tire and recently tightened wheel bolts. The mountain path leads down toward the road. 앰버: She is on the mountain path, using her whole hand rather than a finger for the promise gesture. 찰리: He is on the mountain path with a hand extended, retaining the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손을 활짝 펴 찰리의 거대한 금속 손 전체를 따뜻하게 감싸 쥐고 있는 앰버의 찰나.\n\nLOCATION (lock): On the mountain path descending toward the road, beside the route taken by the damaged camper. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move on the established lateral side, positioning the lens just above the joined hands and looking down obliquely; camera distance alone carries the emphasis into the promise. 앰버's two hands enter from the left to enclose 찰리's larger metal hand extending from the right, with the clasp centered and their forearms and small torso margins preserving believable scale. Keep their faces outside the crop and their attention directed down toward the shared grip, holding the moment after contact without adding a squeeze, shake, or other new action.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain track beneath their hands (The same descent route); used as A softly resolved contextual strip around the forearms prevents the clasp from becoming a disconnected object study.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight and gentle contrast, letting the contact between human hands and precisely rendered metal communicate warmth without a color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt path, surrounding hillside, and daylight colors from the reference. Exclude checkpoint barriers and road fixtures from the separate main-road location.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains low on fuel with a punctured tire and recently tightened wheel bolts. The mountain path leads down toward the road. 앰버: She is on the mountain path, using her whole hand rather than a finger for the promise gesture. 찰리: He is on the mountain path with a hand extended, retaining the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손을 활짝 펴 찰리의 거대한 금속 손 전체를 따뜻하게 감싸 쥐고 있는 앰버의 찰나.\n\nLOCATION (lock): On the mountain path descending toward the road, beside the route taken by the damaged camper. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward move on the established lateral side, positioning the lens just above the joined hands and looking down obliquely; camera distance alone carries the emphasis into the promise. 앰버's two hands enter from the left to enclose 찰리's larger metal hand extending from the right, with the clasp centered and their forearms and small torso margins preserving believable scale. Keep their faces outside the crop and their attention directed down toward the shared grip, holding the moment after contact without adding a squeeze, shake, or other new action.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Mountain track beneath their hands (The same descent route); used as A softly resolved contextual strip around the forearms prevents the clasp from becoming a disconnected object study.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight and gentle contrast, letting the contact between human hands and precisely rendered metal communicate warmth without a color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt path, surrounding hillside, and daylight colors from the reference. Exclude checkpoint barriers and road fixtures from the separate main-road location.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains low on fuel with a punctured tire and recently tightened wheel bolts. The mountain path leads down toward the road. 앰버: She is on the mountain path, using her whole hand rather than a finger for the promise gesture. 찰리: He is on the mountain path with a hand extended, retaining the shop blanket.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "B",
    "direction": "앰버의 두 팔은 왼쪽에서 중앙으로, 찰리의 금속 팔은 오른쪽 위에서 중앙 아래로 향한다. 두 인간 손이 금속 손의 위쪽 가장자리와 아래쪽 손가락 부분을 감싼다. 얼굴과 눈은 잘려 있어 실제 시선 방향은 확인할 수 없다.",
    "built_space": "손 아래와 팔 사이에 흙, 자갈, 작은 바위로 된 산길이 이어진다. 인공 구조물과 도로 시설은 보이지 않아 제외 지시를 따른다. 카메라는 맞잡은 손 위에서 비스듬히 내려다본다. 손은 중앙 부근에 크게 놓이지만 앰버의 몸통과 허벅지, 찰리의 몸체가 차지하는 가장자리 면적은 작은 몸통 여백이라는 지시보다 넓다.",
    "entities": "앰버에 해당하는 밝은 피부의 작은 손 두 개, 금발 일부, 때 묻은 카키 작업복, 목에 걸린 방진 마스크와 가죽 공구 벨트가 보인다. 손과 팔은 어린아이의 체격에 부합하지만 얼굴이 없어 정확한 나이와 혼혈 정체성은 검증할 수 없다. 찰리는 마모된 샌드 베이지 장갑판과 관절을 갖춘 큰 기계 손과 팔로 표현된다. 오른쪽에는 참고 이미지와 유사한 말린 줄무늬 담요가 있다. 캠퍼와 얼굴의 세부는 구도 밖이며 추가 인물이나 문자는 없다.",
    "hard_violations": [],
    "physics": "앰버의 두 손은 각각 손목과 팔에 자연스럽게 연결되고 금속 손 표면에 접촉한다. 찰리의 손도 손목 관절과 장갑 팔에 연결되어 지지된다. 담요는 오른쪽 몸체와 팔 쪽에 밀착되어 있으며 공중에 따로 떠 있지 않다. 흔들거나 새로 움켜쥐는 동작보다 접촉 후 정지한 순간으로 읽힌다."
   },
   {
    "label": "A",
    "direction": "앰버의 두 손이 왼쪽에서 들어와 오른쪽에서 뻗은 찰리의 손 양옆을 감싼다. 앰버의 고개는 맞잡은 손 쪽으로 숙여져 있지만 눈은 보이지 않는다. 찰리의 얼굴은 구도 밖이다. 손의 진입 방향과 접촉 대상은 지시와 맞는다.",
    "built_space": "배경은 자갈과 흙, 마른 풀로 된 산길이며 도로 시설이나 인공 구조물은 없다. 손 위에서 비스듬히 내려다보는 시점이고 접촉 지점도 중앙 부근이다. 다만 왼쪽 위에 앰버의 코와 볼을 포함한 얼굴 일부가 들어와 얼굴을 제외하라는 지시를 어긴다. 양옆 몸통과 찰리의 가슴 일부까지 보여 여백도 넓다.",
    "entities": "앰버의 금발, 어린 얼굴 일부, 밝은 피부의 손 두 개, 카키 작업복, 목의 방진 마스크와 가죽 공구 벨트가 참고 외형에 부합한다. 부분 얼굴만으로 정확한 혼혈 정체성은 단정할 수 없다. 찰리의 큰 기계 손, 샌드 베이지 장갑 팔, 오른쪽 위의 푸른 원자로 일부와 오른쪽의 말린 줄무늬 담요가 보인다. 금속 손은 크지만 앰버의 손이 주로 양옆을 잡아 손 전체를 넓게 감싸는 인상은 상대적으로 약하다. 추가 인물이나 문자는 없다.",
    "hard_violations": [],
    "physics": "두 인간 손은 팔과 정상적으로 이어져 금속 손의 양옆에 닿아 있고, 기계 손은 손목과 전완 관절에 연결되어 있다. 담요는 찰리의 몸과 반대쪽 팔 부근에 붙어 지지되는 상태로 보인다. 부유하는 물체나 불가능한 관절 자세는 없으며 접촉을 유지하는 정적인 순간으로 읽힌다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 사선 하향 클로즈업에서 왼쪽의 두 손이 오른쪽의 거대한 금속 손을 넓게 감싸 핵심 동작을 잘 구현하나, 몸통 여백은 지정보다 넓다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "손의 진입 방향과 접촉은 맞지만 앰버의 얼굴 일부가 프레임에 들어오며, 두 손을 활짝 펴 금속 손 전체를 감싸는 모습도 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 두 팔은 왼쪽에서 중앙으로, 찰리의 금속 팔은 오른쪽 위에서 중앙 아래로 향한다. 두 인간 손이 금속 손의 위쪽 가장자리와 아래쪽 손가락 부분을 감싼다. 얼굴과 눈은 잘려 있어 실제 시선 방향은 확인할 수 없다.",
        "built_space": "손 아래와 팔 사이에 흙, 자갈, 작은 바위로 된 산길이 이어진다. 인공 구조물과 도로 시설은 보이지 않아 제외 지시를 따른다. 카메라는 맞잡은 손 위에서 비스듬히 내려다본다. 손은 중앙 부근에 크게 놓이지만 앰버의 몸통과 허벅지, 찰리의 몸체가 차지하는 가장자리 면적은 작은 몸통 여백이라는 지시보다 넓다.",
        "entities": "앰버에 해당하는 밝은 피부의 작은 손 두 개, 금발 일부, 때 묻은 카키 작업복, 목에 걸린 방진 마스크와 가죽 공구 벨트가 보인다. 손과 팔은 어린아이의 체격에 부합하지만 얼굴이 없어 정확한 나이와 혼혈 정체성은 검증할 수 없다. 찰리는 마모된 샌드 베이지 장갑판과 관절을 갖춘 큰 기계 손과 팔로 표현된다. 오른쪽에는 참고 이미지와 유사한 말린 줄무늬 담요가 있다. 캠퍼와 얼굴의 세부는 구도 밖이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 두 손은 각각 손목과 팔에 자연스럽게 연결되고 금속 손 표면에 접촉한다. 찰리의 손도 손목 관절과 장갑 팔에 연결되어 지지된다. 담요는 오른쪽 몸체와 팔 쪽에 밀착되어 있으며 공중에 따로 떠 있지 않다. 흔들거나 새로 움켜쥐는 동작보다 접촉 후 정지한 순간으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "앰버의 두 손이 왼쪽에서 들어와 오른쪽에서 뻗은 찰리의 손 양옆을 감싼다. 앰버의 고개는 맞잡은 손 쪽으로 숙여져 있지만 눈은 보이지 않는다. 찰리의 얼굴은 구도 밖이다. 손의 진입 방향과 접촉 대상은 지시와 맞는다.",
        "built_space": "배경은 자갈과 흙, 마른 풀로 된 산길이며 도로 시설이나 인공 구조물은 없다. 손 위에서 비스듬히 내려다보는 시점이고 접촉 지점도 중앙 부근이다. 다만 왼쪽 위에 앰버의 코와 볼을 포함한 얼굴 일부가 들어와 얼굴을 제외하라는 지시를 어긴다. 양옆 몸통과 찰리의 가슴 일부까지 보여 여백도 넓다.",
        "entities": "앰버의 금발, 어린 얼굴 일부, 밝은 피부의 손 두 개, 카키 작업복, 목의 방진 마스크와 가죽 공구 벨트가 참고 외형에 부합한다. 부분 얼굴만으로 정확한 혼혈 정체성은 단정할 수 없다. 찰리의 큰 기계 손, 샌드 베이지 장갑 팔, 오른쪽 위의 푸른 원자로 일부와 오른쪽의 말린 줄무늬 담요가 보인다. 금속 손은 크지만 앰버의 손이 주로 양옆을 잡아 손 전체를 넓게 감싸는 인상은 상대적으로 약하다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 인간 손은 팔과 정상적으로 이어져 금속 손의 양옆에 닿아 있고, 기계 손은 손목과 전완 관절에 연결되어 있다. 담요는 찰리의 몸과 반대쪽 팔 부근에 붙어 지지되는 상태로 보인다. 부유하는 물체나 불가능한 관절 자세는 없으며 접촉을 유지하는 정적인 순간으로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 사선 하향 클로즈업에서 왼쪽의 두 손이 오른쪽의 거대한 금속 손을 넓게 감싸 핵심 동작을 잘 구현하나, 몸통 여백은 지정보다 넓다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "손의 진입 방향과 접촉은 맞지만 앰버의 얼굴 일부가 프레임에 들어오며, 두 손을 활짝 펴 금속 손 전체를 감싸는 모습도 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 두 팔은 왼쪽에서 중앙으로, 찰리의 금속 팔은 오른쪽 위에서 중앙 아래로 향한다. 두 인간 손이 금속 손의 위쪽 가장자리와 아래쪽 손가락 부분을 감싼다. 얼굴과 눈은 잘려 있어 실제 시선 방향은 확인할 수 없다.",
        "built_space": "손 아래와 팔 사이에 흙, 자갈, 작은 바위로 된 산길이 이어진다. 인공 구조물과 도로 시설은 보이지 않아 제외 지시를 따른다. 카메라는 맞잡은 손 위에서 비스듬히 내려다본다. 손은 중앙 부근에 크게 놓이지만 앰버의 몸통과 허벅지, 찰리의 몸체가 차지하는 가장자리 면적은 작은 몸통 여백이라는 지시보다 넓다.",
        "entities": "앰버에 해당하는 밝은 피부의 작은 손 두 개, 금발 일부, 때 묻은 카키 작업복, 목에 걸린 방진 마스크와 가죽 공구 벨트가 보인다. 손과 팔은 어린아이의 체격에 부합하지만 얼굴이 없어 정확한 나이와 혼혈 정체성은 검증할 수 없다. 찰리는 마모된 샌드 베이지 장갑판과 관절을 갖춘 큰 기계 손과 팔로 표현된다. 오른쪽에는 참고 이미지와 유사한 말린 줄무늬 담요가 있다. 캠퍼와 얼굴의 세부는 구도 밖이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "앰버의 두 손은 각각 손목과 팔에 자연스럽게 연결되고 금속 손 표면에 접촉한다. 찰리의 손도 손목 관절과 장갑 팔에 연결되어 지지된다. 담요는 오른쪽 몸체와 팔 쪽에 밀착되어 있으며 공중에 따로 떠 있지 않다. 흔들거나 새로 움켜쥐는 동작보다 접촉 후 정지한 순간으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "앰버의 두 손이 왼쪽에서 들어와 오른쪽에서 뻗은 찰리의 손 양옆을 감싼다. 앰버의 고개는 맞잡은 손 쪽으로 숙여져 있지만 눈은 보이지 않는다. 찰리의 얼굴은 구도 밖이다. 손의 진입 방향과 접촉 대상은 지시와 맞는다.",
        "built_space": "배경은 자갈과 흙, 마른 풀로 된 산길이며 도로 시설이나 인공 구조물은 없다. 손 위에서 비스듬히 내려다보는 시점이고 접촉 지점도 중앙 부근이다. 다만 왼쪽 위에 앰버의 코와 볼을 포함한 얼굴 일부가 들어와 얼굴을 제외하라는 지시를 어긴다. 양옆 몸통과 찰리의 가슴 일부까지 보여 여백도 넓다.",
        "entities": "앰버의 금발, 어린 얼굴 일부, 밝은 피부의 손 두 개, 카키 작업복, 목의 방진 마스크와 가죽 공구 벨트가 참고 외형에 부합한다. 부분 얼굴만으로 정확한 혼혈 정체성은 단정할 수 없다. 찰리의 큰 기계 손, 샌드 베이지 장갑 팔, 오른쪽 위의 푸른 원자로 일부와 오른쪽의 말린 줄무늬 담요가 보인다. 금속 손은 크지만 앰버의 손이 주로 양옆을 잡아 손 전체를 넓게 감싸는 인상은 상대적으로 약하다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 인간 손은 팔과 정상적으로 이어져 금속 손의 양옆에 닿아 있고, 기계 손은 손목과 전완 관절에 연결되어 있다. 담요는 찰리의 몸과 반대쪽 팔 부근에 붙어 지지되는 상태로 보인다. 부유하는 물체나 불가능한 관절 자세는 없으며 접촉을 유지하는 정적인 순간으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 9,
   "A": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "얼굴을 제외한 사선 하향 클로즈업에서 왼쪽의 두 손이 오른쪽의 거대한 금속 손을 넓게 감싸 핵심 동작을 잘 구현하나, 몸통 여백은 지정보다 넓다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "손의 진입 방향과 접촉은 맞지만 앰버의 얼굴 일부가 프레임에 들어오며, 두 손을 활짝 펴 금속 손 전체를 감싸는 모습도 A보다 약하다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 앰버, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S48sh24_sel.png",
    "asset_id": "c847d557-d0d9-4e8d-98d3-754c0e01c925",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b47-ea2d-7106-b23e-cbda865d6e8a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh24"
  }
 },
 "S49sh16::signage": {
  "fp": "d478e4d81bf4e591",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b6ef72882b262890": {
  "subjects": [],
  "subject_text": "휴게소 주유소 주유 구역\n도로에서 진입하는 야외 주유 공간. 주유기와 호스, 숫자 표시 계량기가 있으며, 안쪽으로 마트 출입구가 보인다.",
  "identity": "canonical",
  "scope_id": "L67",
  "scope_role": "location_exterior",
  "scope_sha": "18428e216e7a1333"
 },
 "S49sh16::bgfirst_bg": {
  "input_fingerprint": "7c16f8019322cbba",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리, 앰버, 라울을 향해 샷건의 총구를 정면으로 들이민 조일광의 위협적인 구도.\n\nLOCATION (lock): Inside the service-station convenience store, at the medicine aisle facing the armed proprietor. Daylight comes through the storefront.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track behind and beside 찰리 at his lower shoulder height, looking slightly upward past his shoulder toward 조일광 in the center-right midground. Keep 찰리 as a narrow left-edge foreground presence, with 앰버 and 라울 partially visible beside him, their heads lifted toward 조일광 while his attention angles down toward the group. The shotgun runs diagonally toward the group rather than into the lens; emphasize their redirected attention without changing the established distance or lighting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Medicine shelving (Stocked with medicine containers) — The aisle-facing shelves recede beside the group; used as Establish the interrupted medicine search and constrain the confrontation spatially; Shotgun (Raised and aimed at the group) — Seen obliquely from beside its firing axis, with the muzzle directed away from the lens toward the foreground group; used as Connect the threatening figure to the intended targets without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable expressions and precise robot contours without stylizing the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리, 앰버, 라울을 향해 샷건의 총구를 정면으로 들이민 조일광의 위협적인 구도.\n\nLOCATION (lock): Inside the service-station convenience store, at the medicine aisle facing the armed proprietor. Daylight comes through the storefront.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track behind and beside 찰리 at his lower shoulder height, looking slightly upward past his shoulder toward 조일광 in the center-right midground. Keep 찰리 as a narrow left-edge foreground presence, with 앰버 and 라울 partially visible beside him, their heads lifted toward 조일광 while his attention angles down toward the group. The shotgun runs diagonally toward the group rather than into the lens; emphasize their redirected attention without changing the established distance or lighting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Medicine shelving (Stocked with medicine containers) — The aisle-facing shelves recede beside the group; used as Establish the interrupted medicine search and constrain the confrontation spatially; Shotgun (Raised and aimed at the group) — Seen obliquely from beside its firing axis, with the muzzle directed away from the lens toward the foreground group; used as Connect the threatening figure to the intended targets without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable expressions and precise robot contours without stylizing the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh16__bgfirst_bg.png",
  "asset_id": "8db851c9-1539-46b8-9587-f6cc72e980e1",
  "input_asset_ids": [
   "16aa5388-fb84-4348-bd24-c24ddc8d6fed",
   "31ae9965-7777-439a-9c7b-3bc6c7cd5b19"
  ]
 },
 "S49sh16": {
  "input_fingerprint": "125d0b7992f5960f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리, 앰버, 라울을 향해 샷건의 총구를 정면으로 들이민 조일광의 위협적인 구도.\n\nLOCATION (lock): Inside the service-station convenience store, at the medicine aisle facing the armed proprietor. Daylight comes through the storefront. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track behind and beside 찰리 at his lower shoulder height, looking slightly upward past his shoulder toward 조일광 in the center-right midground. Keep 찰리 as a narrow left-edge foreground presence, with 앰버 and 라울 partially visible beside him, their heads lifted toward 조일광 while his attention angles down toward the group. The shotgun runs diagonally toward the group rather than into the lens; emphasize their redirected attention without changing the established distance or lighting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Medicine shelving (Stocked with medicine containers) — The aisle-facing shelves recede beside the group; used as Establish the interrupted medicine search and constrain the confrontation spatially; Shotgun (Raised and aimed at the group) — Seen obliquely from beside its firing axis, with the muzzle directed away from the lens toward the foreground group; used as Connect the threatening figure to the intended targets without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable expressions and precise robot contours without stylizing the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop's television has displayed wanted images of the youths and robot, and a corner mirror overlooks the aisles. Outside, the camper is at the fuel pump with its luggage loaded back aboard. 조일광: He wears a military uniform, Marine Corps cap and medals, and holds a loaded shotgun at the ready. 찰리: He is at the medicine section holding the medicine bottle he selected, still with the shop blanket. 앰버: She is inside the shop at the medicine section. 라울: He is inside the shop at the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조일광 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리, 앰버, 라울을 향해 샷건의 총구를 정면으로 들이민 조일광의 위협적인 구도.\n\nLOCATION (lock): Inside the service-station convenience store, at the medicine aisle facing the armed proprietor. Daylight comes through the storefront. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track behind and beside 찰리 at his lower shoulder height, looking slightly upward past his shoulder toward 조일광 in the center-right midground. Keep 찰리 as a narrow left-edge foreground presence, with 앰버 and 라울 partially visible beside him, their heads lifted toward 조일광 while his attention angles down toward the group. The shotgun runs diagonally toward the group rather than into the lens; emphasize their redirected attention without changing the established distance or lighting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Medicine shelving (Stocked with medicine containers) — The aisle-facing shelves recede beside the group; used as Establish the interrupted medicine search and constrain the confrontation spatially; Shotgun (Raised and aimed at the group) — Seen obliquely from beside its firing axis, with the muzzle directed away from the lens toward the foreground group; used as Connect the threatening figure to the intended targets without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable expressions and precise robot contours without stylizing the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop's television has displayed wanted images of the youths and robot, and a corner mirror overlooks the aisles. Outside, the camper is at the fuel pump with its luggage loaded back aboard. 조일광: He wears a military uniform, Marine Corps cap and medals, and holds a loaded shotgun at the ready. 찰리: He is at the medicine section holding the medicine bottle he selected, still with the shop blanket. 앰버: She is inside the shop at the medicine section. 라울: He is inside the shop at the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조일광 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리, 앰버, 라울을 향해 샷건의 총구를 정면으로 들이민 조일광의 위협적인 구도.\n\nLOCATION (lock): Inside the service-station convenience store, at the medicine aisle facing the armed proprietor. Daylight comes through the storefront. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track behind and beside 찰리 at his lower shoulder height, looking slightly upward past his shoulder toward 조일광 in the center-right midground. Keep 찰리 as a narrow left-edge foreground presence, with 앰버 and 라울 partially visible beside him, their heads lifted toward 조일광 while his attention angles down toward the group. The shotgun runs diagonally toward the group rather than into the lens; emphasize their redirected attention without changing the established distance or lighting.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Medicine shelving (Stocked with medicine containers) — The aisle-facing shelves recede beside the group; used as Establish the interrupted medicine search and constrain the confrontation spatially; Shotgun (Raised and aimed at the group) — Seen obliquely from beside its firing axis, with the muzzle directed away from the lens toward the foreground group; used as Connect the threatening figure to the intended targets without exaggerated muzzle perspective.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve readable expressions and precise robot contours without stylizing the threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shop's television has displayed wanted images of the youths and robot, and a corner mirror overlooks the aisles. Outside, the camper is at the fuel pump with its luggage loaded back aboard. 조일광: He wears a military uniform, Marine Corps cap and medals, and holds a loaded shotgun at the ready. 찰리: He is at the medicine section holding the medicine bottle he selected, still with the shop blanket. 앰버: She is inside the shop at the medicine section. 라울: He is inside the shop at the medicine section.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조일광 right now, so 조일광's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조일광: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조일광 (한국인 남성, 50대의 얼굴, 짧은 검은 머리); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh16__bgfirst_bg.png",
     "asset_id": "8db851c9-1539-46b8-9587-f6cc72e980e1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S49sh16.png",
     "asset_id": "16aa5388-fb84-4348-bd24-c24ddc8d6fed",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 조일광: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1399968>",
     "asset_id": "d81b8d5c-ffe5-4291-a480-6fbe062a1a8c",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1305657>",
     "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B02.png",
     "asset_id": "31ae9965-7777-439a-9c7b-3bc6c7cd5b19",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 조일광: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1399968>",
     "asset_id": "d81b8d5c-ffe5-4291-a480-6fbe062a1a8c",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1305657>",
     "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "인물 간 시선은 서로를 향해 있으나, 샷건이 아예 묘사되지 않아 지시된 총구 방향을 확인할 수 없음.",
    "built_space": "편의점 약품 코너. 창밖 채광과 내부 진열대 구조가 레퍼런스와 일치함.",
    "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 외형 레퍼런스 일치.",
    "hard_violations": [],
    "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
   },
   {
    "label": "B",
    "direction": "시선 교환은 이루어지고 있으나, 샷건이 없어 무기를 겨누는 방향 묘사 불가.",
    "built_space": "편의점 내부 약품 코너 배경 및 구조물이 레퍼런스 사진과 일치함.",
    "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 인물 레퍼런스와 외형 일치.",
    "hard_violations": [],
    "physics": "등장인물 모두 바닥에 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 앵글과 구도를 어느 정도 따랐으나, 프롬프트의 핵심인 조일광이 샷건을 겨누는 동작과 무기가 완전히 누락되어 주요 상황을 묘사하지 못했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "조일광의 손에 샷건이 들려 있지 않아 무장한 위협이라는 중심 묘사에 완전히 실패했으며, 로봇이 전경을 너무 많이 차지합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물 간 시선은 서로를 향해 있으나, 샷건이 아예 묘사되지 않아 지시된 총구 방향을 확인할 수 없음.",
        "built_space": "편의점 약품 코너. 창밖 채광과 내부 진열대 구조가 레퍼런스와 일치함.",
        "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 외형 레퍼런스 일치.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "시선 교환은 이루어지고 있으나, 샷건이 없어 무기를 겨누는 방향 묘사 불가.",
        "built_space": "편의점 내부 약품 코너 배경 및 구조물이 레퍼런스 사진과 일치함.",
        "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 인물 레퍼런스와 외형 일치.",
        "hard_violations": [],
        "physics": "등장인물 모두 바닥에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 앵글과 구도를 어느 정도 따랐으나, 프롬프트의 핵심인 조일광이 샷건을 겨누는 동작과 무기가 완전히 누락되어 주요 상황을 묘사하지 못했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "조일광의 손에 샷건이 들려 있지 않아 무장한 위협이라는 중심 묘사에 완전히 실패했으며, 로봇이 전경을 너무 많이 차지합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물 간 시선은 서로를 향해 있으나, 샷건이 아예 묘사되지 않아 지시된 총구 방향을 확인할 수 없음.",
        "built_space": "편의점 약품 코너. 창밖 채광과 내부 진열대 구조가 레퍼런스와 일치함.",
        "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 외형 레퍼런스 일치.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 발을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "시선 교환은 이루어지고 있으나, 샷건이 없어 무기를 겨누는 방향 묘사 불가.",
        "built_space": "편의점 내부 약품 코너 배경 및 구조물이 레퍼런스 사진과 일치함.",
        "entities": "조일광(군복 착용, 샷건 누락), 찰리(약통과 담요 소지), 앰버, 라울 모두 인물 레퍼런스와 외형 일치.",
        "hard_violations": [],
        "physics": "등장인물 모두 바닥에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷건과 조준 동작이 없어 핵심 위협 순간을 구현하지 못했고, 찰리가 왼쪽 전경을 지나치게 넓게 차지한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "샷건을 겨누는 핵심 행동은 실패했지만, 찰리를 더 좁은 왼쪽 전경에 두고 중앙 오른쪽 조일광과 일행의 시선을 연결한 구도가 A보다 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 라울은 오른쪽 위의 조일광을 올려다보고, 찰리의 얼굴도 그쪽을 향한다. 조일광은 왼쪽의 찰리 쪽을 바라보지만 아이들을 내려다보는 시선은 약하다. 샷건 자체가 없어 일행을 향해야 할 총구 방향과 사선이 전혀 구현되지 않았다.",
        "built_space": "오른쪽에는 약통과 약 상자가 놓인 긴 선반열, 뒤쪽에는 상품 진열대 두 면과 유리 출입문 하나가 보인다. 왼쪽 위에는 벽걸이 텔레비전 하나, 창가 모서리에는 볼록거울 하나가 있다. 낮빛이 들어오는 창호와 진열대 배치는 장소 참조와 대체로 맞으며, 거울에 명백히 불가능한 반사는 보이지 않는다. 조일광은 오른쪽 선반 옆 통로, 일행은 왼쪽 전경에 있으나 찰리의 어깨와 몸통이 화면 약 3분의 1을 차지해 좁은 가장자리 배치와 다르다.",
        "entities": "지정된 네 인물만 보인다. 조일광은 참조와 유사한 중년 한국인 남성 얼굴에 위장 군복과 해병대 모자를 착용했으며 가슴 표장은 있으나 메달은 분명하지 않다. 찰리는 베이지색 장갑판, 흰 얼굴 측면과 주황색 기계 눈을 갖추고 약병과 담요를 들고 있다. 앰버는 금발의 어린 여자아이, 라울은 갈색 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보인다. 아이들의 짙은 후드 의상은 참조의 반소매 티셔츠와 다르다. 핵심 소품인 샷건은 없고 텔레비전 화면에서 지정된 수배 영상도 확인되지 않는다.",
        "hard_violations": [],
        "physics": "인물들은 통로에 서 있는 자세이며 발은 화면 밖에 있다. 공중에 뜬 몸이나 불가능한 관절은 보이지 않는다. 약병은 찰리의 기계 손에 잡혀 있고 담요는 손과 팔에 걸쳐 지지된다. 조일광의 두 손은 빈 채 아래로 내려와 있어 샷건을 들거나 조준하는 동작이 아니다."
       },
       {
        "label": "B",
        "direction": "앰버와 라울의 고개는 오른쪽 위의 조일광을 향하고 찰리도 같은 방향을 본다. 조일광은 왼쪽 전경 일행 쪽으로 시선을 낮추고 있다. 그러나 총과 총구가 없어 렌즈를 비껴 일행에게 대각선으로 향하는 조준은 구현되지 않았다.",
        "built_space": "오른쪽 약품 선반열 하나, 뒤쪽 상품 진열대 두 면, 중앙 뒤 유리 출입문 하나와 왼쪽 창, 모서리 볼록거울 하나가 보인다. 텔레비전은 이 구도에서 확인되지 않는다. 거울의 작은 매장 반사는 명백한 광학적 모순이 없다. 조일광은 중앙 오른쪽 통로에, 찰리는 잘린 왼쪽 전경에, 두 아이는 그 옆에 배치되어 A보다 지정된 화면 관계에 가깝다. 다만 조일광의 허벅지까지 보이는 구도이며 낮은 어깨 높이에서 올려다보는 각도는 뚜렷하지 않다.",
        "entities": "지정된 네 인물이 보이며 추가 인물은 없다. 조일광의 중년 한국인 남성 얼굴, 군복과 해병대 모자는 대체로 부합하지만 메달은 명확하지 않다. 찰리의 베이지색 기계 장갑, 흰 얼굴 측면, 주황색 눈과 손에 든 약병 및 담요가 보인다. 앰버의 금발과 어린 외형, 라울의 갈색 피부와 묶은 곱슬머리가 부합하며 뒷모습 위주라 정확한 얼굴 일치는 판단하기 어렵다. 라울의 파란 티셔츠는 참조에 가깝다. 샷건은 완전히 누락되었다.",
        "hard_violations": [],
        "physics": "네 인물 모두 자연스럽게 서 있는 상체 자세이며 하단에서 발이 잘려 있다. 찰리는 기계 손가락으로 약병을 잡고 담요를 팔과 손 위에 걸치고 있어 지지가 확인된다. 조일광은 배 앞에서 두 손을 모으고 있으며 손과 팔의 연결은 자연스럽지만 총을 받치거나 방아쇠를 잡는 동작은 아니다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "샷건과 조준 동작이 없어 핵심 위협 순간을 구현하지 못했고, 찰리가 왼쪽 전경을 지나치게 넓게 차지한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "샷건을 겨누는 핵심 행동은 실패했지만, 찰리를 더 좁은 왼쪽 전경에 두고 중앙 오른쪽 조일광과 일행의 시선을 연결한 구도가 A보다 가깝다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 라울은 오른쪽 위의 조일광을 올려다보고, 찰리의 얼굴도 그쪽을 향한다. 조일광은 왼쪽의 찰리 쪽을 바라보지만 아이들을 내려다보는 시선은 약하다. 샷건 자체가 없어 일행을 향해야 할 총구 방향과 사선이 전혀 구현되지 않았다.",
        "built_space": "오른쪽에는 약통과 약 상자가 놓인 긴 선반열, 뒤쪽에는 상품 진열대 두 면과 유리 출입문 하나가 보인다. 왼쪽 위에는 벽걸이 텔레비전 하나, 창가 모서리에는 볼록거울 하나가 있다. 낮빛이 들어오는 창호와 진열대 배치는 장소 참조와 대체로 맞으며, 거울에 명백히 불가능한 반사는 보이지 않는다. 조일광은 오른쪽 선반 옆 통로, 일행은 왼쪽 전경에 있으나 찰리의 어깨와 몸통이 화면 약 3분의 1을 차지해 좁은 가장자리 배치와 다르다.",
        "entities": "지정된 네 인물만 보인다. 조일광은 참조와 유사한 중년 한국인 남성 얼굴에 위장 군복과 해병대 모자를 착용했으며 가슴 표장은 있으나 메달은 분명하지 않다. 찰리는 베이지색 장갑판, 흰 얼굴 측면과 주황색 기계 눈을 갖추고 약병과 담요를 들고 있다. 앰버는 금발의 어린 여자아이, 라울은 갈색 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보인다. 아이들의 짙은 후드 의상은 참조의 반소매 티셔츠와 다르다. 핵심 소품인 샷건은 없고 텔레비전 화면에서 지정된 수배 영상도 확인되지 않는다.",
        "hard_violations": [],
        "physics": "인물들은 통로에 서 있는 자세이며 발은 화면 밖에 있다. 공중에 뜬 몸이나 불가능한 관절은 보이지 않는다. 약병은 찰리의 기계 손에 잡혀 있고 담요는 손과 팔에 걸쳐 지지된다. 조일광의 두 손은 빈 채 아래로 내려와 있어 샷건을 들거나 조준하는 동작이 아니다."
       },
       {
        "label": "A",
        "direction": "앰버와 라울의 고개는 오른쪽 위의 조일광을 향하고 찰리도 같은 방향을 본다. 조일광은 왼쪽 전경 일행 쪽으로 시선을 낮추고 있다. 그러나 총과 총구가 없어 렌즈를 비껴 일행에게 대각선으로 향하는 조준은 구현되지 않았다.",
        "built_space": "오른쪽 약품 선반열 하나, 뒤쪽 상품 진열대 두 면, 중앙 뒤 유리 출입문 하나와 왼쪽 창, 모서리 볼록거울 하나가 보인다. 텔레비전은 이 구도에서 확인되지 않는다. 거울의 작은 매장 반사는 명백한 광학적 모순이 없다. 조일광은 중앙 오른쪽 통로에, 찰리는 잘린 왼쪽 전경에, 두 아이는 그 옆에 배치되어 A보다 지정된 화면 관계에 가깝다. 다만 조일광의 허벅지까지 보이는 구도이며 낮은 어깨 높이에서 올려다보는 각도는 뚜렷하지 않다.",
        "entities": "지정된 네 인물이 보이며 추가 인물은 없다. 조일광의 중년 한국인 남성 얼굴, 군복과 해병대 모자는 대체로 부합하지만 메달은 명확하지 않다. 찰리의 베이지색 기계 장갑, 흰 얼굴 측면, 주황색 눈과 손에 든 약병 및 담요가 보인다. 앰버의 금발과 어린 외형, 라울의 갈색 피부와 묶은 곱슬머리가 부합하며 뒷모습 위주라 정확한 얼굴 일치는 판단하기 어렵다. 라울의 파란 티셔츠는 참조에 가깝다. 샷건은 완전히 누락되었다.",
        "hard_violations": [],
        "physics": "네 인물 모두 자연스럽게 서 있는 상체 자세이며 하단에서 발이 잘려 있다. 찰리는 기계 손가락으로 약병을 잡고 담요를 팔과 손 위에 걸치고 있어 지지가 확인된다. 조일광은 배 앞에서 두 손을 모으고 있으며 손과 팔의 연결은 자연스럽지만 총을 받치거나 방아쇠를 잡는 동작은 아니다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.75
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 앵글과 구도를 어느 정도 따랐으나, 프롬프트의 핵심인 조일광이 샷건을 겨누는 동작과 무기가 완전히 누락되어 주요 상황을 묘사하지 못했습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "조일광의 손에 샷건이 들려 있지 않아 무장한 위협이라는 중심 묘사에 완전히 실패했으며, 로봇이 전경을 너무 많이 차지합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B02.png",
    "asset_id": "31ae9965-7777-439a-9c7b-3bc6c7cd5b19",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 조일광: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1399968>",
    "asset_id": "d81b8d5c-ffe5-4291-a480-6fbe062a1a8c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1305657>",
    "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1334467>",
    "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b4c-9903-799b-8664-d7815a7f9b3e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh16__bgfirst_bg.png",
   "bg_asset_id": "8db851c9-1539-46b8-9587-f6cc72e980e1",
   "bg_record_key": "S49sh16::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C01",
   "C05",
   "C06"
  ]
 },
 "S49sh40::signage": {
  "fp": "43b56ba7be45b11a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh40::bgfirst_bg": {
  "input_fingerprint": "ef9a8a6d8eb43707",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨진 창문 밖으로 상반신을 내민 찰리의 가슴 중앙에서 눈부신 에너지가 강한 빛을 뿜어내며 맺혀 있는 근접 구도.\n\nLOCATION (lock): At the shattered front passenger window of the moving camper, straddling the cab and the open roadside air. Daylight floods the broken window opening.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track alongside the passenger window from outside and slightly toward the vehicle's rear, maintaining the oblique upward view from below 찰리's chest. Tighten only the distance so his exposed upper body occupies roughly two-thirds of the frame, with the rotating chest source near center and a narrow section of the broken window opening at lower right. Keep his head turned toward the pursuing drone off-screen, making his protective intervention readable through the forward lean rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Broken, with 찰리 leaning through it) — The opening is seen obliquely from outside and toward the rear of the vehicle; used as Retain a lower-right spatial reference that explains the body's support and position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intense chest emission provides a localized brightness peak while controlled exposure retains the rotating mechanism and surrounding hard-surface detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨진 창문 밖으로 상반신을 내민 찰리의 가슴 중앙에서 눈부신 에너지가 강한 빛을 뿜어내며 맺혀 있는 근접 구도.\n\nLOCATION (lock): At the shattered front passenger window of the moving camper, straddling the cab and the open roadside air. Daylight floods the broken window opening.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track alongside the passenger window from outside and slightly toward the vehicle's rear, maintaining the oblique upward view from below 찰리's chest. Tighten only the distance so his exposed upper body occupies roughly two-thirds of the frame, with the rotating chest source near center and a narrow section of the broken window opening at lower right. Keep his head turned toward the pursuing drone off-screen, making his protective intervention readable through the forward lean rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Broken, with 찰리 leaning through it) — The opening is seen obliquely from outside and toward the rear of the vehicle; used as Retain a lower-right spatial reference that explains the body's support and position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intense chest emission provides a localized brightness peak while controlled exposure retains the rotating mechanism and surrounding hard-surface detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh40__bgfirst_bg.png",
  "asset_id": "c13c08a6-fe66-4f2d-8b37-16878c7c7b01",
  "input_asset_ids": [
   "b4011ab7-3f3c-4cfd-b878-f9a1ab3bfa6e",
   "a0a95db9-7bd7-4a4d-844b-3f58e0713fe5",
   "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2"
  ]
 },
 "S49sh40": {
  "input_fingerprint": "abd9d3477ba68e1d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨진 창문 밖으로 상반신을 내민 찰리의 가슴 중앙에서 눈부신 에너지가 강한 빛을 뿜어내며 맺혀 있는 근접 구도.\n\nLOCATION (lock): At the shattered front passenger window of the moving camper, straddling the cab and the open roadside air. Daylight floods the broken window opening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track alongside the passenger window from outside and slightly toward the vehicle's rear, maintaining the oblique upward view from below 찰리's chest. Tighten only the distance so his exposed upper body occupies roughly two-thirds of the frame, with the rotating chest source near center and a narrow section of the broken window opening at lower right. Keep his head turned toward the pursuing drone off-screen, making his protective intervention readable through the forward lean rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Broken, with 찰리 leaning through it) — The opening is seen obliquely from outside and toward the rear of the vehicle; used as Retain a lower-right spatial reference that explains the body's support and position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intense chest emission provides a localized brightness peak while controlled exposure retains the rotating mechanism and surrounding hard-surface detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fleeing camper has a shotgun-shattered window and a separately smashed front passenger window, with broken glass inside. Its dashboard RPM is climbing rapidly and its interior bulbs are growing brighter. 찰리: His shoulder is damaged and sparking, and his upper body is halfway outside the broken passenger window. The ring in his chest is spinning and emitting energy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨진 창문 밖으로 상반신을 내민 찰리의 가슴 중앙에서 눈부신 에너지가 강한 빛을 뿜어내며 맺혀 있는 근접 구도.\n\nLOCATION (lock): At the shattered front passenger window of the moving camper, straddling the cab and the open roadside air. Daylight floods the broken window opening. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track alongside the passenger window from outside and slightly toward the vehicle's rear, maintaining the oblique upward view from below 찰리's chest. Tighten only the distance so his exposed upper body occupies roughly two-thirds of the frame, with the rotating chest source near center and a narrow section of the broken window opening at lower right. Keep his head turned toward the pursuing drone off-screen, making his protective intervention readable through the forward lean rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Broken, with 찰리 leaning through it) — The opening is seen obliquely from outside and toward the rear of the vehicle; used as Retain a lower-right spatial reference that explains the body's support and position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intense chest emission provides a localized brightness peak while controlled exposure retains the rotating mechanism and surrounding hard-surface detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fleeing camper has a shotgun-shattered window and a separately smashed front passenger window, with broken glass inside. Its dashboard RPM is climbing rapidly and its interior bulbs are growing brighter. 찰리: His shoulder is damaged and sparking, and his upper body is halfway outside the broken passenger window. The ring in his chest is spinning and emitting energy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깨진 창문 밖으로 상반신을 내민 찰리의 가슴 중앙에서 눈부신 에너지가 강한 빛을 뿜어내며 맺혀 있는 근접 구도.\n\nLOCATION (lock): At the shattered front passenger window of the moving camper, straddling the cab and the open roadside air. Daylight floods the broken window opening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track alongside the passenger window from outside and slightly toward the vehicle's rear, maintaining the oblique upward view from below 찰리's chest. Tighten only the distance so his exposed upper body occupies roughly two-thirds of the frame, with the rotating chest source near center and a narrow section of the broken window opening at lower right. Keep his head turned toward the pursuing drone off-screen, making his protective intervention readable through the forward lean rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Passenger-side window opening (Broken, with 찰리 leaning through it) — The opening is seen obliquely from outside and toward the rear of the vehicle; used as Retain a lower-right spatial reference that explains the body's support and position.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The intense chest emission provides a localized brightness peak while controlled exposure retains the rotating mechanism and surrounding hard-surface detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fleeing camper has a shotgun-shattered window and a separately smashed front passenger window, with broken glass inside. Its dashboard RPM is climbing rapidly and its interior bulbs are growing brighter. 찰리: His shoulder is damaged and sparking, and his upper body is halfway outside the broken passenger window. The ring in his chest is spinning and emitting energy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh40__bgfirst_bg.png",
     "asset_id": "c13c08a6-fe66-4f2d-8b37-16878c7c7b01",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S49sh40.png",
     "asset_id": "b4011ab7-3f3c-4cfd-b878-f9a1ab3bfa6e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B03.png",
     "asset_id": "a0a95db9-7bd7-4a4d-844b-3f58e0713fe5",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_roadside_service_station_sel.png",
     "asset_id": "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 고개는 화면 밖 전방 상단(추격하는 드론 방향)을 향하고 있으며, 카메라는 차량 우측면을 따라 전방을 향해 위치함.",
    "built_space": "주행 중인 캠핑카의 우측 조수석 창문 밖. 깨진 유리창 프레임이 하단에 위치하며, 원경의 도로 우측에는 구조물 레퍼런스에 지정된 SK 주유소와 로케이션 레퍼런스의 전신주가 보임.",
    "entities": "찰리(샌드 베이지색 장갑, 흰색 마스크, 두 눈과 선형 입, 파란색 회전형 가슴 원자로)의 외형이 레퍼런스와 정확히 일치함. 우측 어깨 장갑 파손 및 스파크 묘사됨.",
    "hard_violations": [],
    "physics": "찰리의 상반신은 차량의 창틀에 지지되어 밖으로 기울어져 있으며, 차량과 배경의 모션 블러를 통해 고속 주행 중임이 물리적으로 타당하게 표현됨."
   },
   {
    "label": "B",
    "direction": "찰리의 고개는 전방을 향해 살짝 위를 응시하고 있으며, 카메라는 차량 우측 사이드미러 뒤에서 전방을 향하고 있음.",
    "built_space": "주행 중인 차량의 깨진 우측 창문. 하단에 깨진 유리가 있으며 전방에 사이드미러가 보임. 배경은 로케이션 레퍼런스의 언덕과 나무, 전신주로 구성됨.",
    "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크, 가슴의 푸른 에너지)이 레퍼런스와 일치하며, 우측 어깨에서 스파크가 발생하고 있음.",
    "hard_violations": [],
    "physics": "찰리가 창문 밖으로 상체를 내밀고 있으며 창틀이 이를 지지함. 배경의 모션 블러로 차량의 이동 상태가 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로우 앵글의 근접 구도와 찰리의 기계적 디테일(어깨 스파크, 가슴 에너지)을 완벽히 구현했으며, 배경에 로케이션과 구조물 레퍼런스를 자연스럽게 합성해 지시문 충실도가 가장 높습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "캐릭터 디자인과 손상된 어깨의 스파크 묘사는 우수하나, 카메라의 상향 앵글이 A에 비해 덜 극적이며 구조물 레퍼런스가 배경에 반영되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 고개는 화면 밖 전방 상단(추격하는 드론 방향)을 향하고 있으며, 카메라는 차량 우측면을 따라 전방을 향해 위치함.",
        "built_space": "주행 중인 캠핑카의 우측 조수석 창문 밖. 깨진 유리창 프레임이 하단에 위치하며, 원경의 도로 우측에는 구조물 레퍼런스에 지정된 SK 주유소와 로케이션 레퍼런스의 전신주가 보임.",
        "entities": "찰리(샌드 베이지색 장갑, 흰색 마스크, 두 눈과 선형 입, 파란색 회전형 가슴 원자로)의 외형이 레퍼런스와 정확히 일치함. 우측 어깨 장갑 파손 및 스파크 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 상반신은 차량의 창틀에 지지되어 밖으로 기울어져 있으며, 차량과 배경의 모션 블러를 통해 고속 주행 중임이 물리적으로 타당하게 표현됨."
       },
       {
        "label": "B",
        "direction": "찰리의 고개는 전방을 향해 살짝 위를 응시하고 있으며, 카메라는 차량 우측 사이드미러 뒤에서 전방을 향하고 있음.",
        "built_space": "주행 중인 차량의 깨진 우측 창문. 하단에 깨진 유리가 있으며 전방에 사이드미러가 보임. 배경은 로케이션 레퍼런스의 언덕과 나무, 전신주로 구성됨.",
        "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크, 가슴의 푸른 에너지)이 레퍼런스와 일치하며, 우측 어깨에서 스파크가 발생하고 있음.",
        "hard_violations": [],
        "physics": "찰리가 창문 밖으로 상체를 내밀고 있으며 창틀이 이를 지지함. 배경의 모션 블러로 차량의 이동 상태가 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로우 앵글의 근접 구도와 찰리의 기계적 디테일(어깨 스파크, 가슴 에너지)을 완벽히 구현했으며, 배경에 로케이션과 구조물 레퍼런스를 자연스럽게 합성해 지시문 충실도가 가장 높습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "캐릭터 디자인과 손상된 어깨의 스파크 묘사는 우수하나, 카메라의 상향 앵글이 A에 비해 덜 극적이며 구조물 레퍼런스가 배경에 반영되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 고개는 화면 밖 전방 상단(추격하는 드론 방향)을 향하고 있으며, 카메라는 차량 우측면을 따라 전방을 향해 위치함.",
        "built_space": "주행 중인 캠핑카의 우측 조수석 창문 밖. 깨진 유리창 프레임이 하단에 위치하며, 원경의 도로 우측에는 구조물 레퍼런스에 지정된 SK 주유소와 로케이션 레퍼런스의 전신주가 보임.",
        "entities": "찰리(샌드 베이지색 장갑, 흰색 마스크, 두 눈과 선형 입, 파란색 회전형 가슴 원자로)의 외형이 레퍼런스와 정확히 일치함. 우측 어깨 장갑 파손 및 스파크 묘사됨.",
        "hard_violations": [],
        "physics": "찰리의 상반신은 차량의 창틀에 지지되어 밖으로 기울어져 있으며, 차량과 배경의 모션 블러를 통해 고속 주행 중임이 물리적으로 타당하게 표현됨."
       },
       {
        "label": "B",
        "direction": "찰리의 고개는 전방을 향해 살짝 위를 응시하고 있으며, 카메라는 차량 우측 사이드미러 뒤에서 전방을 향하고 있음.",
        "built_space": "주행 중인 차량의 깨진 우측 창문. 하단에 깨진 유리가 있으며 전방에 사이드미러가 보임. 배경은 로케이션 레퍼런스의 언덕과 나무, 전신주로 구성됨.",
        "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크, 가슴의 푸른 에너지)이 레퍼런스와 일치하며, 우측 어깨에서 스파크가 발생하고 있음.",
        "hard_violations": [],
        "physics": "찰리가 창문 밖으로 상체를 내밀고 있으며 창틀이 이를 지지함. 배경의 모션 블러로 차량의 이동 상태가 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "강렬한 가슴 발광과 밀착된 상반신 구도는 좋지만, 차량 뒤쪽에서 보는 지정 시점과 달라 보이고 원자로가 중앙에서 왼쪽으로 치우칩니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "차량 뒤쪽에서 올려다보는 사선 시점, 중앙의 원자로와 화면 밖을 향한 머리 방향이 더 정확하지만, 깨진 창틀과 허리 노출은 요구보다 넓습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 화면 오른쪽 바깥을 향하며 렌즈를 정면으로 보지 않습니다. 추격 드론 자체는 보이지 않으므로 실제 표적과의 정렬은 확인할 수 없습니다. 몸통은 창밖으로 기울어 있고 가슴 에너지는 원형 장치에서 퍼지며 특정 표적을 향한 광선은 없습니다.",
        "built_space": "깨진 창 개구부 하나가 왼쪽과 하단에 크게 걸쳐 있고, 오른쪽에는 사이드미러 하나와 차체의 다른 창 일부가 보입니다. 미러와 뒤로 이어지는 캠퍼 벽면의 배치는 지정된 후방 쪽 시점보다는 차량 앞쪽에서 뒤를 보는 시점에 가깝습니다. 창문을 오른쪽 아래에 좁게 남기라는 배치는 지켜지지 않았습니다. 낮의 가드레일, 전신주와 산지는 장소 참고와 부합하며, 주유소 구조는 이 크롭에서 보이지 않습니다.",
        "entities": "등장 개체는 찰리 하나뿐입니다. 샌드 베이지 기계 장갑, 흰 마스크, 주황색 원형 눈 두 개와 검은 입 선, 육중한 팔, 푸른 가슴 원자로가 참고 정체성과 맞습니다. 화면 왼쪽 어깨에는 파손된 장갑과 불꽃이 있습니다. 원자로의 회전성 빛 무늬와 주변 금속 세부가 함께 보입니다. 하체와 계기판 등은 크롭 밖이므로 평가하지 않습니다.",
        "hard_violations": [],
        "physics": "복부는 창 개구부 안쪽으로 이어지고 하단 창턱에 걸쳐 있으며, 화면 오른쪽 팔도 창턱 바깥으로 내려와 있습니다. 창턱과 실내에 남은 몸이 상체의 지지 관계를 설명하므로 공중에 떠 있는 몸은 아닙니다. 유리 조각은 창 가장자리에 붙어 있고 불꽃은 손상된 어깨에서 발생합니다."
       },
       {
        "label": "B",
        "direction": "찰리는 머리를 화면 오른쪽 바깥으로 돌리고 상체를 창밖으로 내밉니다. 렌즈를 보는 자세가 아니며 화면 밖 추격 드론을 향한다는 연출과 양립합니다. 드론은 보이지 않아 정확한 목표 정렬은 검증할 수 없습니다. 가슴 장치는 카메라 쪽으로 비스듬히 노출되고 중심에서 푸른 빛이 방사됩니다.",
        "built_space": "깨진 조수석 창 개구부 하나와 그 오른쪽의 좁은 유리·차체 부분이 보입니다. 차량 뒤쪽에서 앞쪽 창 기둥을 바라보는 사선 구도로 읽히며 카메라도 가슴보다 낮습니다. 원자로는 화면 중앙 가까이에 있고 상체가 화면 대부분을 차지하지만, 창틀이 하단 전체와 왼쪽까지 넓게 노출됩니다. 오른쪽 먼 배경에는 주유소 캐노피 하나와 독립 간판 하나가 작게 보이며 참고의 회색·빨강·주황 계열을 따릅니다. 먼 거리와 크롭 때문에 주유기 수나 건물 전체 형상까지 확인할 수는 없습니다.",
        "entities": "찰리 한 개체만 보이며 흰 마스크, 주황색 눈 두 개, 검은 입 선과 베이지색 중장갑이 참고와 일치합니다. 큰 어깨와 긴 기계 팔도 정체성을 유지합니다. 화면 왼쪽 어깨는 깨져 내부 부품과 불꽃이 드러납니다. 중앙 원자로에는 푸른 발광 고리와 회전 궤적이 있고 주변 장갑의 재질도 남아 있습니다. 깨진 유리는 창 가장자리에 보이며 별도 인물이나 드론은 추가되지 않았습니다.",
        "hard_violations": [],
        "physics": "허리와 골반 일부가 개구부 안에 남아 있고 하복부가 하단 창턱에 걸쳐 있어 앞으로 내민 상체의 받침이 읽힙니다. 팔은 어깨 관절에서 자연스럽게 이어져 창 가장자리 쪽으로 내려갑니다. 몸이 지지 없이 떠 있지 않으며, 어깨 불꽃은 노출된 파손 부위에서 튀고 유리 잔편은 창틀에 붙어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "강렬한 가슴 발광과 밀착된 상반신 구도는 좋지만, 차량 뒤쪽에서 보는 지정 시점과 달라 보이고 원자로가 중앙에서 왼쪽으로 치우칩니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "차량 뒤쪽에서 올려다보는 사선 시점, 중앙의 원자로와 화면 밖을 향한 머리 방향이 더 정확하지만, 깨진 창틀과 허리 노출은 요구보다 넓습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 화면 오른쪽 바깥을 향하며 렌즈를 정면으로 보지 않습니다. 추격 드론 자체는 보이지 않으므로 실제 표적과의 정렬은 확인할 수 없습니다. 몸통은 창밖으로 기울어 있고 가슴 에너지는 원형 장치에서 퍼지며 특정 표적을 향한 광선은 없습니다.",
        "built_space": "깨진 창 개구부 하나가 왼쪽과 하단에 크게 걸쳐 있고, 오른쪽에는 사이드미러 하나와 차체의 다른 창 일부가 보입니다. 미러와 뒤로 이어지는 캠퍼 벽면의 배치는 지정된 후방 쪽 시점보다는 차량 앞쪽에서 뒤를 보는 시점에 가깝습니다. 창문을 오른쪽 아래에 좁게 남기라는 배치는 지켜지지 않았습니다. 낮의 가드레일, 전신주와 산지는 장소 참고와 부합하며, 주유소 구조는 이 크롭에서 보이지 않습니다.",
        "entities": "등장 개체는 찰리 하나뿐입니다. 샌드 베이지 기계 장갑, 흰 마스크, 주황색 원형 눈 두 개와 검은 입 선, 육중한 팔, 푸른 가슴 원자로가 참고 정체성과 맞습니다. 화면 왼쪽 어깨에는 파손된 장갑과 불꽃이 있습니다. 원자로의 회전성 빛 무늬와 주변 금속 세부가 함께 보입니다. 하체와 계기판 등은 크롭 밖이므로 평가하지 않습니다.",
        "hard_violations": [],
        "physics": "복부는 창 개구부 안쪽으로 이어지고 하단 창턱에 걸쳐 있으며, 화면 오른쪽 팔도 창턱 바깥으로 내려와 있습니다. 창턱과 실내에 남은 몸이 상체의 지지 관계를 설명하므로 공중에 떠 있는 몸은 아닙니다. 유리 조각은 창 가장자리에 붙어 있고 불꽃은 손상된 어깨에서 발생합니다."
       },
       {
        "label": "A",
        "direction": "찰리는 머리를 화면 오른쪽 바깥으로 돌리고 상체를 창밖으로 내밉니다. 렌즈를 보는 자세가 아니며 화면 밖 추격 드론을 향한다는 연출과 양립합니다. 드론은 보이지 않아 정확한 목표 정렬은 검증할 수 없습니다. 가슴 장치는 카메라 쪽으로 비스듬히 노출되고 중심에서 푸른 빛이 방사됩니다.",
        "built_space": "깨진 조수석 창 개구부 하나와 그 오른쪽의 좁은 유리·차체 부분이 보입니다. 차량 뒤쪽에서 앞쪽 창 기둥을 바라보는 사선 구도로 읽히며 카메라도 가슴보다 낮습니다. 원자로는 화면 중앙 가까이에 있고 상체가 화면 대부분을 차지하지만, 창틀이 하단 전체와 왼쪽까지 넓게 노출됩니다. 오른쪽 먼 배경에는 주유소 캐노피 하나와 독립 간판 하나가 작게 보이며 참고의 회색·빨강·주황 계열을 따릅니다. 먼 거리와 크롭 때문에 주유기 수나 건물 전체 형상까지 확인할 수는 없습니다.",
        "entities": "찰리 한 개체만 보이며 흰 마스크, 주황색 눈 두 개, 검은 입 선과 베이지색 중장갑이 참고와 일치합니다. 큰 어깨와 긴 기계 팔도 정체성을 유지합니다. 화면 왼쪽 어깨는 깨져 내부 부품과 불꽃이 드러납니다. 중앙 원자로에는 푸른 발광 고리와 회전 궤적이 있고 주변 장갑의 재질도 남아 있습니다. 깨진 유리는 창 가장자리에 보이며 별도 인물이나 드론은 추가되지 않았습니다.",
        "hard_violations": [],
        "physics": "허리와 골반 일부가 개구부 안에 남아 있고 하복부가 하단 창턱에 걸쳐 있어 앞으로 내민 상체의 받침이 읽힙니다. 팔은 어깨 관절에서 자연스럽게 이어져 창 가장자리 쪽으로 내려갑니다. 몸이 지지 없이 떠 있지 않으며, 어깨 불꽃은 노출된 파손 부위에서 튀고 유리 잔편은 창틀에 붙어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.5
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.5
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 로우 앵글의 근접 구도와 찰리의 기계적 디테일(어깨 스파크, 가슴 에너지)을 완벽히 구현했으며, 배경에 로케이션과 구조물 레퍼런스를 자연스럽게 합성해 지시문 충실도가 가장 높습니다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "캐릭터 디자인과 손상된 어깨의 스파크 묘사는 우수하나, 카메라의 상향 앵글이 A에 비해 덜 극적이며 구조물 레퍼런스가 배경에 반영되지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B03.png",
    "asset_id": "a0a95db9-7bd7-4a4d-844b-3f58e0713fe5",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_roadside_service_station_sel.png",
    "asset_id": "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b62-9bb6-7ad7-96cf-a9c45d1e7090",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh40__bgfirst_bg.png",
   "bg_asset_id": "c13c08a6-fe66-4f2d-8b37-16878c7c7b01",
   "bg_record_key": "S49sh40::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S49sh47::confined_fp_apt": {
  "applies": false,
  "reason_ko": "캠핑카 운전석 내부에서 진행되는 숏이지만, 주인공의 1인칭 시점으로 깨진 창문 밖의 인물들을 바라보는 장면이므로 차량 내부의 복잡한 조작부나 좌석 배치 등 공간적 구조가 화면에 주요하게 드러나지 않습니다. 따라서 평면도 보조가 필수적이지 않습니다.",
  "input_fingerprint": "b36e8e68a997abab"
 },
 "S49sh47::signage": {
  "fp": "1820bd4cca0ab66b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh47::bgfirst_bg": {
  "input_fingerprint": "5d0705bfd3093b95",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 흐릿한 시야 너머로 깨진 창문 밖에서 차 안을 들여다보는 태진과 은영의 실루엣.\n\nLOCATION (lock): Inside the overturned camper's driving cab, looking through a broken window toward the rescuers outside. Smoke hangs in the cab and an interior bulb flickers weakly.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain motionless at 이현우's resting eye position inside the overturned camper, looking obliquely upward through the broken window with the established canted horizon. Place 태진 beyond the upper-left portion of the opening and 은영 farther right, their upper bodies leaning down at different angles as they inspect 이현우 below them rather than presenting frontal poses. Let failing optical focus soften their silhouettes while preserving the window boundary; 이현우 remains entirely outside the image, and neither woman has yet transformed into 미연.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening with the women outside in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Broken camper window opening (Broken and accessible from outside the overturned vehicle) — Seen diagonally upward from the interior, with the women beyond its boundary; used as Separate the trapped subjective viewpoint from the rescuers outside.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dim, intermittent interior illumination and restrained exterior exposure leave the women softly silhouetted, with blurred edges expressing failing focus rather than hallucination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 흐릿한 시야 너머로 깨진 창문 밖에서 차 안을 들여다보는 태진과 은영의 실루엣.\n\nLOCATION (lock): Inside the overturned camper's driving cab, looking through a broken window toward the rescuers outside. Smoke hangs in the cab and an interior bulb flickers weakly.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain motionless at 이현우's resting eye position inside the overturned camper, looking obliquely upward through the broken window with the established canted horizon. Place 태진 beyond the upper-left portion of the opening and 은영 farther right, their upper bodies leaning down at different angles as they inspect 이현우 below them rather than presenting frontal poses. Let failing optical focus soften their silhouettes while preserving the window boundary; 이현우 remains entirely outside the image, and neither woman has yet transformed into 미연.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening with the women outside in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Broken camper window opening (Broken and accessible from outside the overturned vehicle) — Seen diagonally upward from the interior, with the women beyond its boundary; used as Separate the trapped subjective viewpoint from the rescuers outside.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dim, intermittent interior illumination and restrained exterior exposure leave the women softly silhouetted, with blurred edges expressing failing focus rather than hallucination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh47__bgfirst_bg.png",
  "asset_id": "645528b1-cd2a-4fd6-ba5a-d94f6f395421",
  "input_asset_ids": [
   "8d35d5c1-2c31-43ed-a124-edb6271a92c7",
   "a9202095-b456-4be5-85cc-eab29211f7df",
   "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2"
  ]
 },
 "S49sh47": {
  "input_fingerprint": "d47dde917bef1a9a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 흐릿한 시야 너머로 깨진 창문 밖에서 차 안을 들여다보는 태진과 은영의 실루엣.\n\nLOCATION (lock): Inside the overturned camper's driving cab, looking through a broken window toward the rescuers outside. Smoke hangs in the cab and an interior bulb flickers weakly. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain motionless at 이현우's resting eye position inside the overturned camper, looking obliquely upward through the broken window with the established canted horizon. Place 태진 beyond the upper-left portion of the opening and 은영 farther right, their upper bodies leaning down at different angles as they inspect 이현우 below them rather than presenting frontal poses. Let failing optical focus soften their silhouettes while preserving the window boundary; 이현우 remains entirely outside the image, and neither woman has yet transformed into 미연.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening with the women outside in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Broken camper window opening (Broken and accessible from outside the overturned vehicle) — Seen diagonally upward from the interior, with the women beyond its boundary; used as Separate the trapped subjective viewpoint from the rescuers outside.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dim, intermittent interior illumination and restrained exterior exposure leave the women softly silhouetted, with blurred edges expressing failing focus rather than hallucination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper lies completely overturned after striking the trash embankment, with shattered windows, smoke inside and an interior bulb flickering weakly. The hunting drone has exploded. 태진: She is outside the wreck, looking into the camper. 은영: She is outside the wreck, looking into the camper.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음.; 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective and every fixed fitting it has stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: every fixed fitting this space has, how many of each, which way each faces, and what any reflective surface among them could return from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second one of anything, moving a fitting or growing a new surface is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: where the body rests and on what, which side of a fitting the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 흐릿한 시야 너머로 깨진 창문 밖에서 차 안을 들여다보는 태진과 은영의 실루엣.\n\nLOCATION (lock): Inside the overturned camper's driving cab, looking through a broken window toward the rescuers outside. Smoke hangs in the cab and an interior bulb flickers weakly. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain motionless at 이현우's resting eye position inside the overturned camper, looking obliquely upward through the broken window with the established canted horizon. Place 태진 beyond the upper-left portion of the opening and 은영 farther right, their upper bodies leaning down at different angles as they inspect 이현우 below them rather than presenting frontal poses. Let failing optical focus soften their silhouettes while preserving the window boundary; 이현우 remains entirely outside the image, and neither woman has yet transformed into 미연.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening with the women outside in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Broken camper window opening (Broken and accessible from outside the overturned vehicle) — Seen diagonally upward from the interior, with the women beyond its boundary; used as Separate the trapped subjective viewpoint from the rescuers outside.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dim, intermittent interior illumination and restrained exterior exposure leave the women softly silhouetted, with blurred edges expressing failing focus rather than hallucination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper lies completely overturned after striking the trash embankment, with shattered windows, smoke inside and an interior bulb flickering weakly. The hunting drone has exploded. 태진: She is outside the wreck, looking into the camper. 은영: She is outside the wreck, looking into the camper.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음.; 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 흐릿한 시야 너머로 깨진 창문 밖에서 차 안을 들여다보는 태진과 은영의 실루엣.\n\nLOCATION (lock): Inside the overturned camper's driving cab, looking through a broken window toward the rescuers outside. Smoke hangs in the cab and an interior bulb flickers weakly. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain motionless at 이현우's resting eye position inside the overturned camper, looking obliquely upward through the broken window with the established canted horizon. Place 태진 beyond the upper-left portion of the opening and 은영 farther right, their upper bodies leaning down at different angles as they inspect 이현우 below them rather than presenting frontal poses. Let failing optical focus soften their silhouettes while preserving the window boundary; 이현우 remains entirely outside the image, and neither woman has yet transformed into 미연.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Broken window opening with the women outside in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Broken camper window opening (Broken and accessible from outside the overturned vehicle) — Seen diagonally upward from the interior, with the women beyond its boundary; used as Separate the trapped subjective viewpoint from the rescuers outside.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dim, intermittent interior illumination and restrained exterior exposure leave the women softly silhouetted, with blurred edges expressing failing focus rather than hallucination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper lies completely overturned after striking the trash embankment, with shattered windows, smoke inside and an interior bulb flickering weakly. The hunting drone has exploded. 태진: She is outside the wreck, looking into the camper. 은영: She is outside the wreck, looking into the camper.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음.; 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh47__bgfirst_bg.png",
     "asset_id": "645528b1-cd2a-4fd6-ba5a-d94f6f395421",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched fixed fittings and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S49sh47.png",
     "asset_id": "8d35d5c1-2c31-43ed-a124-edb6271a92c7",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:769818>",
     "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 은영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:981336>",
     "asset_id": "c128fa4a-1680-4e5b-a9d7-a21abf9186af",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B01.png",
     "asset_id": "a9202095-b456-4be5-85cc-eab29211f7df",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_roadside_service_station_sel.png",
     "asset_id": "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:769818>",
     "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 은영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:981336>",
     "asset_id": "c128fa4a-1680-4e5b-a9d7-a21abf9186af",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 여성 모두 차량 내부의 관찰자가 있는 아래쪽을 정확히 내려다보고 있음.",
    "built_space": "기울어진 차량 내부 창틀, 깨진 유리조각, 좌측 상단의 실내등이 보이며, 창밖 배경에는 주유소와 쓰레기 더미가 위치함.",
    "entities": "태진(좌)은 작업복과 모자, 은영(우)은 검은 티와 멜빵바지를 착용해 참고 이미지와 일치함. 배경의 SK 주유소 구조물도 확인됨.",
    "hard_violations": [
     "[gemini-pro] 은영의 멜빵바지 주머니에 실제 공구가 아닌 하얀색 평면 렌치 그래픽이 인쇄되어 있어 재질 사실성 규정 위반."
    ],
    "physics": "외부에서 차 안을 들여다보기 위해 몸을 숙인 자세가 자연스럽고 안정적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 창문 밖에서 차량 내부의 아래쪽을 향해 시선을 고정하고 있음.",
    "built_space": "차량 내부의 기울어진 프레임과 실내등이 존재하며, 창 너머로 참고 사진의 주유소와 쓰레기장이 적절한 비율로 배치됨.",
    "entities": "태진과 은영의 인상착의 및 의상이 참고 자료와 정확히 일치하며, 의상의 공구 디테일도 입체적으로 묘사됨.",
    "hard_violations": [],
    "physics": "기울어진 창틀 너머로 허리를 굽히고 안을 살피는 자세가 물리적 모순 없이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 어둡고 흐릿한 조명 분위기에 조금 더 부합하며, 인물들의 의상과 장비 디테일이 사실적으로 구현되어 프롬프트에 충실함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 시선 처리는 훌륭하지만, 요구된 실루엣과 흐릿한 초점 효과가 부족하고 은영의 멜빵에 그려진 평면적인 그래픽이 사실성을 떨어뜨림."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 여성 모두 차량 내부의 관찰자가 있는 아래쪽을 정확히 내려다보고 있음.",
        "built_space": "기울어진 차량 내부 창틀, 깨진 유리조각, 좌측 상단의 실내등이 보이며, 창밖 배경에는 주유소와 쓰레기 더미가 위치함.",
        "entities": "태진(좌)은 작업복과 모자, 은영(우)은 검은 티와 멜빵바지를 착용해 참고 이미지와 일치함. 배경의 SK 주유소 구조물도 확인됨.",
        "hard_violations": [
         "은영의 멜빵바지 주머니에 실제 공구가 아닌 하얀색 평면 렌치 그래픽이 인쇄되어 있어 재질 사실성 규정 위반."
        ],
        "physics": "외부에서 차 안을 들여다보기 위해 몸을 숙인 자세가 자연스럽고 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 창문 밖에서 차량 내부의 아래쪽을 향해 시선을 고정하고 있음.",
        "built_space": "차량 내부의 기울어진 프레임과 실내등이 존재하며, 창 너머로 참고 사진의 주유소와 쓰레기장이 적절한 비율로 배치됨.",
        "entities": "태진과 은영의 인상착의 및 의상이 참고 자료와 정확히 일치하며, 의상의 공구 디테일도 입체적으로 묘사됨.",
        "hard_violations": [],
        "physics": "기울어진 창틀 너머로 허리를 굽히고 안을 살피는 자세가 물리적 모순 없이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 어둡고 흐릿한 조명 분위기에 조금 더 부합하며, 인물들의 의상과 장비 디테일이 사실적으로 구현되어 프롬프트에 충실함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 시선 처리는 훌륭하지만, 요구된 실루엣과 흐릿한 초점 효과가 부족하고 은영의 멜빵에 그려진 평면적인 그래픽이 사실성을 떨어뜨림."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 여성 모두 차량 내부의 관찰자가 있는 아래쪽을 정확히 내려다보고 있음.",
        "built_space": "기울어진 차량 내부 창틀, 깨진 유리조각, 좌측 상단의 실내등이 보이며, 창밖 배경에는 주유소와 쓰레기 더미가 위치함.",
        "entities": "태진(좌)은 작업복과 모자, 은영(우)은 검은 티와 멜빵바지를 착용해 참고 이미지와 일치함. 배경의 SK 주유소 구조물도 확인됨.",
        "hard_violations": [
         "은영의 멜빵바지 주머니에 실제 공구가 아닌 하얀색 평면 렌치 그래픽이 인쇄되어 있어 재질 사실성 규정 위반."
        ],
        "physics": "외부에서 차 안을 들여다보기 위해 몸을 숙인 자세가 자연스럽고 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 창문 밖에서 차량 내부의 아래쪽을 향해 시선을 고정하고 있음.",
        "built_space": "차량 내부의 기울어진 프레임과 실내등이 존재하며, 창 너머로 참고 사진의 주유소와 쓰레기장이 적절한 비율로 배치됨.",
        "entities": "태진과 은영의 인상착의 및 의상이 참고 자료와 정확히 일치하며, 의상의 공구 디테일도 입체적으로 묘사됨.",
        "hard_violations": [],
        "physics": "기울어진 창틀 너머로 허리를 굽히고 안을 살피는 자세가 물리적 모순 없이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "차 안에서 비스듬히 올려다보는 구도와 두 여성의 배치는 맞지만, 배경 주유소가 구조 참조의 은회색 곡면 캐노피·짙은 외벽과 다르고 흐릿한 실루엣 표현도 약합니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "깨진 창 너머 서로 다른 각도로 숙여 보는 두 여성과 낮은 주관 시점을 구현하고, 구조 참조의 주유소까지 더 충실히 유지하지만 얼굴은 요구보다 다소 또렷합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 태진과 오른쪽 은영 모두 고개와 시선을 창 아래의 카메라, 즉 화면 밖 이현우의 위치로 향한다. 태진은 왼팔을 창 왼쪽 방향으로 뻗고, 은영은 상체를 앞으로 숙인다. 둘 다 똑바로 서서 정면을 제시하는 자세는 아니다.",
        "built_space": "전경에 큰 깨진 창 개구부 하나, 그 가장자리의 유리와 검은 고무 몰딩, 왼쪽 위 실내등 하나, 왼쪽 아래 스피커 그릴 하나와 좌석 일부가 보인다. 두 여성은 창틀 밖에 있고 카메라는 어두운 실내에서 비스듬히 위를 본다. 쓰레기 둔덕과 먼 주유소는 위치 참조에 가깝지만, 주유소는 작은 직선형 적백색 캐노피와 밝은 건물로 보여 구조 참조의 은회색 곡면 캐노피와 짙은 외벽을 따르지 않는다. 반사는 보이지 않는다.",
        "entities": "여성은 정확히 두 명이며 이현우나 추가 인물은 없다. 왼쪽은 검은 단발과 모자, 오염된 짙은 파란 작업복의 태진이고 오른쪽은 검은 단발, 검은 반팔과 갈색 멜빵의 데님 작업복을 입은 은영으로 읽힌다. 두 사람 모두 참조와 대체로 부합하는 30대 안팎의 동아시아계 여성 외형이다. 연기와 따뜻한 실내등, 낮의 하늘이 보인다. 얼굴과 옷의 정보가 상당히 남아 있어 초점을 잃은 부드러운 실루엣 효과는 제한적이다.",
        "hard_violations": [],
        "physics": "두 사람의 하체와 발은 창틀에 가려 지면 접촉을 직접 확인할 수 없지만, 밖에서 허리를 굽혀 들여다보는 자세는 자연스럽다. 태진의 뻗은 팔 끝은 가려져 지지 접촉을 확정할 수 없고, 은영의 내려간 팔은 하단 창틀 부근에 닿는 듯 보인다. 공중에 뜬 몸으로 보이지 않는다. 유리 파편은 창틀에 붙거나 아래 홈에 놓여 있고 실내등은 내장재에 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "태진은 개구부 왼쪽 위에서 몸을 기울여 아래의 이현우 쪽을 보고, 은영은 더 오른쪽에서 머리를 낮추고 상체를 더 깊게 숙여 같은 실내 지점을 살핀다. 두 사람의 기울기와 머리 높이가 달라 구조 요청의 순간으로 읽힌다.",
        "built_space": "깨진 창 개구부 하나가 대각선 경계로 유지되고, 왼쪽 위 실내등 하나와 왼쪽 아래 스피커 그릴 하나, 좌석 일부가 보인다. 카메라는 창보다 낮은 실내에 있으며 여성 둘은 창 경계 밖에 있다. 바깥에는 쓰레기 둔덕과 주유소가 보이고, 주유소의 은회색 둥근 캐노피, 적색·주황색 띠, 기둥과 짙은 회색 건물이 구조 참조에 더 가깝다. 배경 일부는 인물에 가려져 전체 주유기 수를 확정할 수 없다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "태진과 은영으로 읽히는 여성 두 명만 보이고 이현우는 완전히 화면 밖이다. 태진의 모자와 검은 단발, 기름때 묻은 짙은 작업복, 은영의 검은 단발과 검은 반팔·데님 멜빵 작업복이 참조에 부합한다. 얼굴은 두 참조 인물의 연령대와 외형에 대체로 맞으며 변신이나 비정상적인 눈 표현은 없다. 실내 연기, 작은 실내등, 흐린 낮빛이 보인다. 인물은 창틀보다 부드럽지만 요구한 흐릿한 실루엣에 비하면 얼굴 식별성이 높다.",
        "hard_violations": [],
        "physics": "두 여성은 창 밖에서 상체를 숙이고 있으며 발과 지면은 프레임 아래에 가려져 있다. 은영의 팔은 하단 창 가장자리 쪽으로 내려가고 손의 접촉점은 가려져 있으며, 태진도 하체가 창 밖 아래로 이어지는 정상적인 굽힘 자세다. 보이는 범위에서 부유하거나 불가능하게 꺾인 신체는 없다. 파편은 창 가장자리와 하단 홈에 지지되고, 실내등은 패널에 고정되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "차 안에서 비스듬히 올려다보는 구도와 두 여성의 배치는 맞지만, 배경 주유소가 구조 참조의 은회색 곡면 캐노피·짙은 외벽과 다르고 흐릿한 실루엣 표현도 약합니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "깨진 창 너머 서로 다른 각도로 숙여 보는 두 여성과 낮은 주관 시점을 구현하고, 구조 참조의 주유소까지 더 충실히 유지하지만 얼굴은 요구보다 다소 또렷합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 태진과 오른쪽 은영 모두 고개와 시선을 창 아래의 카메라, 즉 화면 밖 이현우의 위치로 향한다. 태진은 왼팔을 창 왼쪽 방향으로 뻗고, 은영은 상체를 앞으로 숙인다. 둘 다 똑바로 서서 정면을 제시하는 자세는 아니다.",
        "built_space": "전경에 큰 깨진 창 개구부 하나, 그 가장자리의 유리와 검은 고무 몰딩, 왼쪽 위 실내등 하나, 왼쪽 아래 스피커 그릴 하나와 좌석 일부가 보인다. 두 여성은 창틀 밖에 있고 카메라는 어두운 실내에서 비스듬히 위를 본다. 쓰레기 둔덕과 먼 주유소는 위치 참조에 가깝지만, 주유소는 작은 직선형 적백색 캐노피와 밝은 건물로 보여 구조 참조의 은회색 곡면 캐노피와 짙은 외벽을 따르지 않는다. 반사는 보이지 않는다.",
        "entities": "여성은 정확히 두 명이며 이현우나 추가 인물은 없다. 왼쪽은 검은 단발과 모자, 오염된 짙은 파란 작업복의 태진이고 오른쪽은 검은 단발, 검은 반팔과 갈색 멜빵의 데님 작업복을 입은 은영으로 읽힌다. 두 사람 모두 참조와 대체로 부합하는 30대 안팎의 동아시아계 여성 외형이다. 연기와 따뜻한 실내등, 낮의 하늘이 보인다. 얼굴과 옷의 정보가 상당히 남아 있어 초점을 잃은 부드러운 실루엣 효과는 제한적이다.",
        "hard_violations": [],
        "physics": "두 사람의 하체와 발은 창틀에 가려 지면 접촉을 직접 확인할 수 없지만, 밖에서 허리를 굽혀 들여다보는 자세는 자연스럽다. 태진의 뻗은 팔 끝은 가려져 지지 접촉을 확정할 수 없고, 은영의 내려간 팔은 하단 창틀 부근에 닿는 듯 보인다. 공중에 뜬 몸으로 보이지 않는다. 유리 파편은 창틀에 붙거나 아래 홈에 놓여 있고 실내등은 내장재에 고정되어 있다."
       },
       {
        "label": "A",
        "direction": "태진은 개구부 왼쪽 위에서 몸을 기울여 아래의 이현우 쪽을 보고, 은영은 더 오른쪽에서 머리를 낮추고 상체를 더 깊게 숙여 같은 실내 지점을 살핀다. 두 사람의 기울기와 머리 높이가 달라 구조 요청의 순간으로 읽힌다.",
        "built_space": "깨진 창 개구부 하나가 대각선 경계로 유지되고, 왼쪽 위 실내등 하나와 왼쪽 아래 스피커 그릴 하나, 좌석 일부가 보인다. 카메라는 창보다 낮은 실내에 있으며 여성 둘은 창 경계 밖에 있다. 바깥에는 쓰레기 둔덕과 주유소가 보이고, 주유소의 은회색 둥근 캐노피, 적색·주황색 띠, 기둥과 짙은 회색 건물이 구조 참조에 더 가깝다. 배경 일부는 인물에 가려져 전체 주유기 수를 확정할 수 없다. 불가능한 반사나 중복 설비는 보이지 않는다.",
        "entities": "태진과 은영으로 읽히는 여성 두 명만 보이고 이현우는 완전히 화면 밖이다. 태진의 모자와 검은 단발, 기름때 묻은 짙은 작업복, 은영의 검은 단발과 검은 반팔·데님 멜빵 작업복이 참조에 부합한다. 얼굴은 두 참조 인물의 연령대와 외형에 대체로 맞으며 변신이나 비정상적인 눈 표현은 없다. 실내 연기, 작은 실내등, 흐린 낮빛이 보인다. 인물은 창틀보다 부드럽지만 요구한 흐릿한 실루엣에 비하면 얼굴 식별성이 높다.",
        "hard_violations": [],
        "physics": "두 여성은 창 밖에서 상체를 숙이고 있으며 발과 지면은 프레임 아래에 가려져 있다. 은영의 팔은 하단 창 가장자리 쪽으로 내려가고 손의 접촉점은 가려져 있으며, 태진도 하체가 창 밖 아래로 이어지는 정상적인 굽힘 자세다. 보이는 범위에서 부유하거나 불가능하게 꺾인 신체는 없다. 파편은 창 가장자리와 하단 홈에 지지되고, 실내등은 패널에 고정되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.607,
    "B": 1.778
   },
   "violations": {
    "A": [
     "[gemini-pro] 은영의 멜빵바지 주머니에 실제 공구가 아닌 하얀색 평면 렌치 그래픽이 인쇄되어 있어 재질 사실성 규정 위반."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1778,
   "A": 1607
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1778,
    "verdict_ko": "요구된 어둡고 흐릿한 조명 분위기에 조금 더 부합하며, 인물들의 의상과 장비 디테일이 사실적으로 구현되어 프롬프트에 충실함."
   },
   {
    "label": "A",
    "score": 1607,
    "verdict_ko": "구도와 시선 처리는 훌륭하지만, 요구된 실루엣과 흐릿한 초점 효과가 부족하고 은영의 멜빵에 그려진 평면적인 그래픽이 사실성을 떨어뜨림.  ★위반: [gemini-pro] 은영의 멜빵바지 주머니에 실제 공구가 아닌 하얀색 평면 렌치 그래픽이 인쇄되어 있어 재질 사실성 규정 위반."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L67B01.png",
    "asset_id": "a9202095-b456-4be5-85cc-eab29211f7df",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_roadside_service_station_sel.png",
    "asset_id": "c3c4d03b-d9f9-4bfc-bc05-d962db6ab9f2",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:769818>",
    "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 은영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:981336>",
    "asset_id": "c128fa4a-1680-4e5b-a9d7-a21abf9186af",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b6a-2523-7337-b175-388610f0024b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S49sh47__bgfirst_bg.png",
   "bg_asset_id": "645528b1-cd2a-4fd6-ba5a-d94f6f395421",
   "bg_record_key": "S49sh47::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S50sh2::signage": {
  "fp": "86dccd47e922cdb4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::865f1ae3d3bda5dc": {
  "subjects": [],
  "subject_text": "카센터 창고·정비 작업 공간\n벽면에 여러 록밴드 포스터가 붙은 창고형 정비 공간. 커다란 스피커와 펜치 등 정비 도구, 수리용 용액 병이 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L70",
  "scope_role": "location_interior",
  "scope_sha": "6061c82d4ecdc95a"
 },
 "S50sh2::bgfirst_bg": {
  "input_fingerprint": "2b15d04d5e887c36",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낯선 방 안에서 상체를 반쯤 일으킨 자세로 자신의 다리 쪽을 멍하니 내려다보고 있는 이현우의 상반신.\n\nLOCATION (lock): In the recovery corner of an auto-repair workshop's storage room, where rock-band posters cover the walls. Daytime light reaches the interior from the workshop opening.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin beside 이현우 and above his half-raised eye line, looking diagonally down across his upper body toward his legs before the dolly-out opens the room. Place his torso left of center and let the lower frame follow his downward gaze toward the bandaged legs just beyond the crop, catching the effort of lifting himself rather than a settled seated pose. Keep a limited patch of poster-covered wall behind him, maintaining attention on his bewildered inspection of his own injuries.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached around the room) — Printed band imagery faces the camera obliquely, without requiring individual text to be readable; used as Provide a restrained unfamiliar-room cue behind the upper body.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and gentle tonal separation keep the awakening intimate and observational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 낯선 방 안에서 상체를 반쯤 일으킨 자세로 자신의 다리 쪽을 멍하니 내려다보고 있는 이현우의 상반신.\n\nLOCATION (lock): In the recovery corner of an auto-repair workshop's storage room, where rock-band posters cover the walls. Daytime light reaches the interior from the workshop opening.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin beside 이현우 and above his half-raised eye line, looking diagonally down across his upper body toward his legs before the dolly-out opens the room. Place his torso left of center and let the lower frame follow his downward gaze toward the bandaged legs just beyond the crop, catching the effort of lifting himself rather than a settled seated pose. Keep a limited patch of poster-covered wall behind him, maintaining attention on his bewildered inspection of his own injuries.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached around the room) — Printed band imagery faces the camera obliquely, without requiring individual text to be readable; used as Provide a restrained unfamiliar-room cue behind the upper body.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and gentle tonal separation keep the awakening intimate and observational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh2__bgfirst_bg.png",
  "asset_id": "a9cb0a55-2f56-4c69-a68f-b5c638b6a8ba",
  "input_asset_ids": [
   "2bc2d940-b53b-41e2-ba58-940827f11039",
   "2445fcea-a876-4f26-a752-65f6630fcd6f"
  ]
 },
 "S50sh2": {
  "input_fingerprint": "75abad573a196ccd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낯선 방 안에서 상체를 반쯤 일으킨 자세로 자신의 다리 쪽을 멍하니 내려다보고 있는 이현우의 상반신.\n\nLOCATION (lock): In the recovery corner of an auto-repair workshop's storage room, where rock-band posters cover the walls. Daytime light reaches the interior from the workshop opening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin beside 이현우 and above his half-raised eye line, looking diagonally down across his upper body toward his legs before the dolly-out opens the room. Place his torso left of center and let the lower frame follow his downward gaze toward the bandaged legs just beyond the crop, catching the effort of lifting himself rather than a settled seated pose. Keep a limited patch of poster-covered wall behind him, maintaining attention on his bewildered inspection of his own injuries.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached around the room) — Printed band imagery faces the camera obliquely, without requiring individual text to be readable; used as Provide a restrained unfamiliar-room cue behind the upper body.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and gentle tonal separation keep the awakening intimate and observational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the room, and a large pair of pliers is within reach. The recovered camper is under repair at the garage, with its engine and brakes badly damaged. 이현우: He has awakened and is raising his torso to inspect his bandaged leg. His body bears visible signs of medical treatment.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낯선 방 안에서 상체를 반쯤 일으킨 자세로 자신의 다리 쪽을 멍하니 내려다보고 있는 이현우의 상반신.\n\nLOCATION (lock): In the recovery corner of an auto-repair workshop's storage room, where rock-band posters cover the walls. Daytime light reaches the interior from the workshop opening. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin beside 이현우 and above his half-raised eye line, looking diagonally down across his upper body toward his legs before the dolly-out opens the room. Place his torso left of center and let the lower frame follow his downward gaze toward the bandaged legs just beyond the crop, catching the effort of lifting himself rather than a settled seated pose. Keep a limited patch of poster-covered wall behind him, maintaining attention on his bewildered inspection of his own injuries.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached around the room) — Printed band imagery faces the camera obliquely, without requiring individual text to be readable; used as Provide a restrained unfamiliar-room cue behind the upper body.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and gentle tonal separation keep the awakening intimate and observational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the room, and a large pair of pliers is within reach. The recovered camper is under repair at the garage, with its engine and brakes badly damaged. 이현우: He has awakened and is raising his torso to inspect his bandaged leg. His body bears visible signs of medical treatment.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낯선 방 안에서 상체를 반쯤 일으킨 자세로 자신의 다리 쪽을 멍하니 내려다보고 있는 이현우의 상반신.\n\nLOCATION (lock): In the recovery corner of an auto-repair workshop's storage room, where rock-band posters cover the walls. Daytime light reaches the interior from the workshop opening. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin beside 이현우 and above his half-raised eye line, looking diagonally down across his upper body toward his legs before the dolly-out opens the room. Place his torso left of center and let the lower frame follow his downward gaze toward the bandaged legs just beyond the crop, catching the effort of lifting himself rather than a settled seated pose. Keep a limited patch of poster-covered wall behind him, maintaining attention on his bewildered inspection of his own injuries.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rock-band posters (Attached around the room) — Printed band imagery faces the camera obliquely, without requiring individual text to be readable; used as Provide a restrained unfamiliar-room cue behind the upper body.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and gentle tonal separation keep the awakening intimate and observational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rock-band posters cover the room, and a large pair of pliers is within reach. The recovered camper is under repair at the garage, with its engine and brakes badly damaged. 이현우: He has awakened and is raising his torso to inspect his bandaged leg. His body bears visible signs of medical treatment.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh2__bgfirst_bg.png",
     "asset_id": "a9cb0a55-2f56-4c69-a68f-b5c638b6a8ba",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S50sh2.png",
     "asset_id": "2bc2d940-b53b-41e2-ba58-940827f11039",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L70B01.png",
     "asset_id": "2445fcea-a876-4f26-a752-65f6630fcd6f",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "고개는 조금 숙였지만 눈은 화면 오른쪽 전방의 작업장 공간을 향한다. 아래쪽에 보이는 자신의 붕대 감긴 다리에 시선이 닿지 않는다. 카메라도 상체를 대각선 아래로 따라 내려다보기보다 옆에서 넓게 관찰한다.",
    "built_space": "왼쪽 벽을 따라 침상 하나가 있고, 인물은 그 가장자리에 걸터앉아 양다리를 바깥으로 내렸다. 뒤쪽에는 베개 하나, 침상 옆 작업대 하나, 높은 수납장과 좁은 수납장 각 하나, 공구 작업대 하나가 보이며 오른쪽에는 캠핑카 한 대의 일부가 보인다. 낡은 벽과 바닥, 수납장 및 작업대의 관계는 장소 참조와 대체로 이어지지만, 포스터 벽과 작업장 전체를 넓게 보여 제한된 상반신 배경이라는 지시에서 벗어난다.",
    "entities": "실제 인물은 한 명이며, 포스터 속 인물들은 인쇄물이다. 짧고 헝클어진 검은 머리의 마른 동아시아계 청년 남성으로 참조 인물과 대체로 부합하고, 귀에 소형 인이어 장치가 있다. 어두운 낡은 셔츠와 흙먼지 묻은 바지, 양쪽 다리의 붕대가 보이나 셔츠의 핏자국은 뚜렷하지 않다. 작업대에는 큰 빨간 손잡이 펜치와 금속 컵이 있다. 포스터에는 참조에 없는 'TUAPUS' 등의 문구가 읽힌다.",
    "hard_violations": [
     "참조에 없는 판독 가능한 밴드명 형태의 포스터 문구를 생성하여, 새로운 문구를 발명하지 말라는 명시적 금지 조건을 위반했다."
    ],
    "physics": "골반과 허벅지는 침상 가장자리에 지지되고, 보이는 손은 담요 위를 짚고 있다. 다리는 무릎에서 굽어 아래로 내려가므로 떠 있는 신체는 없다. 다만 이미 가장자리에 앉아 몸을 앞으로 기울인 상태여서, 누운 상태에서 상체를 반쯤 들어 올리는 동작의 지지 관계로는 읽히지 않는다. 공구와 컵은 작업대 위에 놓여 있다."
   },
   {
    "label": "A",
    "direction": "고개와 눈이 화면 아래 오른쪽의 자신의 붕대 감긴 무릎 쪽으로 향하고, 손도 그 부위에 놓여 있어 부상을 살펴보는 대상이 명확하다. 카메라는 숙인 눈높이보다 높지만, 상체를 가로질러 다리 방향으로 내려다보는 측면 대각선보다는 앞쪽에서 보는 인상이 강하다.",
    "built_space": "왼쪽 포스터 벽을 따라 침상 하나가 있고, 인물의 골반과 뻗은 다리가 침상 위에 놓인다. 뒤에는 베개 하나, 오른쪽에는 바퀴 달린 작업대 하나가 있으며, 배경에 수납장 두 개, 공구 작업대 하나, 캠핑카 한 대와 오른쪽 작업장 개구부가 보인다. 참조의 주요 공간 관계와 낡은 재료는 잘 유지된다. 그러나 캠핑카와 작업장 내부가 화면의 큰 부분을 차지해 제한된 포스터 벽만 남기라는 구도와 다르다.",
    "entities": "실제 인물은 한 명이다. 헝클어진 짧은 검은 머리, 마른 체격과 젊은 동아시아계 남성의 외형이 참조와 대체로 맞지만, 숙인 얼굴 때문에 세부 동일성 판단은 제한된다. 귀의 인이어 장치, 핏자국과 먼지가 묻은 어두운 셔츠 및 바지가 확인된다. 한쪽 무릎 주변에 피가 밴 붕대가 있고, 손이 닿을 만한 옆 작업대에는 큰 빨간 손잡이 펜치가 있다. 록밴드 포스터와 수리 중인 캠핑카도 보인다.",
    "hard_violations": [],
    "physics": "골반과 다리는 매트리스가 받치고, 화면 왼쪽 손바닥은 침상을 눌러 상체를 지지한다. 다른 손은 붕대 감긴 무릎 위에 자연스럽게 놓여 있다. 지지 없는 신체나 물체는 없다. 한 손으로 몸을 지탱하는 긴장은 보이지만, 상체는 이미 상당히 세워져 있어 반쯤 일어나는 순간보다는 앉아서 부상을 확인하는 자세에 가깝다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "시선이 자신의 다리가 아닌 오른쪽 전방을 향하고, 침상 가장자리에 이미 앉은 넓은 구도이며, 참조에 없는 포스터 문구까지 생성했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "자신의 붕대 감긴 다리를 내려다보는 시선과 왼쪽 상체 배치는 더 정확하지만, 반쯤 몸을 일으키는 순간보다 앉아 있는 자세이며 다리와 작업장 배경을 지나치게 드러낸다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개는 조금 숙였지만 눈은 화면 오른쪽 전방의 작업장 공간을 향한다. 아래쪽에 보이는 자신의 붕대 감긴 다리에 시선이 닿지 않는다. 카메라도 상체를 대각선 아래로 따라 내려다보기보다 옆에서 넓게 관찰한다.",
        "built_space": "왼쪽 벽을 따라 침상 하나가 있고, 인물은 그 가장자리에 걸터앉아 양다리를 바깥으로 내렸다. 뒤쪽에는 베개 하나, 침상 옆 작업대 하나, 높은 수납장과 좁은 수납장 각 하나, 공구 작업대 하나가 보이며 오른쪽에는 캠핑카 한 대의 일부가 보인다. 낡은 벽과 바닥, 수납장 및 작업대의 관계는 장소 참조와 대체로 이어지지만, 포스터 벽과 작업장 전체를 넓게 보여 제한된 상반신 배경이라는 지시에서 벗어난다.",
        "entities": "실제 인물은 한 명이며, 포스터 속 인물들은 인쇄물이다. 짧고 헝클어진 검은 머리의 마른 동아시아계 청년 남성으로 참조 인물과 대체로 부합하고, 귀에 소형 인이어 장치가 있다. 어두운 낡은 셔츠와 흙먼지 묻은 바지, 양쪽 다리의 붕대가 보이나 셔츠의 핏자국은 뚜렷하지 않다. 작업대에는 큰 빨간 손잡이 펜치와 금속 컵이 있다. 포스터에는 참조에 없는 'TUAPUS' 등의 문구가 읽힌다.",
        "hard_violations": [
         "참조에 없는 판독 가능한 밴드명 형태의 포스터 문구를 생성하여, 새로운 문구를 발명하지 말라는 명시적 금지 조건을 위반했다."
        ],
        "physics": "골반과 허벅지는 침상 가장자리에 지지되고, 보이는 손은 담요 위를 짚고 있다. 다리는 무릎에서 굽어 아래로 내려가므로 떠 있는 신체는 없다. 다만 이미 가장자리에 앉아 몸을 앞으로 기울인 상태여서, 누운 상태에서 상체를 반쯤 들어 올리는 동작의 지지 관계로는 읽히지 않는다. 공구와 컵은 작업대 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "고개와 눈이 화면 아래 오른쪽의 자신의 붕대 감긴 무릎 쪽으로 향하고, 손도 그 부위에 놓여 있어 부상을 살펴보는 대상이 명확하다. 카메라는 숙인 눈높이보다 높지만, 상체를 가로질러 다리 방향으로 내려다보는 측면 대각선보다는 앞쪽에서 보는 인상이 강하다.",
        "built_space": "왼쪽 포스터 벽을 따라 침상 하나가 있고, 인물의 골반과 뻗은 다리가 침상 위에 놓인다. 뒤에는 베개 하나, 오른쪽에는 바퀴 달린 작업대 하나가 있으며, 배경에 수납장 두 개, 공구 작업대 하나, 캠핑카 한 대와 오른쪽 작업장 개구부가 보인다. 참조의 주요 공간 관계와 낡은 재료는 잘 유지된다. 그러나 캠핑카와 작업장 내부가 화면의 큰 부분을 차지해 제한된 포스터 벽만 남기라는 구도와 다르다.",
        "entities": "실제 인물은 한 명이다. 헝클어진 짧은 검은 머리, 마른 체격과 젊은 동아시아계 남성의 외형이 참조와 대체로 맞지만, 숙인 얼굴 때문에 세부 동일성 판단은 제한된다. 귀의 인이어 장치, 핏자국과 먼지가 묻은 어두운 셔츠 및 바지가 확인된다. 한쪽 무릎 주변에 피가 밴 붕대가 있고, 손이 닿을 만한 옆 작업대에는 큰 빨간 손잡이 펜치가 있다. 록밴드 포스터와 수리 중인 캠핑카도 보인다.",
        "hard_violations": [],
        "physics": "골반과 다리는 매트리스가 받치고, 화면 왼쪽 손바닥은 침상을 눌러 상체를 지지한다. 다른 손은 붕대 감긴 무릎 위에 자연스럽게 놓여 있다. 지지 없는 신체나 물체는 없다. 한 손으로 몸을 지탱하는 긴장은 보이지만, 상체는 이미 상당히 세워져 있어 반쯤 일어나는 순간보다는 앉아서 부상을 확인하는 자세에 가깝다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "시선이 자신의 다리가 아닌 오른쪽 전방을 향하고, 침상 가장자리에 이미 앉은 넓은 구도이며, 참조에 없는 포스터 문구까지 생성했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "자신의 붕대 감긴 다리를 내려다보는 시선과 왼쪽 상체 배치는 더 정확하지만, 반쯤 몸을 일으키는 순간보다 앉아 있는 자세이며 다리와 작업장 배경을 지나치게 드러낸다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개는 조금 숙였지만 눈은 화면 오른쪽 전방의 작업장 공간을 향한다. 아래쪽에 보이는 자신의 붕대 감긴 다리에 시선이 닿지 않는다. 카메라도 상체를 대각선 아래로 따라 내려다보기보다 옆에서 넓게 관찰한다.",
        "built_space": "왼쪽 벽을 따라 침상 하나가 있고, 인물은 그 가장자리에 걸터앉아 양다리를 바깥으로 내렸다. 뒤쪽에는 베개 하나, 침상 옆 작업대 하나, 높은 수납장과 좁은 수납장 각 하나, 공구 작업대 하나가 보이며 오른쪽에는 캠핑카 한 대의 일부가 보인다. 낡은 벽과 바닥, 수납장 및 작업대의 관계는 장소 참조와 대체로 이어지지만, 포스터 벽과 작업장 전체를 넓게 보여 제한된 상반신 배경이라는 지시에서 벗어난다.",
        "entities": "실제 인물은 한 명이며, 포스터 속 인물들은 인쇄물이다. 짧고 헝클어진 검은 머리의 마른 동아시아계 청년 남성으로 참조 인물과 대체로 부합하고, 귀에 소형 인이어 장치가 있다. 어두운 낡은 셔츠와 흙먼지 묻은 바지, 양쪽 다리의 붕대가 보이나 셔츠의 핏자국은 뚜렷하지 않다. 작업대에는 큰 빨간 손잡이 펜치와 금속 컵이 있다. 포스터에는 참조에 없는 'TUAPUS' 등의 문구가 읽힌다.",
        "hard_violations": [
         "참조에 없는 판독 가능한 밴드명 형태의 포스터 문구를 생성하여, 새로운 문구를 발명하지 말라는 명시적 금지 조건을 위반했다."
        ],
        "physics": "골반과 허벅지는 침상 가장자리에 지지되고, 보이는 손은 담요 위를 짚고 있다. 다리는 무릎에서 굽어 아래로 내려가므로 떠 있는 신체는 없다. 다만 이미 가장자리에 앉아 몸을 앞으로 기울인 상태여서, 누운 상태에서 상체를 반쯤 들어 올리는 동작의 지지 관계로는 읽히지 않는다. 공구와 컵은 작업대 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "고개와 눈이 화면 아래 오른쪽의 자신의 붕대 감긴 무릎 쪽으로 향하고, 손도 그 부위에 놓여 있어 부상을 살펴보는 대상이 명확하다. 카메라는 숙인 눈높이보다 높지만, 상체를 가로질러 다리 방향으로 내려다보는 측면 대각선보다는 앞쪽에서 보는 인상이 강하다.",
        "built_space": "왼쪽 포스터 벽을 따라 침상 하나가 있고, 인물의 골반과 뻗은 다리가 침상 위에 놓인다. 뒤에는 베개 하나, 오른쪽에는 바퀴 달린 작업대 하나가 있으며, 배경에 수납장 두 개, 공구 작업대 하나, 캠핑카 한 대와 오른쪽 작업장 개구부가 보인다. 참조의 주요 공간 관계와 낡은 재료는 잘 유지된다. 그러나 캠핑카와 작업장 내부가 화면의 큰 부분을 차지해 제한된 포스터 벽만 남기라는 구도와 다르다.",
        "entities": "실제 인물은 한 명이다. 헝클어진 짧은 검은 머리, 마른 체격과 젊은 동아시아계 남성의 외형이 참조와 대체로 맞지만, 숙인 얼굴 때문에 세부 동일성 판단은 제한된다. 귀의 인이어 장치, 핏자국과 먼지가 묻은 어두운 셔츠 및 바지가 확인된다. 한쪽 무릎 주변에 피가 밴 붕대가 있고, 손이 닿을 만한 옆 작업대에는 큰 빨간 손잡이 펜치가 있다. 록밴드 포스터와 수리 중인 캠핑카도 보인다.",
        "hard_violations": [],
        "physics": "골반과 다리는 매트리스가 받치고, 화면 왼쪽 손바닥은 침상을 눌러 상체를 지지한다. 다른 손은 붕대 감긴 무릎 위에 자연스럽게 놓여 있다. 지지 없는 신체나 물체는 없다. 한 손으로 몸을 지탱하는 긴장은 보이지만, 상체는 이미 상당히 세워져 있어 반쯤 일어나는 순간보다는 앉아서 부상을 확인하는 자세에 가깝다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2,
   "A": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "시선이 자신의 다리가 아닌 오른쪽 전방을 향하고, 침상 가장자리에 이미 앉은 넓은 구도이며, 참조에 없는 포스터 문구까지 생성했다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "자신의 붕대 감긴 다리를 내려다보는 시선과 왼쪽 상체 배치는 더 정확하지만, 반쯤 몸을 일으키는 순간보다 앉아 있는 자세이며 다리와 작업장 배경을 지나치게 드러낸다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L70B01.png",
    "asset_id": "2445fcea-a876-4f26-a752-65f6630fcd6f",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b71-a8db-750f-9b3d-bb29100d2e85",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh2__bgfirst_bg.png",
   "bg_asset_id": "a9cb0a55-2f56-4c69-a68f-b5c638b6a8ba",
   "bg_record_key": "S50sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh7::signage": {
  "fp": "f89ee0a809520b2f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh7::bgfirst_bg": {
  "input_fingerprint": "b2f253ef74bc2df7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 펜치를 잡으려는 이현우를 향해 고개를 돌린 채 아니꼽다는 듯 비웃는 표정을 짓고 있는 태진의 상체.\n\nLOCATION (lock): Beside the damaged camper in the auto shop's repair bay, facing the adjoining recovery area. Daylight enters through the workshop entrance.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut on 이현우's reach, hold beside the vehicle at 태진's upper-chest height, slightly off the line between them and looking gently upward. Frame 태진's upper body on the right, her torso still oriented toward the repair while her head turns left toward 이현우 reaching for the pliers off-screen. Keep a modest section of the vehicle at the lower edge and negative space toward him, allowing the dismissive expression—not an additional camera move—to carry the interruption.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Damaged and being repaired) — Only an oblique portion beside 태진 is visible along the lower edge; used as Anchor her interrupted task without exposing an invented repair configuration.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained ambient illumination, with enough facial separation to read 태진's dismissive expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 펜치를 잡으려는 이현우를 향해 고개를 돌린 채 아니꼽다는 듯 비웃는 표정을 짓고 있는 태진의 상체.\n\nLOCATION (lock): Beside the damaged camper in the auto shop's repair bay, facing the adjoining recovery area. Daylight enters through the workshop entrance.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut on 이현우's reach, hold beside the vehicle at 태진's upper-chest height, slightly off the line between them and looking gently upward. Frame 태진's upper body on the right, her torso still oriented toward the repair while her head turns left toward 이현우 reaching for the pliers off-screen. Keep a modest section of the vehicle at the lower edge and negative space toward him, allowing the dismissive expression—not an additional camera move—to carry the interruption.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Damaged and being repaired) — Only an oblique portion beside 태진 is visible along the lower edge; used as Anchor her interrupted task without exposing an invented repair configuration.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained ambient illumination, with enough facial separation to read 태진's dismissive expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh7__bgfirst_bg.png",
  "asset_id": "ba89ad40-df97-4eaf-bd86-34cc7ee6ae42",
  "input_asset_ids": [
   "59c59c56-933a-4ce9-8d9e-2d34328eab69",
   "2445fcea-a876-4f26-a752-65f6630fcd6f"
  ]
 },
 "S50sh7": {
  "input_fingerprint": "1c8adc1eb9ba5733",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 펜치를 잡으려는 이현우를 향해 고개를 돌린 채 아니꼽다는 듯 비웃는 표정을 짓고 있는 태진의 상체.\n\nLOCATION (lock): Beside the damaged camper in the auto shop's repair bay, facing the adjoining recovery area. Daylight enters through the workshop entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut on 이현우's reach, hold beside the vehicle at 태진's upper-chest height, slightly off the line between them and looking gently upward. Frame 태진's upper body on the right, her torso still oriented toward the repair while her head turns left toward 이현우 reaching for the pliers off-screen. Keep a modest section of the vehicle at the lower edge and negative space toward him, allowing the dismissive expression—not an additional camera move—to carry the interruption.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Damaged and being repaired) — Only an oblique portion beside 태진 is visible along the lower edge; used as Anchor her interrupted task without exposing an invented repair configuration.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained ambient illumination, with enough facial separation to read 태진's dismissive expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rock-band posters and large pliers remain in the room. The camper is still under repair, with extensive engine and brake damage, beside a large music speaker. 태진: She remains at the camper repair area, with her bobbed hair unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 펜치를 잡으려는 이현우를 향해 고개를 돌린 채 아니꼽다는 듯 비웃는 표정을 짓고 있는 태진의 상체.\n\nLOCATION (lock): Beside the damaged camper in the auto shop's repair bay, facing the adjoining recovery area. Daylight enters through the workshop entrance. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut on 이현우's reach, hold beside the vehicle at 태진's upper-chest height, slightly off the line between them and looking gently upward. Frame 태진's upper body on the right, her torso still oriented toward the repair while her head turns left toward 이현우 reaching for the pliers off-screen. Keep a modest section of the vehicle at the lower edge and negative space toward him, allowing the dismissive expression—not an additional camera move—to carry the interruption.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Damaged and being repaired) — Only an oblique portion beside 태진 is visible along the lower edge; used as Anchor her interrupted task without exposing an invented repair configuration.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained ambient illumination, with enough facial separation to read 태진's dismissive expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rock-band posters and large pliers remain in the room. The camper is still under repair, with extensive engine and brake damage, beside a large music speaker. 태진: She remains at the camper repair area, with her bobbed hair unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 펜치를 잡으려는 이현우를 향해 고개를 돌린 채 아니꼽다는 듯 비웃는 표정을 짓고 있는 태진의 상체.\n\nLOCATION (lock): Beside the damaged camper in the auto shop's repair bay, facing the adjoining recovery area. Daylight enters through the workshop entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit cut on 이현우's reach, hold beside the vehicle at 태진's upper-chest height, slightly off the line between them and looking gently upward. Frame 태진's upper body on the right, her torso still oriented toward the repair while her head turns left toward 이현우 reaching for the pliers off-screen. Keep a modest section of the vehicle at the lower edge and negative space toward him, allowing the dismissive expression—not an additional camera move—to carry the interruption.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Damaged and being repaired) — Only an oblique portion beside 태진 is visible along the lower edge; used as Anchor her interrupted task without exposing an invented repair configuration.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established restrained ambient illumination, with enough facial separation to read 태진's dismissive expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rock-band posters and large pliers remain in the room. The camper is still under repair, with extensive engine and brake damage, beside a large music speaker. 태진: She remains at the camper repair area, with her bobbed hair unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 태진 (한국인 여성, 30대 초반의 얼굴, 검은 단발머리) — wearing: 짙은 파란색 캔버스 재질의 튼튼한 커버올 작업복으로 기름때가 잔뜩 묻어 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh7__bgfirst_bg.png",
     "asset_id": "ba89ad40-df97-4eaf-bd86-34cc7ee6ae42",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S50sh7.png",
     "asset_id": "59c59c56-933a-4ce9-8d9e-2d34328eab69",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:769818>",
     "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L70B01.png",
     "asset_id": "2445fcea-a876-4f26-a752-65f6630fcd6f",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:769818>",
     "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "태진의 고개는 화면 왼쪽 오프스크린을 향해 비웃는 표정을 짓고 있으며, 상체는 우측 하단의 수리 영역을 향해 있음.",
    "built_space": "카메라는 지시대로 좌측의 휴게 공간(포스터 벽)을 향하고 있음. 다만 전경에 금속 테이블이 위치함.",
    "entities": "태진의 인상착의(작업복, 모자, 단발머리)는 참조와 일치함. 벽면 포스터의 한글 텍스트도 정확히 일치함. 그러나 캠퍼 측면과 별개의 자동차 엔진룸이 동시에 존재함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에서 명시적으로 금지한 '창작된 수리 형태(노출된 엔진룸)' 등장 및 추가 차량(엔진룸) 혼합",
     "[gemini-pro] 테이블 위 펜치(플라이어) 중복 생성"
    ],
    "physics": "오른팔을 하단 엔진룸 쪽으로 내린 채 서 있으며 체중 지지에 어색함이 없음."
   },
   {
    "label": "B",
    "direction": "태진의 고개는 왼쪽을 향해 비웃는 듯한 표정을 짓고 있으며, 상체는 우측에 있는 캠퍼 정면을 향함.",
    "built_space": "지정된 휴게 공간(포스터 벽)이 아닌 후면 공구 벽과 열린 문을 향해 카메라가 배치되어 공간 방향이 틀림. 캠퍼의 주차 방향도 참조 사진과 90도 다름.",
    "entities": "태진의 착장과 얼굴은 참조와 일치하나, 캠퍼는 오버캡이 없는 일반 밴 형태로 변경됨.",
    "hard_violations": [
     "[gemini-pro] 캠퍼가 원본 참조 사진의 공간과 어긋나게 90도 회전되어 배치된 공간 왜곡"
    ],
    "physics": "차량 옆에 서 있는 자세로 발 아래 지지는 보이지 않으나 상체 자세에 물리적 오류는 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 프레이밍(우측 배치, 좌측 여백)과 포스터의 한글 텍스트까지 완벽히 구현했으나, 프롬프트에서 금지한 '창작된 수리 형태(엔진룸)'를 노출하고 두 대의 차량이 섞인 듯한 물리적 오류가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물을 우측에 배치하라는 프레이밍 지시를 반대로 적용(좌측 배치)하였으며, 카메라 방향도 휴게 공간이 아닌 공구 벽을 향하고 있어 연출과 공간 구성을 모두 실패함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "태진의 고개는 화면 왼쪽 오프스크린을 향해 비웃는 표정을 짓고 있으며, 상체는 우측 하단의 수리 영역을 향해 있음.",
        "built_space": "카메라는 지시대로 좌측의 휴게 공간(포스터 벽)을 향하고 있음. 다만 전경에 금속 테이블이 위치함.",
        "entities": "태진의 인상착의(작업복, 모자, 단발머리)는 참조와 일치함. 벽면 포스터의 한글 텍스트도 정확히 일치함. 그러나 캠퍼 측면과 별개의 자동차 엔진룸이 동시에 존재함.",
        "hard_violations": [
         "프롬프트에서 명시적으로 금지한 '창작된 수리 형태(노출된 엔진룸)' 등장 및 추가 차량(엔진룸) 혼합",
         "테이블 위 펜치(플라이어) 중복 생성"
        ],
        "physics": "오른팔을 하단 엔진룸 쪽으로 내린 채 서 있으며 체중 지지에 어색함이 없음."
       },
       {
        "label": "B",
        "direction": "태진의 고개는 왼쪽을 향해 비웃는 듯한 표정을 짓고 있으며, 상체는 우측에 있는 캠퍼 정면을 향함.",
        "built_space": "지정된 휴게 공간(포스터 벽)이 아닌 후면 공구 벽과 열린 문을 향해 카메라가 배치되어 공간 방향이 틀림. 캠퍼의 주차 방향도 참조 사진과 90도 다름.",
        "entities": "태진의 착장과 얼굴은 참조와 일치하나, 캠퍼는 오버캡이 없는 일반 밴 형태로 변경됨.",
        "hard_violations": [
         "캠퍼가 원본 참조 사진의 공간과 어긋나게 90도 회전되어 배치된 공간 왜곡"
        ],
        "physics": "차량 옆에 서 있는 자세로 발 아래 지지는 보이지 않으나 상체 자세에 물리적 오류는 없음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 프레이밍(우측 배치, 좌측 여백)과 포스터의 한글 텍스트까지 완벽히 구현했으나, 프롬프트에서 금지한 '창작된 수리 형태(엔진룸)'를 노출하고 두 대의 차량이 섞인 듯한 물리적 오류가 발생함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물을 우측에 배치하라는 프레이밍 지시를 반대로 적용(좌측 배치)하였으며, 카메라 방향도 휴게 공간이 아닌 공구 벽을 향하고 있어 연출과 공간 구성을 모두 실패함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "태진의 고개는 화면 왼쪽 오프스크린을 향해 비웃는 표정을 짓고 있으며, 상체는 우측 하단의 수리 영역을 향해 있음.",
        "built_space": "카메라는 지시대로 좌측의 휴게 공간(포스터 벽)을 향하고 있음. 다만 전경에 금속 테이블이 위치함.",
        "entities": "태진의 인상착의(작업복, 모자, 단발머리)는 참조와 일치함. 벽면 포스터의 한글 텍스트도 정확히 일치함. 그러나 캠퍼 측면과 별개의 자동차 엔진룸이 동시에 존재함.",
        "hard_violations": [
         "프롬프트에서 명시적으로 금지한 '창작된 수리 형태(노출된 엔진룸)' 등장 및 추가 차량(엔진룸) 혼합",
         "테이블 위 펜치(플라이어) 중복 생성"
        ],
        "physics": "오른팔을 하단 엔진룸 쪽으로 내린 채 서 있으며 체중 지지에 어색함이 없음."
       },
       {
        "label": "B",
        "direction": "태진의 고개는 왼쪽을 향해 비웃는 듯한 표정을 짓고 있으며, 상체는 우측에 있는 캠퍼 정면을 향함.",
        "built_space": "지정된 휴게 공간(포스터 벽)이 아닌 후면 공구 벽과 열린 문을 향해 카메라가 배치되어 공간 방향이 틀림. 캠퍼의 주차 방향도 참조 사진과 90도 다름.",
        "entities": "태진의 착장과 얼굴은 참조와 일치하나, 캠퍼는 오버캡이 없는 일반 밴 형태로 변경됨.",
        "hard_violations": [
         "캠퍼가 원본 참조 사진의 공간과 어긋나게 90도 회전되어 배치된 공간 왜곡"
        ],
        "physics": "차량 옆에 서 있는 자세로 발 아래 지지는 보이지 않으나 상체 자세에 물리적 오류는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "오른쪽 상체 배치와 왼쪽 곁눈질은 맞지만, 캠퍼가 화면 오른쪽을 크게 차지하고 회복 구역을 향한 시점과 고개를 돌려 비웃는 연기가 약하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수리 쪽에 남은 몸통과 왼쪽으로 돌아간 얼굴, 아니꼬운 표정 및 장소 재현이 더 정확하지만, 캠퍼 측면과 노출된 정비 부위를 지나치게 많이 보여준다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "태진의 몸통과 두 팔은 오른쪽 아래 차량 작업 부위를 향하고, 눈은 화면 밖 왼쪽을 본다. 이현우를 바라보는 방향은 맞지만 얼굴 자체의 회전은 작고 곁눈질이 더 두드러진다. 입술을 비틀고 있어 불쾌함은 읽히나 비웃음은 비교적 약하다.",
        "built_space": "왼쪽에 큰 양문 수납장 하나와 좁은 수납장 하나, 뒤쪽에 공구판과 작업대 각 하나, 바퀴 달린 의자 하나, 입구 근처에 세로형 스피커 하나가 보인다. 낡은 콘크리트와 수납 설비는 참조 장소와 유사하다. 다만 회복 구역 대신 공구 작업대와 출입구를 배경으로 삼았고, 캠퍼의 창문·거울·차체가 오른쪽 가장자리를 위아래로 크게 채워 하단의 작은 사선 부분만 보이라는 구도를 따르지 않는다.",
        "entities": "보이는 사람은 태진 한 명뿐이며, 참조와 유사한 성인 한국인 여성의 얼굴, 검은 단발, 남색 작업모와 때 묻은 짙은 남색 작업복이다. 캠퍼와 큰 스피커는 보인다. 포스터와 큰 펜치는 프레임에서 확인되지 않으며, 화면 밖 이현우나 그의 신체를 추가하지 않았다. 입구의 밝은 자연광은 낮 설정에 맞는다.",
        "hard_violations": [],
        "physics": "태진은 서서 차량 쪽으로 팔을 내민 자세이며 하체와 손은 프레임 아래로 잘려 있다. 보이지 않는 발이나 손의 접촉을 확정할 수는 없지만, 상체 자세에 부유나 불가능한 관절은 없다. 수납장과 의자는 바닥에, 용기들은 선반과 작업대에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "몸통과 팔은 오른쪽 아래 정비 부위를 향한 채 얼굴과 눈을 화면 밖 왼쪽으로 돌렸다. 보이지 않는 이현우 쪽을 돌아보는 관계가 분명하고, 치켜든 턱과 비틀린 입꼬리가 아니꼽게 비웃는 순간을 잘 드러낸다. 작업대의 펜치는 왼쪽을 향해 놓여 있으며 누구도 들고 있지 않다.",
        "built_space": "왼쪽 벽에 참조의 포스터들이 있고, 큰 양문 수납장 하나와 좁은 수납장 하나, 뒤쪽 공구판·작업대·의자 각 하나, 오른쪽 선반과 큰 스피커 하나, 밝은 출입구가 보인다. 왼쪽 전경에는 펜치와 공구함을 올린 금속 작업대가 있다. 태진은 오른쪽에 허리까지 보이고 왼쪽 여백도 확보되어 있다. 다만 캠퍼 측면이 중앙 배경까지 드러나며 오른쪽 하단의 노출된 정비 부위도 커서, 차량을 작은 하단 단서로만 제한하라는 요구에는 미달한다.",
        "entities": "태진 한 명만 보이며, 성인 한국인 여성의 외모와 검은 단발, 작업모, 기름때가 묻은 남색 커버올이 참조와 가깝다. 캠퍼, 큰 스피커, 록 포스터, 붉은 손잡이의 큰 펜치가 확인된다. 포스터의 한글은 장소 참조에 있는 문구를 유지한다. 이현우는 화면 밖에 남아 있고 추가 인물은 없다. 엔진·브레이크의 광범위한 손상 정도는 이 화면만으로 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "태진의 상체는 서 있는 몸에서 자연스럽게 이어지고, 두 팔은 전경 차량 부위 뒤로 내려간다. 손의 실제 접촉은 가려져 있으나 들고 있는 물체나 지지 없이 떠 있는 신체는 없다. 펜치와 공구함, 금속 용기는 작업대 위에 놓였고, 캠퍼의 보이는 바퀴는 바닥에 닿아 있다. 머리만 왼쪽으로 돌린 자세는 수리를 중단하고 반응하는 동작으로 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "오른쪽 상체 배치와 왼쪽 곁눈질은 맞지만, 캠퍼가 화면 오른쪽을 크게 차지하고 회복 구역을 향한 시점과 고개를 돌려 비웃는 연기가 약하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "수리 쪽에 남은 몸통과 왼쪽으로 돌아간 얼굴, 아니꼬운 표정 및 장소 재현이 더 정확하지만, 캠퍼 측면과 노출된 정비 부위를 지나치게 많이 보여준다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "태진의 몸통과 두 팔은 오른쪽 아래 차량 작업 부위를 향하고, 눈은 화면 밖 왼쪽을 본다. 이현우를 바라보는 방향은 맞지만 얼굴 자체의 회전은 작고 곁눈질이 더 두드러진다. 입술을 비틀고 있어 불쾌함은 읽히나 비웃음은 비교적 약하다.",
        "built_space": "왼쪽에 큰 양문 수납장 하나와 좁은 수납장 하나, 뒤쪽에 공구판과 작업대 각 하나, 바퀴 달린 의자 하나, 입구 근처에 세로형 스피커 하나가 보인다. 낡은 콘크리트와 수납 설비는 참조 장소와 유사하다. 다만 회복 구역 대신 공구 작업대와 출입구를 배경으로 삼았고, 캠퍼의 창문·거울·차체가 오른쪽 가장자리를 위아래로 크게 채워 하단의 작은 사선 부분만 보이라는 구도를 따르지 않는다.",
        "entities": "보이는 사람은 태진 한 명뿐이며, 참조와 유사한 성인 한국인 여성의 얼굴, 검은 단발, 남색 작업모와 때 묻은 짙은 남색 작업복이다. 캠퍼와 큰 스피커는 보인다. 포스터와 큰 펜치는 프레임에서 확인되지 않으며, 화면 밖 이현우나 그의 신체를 추가하지 않았다. 입구의 밝은 자연광은 낮 설정에 맞는다.",
        "hard_violations": [],
        "physics": "태진은 서서 차량 쪽으로 팔을 내민 자세이며 하체와 손은 프레임 아래로 잘려 있다. 보이지 않는 발이나 손의 접촉을 확정할 수는 없지만, 상체 자세에 부유나 불가능한 관절은 없다. 수납장과 의자는 바닥에, 용기들은 선반과 작업대에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "몸통과 팔은 오른쪽 아래 정비 부위를 향한 채 얼굴과 눈을 화면 밖 왼쪽으로 돌렸다. 보이지 않는 이현우 쪽을 돌아보는 관계가 분명하고, 치켜든 턱과 비틀린 입꼬리가 아니꼽게 비웃는 순간을 잘 드러낸다. 작업대의 펜치는 왼쪽을 향해 놓여 있으며 누구도 들고 있지 않다.",
        "built_space": "왼쪽 벽에 참조의 포스터들이 있고, 큰 양문 수납장 하나와 좁은 수납장 하나, 뒤쪽 공구판·작업대·의자 각 하나, 오른쪽 선반과 큰 스피커 하나, 밝은 출입구가 보인다. 왼쪽 전경에는 펜치와 공구함을 올린 금속 작업대가 있다. 태진은 오른쪽에 허리까지 보이고 왼쪽 여백도 확보되어 있다. 다만 캠퍼 측면이 중앙 배경까지 드러나며 오른쪽 하단의 노출된 정비 부위도 커서, 차량을 작은 하단 단서로만 제한하라는 요구에는 미달한다.",
        "entities": "태진 한 명만 보이며, 성인 한국인 여성의 외모와 검은 단발, 작업모, 기름때가 묻은 남색 커버올이 참조와 가깝다. 캠퍼, 큰 스피커, 록 포스터, 붉은 손잡이의 큰 펜치가 확인된다. 포스터의 한글은 장소 참조에 있는 문구를 유지한다. 이현우는 화면 밖에 남아 있고 추가 인물은 없다. 엔진·브레이크의 광범위한 손상 정도는 이 화면만으로 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "태진의 상체는 서 있는 몸에서 자연스럽게 이어지고, 두 팔은 전경 차량 부위 뒤로 내려간다. 손의 실제 접촉은 가려져 있으나 들고 있는 물체나 지지 없이 떠 있는 신체는 없다. 펜치와 공구함, 금속 용기는 작업대 위에 놓였고, 캠퍼의 보이는 바퀴는 바닥에 닿아 있다. 머리만 왼쪽으로 돌린 자세는 수리를 중단하고 반응하는 동작으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.314
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.064
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에서 명시적으로 금지한 '창작된 수리 형태(노출된 엔진룸)' 등장 및 추가 차량(엔진룸) 혼합",
     "[gemini-pro] 테이블 위 펜치(플라이어) 중복 생성"
    ],
    "B": [
     "[gemini-pro] 캠퍼가 원본 참조 사진의 공간과 어긋나게 90도 회전되어 배치된 공간 왜곡"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1064
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 프레이밍(우측 배치, 좌측 여백)과 포스터의 한글 텍스트까지 완벽히 구현했으나, 프롬프트에서 금지한 '창작된 수리 형태(엔진룸)'를 노출하고 두 대의 차량이 섞인 듯한 물리적 오류가 발생함.  ★위반: [gemini-pro] 프롬프트에서 명시적으로 금지한 '창작된 수리 형태(노출된 엔진룸)' 등장 및 추가 차량(엔진룸) 혼합 / [gemini-pro] 테이블 위 펜치(플라이어) 중복 생성"
   },
   {
    "label": "B",
    "score": 1064,
    "verdict_ko": "인물을 우측에 배치하라는 프레이밍 지시를 반대로 적용(좌측 배치)하였으며, 카메라 방향도 휴게 공간이 아닌 공구 벽을 향하고 있어 연출과 공간 구성을 모두 실패함.  ★위반: [gemini-pro] 캠퍼가 원본 참조 사진의 공간과 어긋나게 90도 회전되어 배치된 공간 왜곡"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L70B01.png",
    "asset_id": "2445fcea-a876-4f26-a752-65f6630fcd6f",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 태진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:769818>",
    "asset_id": "aa53b149-662e-4d5f-a036-4225a0b50eb0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b7a-d6e9-7b38-85e3-0f68818dbc66",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh7__bgfirst_bg.png",
   "bg_asset_id": "ba89ad40-df97-4eaf-bd86-34cc7ee6ae42",
   "bg_record_key": "S50sh7::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh14::signage": {
  "fp": "423ad2ebd75c7605",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh14": {
  "input_fingerprint": "da92f1d207611aee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 금발머리 여자애를 떠올린 듯 고개를 끄덕인 채 차분하게 입을 열어 대답하는 은영의 상반신.\n\nLOCATION (lock): In the auto shop's repair bay beside the camper, near the entrance through which the second mechanic arrived. Daylight reaches the work area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the established lateral move beside the vehicle settle just below 은영's eye level, ending at an oblique angle to her exchange with 이현우. Frame her upper body on the right with open space to the left toward 이현우 off-screen, capturing the small downward completion of her nod as she answers him. Keep the vehicle softly present behind her and emphasize only the settled camera distance, without reframing her response as a visual memory of 앰버.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Still positioned in the repair space) — A partial side view remains behind 은영; used as Preserve the location of the preceding conversation without competing with her response.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent subdued ambient light and soft tonal separation support 은영's calm recognition without a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the repair area, damaged camper, tools, and established daylight appearance from the reference. Exclude the separate recovery room's bedding and any furnishings from the container dining room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains badly damaged and under repair, with repair liquid just poured into its inlet. The room's rock-band posters, large speaker and pliers remain in place. 은영: She is inside the garage after handing over the bottle of repair liquid.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 금발머리 여자애를 떠올린 듯 고개를 끄덕인 채 차분하게 입을 열어 대답하는 은영의 상반신.\n\nLOCATION (lock): In the auto shop's repair bay beside the camper, near the entrance through which the second mechanic arrived. Daylight reaches the work area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the established lateral move beside the vehicle settle just below 은영's eye level, ending at an oblique angle to her exchange with 이현우. Frame her upper body on the right with open space to the left toward 이현우 off-screen, capturing the small downward completion of her nod as she answers him. Keep the vehicle softly present behind her and emphasize only the settled camera distance, without reframing her response as a visual memory of 앰버.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Still positioned in the repair space) — A partial side view remains behind 은영; used as Preserve the location of the preceding conversation without competing with her response.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent subdued ambient light and soft tonal separation support 은영's calm recognition without a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the repair area, damaged camper, tools, and established daylight appearance from the reference. Exclude the separate recovery room's bedding and any furnishings from the container dining room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains badly damaged and under repair, with repair liquid just poured into its inlet. The room's rock-band posters, large speaker and pliers remain in place. 은영: She is inside the garage after handing over the bottle of repair liquid.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 금발머리 여자애를 떠올린 듯 고개를 끄덕인 채 차분하게 입을 열어 대답하는 은영의 상반신.\n\nLOCATION (lock): In the auto shop's repair bay beside the camper, near the entrance through which the second mechanic arrived. Daylight reaches the work area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the established lateral move beside the vehicle settle just below 은영's eye level, ending at an oblique angle to her exchange with 이현우. Frame her upper body on the right with open space to the left toward 이현우 off-screen, capturing the small downward completion of her nod as she answers him. Keep the vehicle softly present behind her and emphasize only the settled camera distance, without reframing her response as a visual memory of 앰버.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Camper under repair (Still positioned in the repair space) — A partial side view remains behind 은영; used as Preserve the location of the preceding conversation without competing with her response.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent subdued ambient light and soft tonal separation support 은영's calm recognition without a lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the repair area, damaged camper, tools, and established daylight appearance from the reference. Exclude the separate recovery room's bedding and any furnishings from the container dining room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains badly damaged and under repair, with repair liquid just poured into its inlet. The room's rock-band posters, large speaker and pliers remain in place. 은영: She is inside the garage after handing over the bottle of repair liquid.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 은영 (한국인 여성, 30대 초반의 얼굴, 검은 머리카락) — wearing: 검은색 반팔 티셔츠 위에 멜빵을 걸친 기름때 묻은 작업용 데님 팬츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "은영의 시선이 화면 왼쪽 오프스크린을 향함.",
    "built_space": "정비소 내부. 왼쪽 작업대와 오른쪽 캠핑카의 배치 및 조명이 이전 샷의 환경과 일치함.",
    "entities": "은영의 외모는 일치하나, 프롬프트의 복장 지시를 무시하고 레퍼런스 이미지의 파란색 작업복을 그대로 입고 있음.",
    "hard_violations": [],
    "physics": "자연스럽게 서서 기댄 자세로 물리적 오류 없이 지지됨."
   },
   {
    "label": "B",
    "direction": "은영의 시선이 화면 왼쪽 오프스크린(이현우가 있는 방향)을 향하고 있음.",
    "built_space": "정비소 내부. 왼쪽에 공구가 놓인 작업대, 오른쪽에 수리 중인 캠핑카가 이전 샷과 동일한 위치에 있음.",
    "entities": "은영의 외모가 캐릭터 레퍼런스와 일치하며, 프롬프트에 명시된 검은 티셔츠와 데님 오버올을 착용함(모자 추가됨).",
    "hard_violations": [],
    "physics": "자연스럽게 서서 캠핑카 쪽에 몸을 의지한 채 안정적인 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 멜빵 데님 팬츠와 검은색 티셔츠 복장을 잘 반영했으며, 대답하는 순간의 표정과 구도를 성공적으로 연출함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "이전 샷의 파란색 작업복을 그대로 가져와 명시된 인물 복장 지침을 명백히 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "은영의 시선이 화면 왼쪽 오프스크린을 향함.",
        "built_space": "정비소 내부. 왼쪽 작업대와 오른쪽 캠핑카의 배치 및 조명이 이전 샷의 환경과 일치함.",
        "entities": "은영의 외모는 일치하나, 프롬프트의 복장 지시를 무시하고 레퍼런스 이미지의 파란색 작업복을 그대로 입고 있음.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 기댄 자세로 물리적 오류 없이 지지됨."
       },
       {
        "label": "B",
        "direction": "은영의 시선이 화면 왼쪽 오프스크린(이현우가 있는 방향)을 향하고 있음.",
        "built_space": "정비소 내부. 왼쪽에 공구가 놓인 작업대, 오른쪽에 수리 중인 캠핑카가 이전 샷과 동일한 위치에 있음.",
        "entities": "은영의 외모가 캐릭터 레퍼런스와 일치하며, 프롬프트에 명시된 검은 티셔츠와 데님 오버올을 착용함(모자 추가됨).",
        "hard_violations": [],
        "physics": "자연스럽게 서서 캠핑카 쪽에 몸을 의지한 채 안정적인 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 멜빵 데님 팬츠와 검은색 티셔츠 복장을 잘 반영했으며, 대답하는 순간의 표정과 구도를 성공적으로 연출함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "이전 샷의 파란색 작업복을 그대로 가져와 명시된 인물 복장 지침을 명백히 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "은영의 시선이 화면 왼쪽 오프스크린을 향함.",
        "built_space": "정비소 내부. 왼쪽 작업대와 오른쪽 캠핑카의 배치 및 조명이 이전 샷의 환경과 일치함.",
        "entities": "은영의 외모는 일치하나, 프롬프트의 복장 지시를 무시하고 레퍼런스 이미지의 파란색 작업복을 그대로 입고 있음.",
        "hard_violations": [],
        "physics": "자연스럽게 서서 기댄 자세로 물리적 오류 없이 지지됨."
       },
       {
        "label": "B",
        "direction": "은영의 시선이 화면 왼쪽 오프스크린(이현우가 있는 방향)을 향하고 있음.",
        "built_space": "정비소 내부. 왼쪽에 공구가 놓인 작업대, 오른쪽에 수리 중인 캠핑카가 이전 샷과 동일한 위치에 있음.",
        "entities": "은영의 외모가 캐릭터 레퍼런스와 일치하며, 프롬프트에 명시된 검은 티셔츠와 데님 오버올을 착용함(모자 추가됨).",
        "hard_violations": [],
        "physics": "자연스럽게 서서 캠핑카 쪽에 몸을 의지한 채 안정적인 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "오른쪽 상반신 구도와 검은 반팔·데님 멜빵은 맞지만, 이전 장면 인물의 얼굴과 모자를 옮겼으며 고개를 끄덕여 내리는 순간이 드러나지 않는다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "공간과 왼쪽 상대를 향한 응답 구도는 맞지만, 이전 장면 인물의 얼굴·모자·긴팔 작업복을 그대로 이어 받아 은영의 인물 및 의상 지정을 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈은 화면 밖 왼쪽, 약간 위를 향해 있어 이현우를 바라보는 대화 방향으로 읽힌다. 입은 조금 열려 있지만 턱이 내려오는 끄덕임의 마무리보다는 고개를 들고 답하는 모습이다. 겨누거나 움직이는 물체는 없다.",
        "built_space": "오른쪽에 여성의 상반신, 그 뒤 중앙에 캠퍼 한 대의 측면이 있다. 왼쪽에는 금속 작업대 하나와 큰 수납장 하나, 뒤쪽에는 공구판과 작업대, 오른쪽 출입구 옆에는 선반과 대형 스피커 한 대가 보인다. 왼쪽 벽의 낡은 포스터와 출입구의 주광도 유지된다. 전경 오른쪽 아래 차량 정비 부위가 팔 일부를 가린다. 중복 설비나 불가능한 반사는 보이지 않지만 이전 사진의 시점과 배치를 매우 가깝게 반복한다.",
        "entities": "보이는 사람은 검은 단발의 성인 여성 한 명이며 외관상 지정 연령대와 대체로 맞는다. 그러나 얼굴과 앞머리, 작업 모자는 은영의 인물 참조보다 이전 장면 여성에 가깝다. 검은 반팔 티셔츠와 갈색 끈의 오염된 데님 멜빵은 지정 의상에 부합한다. 가슴 주머니의 공구도 인물 참조와 대응한다. 캠퍼, 포스터, 스피커, 작업대 위 펜치 한 개와 드라이버 한 개가 보인다. 수리액 병과 주입구 상태는 확인되지 않으며 금발 소녀나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "상체는 똑바로 선 자세로 화면 아래의 하체에 이어지며 발은 구도 밖이다. 팔은 아래로 내려가 전경 차량에 가려져 손의 접촉 여부는 확인할 수 없다. 공구는 가슴 주머니에 꽂혀 있고 작업대의 물건은 상판에 놓여 있다. 떠 있는 신체나 지지 없는 물체, 명백한 해부학적 불가능성은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "시선과 얼굴이 화면 밖 왼쪽 위를 향하고 입이 조금 열려 있어 왼쪽의 이현우에게 답하는 방향은 맞는다. 다만 머리는 옆으로 기울어진 채 들려 있으며, 작게 끄덕인 뒤 아래로 내려오는 순간은 분명하지 않다. 팔은 왼쪽 상대가 아니라 앞쪽 차량 정비 부위를 향한다.",
        "built_space": "여성이 화면 오른쪽에 있고 캠퍼 한 대가 중앙 뒤로 이어진다. 왼쪽 작업대 하나와 큰 수납장 하나, 후면 공구판과 작업대, 오른쪽 선반 및 대형 스피커 한 대, 밝은 출입구가 보인다. 낡은 포스터와 금속·콘크리트 재질도 이전 공간을 따른다. 전경 오른쪽 아래에는 차량 정비 부위가 있으며 여성은 그 바로 뒤에 서 있다. 설비 중복이나 불가능한 반사는 없으나 이전 사진의 정비 자세와 카메라 구도를 거의 그대로 반복한다.",
        "entities": "검은 단발의 성인 여성 한 명만 보인다. 얼굴·앞머리·작업 모자는 지정된 은영 참조가 아니라 이전 장면 여성의 외형을 따른다. 의상도 검은 반팔과 데님 멜빵이 아닌 기름때 묻은 긴팔 일체형 작업복으로, 이전 사진의 옷을 이어 받았다. 캠퍼, 벽 포스터, 대형 스피커, 작업대 위 펜치 한 개와 드라이버 한 개는 보인다. 수리액 병이나 주입 상태는 확인할 수 없으며 금발 소녀와 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "몸통은 화면 아래의 골반과 하체로 자연스럽게 이어진 서 있는 자세다. 두 팔은 앞쪽 차량으로 내려가며 손과 정확한 접촉점은 전경에 대부분 가려져 있다. 작업대 위 물건과 선반의 용기는 각각 상판과 선반에 지지된다. 부유하거나 지지 없이 매달린 물체는 없고 자세 자체는 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "오른쪽 상반신 구도와 검은 반팔·데님 멜빵은 맞지만, 이전 장면 인물의 얼굴과 모자를 옮겼으며 고개를 끄덕여 내리는 순간이 드러나지 않는다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "공간과 왼쪽 상대를 향한 응답 구도는 맞지만, 이전 장면 인물의 얼굴·모자·긴팔 작업복을 그대로 이어 받아 은영의 인물 및 의상 지정을 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈은 화면 밖 왼쪽, 약간 위를 향해 있어 이현우를 바라보는 대화 방향으로 읽힌다. 입은 조금 열려 있지만 턱이 내려오는 끄덕임의 마무리보다는 고개를 들고 답하는 모습이다. 겨누거나 움직이는 물체는 없다.",
        "built_space": "오른쪽에 여성의 상반신, 그 뒤 중앙에 캠퍼 한 대의 측면이 있다. 왼쪽에는 금속 작업대 하나와 큰 수납장 하나, 뒤쪽에는 공구판과 작업대, 오른쪽 출입구 옆에는 선반과 대형 스피커 한 대가 보인다. 왼쪽 벽의 낡은 포스터와 출입구의 주광도 유지된다. 전경 오른쪽 아래 차량 정비 부위가 팔 일부를 가린다. 중복 설비나 불가능한 반사는 보이지 않지만 이전 사진의 시점과 배치를 매우 가깝게 반복한다.",
        "entities": "보이는 사람은 검은 단발의 성인 여성 한 명이며 외관상 지정 연령대와 대체로 맞는다. 그러나 얼굴과 앞머리, 작업 모자는 은영의 인물 참조보다 이전 장면 여성에 가깝다. 검은 반팔 티셔츠와 갈색 끈의 오염된 데님 멜빵은 지정 의상에 부합한다. 가슴 주머니의 공구도 인물 참조와 대응한다. 캠퍼, 포스터, 스피커, 작업대 위 펜치 한 개와 드라이버 한 개가 보인다. 수리액 병과 주입구 상태는 확인되지 않으며 금발 소녀나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "상체는 똑바로 선 자세로 화면 아래의 하체에 이어지며 발은 구도 밖이다. 팔은 아래로 내려가 전경 차량에 가려져 손의 접촉 여부는 확인할 수 없다. 공구는 가슴 주머니에 꽂혀 있고 작업대의 물건은 상판에 놓여 있다. 떠 있는 신체나 지지 없는 물체, 명백한 해부학적 불가능성은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "시선과 얼굴이 화면 밖 왼쪽 위를 향하고 입이 조금 열려 있어 왼쪽의 이현우에게 답하는 방향은 맞는다. 다만 머리는 옆으로 기울어진 채 들려 있으며, 작게 끄덕인 뒤 아래로 내려오는 순간은 분명하지 않다. 팔은 왼쪽 상대가 아니라 앞쪽 차량 정비 부위를 향한다.",
        "built_space": "여성이 화면 오른쪽에 있고 캠퍼 한 대가 중앙 뒤로 이어진다. 왼쪽 작업대 하나와 큰 수납장 하나, 후면 공구판과 작업대, 오른쪽 선반 및 대형 스피커 한 대, 밝은 출입구가 보인다. 낡은 포스터와 금속·콘크리트 재질도 이전 공간을 따른다. 전경 오른쪽 아래에는 차량 정비 부위가 있으며 여성은 그 바로 뒤에 서 있다. 설비 중복이나 불가능한 반사는 없으나 이전 사진의 정비 자세와 카메라 구도를 거의 그대로 반복한다.",
        "entities": "검은 단발의 성인 여성 한 명만 보인다. 얼굴·앞머리·작업 모자는 지정된 은영 참조가 아니라 이전 장면 여성의 외형을 따른다. 의상도 검은 반팔과 데님 멜빵이 아닌 기름때 묻은 긴팔 일체형 작업복으로, 이전 사진의 옷을 이어 받았다. 캠퍼, 벽 포스터, 대형 스피커, 작업대 위 펜치 한 개와 드라이버 한 개는 보인다. 수리액 병이나 주입 상태는 확인할 수 없으며 금발 소녀와 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "몸통은 화면 아래의 골반과 하체로 자연스럽게 이어진 서 있는 자세다. 두 팔은 앞쪽 차량으로 내려가며 손과 정확한 접촉점은 전경에 대부분 가려져 있다. 작업대 위 물건과 선반의 용기는 각각 상판과 선반에 지지된다. 부유하거나 지지 없이 매달린 물체는 없고 자세 자체는 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.171,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.171,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1171
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 멜빵 데님 팬츠와 검은색 티셔츠 복장을 잘 반영했으며, 대답하는 순간의 표정과 구도를 성공적으로 연출함."
   },
   {
    "label": "A",
    "score": 1171,
    "verdict_ko": "이전 샷의 파란색 작업복을 그대로 가져와 명시된 인물 복장 지침을 명백히 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh7_sel.png",
    "asset_id": "00aa0ef7-38db-4c5d-a949-e168607e0dd1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 은영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:981336>",
    "asset_id": "c128fa4a-1680-4e5b-a9d7-a21abf9186af",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b84-b855-7cd9-98f6-20267076c8ff",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh7"
  }
 },
 "S51sh4::signage": {
  "fp": "38fe9d986045e3c0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::83409c94d221ffc0": {
  "subjects": [],
  "subject_text": "카센터 옆 컨테이너 간이식당 내부\n철제 주거동을 활용한 작은 간이식당. 출입문 안쪽에 식탁과 의자가 놓여 있고, 식탁 위로 국밥 그릇과 식기가 보인다.",
  "identity": "canonical",
  "scope_id": "L71",
  "scope_role": "location_interior",
  "scope_sha": "6c9d5eff0e5a2188"
 },
 "S51sh4::bgfirst_bg": {
  "input_fingerprint": "718dd332eaae00ea",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달려온 추진력이 남은 채 두 팔로 이현우의 허리를 와락 껴안아 밀착된 충돌 순간의 앰버.\n\nLOCATION (lock): Just inside the entrance to the container dining room beside the auto-repair shop. Daylight enters through the opened door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track at contact from the established three-quarter side position, outside 앰버's running line and below 이현우's eye level. Place 앰버 on the left leaning into 이현우 on the right, framing from her wrapping arms to his face so her remaining forward momentum and his slight backward weight transfer are visible together. Her face tips up toward him while he looks down at her, with the opened entrance behind them establishing that the embrace interrupts his arrival.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-area entrance (Opened for 이현우's arrival) — Seen obliquely behind 이현우 rather than symmetrically surrounding the pair; used as Anchor the meeting point and distinguish the embrace from a posed two-shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light and gentle facial contrast convey warmth through proximity and expression rather than a change in color or brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달려온 추진력이 남은 채 두 팔로 이현우의 허리를 와락 껴안아 밀착된 충돌 순간의 앰버.\n\nLOCATION (lock): Just inside the entrance to the container dining room beside the auto-repair shop. Daylight enters through the opened door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track at contact from the established three-quarter side position, outside 앰버's running line and below 이현우's eye level. Place 앰버 on the left leaning into 이현우 on the right, framing from her wrapping arms to his face so her remaining forward momentum and his slight backward weight transfer are visible together. Her face tips up toward him while he looks down at her, with the opened entrance behind them establishing that the embrace interrupts his arrival.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-area entrance (Opened for 이현우's arrival) — Seen obliquely behind 이현우 rather than symmetrically surrounding the pair; used as Anchor the meeting point and distinguish the embrace from a posed two-shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light and gentle facial contrast convey warmth through proximity and expression rather than a change in color or brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S51sh4__bgfirst_bg.png",
  "asset_id": "ef9cec48-bf79-4ea5-bbc0-dd10e7b64542",
  "input_asset_ids": [
   "5b63c85d-7e6c-4b4f-a0ba-609025bc893e",
   "d0e03058-517b-45c2-ae63-97044a0c4211"
  ]
 },
 "S51sh4": {
  "input_fingerprint": "19a7803784d85af2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달려온 추진력이 남은 채 두 팔로 이현우의 허리를 와락 껴안아 밀착된 충돌 순간의 앰버.\n\nLOCATION (lock): Just inside the entrance to the container dining room beside the auto-repair shop. Daylight enters through the opened door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track at contact from the established three-quarter side position, outside 앰버's running line and below 이현우's eye level. Place 앰버 on the left leaning into 이현우 on the right, framing from her wrapping arms to his face so her remaining forward momentum and his slight backward weight transfer are visible together. Her face tips up toward him while he looks down at her, with the opened entrance behind them establishing that the embrace interrupts his arrival.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-area entrance (Opened for 이현우's arrival) — Seen obliquely behind 이현우 rather than symmetrically surrounding the pair; used as Anchor the meeting point and distinguish the embrace from a posed two-shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light and gentle facial contrast convey warmth through proximity and expression rather than a change in color or brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup are set out in the makeshift container dining room beside the garage. 앰버: She has just risen from her meal and rushed forward, still with a mouthful of rice soup. 이현우: He has entered the dining room with his leg bandaged and his other injuries treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달려온 추진력이 남은 채 두 팔로 이현우의 허리를 와락 껴안아 밀착된 충돌 순간의 앰버.\n\nLOCATION (lock): Just inside the entrance to the container dining room beside the auto-repair shop. Daylight enters through the opened door. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track at contact from the established three-quarter side position, outside 앰버's running line and below 이현우's eye level. Place 앰버 on the left leaning into 이현우 on the right, framing from her wrapping arms to his face so her remaining forward momentum and his slight backward weight transfer are visible together. Her face tips up toward him while he looks down at her, with the opened entrance behind them establishing that the embrace interrupts his arrival.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-area entrance (Opened for 이현우's arrival) — Seen obliquely behind 이현우 rather than symmetrically surrounding the pair; used as Anchor the meeting point and distinguish the embrace from a posed two-shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light and gentle facial contrast convey warmth through proximity and expression rather than a change in color or brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup are set out in the makeshift container dining room beside the garage. 앰버: She has just risen from her meal and rushed forward, still with a mouthful of rice soup. 이현우: He has entered the dining room with his leg bandaged and his other injuries treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달려온 추진력이 남은 채 두 팔로 이현우의 허리를 와락 껴안아 밀착된 충돌 순간의 앰버.\n\nLOCATION (lock): Just inside the entrance to the container dining room beside the auto-repair shop. Daylight enters through the opened door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track at contact from the established three-quarter side position, outside 앰버's running line and below 이현우's eye level. Place 앰버 on the left leaning into 이현우 on the right, framing from her wrapping arms to his face so her remaining forward momentum and his slight backward weight transfer are visible together. Her face tips up toward him while he looks down at her, with the opened entrance behind them establishing that the embrace interrupts his arrival.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining-area entrance (Opened for 이현우's arrival) — Seen obliquely behind 이현우 rather than symmetrically surrounding the pair; used as Anchor the meeting point and distinguish the embrace from a posed two-shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light and gentle facial contrast convey warmth through proximity and expression rather than a change in color or brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls of rice soup are set out in the makeshift container dining room beside the garage. 앰버: She has just risen from her meal and rushed forward, still with a mouthful of rice soup. 이현우: He has entered the dining room with his leg bandaged and his other injuries treated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S51sh4__bgfirst_bg.png",
     "asset_id": "ef9cec48-bf79-4ea5-bbc0-dd10e7b64542",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S51sh4.png",
     "asset_id": "5b63c85d-7e6c-4b4f-a0ba-609025bc893e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L71B01.png",
     "asset_id": "d0e03058-517b-45c2-ae63-97044a0c4211",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "왼쪽 앰버가 오른쪽 이현우의 허리에 밀착하며 얼굴을 들어 그의 눈을 바라보고, 이현우도 앰버의 얼굴을 내려다본다. 앰버의 몸통은 오른쪽으로 기울고 팔은 그의 허리 뒤로 돌아간다. 요구된 접근 방향과 상호 시선이 맞는다.",
    "built_space": "왼쪽에 창문 일부 하나, 검은 벤치 등받이 하나, 식탁 하나와 전경 식탁 모서리가 보인다. 뒤에는 냉장고 하나와 취사기 하나, 오른쪽에는 열린 출입문 하나와 쓰레기통 하나가 있다. 낡은 패널 벽과 좁은 통로가 장소 사진에 부합한다. 출입문은 이현우의 오른쪽 뒤에 비대칭으로 보이며 두 사람은 입구 가까운 실내에 있다. 반사상이나 중복된 고정 설비는 보이지 않는다.",
    "entities": "등장인물은 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리의 마른 동아시아계 청년으로, 참조와 유사한 얼굴 및 피와 먼지가 묻은 어두운 셔츠, 귀의 소형 인이어 장치를 갖췄다. 앰버는 금발과 창백한 피부의 어린 여자아이로 참조의 연령대와 외형에 가깝고, 카키 작업복과 가죽 공구 벨트, 목에 내려 둔 기계식 방진 마스크가 보인다. 부푼 볼과 다문 입은 음식을 머금은 상태를 잘 전달한다. 식탁에는 검은 국그릇이 보이나 내용물은 분명하지 않다. 다리 붕대는 프레임 밖이다.",
    "hard_violations": [],
    "physics": "앰버의 보이는 팔이 이현우의 허리를 감싸고 가슴과 몸통이 접촉한다. 반대쪽 팔의 대부분과 손은 몸 뒤에 가려져 있다. 앰버는 허리부터 앞으로 기울며 이현우는 약하게 뒤로 물러나는 자세다. 발은 화면 밖이므로 접지는 직접 확인되지 않지만 공중에 떠 있는 모습은 아니다. 마스크는 목의 끈, 공구는 벨트, 식기는 식탁이 지지한다. 물리적으로 가능한 포옹이나 달려온 충격은 비교적 절제되어 있다."
   },
   {
    "label": "A",
    "direction": "왼쪽 앰버가 오른쪽 이현우를 향해 뚜렷하게 기울어 허리를 감싼다. 앰버의 눈은 이현우의 얼굴을 올려다보고 이현우의 시선도 앰버에게 내려간다. 뒤로 흐르는 머리카락과 서로 기운 몸통이 접근 및 충돌 방향을 전달한다.",
    "built_space": "왼쪽 창문 하나와 메뉴판 하나, 검은 벤치 등받이 두 구간, 전경 식탁 하나와 뒤쪽 식탁 일부가 보인다. 후면에는 냉장고 하나와 취사기 하나, 오른쪽 뒤에는 열린 출입문 하나와 쓰레기통 하나가 있다. 재료와 설비 배치는 장소 사진과 대체로 맞지만, 두 사람 뒤로 긴 통로가 남아 입구 바로 안쪽보다 식당 깊숙한 곳에서 만난 듯하다. 출입문은 이현우 뒤 오른쪽에 있으나 A보다 작고 멀다. 불가능한 반사나 설비 중복은 없다.",
    "entities": "두 인물 외에 추가 인물은 없다. 이현우의 검은 머리, 젊은 동아시아계 얼굴, 마른 체격, 오염된 어두운 셔츠와 바지, 인이어 장치는 요구에 부합한다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 공구 가죽 벨트, 목에 걸린 방진 마스크를 착용한다. 얼굴은 측면이라 참조 얼굴과의 세부 비교가 제한되며, 다문 입과 볼은 음식을 머금은 상태로 읽힐 수 있다. 전경에는 국그릇과 밥그릇, 반찬이 놓여 있다. 다리 붕대는 프레임 밖이다.",
    "hard_violations": [],
    "physics": "앰버의 팔과 몸통이 이현우의 허리에 붙어 있고, 이현우는 상체를 뒤로 기울여 충격을 받는 자세다. 앰버의 하체는 화면 아래 왼쪽으로 이어지고 양쪽 발은 잘려 있어 지면 지지는 직접 보이지 않지만, 도약이나 무지지 부유로 볼 근거는 없다. 이현우의 펼친 손은 포옹에 반응하는 순간으로 자연스럽다. 목끈과 벨트가 마스크와 공구를 지지하고 식기는 식탁 위에 놓여 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "감싸는 팔부터 이현우의 얼굴까지 집중한 미디엄 구도와 입구 바로 안쪽의 만남을 더 정확히 구현하며, 다만 충돌 직후의 반동은 다소 약하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앰버의 전진과 이현우의 뒤로 기우는 반응은 선명하지만, 허벅지와 식탁까지 넓힌 구도 및 멀어진 출입문이 지정된 프레이밍과 만남의 위치에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 앰버가 오른쪽 이현우의 허리에 밀착하며 얼굴을 들어 그의 눈을 바라보고, 이현우도 앰버의 얼굴을 내려다본다. 앰버의 몸통은 오른쪽으로 기울고 팔은 그의 허리 뒤로 돌아간다. 요구된 접근 방향과 상호 시선이 맞는다.",
        "built_space": "왼쪽에 창문 일부 하나, 검은 벤치 등받이 하나, 식탁 하나와 전경 식탁 모서리가 보인다. 뒤에는 냉장고 하나와 취사기 하나, 오른쪽에는 열린 출입문 하나와 쓰레기통 하나가 있다. 낡은 패널 벽과 좁은 통로가 장소 사진에 부합한다. 출입문은 이현우의 오른쪽 뒤에 비대칭으로 보이며 두 사람은 입구 가까운 실내에 있다. 반사상이나 중복된 고정 설비는 보이지 않는다.",
        "entities": "등장인물은 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리의 마른 동아시아계 청년으로, 참조와 유사한 얼굴 및 피와 먼지가 묻은 어두운 셔츠, 귀의 소형 인이어 장치를 갖췄다. 앰버는 금발과 창백한 피부의 어린 여자아이로 참조의 연령대와 외형에 가깝고, 카키 작업복과 가죽 공구 벨트, 목에 내려 둔 기계식 방진 마스크가 보인다. 부푼 볼과 다문 입은 음식을 머금은 상태를 잘 전달한다. 식탁에는 검은 국그릇이 보이나 내용물은 분명하지 않다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 보이는 팔이 이현우의 허리를 감싸고 가슴과 몸통이 접촉한다. 반대쪽 팔의 대부분과 손은 몸 뒤에 가려져 있다. 앰버는 허리부터 앞으로 기울며 이현우는 약하게 뒤로 물러나는 자세다. 발은 화면 밖이므로 접지는 직접 확인되지 않지만 공중에 떠 있는 모습은 아니다. 마스크는 목의 끈, 공구는 벨트, 식기는 식탁이 지지한다. 물리적으로 가능한 포옹이나 달려온 충격은 비교적 절제되어 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 앰버가 오른쪽 이현우를 향해 뚜렷하게 기울어 허리를 감싼다. 앰버의 눈은 이현우의 얼굴을 올려다보고 이현우의 시선도 앰버에게 내려간다. 뒤로 흐르는 머리카락과 서로 기운 몸통이 접근 및 충돌 방향을 전달한다.",
        "built_space": "왼쪽 창문 하나와 메뉴판 하나, 검은 벤치 등받이 두 구간, 전경 식탁 하나와 뒤쪽 식탁 일부가 보인다. 후면에는 냉장고 하나와 취사기 하나, 오른쪽 뒤에는 열린 출입문 하나와 쓰레기통 하나가 있다. 재료와 설비 배치는 장소 사진과 대체로 맞지만, 두 사람 뒤로 긴 통로가 남아 입구 바로 안쪽보다 식당 깊숙한 곳에서 만난 듯하다. 출입문은 이현우 뒤 오른쪽에 있으나 A보다 작고 멀다. 불가능한 반사나 설비 중복은 없다.",
        "entities": "두 인물 외에 추가 인물은 없다. 이현우의 검은 머리, 젊은 동아시아계 얼굴, 마른 체격, 오염된 어두운 셔츠와 바지, 인이어 장치는 요구에 부합한다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 공구 가죽 벨트, 목에 걸린 방진 마스크를 착용한다. 얼굴은 측면이라 참조 얼굴과의 세부 비교가 제한되며, 다문 입과 볼은 음식을 머금은 상태로 읽힐 수 있다. 전경에는 국그릇과 밥그릇, 반찬이 놓여 있다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 팔과 몸통이 이현우의 허리에 붙어 있고, 이현우는 상체를 뒤로 기울여 충격을 받는 자세다. 앰버의 하체는 화면 아래 왼쪽으로 이어지고 양쪽 발은 잘려 있어 지면 지지는 직접 보이지 않지만, 도약이나 무지지 부유로 볼 근거는 없다. 이현우의 펼친 손은 포옹에 반응하는 순간으로 자연스럽다. 목끈과 벨트가 마스크와 공구를 지지하고 식기는 식탁 위에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "감싸는 팔부터 이현우의 얼굴까지 집중한 미디엄 구도와 입구 바로 안쪽의 만남을 더 정확히 구현하며, 다만 충돌 직후의 반동은 다소 약하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앰버의 전진과 이현우의 뒤로 기우는 반응은 선명하지만, 허벅지와 식탁까지 넓힌 구도 및 멀어진 출입문이 지정된 프레이밍과 만남의 위치에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 앰버가 오른쪽 이현우의 허리에 밀착하며 얼굴을 들어 그의 눈을 바라보고, 이현우도 앰버의 얼굴을 내려다본다. 앰버의 몸통은 오른쪽으로 기울고 팔은 그의 허리 뒤로 돌아간다. 요구된 접근 방향과 상호 시선이 맞는다.",
        "built_space": "왼쪽에 창문 일부 하나, 검은 벤치 등받이 하나, 식탁 하나와 전경 식탁 모서리가 보인다. 뒤에는 냉장고 하나와 취사기 하나, 오른쪽에는 열린 출입문 하나와 쓰레기통 하나가 있다. 낡은 패널 벽과 좁은 통로가 장소 사진에 부합한다. 출입문은 이현우의 오른쪽 뒤에 비대칭으로 보이며 두 사람은 입구 가까운 실내에 있다. 반사상이나 중복된 고정 설비는 보이지 않는다.",
        "entities": "등장인물은 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리의 마른 동아시아계 청년으로, 참조와 유사한 얼굴 및 피와 먼지가 묻은 어두운 셔츠, 귀의 소형 인이어 장치를 갖췄다. 앰버는 금발과 창백한 피부의 어린 여자아이로 참조의 연령대와 외형에 가깝고, 카키 작업복과 가죽 공구 벨트, 목에 내려 둔 기계식 방진 마스크가 보인다. 부푼 볼과 다문 입은 음식을 머금은 상태를 잘 전달한다. 식탁에는 검은 국그릇이 보이나 내용물은 분명하지 않다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 보이는 팔이 이현우의 허리를 감싸고 가슴과 몸통이 접촉한다. 반대쪽 팔의 대부분과 손은 몸 뒤에 가려져 있다. 앰버는 허리부터 앞으로 기울며 이현우는 약하게 뒤로 물러나는 자세다. 발은 화면 밖이므로 접지는 직접 확인되지 않지만 공중에 떠 있는 모습은 아니다. 마스크는 목의 끈, 공구는 벨트, 식기는 식탁이 지지한다. 물리적으로 가능한 포옹이나 달려온 충격은 비교적 절제되어 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 앰버가 오른쪽 이현우를 향해 뚜렷하게 기울어 허리를 감싼다. 앰버의 눈은 이현우의 얼굴을 올려다보고 이현우의 시선도 앰버에게 내려간다. 뒤로 흐르는 머리카락과 서로 기운 몸통이 접근 및 충돌 방향을 전달한다.",
        "built_space": "왼쪽 창문 하나와 메뉴판 하나, 검은 벤치 등받이 두 구간, 전경 식탁 하나와 뒤쪽 식탁 일부가 보인다. 후면에는 냉장고 하나와 취사기 하나, 오른쪽 뒤에는 열린 출입문 하나와 쓰레기통 하나가 있다. 재료와 설비 배치는 장소 사진과 대체로 맞지만, 두 사람 뒤로 긴 통로가 남아 입구 바로 안쪽보다 식당 깊숙한 곳에서 만난 듯하다. 출입문은 이현우 뒤 오른쪽에 있으나 A보다 작고 멀다. 불가능한 반사나 설비 중복은 없다.",
        "entities": "두 인물 외에 추가 인물은 없다. 이현우의 검은 머리, 젊은 동아시아계 얼굴, 마른 체격, 오염된 어두운 셔츠와 바지, 인이어 장치는 요구에 부합한다. 앰버는 금발의 어린 여자아이이며 카키 작업복, 공구 가죽 벨트, 목에 걸린 방진 마스크를 착용한다. 얼굴은 측면이라 참조 얼굴과의 세부 비교가 제한되며, 다문 입과 볼은 음식을 머금은 상태로 읽힐 수 있다. 전경에는 국그릇과 밥그릇, 반찬이 놓여 있다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 팔과 몸통이 이현우의 허리에 붙어 있고, 이현우는 상체를 뒤로 기울여 충격을 받는 자세다. 앰버의 하체는 화면 아래 왼쪽으로 이어지고 양쪽 발은 잘려 있어 지면 지지는 직접 보이지 않지만, 도약이나 무지지 부유로 볼 근거는 없다. 이현우의 펼친 손은 포옹에 반응하는 순간으로 자연스럽다. 목끈과 벨트가 마스크와 공구를 지지하고 식기는 식탁 위에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 9,
   "A": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "감싸는 팔부터 이현우의 얼굴까지 집중한 미디엄 구도와 입구 바로 안쪽의 만남을 더 정확히 구현하며, 다만 충돌 직후의 반동은 다소 약하다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "앰버의 전진과 이현우의 뒤로 기우는 반응은 선명하지만, 허벅지와 식탁까지 넓힌 구도 및 멀어진 출입문이 지정된 프레이밍과 만남의 위치에서 벗어난다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L71B01.png",
    "asset_id": "d0e03058-517b-45c2-ae63-97044a0c4211",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b89-608c-720d-9149-8a6c2430aa67",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S51sh4__bgfirst_bg.png",
   "bg_asset_id": "ef9cec48-bf79-4ea5-bbc0-dd10e7b64542",
   "bg_record_key": "S51sh4::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S51sh22::signage": {
  "fp": "4a4f5e1238bca714",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "종이 편지"
   }
  ],
  "dropped": []
 },
 "S51sh22": {
  "input_fingerprint": "382c4abc115c7de6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막 잠에서 깬 이현우의 얼굴 바로 앞에 종이 편지를 뻗은 손으로 들이민 상태의 앰버 근접 찰나.\n\nLOCATION (lock): In the overnight sleeping area at the auto-repair premises, in morning light. The text does not establish that this is still the dining container. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain just behind and beside 이현우's shoulder at his reclining eye height, shifting only slightly sideways to see obliquely between his face and 앰버's extended hand. Let 앰버's face and reaching upper body occupy the central-right area, retaining 이현우's near cheek at the left edge and the letter across less than a quarter of the lower center, close to him but clear of her face. 앰버 looks down toward the newly awakened 이현우 while his lowered eyes begin to attend to the paper, preserving the urgency of a handoff rather than presenting the letter to the audience.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Extended paper letter below 앰버's face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held out close to 이현우's face) — The message-bearing face points toward 이현우 and is seen obliquely by the camera; it contains the farewell mentioning 해남 and remembering 앰버; used as Bridge the space between the two faces while leaving 앰버's expression unobstructed; Camper interior (Occupied during the next-day awakening); used as Provide a narrow contextual background without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination keeps the faces and paper distinct without turning the letter into a luminous graphic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's handwritten farewell note says thanks, asks for a visit in Haenam and names Amber as a remembered friend. The camper is still at the garage. 이현우: He is being awakened the following morning, with his treated leg still bandaged. 앰버: She is holding out the farewell letter.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막 잠에서 깬 이현우의 얼굴 바로 앞에 종이 편지를 뻗은 손으로 들이민 상태의 앰버 근접 찰나.\n\nLOCATION (lock): In the overnight sleeping area at the auto-repair premises, in morning light. The text does not establish that this is still the dining container. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain just behind and beside 이현우's shoulder at his reclining eye height, shifting only slightly sideways to see obliquely between his face and 앰버's extended hand. Let 앰버's face and reaching upper body occupy the central-right area, retaining 이현우's near cheek at the left edge and the letter across less than a quarter of the lower center, close to him but clear of her face. 앰버 looks down toward the newly awakened 이현우 while his lowered eyes begin to attend to the paper, preserving the urgency of a handoff rather than presenting the letter to the audience.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Extended paper letter below 앰버's face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held out close to 이현우's face) — The message-bearing face points toward 이현우 and is seen obliquely by the camera; it contains the farewell mentioning 해남 and remembering 앰버; used as Bridge the space between the two faces while leaving 앰버's expression unobstructed; Camper interior (Occupied during the next-day awakening); used as Provide a narrow contextual background without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination keeps the faces and paper distinct without turning the letter into a luminous graphic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's handwritten farewell note says thanks, asks for a visit in Haenam and names Amber as a remembered friend. The camper is still at the garage. 이현우: He is being awakened the following morning, with his treated leg still bandaged. 앰버: She is holding out the farewell letter.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 막 잠에서 깬 이현우의 얼굴 바로 앞에 종이 편지를 뻗은 손으로 들이민 상태의 앰버 근접 찰나.\n\nLOCATION (lock): In the overnight sleeping area at the auto-repair premises, in morning light. The text does not establish that this is still the dining container. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain just behind and beside 이현우's shoulder at his reclining eye height, shifting only slightly sideways to see obliquely between his face and 앰버's extended hand. Let 앰버's face and reaching upper body occupy the central-right area, retaining 이현우's near cheek at the left edge and the letter across less than a quarter of the lower center, close to him but clear of her face. 앰버 looks down toward the newly awakened 이현우 while his lowered eyes begin to attend to the paper, preserving the urgency of a handoff rather than presenting the letter to the audience.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Extended paper letter below 앰버's face in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 찰리's paper letter (Held out close to 이현우's face) — The message-bearing face points toward 이현우 and is seen obliquely by the camera; it contains the farewell mentioning 해남 and remembering 앰버; used as Bridge the space between the two faces while leaving 앰버's expression unobstructed; Camper interior (Occupied during the next-day awakening); used as Provide a narrow contextual background without introducing unspecified furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination keeps the faces and paper distinct without turning the letter into a luminous graphic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's handwritten farewell note says thanks, asks for a visit in Haenam and names Amber as a remembered friend. The camper is still at the garage. 이현우: He is being awakened the following morning, with his treated leg still bandaged. 앰버: She is holding out the farewell letter.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "앰버는 이현우를 향해 시선을 내리고 있으며, 이현우는 아래로 향한 시선으로 편지를 보고 있음.",
    "built_space": "이전 샷의 정비소 내부와 캠핑카가 배경에 일관성 있게 배치되어 있음.",
    "entities": "이현우의 측면, 앰버의 복장 및 방진 마스크, 편지의 텍스트 내용 모두 레퍼런스와 일치하게 나타남.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy: 편지를 들고 있는 앰버의 팔이 어깨가 아닌 가슴 중앙에서 뻗어 나옴"
    ],
    "physics": "편지를 들고 있는 앰버의 팔이 인체 구조상 불가능한 위치(가슴 중앙)에서 생성되어 자연스럽게 몸과 연결되지 않음."
   },
   {
    "label": "B",
    "direction": "앰버는 이현우를 내려다보고 있고, 이현우 역시 내민 편지를 향해 시선을 두고 있음.",
    "built_space": "정비소 배경과 뒤편의 캠핑카가 프롬프트 및 이전 샷과 일관성 있게 잘 렌더링됨.",
    "entities": "이현우의 헝클어진 머리와 인이어 무전기, 앰버의 복장과 마스크, 찰리의 편지에 적힌 텍스트 모두 레퍼런스를 정확히 반영함.",
    "hard_violations": [],
    "physics": "앰버의 오른팔이 어깨에서부터 자연스럽게 뻗어 나와 편지를 안정적으로 잡고 있으며 물리적으로 어색한 부분이 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요구된 클로즈업 구도와 인물 간의 시선 처리, 편지의 디테일 및 배경을 해부학적 오류 없이 자연스럽고 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "앰버의 팔이 가슴 중앙에서 뻗어 나오는 치명적인 해부학적 오류가 발생하여 탈락입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 이현우를 향해 시선을 내리고 있으며, 이현우는 아래로 향한 시선으로 편지를 보고 있음.",
        "built_space": "이전 샷의 정비소 내부와 캠핑카가 배경에 일관성 있게 배치되어 있음.",
        "entities": "이현우의 측면, 앰버의 복장 및 방진 마스크, 편지의 텍스트 내용 모두 레퍼런스와 일치하게 나타남.",
        "hard_violations": [
         "physically impossible anatomy: 편지를 들고 있는 앰버의 팔이 어깨가 아닌 가슴 중앙에서 뻗어 나옴"
        ],
        "physics": "편지를 들고 있는 앰버의 팔이 인체 구조상 불가능한 위치(가슴 중앙)에서 생성되어 자연스럽게 몸과 연결되지 않음."
       },
       {
        "label": "B",
        "direction": "앰버는 이현우를 내려다보고 있고, 이현우 역시 내민 편지를 향해 시선을 두고 있음.",
        "built_space": "정비소 배경과 뒤편의 캠핑카가 프롬프트 및 이전 샷과 일관성 있게 잘 렌더링됨.",
        "entities": "이현우의 헝클어진 머리와 인이어 무전기, 앰버의 복장과 마스크, 찰리의 편지에 적힌 텍스트 모두 레퍼런스를 정확히 반영함.",
        "hard_violations": [],
        "physics": "앰버의 오른팔이 어깨에서부터 자연스럽게 뻗어 나와 편지를 안정적으로 잡고 있으며 물리적으로 어색한 부분이 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요구된 클로즈업 구도와 인물 간의 시선 처리, 편지의 디테일 및 배경을 해부학적 오류 없이 자연스럽고 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "앰버의 팔이 가슴 중앙에서 뻗어 나오는 치명적인 해부학적 오류가 발생하여 탈락입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 이현우를 향해 시선을 내리고 있으며, 이현우는 아래로 향한 시선으로 편지를 보고 있음.",
        "built_space": "이전 샷의 정비소 내부와 캠핑카가 배경에 일관성 있게 배치되어 있음.",
        "entities": "이현우의 측면, 앰버의 복장 및 방진 마스크, 편지의 텍스트 내용 모두 레퍼런스와 일치하게 나타남.",
        "hard_violations": [
         "physically impossible anatomy: 편지를 들고 있는 앰버의 팔이 어깨가 아닌 가슴 중앙에서 뻗어 나옴"
        ],
        "physics": "편지를 들고 있는 앰버의 팔이 인체 구조상 불가능한 위치(가슴 중앙)에서 생성되어 자연스럽게 몸과 연결되지 않음."
       },
       {
        "label": "B",
        "direction": "앰버는 이현우를 내려다보고 있고, 이현우 역시 내민 편지를 향해 시선을 두고 있음.",
        "built_space": "정비소 배경과 뒤편의 캠핑카가 프롬프트 및 이전 샷과 일관성 있게 잘 렌더링됨.",
        "entities": "이현우의 헝클어진 머리와 인이어 무전기, 앰버의 복장과 마스크, 찰리의 편지에 적힌 텍스트 모두 레퍼런스를 정확히 반영함.",
        "hard_violations": [],
        "physics": "앰버의 오른팔이 어깨에서부터 자연스럽게 뻗어 나와 편지를 안정적으로 잡고 있으며 물리적으로 어색한 부분이 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "편지를 내미는 손과 두 사람의 시선은 맞지만, 현우가 왼쪽을 과도하게 차지하고 기대어 막 깨어난 자세가 불분명하며 이전 장면의 공간 배치도 B보다 덜 일치한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "침구와 캠핑카 정면, 작업대가 이전 장소를 더 충실히 이어 주고 편지 전달도 자연스럽지만, 요구한 캠핑카 내부 배경과 왼쪽 뺨만 남기는 밀착 구도는 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 왼쪽 아래의 현우 얼굴을 바라보고, 현우는 고개와 눈을 아래쪽 편지로 향한다. 앰버의 팔은 현우 쪽으로 뻗어 있다. 글씨 면은 현우와 카메라가 함께 있는 쪽을 향하므로 앞뒤가 뒤집힌 것은 아니지만, 종이가 카메라에 비교적 정면으로 보여 관객에게 제시하는 느낌도 있다.",
        "built_space": "배경에는 캠핑카 한 대의 옆면, 뒤쪽 작업대 하나와 공구 벽, 왼쪽 아래 보조 작업면, 오른쪽 출입구 하나가 보인다. 캠핑카 내부가 아니라 정비소 작업 공간이다. 이전 장면의 재질과 낮빛은 이어지지만 캠핑카와 작업 공간을 보는 축이 달라졌다. 현우의 머리와 어깨가 화면 왼쪽 약 3분의 1을 차지하여 뺨만 가장자리에 남기라는 지정보다 크다. 앰버는 중앙 오른쪽, 편지는 하단 중앙에 있고 얼굴을 가리지 않는다.",
        "entities": "두 사람만 보인다. 현우의 젊은 동아시아계 남성 외모, 헝클어진 검은 머리, 검은 인이어와 낡은 어두운 셔츠는 참조와 대체로 맞는다. 앰버는 금발의 어린 혼혈 소녀로 보이며 카키 작업복, 목에 걸린 기계식 방진 마스크와 가죽 공구 벨트가 참조에 가깝다. 편지는 누렇게 낡고 접힌 종이 한 장이며 참조의 한글 필체와 행 구성을 따른다. 다만 읽히는 문구에서 요구된 해남 방문과 앰버를 기억한다는 내용을 확인할 수 없다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 엄지와 다른 손가락이 종이 윗부분을 집고 있으며 손목과 팔이 작업복 소매로 자연스럽게 연결된다. 종이에는 접힘과 약한 처짐이 있어 손에 들린 물체로 성립한다. 앰버는 몸을 앞으로 숙였고 하체 지지점은 화면 밖이다. 현우의 좌면도 보이지 않지만 몸이 공중에 떠 있다고 판단할 근거는 없다. 다만 상체가 세워져 보여 기대어 누운 상태에서 막 깨어나는 순간은 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "앰버의 시선은 왼쪽 아래 현우의 얼굴에 닿고, 현우의 내려간 눈은 편지 쪽을 향한다. 뻗은 팔과 손은 종이를 현우 가까이 가져간다. 글씨 면은 현우 쪽을 향하면서 그의 옆에 있는 카메라에도 비스듬히 보이며, 종이는 앰버의 표정을 가리지 않는다.",
        "built_space": "오른쪽 뒤에 캠핑카 한 대의 정면, 중앙 뒤에 공구 벽과 작업대 하나, 위쪽 선반, 왼쪽 아래에 줄무늬 베개와 침구가 보인다. 이전 장면의 캠핑카 정면과 작업대의 관계, 잠자리 흔적이 A보다 잘 이어진다. 그러나 여기도 캠핑카 내부가 아니라 정비소 안의 잠자리다. 카메라는 현우의 어깨 옆뒤에 있지만 그의 머리와 어깨를 크게 포함한다. 앰버의 얼굴과 상체는 중앙 오른쪽에, 편지는 화면 면적 4분의 1보다 작은 하단 중앙에 놓인다.",
        "entities": "현우와 앰버 두 사람만 등장한다. 현우의 검은 흐트러진 머리, 젊은 동아시아계 남성 옆얼굴, 인이어와 먼지 묻은 셔츠가 참조와 부합한다. 앰버의 어린 얼굴, 금발, 밝은 피부, 카키 작업복과 목에 걸린 정교한 마스크가 참조를 따른다. 보이는 손은 앰버의 소매와 연결되고 피부색도 일치한다. 가죽 공구 벨트도 보인다. 편지는 참조와 유사한 낡은 한글 손편지이지만, 요구한 해남과 앰버 관련 작별 내용은 확인되지 않는다. 붕대가 있는 다리는 구도 밖이다.",
        "hard_violations": [],
        "physics": "앰버가 손가락으로 종이 위쪽을 확실히 집고 있고 종이 아래쪽은 중력에 따라 내려오며 휜다. 팔과 손목의 연결 및 앞으로 기운 상체는 편지를 내미는 동작으로 가능하다. 현우 바로 뒤와 아래에 베개와 침구가 있어 잠자리의 지지 맥락이 확인된다. 하체와 직접적인 좌면 접촉은 잘렸으므로 세부 지지는 확인할 수 없지만 부유나 불가능한 자세는 보이지 않는다. 현우는 여전히 비교적 세운 상체로 보여 요구한 누운 눈높이는 충분히 드러나지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "편지를 내미는 손과 두 사람의 시선은 맞지만, 현우가 왼쪽을 과도하게 차지하고 기대어 막 깨어난 자세가 불분명하며 이전 장면의 공간 배치도 B보다 덜 일치한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "침구와 캠핑카 정면, 작업대가 이전 장소를 더 충실히 이어 주고 편지 전달도 자연스럽지만, 요구한 캠핑카 내부 배경과 왼쪽 뺨만 남기는 밀착 구도는 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 왼쪽 아래의 현우 얼굴을 바라보고, 현우는 고개와 눈을 아래쪽 편지로 향한다. 앰버의 팔은 현우 쪽으로 뻗어 있다. 글씨 면은 현우와 카메라가 함께 있는 쪽을 향하므로 앞뒤가 뒤집힌 것은 아니지만, 종이가 카메라에 비교적 정면으로 보여 관객에게 제시하는 느낌도 있다.",
        "built_space": "배경에는 캠핑카 한 대의 옆면, 뒤쪽 작업대 하나와 공구 벽, 왼쪽 아래 보조 작업면, 오른쪽 출입구 하나가 보인다. 캠핑카 내부가 아니라 정비소 작업 공간이다. 이전 장면의 재질과 낮빛은 이어지지만 캠핑카와 작업 공간을 보는 축이 달라졌다. 현우의 머리와 어깨가 화면 왼쪽 약 3분의 1을 차지하여 뺨만 가장자리에 남기라는 지정보다 크다. 앰버는 중앙 오른쪽, 편지는 하단 중앙에 있고 얼굴을 가리지 않는다.",
        "entities": "두 사람만 보인다. 현우의 젊은 동아시아계 남성 외모, 헝클어진 검은 머리, 검은 인이어와 낡은 어두운 셔츠는 참조와 대체로 맞는다. 앰버는 금발의 어린 혼혈 소녀로 보이며 카키 작업복, 목에 걸린 기계식 방진 마스크와 가죽 공구 벨트가 참조에 가깝다. 편지는 누렇게 낡고 접힌 종이 한 장이며 참조의 한글 필체와 행 구성을 따른다. 다만 읽히는 문구에서 요구된 해남 방문과 앰버를 기억한다는 내용을 확인할 수 없다. 다리 붕대는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "앰버의 엄지와 다른 손가락이 종이 윗부분을 집고 있으며 손목과 팔이 작업복 소매로 자연스럽게 연결된다. 종이에는 접힘과 약한 처짐이 있어 손에 들린 물체로 성립한다. 앰버는 몸을 앞으로 숙였고 하체 지지점은 화면 밖이다. 현우의 좌면도 보이지 않지만 몸이 공중에 떠 있다고 판단할 근거는 없다. 다만 상체가 세워져 보여 기대어 누운 상태에서 막 깨어나는 순간은 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "앰버의 시선은 왼쪽 아래 현우의 얼굴에 닿고, 현우의 내려간 눈은 편지 쪽을 향한다. 뻗은 팔과 손은 종이를 현우 가까이 가져간다. 글씨 면은 현우 쪽을 향하면서 그의 옆에 있는 카메라에도 비스듬히 보이며, 종이는 앰버의 표정을 가리지 않는다.",
        "built_space": "오른쪽 뒤에 캠핑카 한 대의 정면, 중앙 뒤에 공구 벽과 작업대 하나, 위쪽 선반, 왼쪽 아래에 줄무늬 베개와 침구가 보인다. 이전 장면의 캠핑카 정면과 작업대의 관계, 잠자리 흔적이 A보다 잘 이어진다. 그러나 여기도 캠핑카 내부가 아니라 정비소 안의 잠자리다. 카메라는 현우의 어깨 옆뒤에 있지만 그의 머리와 어깨를 크게 포함한다. 앰버의 얼굴과 상체는 중앙 오른쪽에, 편지는 화면 면적 4분의 1보다 작은 하단 중앙에 놓인다.",
        "entities": "현우와 앰버 두 사람만 등장한다. 현우의 검은 흐트러진 머리, 젊은 동아시아계 남성 옆얼굴, 인이어와 먼지 묻은 셔츠가 참조와 부합한다. 앰버의 어린 얼굴, 금발, 밝은 피부, 카키 작업복과 목에 걸린 정교한 마스크가 참조를 따른다. 보이는 손은 앰버의 소매와 연결되고 피부색도 일치한다. 가죽 공구 벨트도 보인다. 편지는 참조와 유사한 낡은 한글 손편지이지만, 요구한 해남과 앰버 관련 작별 내용은 확인되지 않는다. 붕대가 있는 다리는 구도 밖이다.",
        "hard_violations": [],
        "physics": "앰버가 손가락으로 종이 위쪽을 확실히 집고 있고 종이 아래쪽은 중력에 따라 내려오며 휜다. 팔과 손목의 연결 및 앞으로 기운 상체는 편지를 내미는 동작으로 가능하다. 현우 바로 뒤와 아래에 베개와 침구가 있어 잠자리의 지지 맥락이 확인된다. 하체와 직접적인 좌면 접촉은 잘렸으므로 세부 지지는 확인할 수 없지만 부유나 불가능한 자세는 보이지 않는다. 현우는 여전히 비교적 세운 상체로 보여 요구한 누운 눈높이는 충분히 드러나지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 1.833
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.833
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy: 편지를 들고 있는 앰버의 팔이 어깨가 아닌 가슴 중앙에서 뻗어 나옴"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1833,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1833,
    "verdict_ko": "요구된 클로즈업 구도와 인물 간의 시선 처리, 편지의 디테일 및 배경을 해부학적 오류 없이 자연스럽고 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "앰버의 팔이 가슴 중앙에서 뻗어 나오는 치명적인 해부학적 오류가 발생하여 탈락입니다.  ★위반: [gemini-pro] physically impossible anatomy: 편지를 들고 있는 앰버의 팔이 어깨가 아닌 가슴 중앙에서 뻗어 나옴"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S50sh2_sel.png",
    "asset_id": "70fef7a7-e255-4d7a-9520-914b4b90df56",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 종이 편지: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1132412>",
    "asset_id": "a63a5a0b-350a-46b3-bbab-fe8ad262a556",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b90-06b4-7890-bc22-1bed1b4b5eac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh2"
  }
 },
 "S51sh32::signage": {
  "fp": "cf10ce6fdad69627",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh32": {
  "input_fingerprint": "42c6e9d1a0b97ea7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바퀴 뒤로 거친 흙먼지를 흩뿌리며 카센터 마당을 전속력으로 달리는 mid-action 순간의 캠핑카 뒷모습 전경.\n\nLOCATION (lock): In the outdoor yard of the isolated auto-repair shop, along the camper's departure route. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a low, static position in the yard, offset from the departure axis and looking along the camper's path at a shallow rear three-quarter angle. Keep the entire vehicle below one-third of the image area as it accelerates from lower center toward upper right, leaving clear space ahead and the described dirt trail behind its wheels. Emphasize only the increasing camera-to-vehicle distance as it recedes, with no visible occupants and no compensating pan.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper crossing the yard in the middle-center of the frame, midground, moves toward Open departure space at upper right; Open departure space in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating away across the yard) — Rear and one side are visible, with the front directed diagonally away toward the upper right; used as Provide the complete receding subject at a realistic scale within the yard; Car-center yard (Being crossed by the departing camper); used as Supply open departure space and a stable reference for acceleration; Thrown dirt behind the wheels (Scattering in the camper's wake); used as Make the departure impulse visible behind, rather than ahead of, the vehicle.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime exposure and controlled contrast keep the departing vehicle and thrown dirt legible without introducing additional atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is running again and leaving the garage with the luggage aboard. Its supplies now include additional food and medicine for Amber.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바퀴 뒤로 거친 흙먼지를 흩뿌리며 카센터 마당을 전속력으로 달리는 mid-action 순간의 캠핑카 뒷모습 전경.\n\nLOCATION (lock): In the outdoor yard of the isolated auto-repair shop, along the camper's departure route. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a low, static position in the yard, offset from the departure axis and looking along the camper's path at a shallow rear three-quarter angle. Keep the entire vehicle below one-third of the image area as it accelerates from lower center toward upper right, leaving clear space ahead and the described dirt trail behind its wheels. Emphasize only the increasing camera-to-vehicle distance as it recedes, with no visible occupants and no compensating pan.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper crossing the yard in the middle-center of the frame, midground, moves toward Open departure space at upper right; Open departure space in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating away across the yard) — Rear and one side are visible, with the front directed diagonally away toward the upper right; used as Provide the complete receding subject at a realistic scale within the yard; Car-center yard (Being crossed by the departing camper); used as Supply open departure space and a stable reference for acceleration; Thrown dirt behind the wheels (Scattering in the camper's wake); used as Make the departure impulse visible behind, rather than ahead of, the vehicle.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime exposure and controlled contrast keep the departing vehicle and thrown dirt legible without introducing additional atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is running again and leaving the garage with the luggage aboard. Its supplies now include additional food and medicine for Amber.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바퀴 뒤로 거친 흙먼지를 흩뿌리며 카센터 마당을 전속력으로 달리는 mid-action 순간의 캠핑카 뒷모습 전경.\n\nLOCATION (lock): In the outdoor yard of the isolated auto-repair shop, along the camper's departure route. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a low, static position in the yard, offset from the departure axis and looking along the camper's path at a shallow rear three-quarter angle. Keep the entire vehicle below one-third of the image area as it accelerates from lower center toward upper right, leaving clear space ahead and the described dirt trail behind its wheels. Emphasize only the increasing camera-to-vehicle distance as it recedes, with no visible occupants and no compensating pan.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Camper crossing the yard in the middle-center of the frame, midground, moves toward Open departure space at upper right; Open departure space in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Departing camper (Accelerating away across the yard) — Rear and one side are visible, with the front directed diagonally away toward the upper right; used as Provide the complete receding subject at a realistic scale within the yard; Car-center yard (Being crossed by the departing camper); used as Supply open departure space and a stable reference for acceleration; Thrown dirt behind the wheels (Scattering in the camper's wake); used as Make the departure impulse visible behind, rather than ahead of, the vehicle.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime exposure and controlled contrast keep the departing vehicle and thrown dirt legible without introducing additional atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is running again and leaving the garage with the luggage aboard. Its supplies now include additional food and medicine for Amber.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카가 프레임 우측 상단을 향해 주행하며 멀어지고 있으며, 차량 뒤쪽으로 흙먼지가 날리고 있습니다.",
    "built_space": "좌측에는 리프트에 차량이 올려진 카센터 건물이, 중앙에는 작은 사무실 건물이 위치합니다. 레퍼런스 이미지와 완벽히 동일한 구도와 높이에서 공간이 렌더링되었습니다.",
    "entities": "후면과 측면이 보이는 흰색 캠핑카가 등장하며, 지시대로 인물은 화면에 나타나지 않습니다.",
    "hard_violations": [],
    "physics": "캠핑카의 바퀴가 지면을 단단히 지지하고 있으며, 흩뿌려지는 흙먼지의 물리적 표현도 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "캠핑카의 전면부가 우측 상단을 향해 멀어지고 있으며, 후방에 흙먼지가 길게 흩뿌려지고 있습니다.",
    "built_space": "원본 레퍼런스의 카센터와 중앙 건물의 구조를 정확히 유지하면서, 지면에 매우 가까운 로우 앵글로 공간을 새롭게 재구성했습니다.",
    "entities": "흰색 캠핑카의 후면과 우측면이 보이며, 주변에 어떠한 인물도 등장하지 않아 지시사항을 잘 따랐습니다.",
    "hard_violations": [],
    "physics": "차량이 흙길 위를 안정적으로 주행 중이며, 바퀴의 회전과 이동 방향에 맞게 흙먼지가 물리적으로 타당하게 발생하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 매우 낮은 카메라 위치(low position)를 성공적으로 구현하였으며, 레퍼런스의 공간적 특징을 잘 유지했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스 이미지의 프레이밍을 절대 복사하지 말라는 지시를 어겼으며, 요구된 낮은 카메라 앵글도 반영하지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 프레임 우측 상단을 향해 주행하며 멀어지고 있으며, 차량 뒤쪽으로 흙먼지가 날리고 있습니다.",
        "built_space": "좌측에는 리프트에 차량이 올려진 카센터 건물이, 중앙에는 작은 사무실 건물이 위치합니다. 레퍼런스 이미지와 완벽히 동일한 구도와 높이에서 공간이 렌더링되었습니다.",
        "entities": "후면과 측면이 보이는 흰색 캠핑카가 등장하며, 지시대로 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면을 단단히 지지하고 있으며, 흩뿌려지는 흙먼지의 물리적 표현도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 전면부가 우측 상단을 향해 멀어지고 있으며, 후방에 흙먼지가 길게 흩뿌려지고 있습니다.",
        "built_space": "원본 레퍼런스의 카센터와 중앙 건물의 구조를 정확히 유지하면서, 지면에 매우 가까운 로우 앵글로 공간을 새롭게 재구성했습니다.",
        "entities": "흰색 캠핑카의 후면과 우측면이 보이며, 주변에 어떠한 인물도 등장하지 않아 지시사항을 잘 따랐습니다.",
        "hard_violations": [],
        "physics": "차량이 흙길 위를 안정적으로 주행 중이며, 바퀴의 회전과 이동 방향에 맞게 흙먼지가 물리적으로 타당하게 발생하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 매우 낮은 카메라 위치(low position)를 성공적으로 구현하였으며, 레퍼런스의 공간적 특징을 잘 유지했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "레퍼런스 이미지의 프레이밍을 절대 복사하지 말라는 지시를 어겼으며, 요구된 낮은 카메라 앵글도 반영하지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카가 프레임 우측 상단을 향해 주행하며 멀어지고 있으며, 차량 뒤쪽으로 흙먼지가 날리고 있습니다.",
        "built_space": "좌측에는 리프트에 차량이 올려진 카센터 건물이, 중앙에는 작은 사무실 건물이 위치합니다. 레퍼런스 이미지와 완벽히 동일한 구도와 높이에서 공간이 렌더링되었습니다.",
        "entities": "후면과 측면이 보이는 흰색 캠핑카가 등장하며, 지시대로 인물은 화면에 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "캠핑카의 바퀴가 지면을 단단히 지지하고 있으며, 흩뿌려지는 흙먼지의 물리적 표현도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 전면부가 우측 상단을 향해 멀어지고 있으며, 후방에 흙먼지가 길게 흩뿌려지고 있습니다.",
        "built_space": "원본 레퍼런스의 카센터와 중앙 건물의 구조를 정확히 유지하면서, 지면에 매우 가까운 로우 앵글로 공간을 새롭게 재구성했습니다.",
        "entities": "흰색 캠핑카의 후면과 우측면이 보이며, 주변에 어떠한 인물도 등장하지 않아 지시사항을 잘 따랐습니다.",
        "hard_violations": [],
        "physics": "차량이 흙길 위를 안정적으로 주행 중이며, 바퀴의 회전과 이동 방향에 맞게 흙먼지가 물리적으로 타당하게 발생하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 마당 시점과 작은 후방 사선의 캠핑카, 바퀴 뒤 흙먼지를 충실히 구현하지만 차량이 지정된 중간 중앙보다 오른쪽에 치우친다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "출발 방향과 장소는 맞지만 참고 사진의 높은 시점과 구도를 거의 답습하여 핵심인 낮은 카메라와 중앙 중경 배치를 놓친다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 뒤와 오른쪽 옆면이 보이며 앞머리는 화면 오른쪽 위의 비포장 출구를 향한다. 흙먼지는 바퀴 뒤에서 왼쪽 아래 방향으로 길게 남아 진행 방향과 맞는다. 사람이나 시선은 보이지 않는다.",
        "built_space": "왼쪽 정비동에는 중앙 기둥으로 나뉜 개방 작업구 두 칸과 내부 리프트 설비가 있고, 검은 차량 한 대가 올라가 있다. 중앙의 낮은 골판 외벽 부속동에는 출입문 하나와 창 세 개가 보이며, 문 양옆에 의자와 작업 가구가 놓였다. 양쪽 가장자리의 타이어·드럼통과 오른쪽 흙길도 참고 장소와 대응한다. 지면 가까운 카메라와 넓은 전경이 구현되었고 캠핑카 전체 면적은 화면의 3분의 1보다 훨씬 작지만, 차량 중심은 중앙이 아니라 오른쪽 중경에 있다.",
        "entities": "흰색 캠핑카 한 대는 참고의 후면 금속 운반대, 작은 후면 창, 세로형 후미등과 측면 창을 유지한다. 정비소, 흙마당, 출구, 타이어 더미와 먼지가 모두 보인다. 살아 있는 사람이나 탑승자는 보이지 않는다. 실내 짐·식량·의약품은 외부 구도에서 확인할 수 없으며 누락으로 판단할 근거가 없다. 낮 시간의 실물 재질로 표현되었고 뚜렷한 추가 문구나 화면 위 표식은 없다.",
        "hard_violations": [],
        "physics": "캠핑카는 타이어로 흙바닥을 딛고 있고 차체 아래 접지 그림자도 보인다. 바퀴 부근에서 시작해 뒤로 낮게 퍼지는 먼지는 가속으로 흙이 튀는 상황과 양립한다. 정비동의 검은 차량은 리프트에 지지되어 있으며, 주변 집기는 지면이나 작업대에 놓여 있다. 근거 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "캠핑카의 후면과 오른쪽 측면이 보이고 앞머리는 오른쪽 위의 열린 흙길을 향한다. 먼지는 차량 뒤에서 왼쪽 아래 마당 쪽으로 이어져 출발 궤적과 일치한다. 보이는 사람이나 시선은 없다.",
        "built_space": "왼쪽 정비동의 개방 작업구 두 칸, 중앙 기둥, 리프트 위 검은 차량 한 대가 참고와 같은 배치다. 중앙 부속동에는 출입문 하나, 정면 창 두 개와 오른쪽 차양 아래 창 부분이 보인다. 의자·작업 가구, 양쪽 타이어와 드럼통, 왼쪽 전경의 잘린 차량과 금속 구조물이 참고 구도와 거의 그대로 대응한다. 다만 지면을 상당히 내려다보는 시점으로 낮은 카메라 지시와 어긋나며, 작은 캠핑카도 중앙 중경보다 오른쪽 출구 가까이에 놓였다.",
        "entities": "흰색 캠핑카 한 대의 후면 운반대, 작은 창, 후미등과 측면 형태는 참고와 잘 맞는다. 정비소와 비포장 마당, 출구, 흙먼지가 보이며 사람이나 얼굴은 없다. 내부에 실린 짐과 추가 식량·의약품은 이 구도에서 확인되지 않는다. 낮의 자연광과 물리적인 표면 질감이 유지되며 추가 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴는 지면에 닿고, 바퀴 뒤로 먼지와 작은 흙 입자가 흩어져 가속 중인 차량의 후류로 읽힌다. 차량이 공중에 떠 있다는 증거는 없다. 정비 중인 검은 차량은 리프트가 받치고 있으며 드럼통·타이어·가구는 지면에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 마당 시점과 작은 후방 사선의 캠핑카, 바퀴 뒤 흙먼지를 충실히 구현하지만 차량이 지정된 중간 중앙보다 오른쪽에 치우친다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "출발 방향과 장소는 맞지만 참고 사진의 높은 시점과 구도를 거의 답습하여 핵심인 낮은 카메라와 중앙 중경 배치를 놓친다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카의 뒤와 오른쪽 옆면이 보이며 앞머리는 화면 오른쪽 위의 비포장 출구를 향한다. 흙먼지는 바퀴 뒤에서 왼쪽 아래 방향으로 길게 남아 진행 방향과 맞는다. 사람이나 시선은 보이지 않는다.",
        "built_space": "왼쪽 정비동에는 중앙 기둥으로 나뉜 개방 작업구 두 칸과 내부 리프트 설비가 있고, 검은 차량 한 대가 올라가 있다. 중앙의 낮은 골판 외벽 부속동에는 출입문 하나와 창 세 개가 보이며, 문 양옆에 의자와 작업 가구가 놓였다. 양쪽 가장자리의 타이어·드럼통과 오른쪽 흙길도 참고 장소와 대응한다. 지면 가까운 카메라와 넓은 전경이 구현되었고 캠핑카 전체 면적은 화면의 3분의 1보다 훨씬 작지만, 차량 중심은 중앙이 아니라 오른쪽 중경에 있다.",
        "entities": "흰색 캠핑카 한 대는 참고의 후면 금속 운반대, 작은 후면 창, 세로형 후미등과 측면 창을 유지한다. 정비소, 흙마당, 출구, 타이어 더미와 먼지가 모두 보인다. 살아 있는 사람이나 탑승자는 보이지 않는다. 실내 짐·식량·의약품은 외부 구도에서 확인할 수 없으며 누락으로 판단할 근거가 없다. 낮 시간의 실물 재질로 표현되었고 뚜렷한 추가 문구나 화면 위 표식은 없다.",
        "hard_violations": [],
        "physics": "캠핑카는 타이어로 흙바닥을 딛고 있고 차체 아래 접지 그림자도 보인다. 바퀴 부근에서 시작해 뒤로 낮게 퍼지는 먼지는 가속으로 흙이 튀는 상황과 양립한다. 정비동의 검은 차량은 리프트에 지지되어 있으며, 주변 집기는 지면이나 작업대에 놓여 있다. 근거 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "캠핑카의 후면과 오른쪽 측면이 보이고 앞머리는 오른쪽 위의 열린 흙길을 향한다. 먼지는 차량 뒤에서 왼쪽 아래 마당 쪽으로 이어져 출발 궤적과 일치한다. 보이는 사람이나 시선은 없다.",
        "built_space": "왼쪽 정비동의 개방 작업구 두 칸, 중앙 기둥, 리프트 위 검은 차량 한 대가 참고와 같은 배치다. 중앙 부속동에는 출입문 하나, 정면 창 두 개와 오른쪽 차양 아래 창 부분이 보인다. 의자·작업 가구, 양쪽 타이어와 드럼통, 왼쪽 전경의 잘린 차량과 금속 구조물이 참고 구도와 거의 그대로 대응한다. 다만 지면을 상당히 내려다보는 시점으로 낮은 카메라 지시와 어긋나며, 작은 캠핑카도 중앙 중경보다 오른쪽 출구 가까이에 놓였다.",
        "entities": "흰색 캠핑카 한 대의 후면 운반대, 작은 창, 후미등과 측면 형태는 참고와 잘 맞는다. 정비소와 비포장 마당, 출구, 흙먼지가 보이며 사람이나 얼굴은 없다. 내부에 실린 짐과 추가 식량·의약품은 이 구도에서 확인되지 않는다. 낮의 자연광과 물리적인 표면 질감이 유지되며 추가 자막이나 도식은 없다.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴는 지면에 닿고, 바퀴 뒤로 먼지와 작은 흙 입자가 흩어져 가속 중인 차량의 후류로 읽힌다. 차량이 공중에 떠 있다는 증거는 없다. 정비 중인 검은 차량은 리프트가 받치고 있으며 드럼통·타이어·가구는 지면에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.179,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.179,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 매우 낮은 카메라 위치(low position)를 성공적으로 구현하였으며, 레퍼런스의 공간적 특징을 잘 유지했습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "레퍼런스 이미지의 프레이밍을 절대 복사하지 말라는 지시를 어겼으며, 요구된 낮은 카메라 앵글도 반영하지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L72B01.png",
    "asset_id": "d16f5ca7-5448-4082-bac5-fae962a039a1",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b95-2e9b-78aa-aab0-1d90c804440c",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S52sh1::signage": {
  "fp": "73cc0ba5098c96c0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3800878e392e40fa": {
  "subjects": [],
  "subject_text": "찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길\n내륙을 잇는 지방도로와 양 갈래 분기점. 길옆 숲으로 샛길이 갈라지고, 안쪽으로 나무 사이 산길이 이어진다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L173",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::solo_country_road": {
  "input_fingerprint": "49f625580cbf2766",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "solo_country_road",
    "tags": [
     "S52sh1"
    ]
   },
   "context_sig": "d790a67865c2a4c8"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the middle of an empty rural road in daylight, away from the auto-repair shop.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 52. 지방도로 – D\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the middle of an empty rural road in daylight, away from the auto-repair shop.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 52. 지방도로 – D\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_solo_country_road_d24ba1.png",
  "asset_id": "70f6fa08-8e0e-4c75-99d5-4f9275a5d0bd",
  "input_asset_ids": [
   "f12b2d5c-4c71-4c55-a156-7891d8ed85b3"
  ],
  "origin_tag": "S52sh1",
  "place_text": "In the middle of an empty rural road in daylight, away from the auto-repair shop.",
  "origin_inputs": {
   "place_text": "In the middle of an empty rural road in daylight, away from the auto-repair shop.",
   "time_of_day_en": "day",
   "conti_asset_id": "f12b2d5c-4c71-4c55-a156-7891d8ed85b3"
  }
 },
 "S52sh1::bgfirst_bg": {
  "input_fingerprint": "6f405a305ad38208",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 아무도 없는 한적한 지방도로 한가운데, 한 발을 앞으로 뻗고 다른 발로 땅을 밀어내는 mid-stride 자세로 묵묵히 나아가는 비장한 기세의 찰리의 넓은 전신 구도.\n\nLOCATION (lock): In the middle of an empty rural road in daylight, away from the auto-repair shop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track ahead of 찰리 and diagonally to one side from slightly above his head, tilting gently downward while retaining his complete body and generous empty road around him. Place him just left of center, one large boot reaching forward as the other pushes off, with the oversized straw hat and colorful raincoat preserving the comic contrast to his solemn expression. His attention stays on the road ahead beyond the frame, not the lens; maintain this forward three-quarter axis before the subsequent gradual approach and lowering.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial road (Empty around 찰리) — The road extends behind him and continues toward his off-screen destination; used as Give the solitary full-body stride spatial weight without adding roadside structures or weather.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves the raincoat's scripted colors within a restrained neutral grade and keeps the robot's visible hard surfaces precise.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 아무도 없는 한적한 지방도로 한가운데, 한 발을 앞으로 뻗고 다른 발로 땅을 밀어내는 mid-stride 자세로 묵묵히 나아가는 비장한 기세의 찰리의 넓은 전신 구도.\n\nLOCATION (lock): In the middle of an empty rural road in daylight, away from the auto-repair shop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track ahead of 찰리 and diagonally to one side from slightly above his head, tilting gently downward while retaining his complete body and generous empty road around him. Place him just left of center, one large boot reaching forward as the other pushes off, with the oversized straw hat and colorful raincoat preserving the comic contrast to his solemn expression. His attention stays on the road ahead beyond the frame, not the lens; maintain this forward three-quarter axis before the subsequent gradual approach and lowering.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial road (Empty around 찰리) — The road extends behind him and continues toward his off-screen destination; used as Give the solitary full-body stride spatial weight without adding roadside structures or weather.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves the raincoat's scripted colors within a restrained neutral grade and keeps the robot's visible hard surfaces precise.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S52sh1__bgfirst_bg.png",
  "asset_id": "7dae5a11-d0cd-4944-9e7a-d897543bb761",
  "input_asset_ids": [
   "f12b2d5c-4c71-4c55-a156-7891d8ed85b3",
   "70f6fa08-8e0e-4c75-99d5-4f9275a5d0bd"
  ]
 },
 "S52sh1": {
  "input_fingerprint": "1766edee4fbf7de4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아무도 없는 한적한 지방도로 한가운데, 한 발을 앞으로 뻗고 다른 발로 땅을 밀어내는 mid-stride 자세로 묵묵히 나아가는 비장한 기세의 찰리의 넓은 전신 구도.\n\nLOCATION (lock): In the middle of an empty rural road in daylight, away from the auto-repair shop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track ahead of 찰리 and diagonally to one side from slightly above his head, tilting gently downward while retaining his complete body and generous empty road around him. Place him just left of center, one large boot reaching forward as the other pushes off, with the oversized straw hat and colorful raincoat preserving the comic contrast to his solemn expression. His attention stays on the road ahead beyond the frame, not the lens; maintain this forward three-quarter axis before the subsequent gradual approach and lowering.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial road (Empty around 찰리) — The road extends behind him and continues toward his off-screen destination; used as Give the solitary full-body stride spatial weight without adding roadside structures or weather.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves the raincoat's scripted colors within a restrained neutral grade and keeps the robot's visible hard surfaces precise.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 찰리: His shoulder is repaired, and he now wears an oversized straw hat, oversized boots and a colorful raincoat while walking along the rural road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아무도 없는 한적한 지방도로 한가운데, 한 발을 앞으로 뻗고 다른 발로 땅을 밀어내는 mid-stride 자세로 묵묵히 나아가는 비장한 기세의 찰리의 넓은 전신 구도.\n\nLOCATION (lock): In the middle of an empty rural road in daylight, away from the auto-repair shop. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track ahead of 찰리 and diagonally to one side from slightly above his head, tilting gently downward while retaining his complete body and generous empty road around him. Place him just left of center, one large boot reaching forward as the other pushes off, with the oversized straw hat and colorful raincoat preserving the comic contrast to his solemn expression. His attention stays on the road ahead beyond the frame, not the lens; maintain this forward three-quarter axis before the subsequent gradual approach and lowering.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial road (Empty around 찰리) — The road extends behind him and continues toward his off-screen destination; used as Give the solitary full-body stride spatial weight without adding roadside structures or weather.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves the raincoat's scripted colors within a restrained neutral grade and keeps the robot's visible hard surfaces precise.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 찰리: His shoulder is repaired, and he now wears an oversized straw hat, oversized boots and a colorful raincoat while walking along the rural road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아무도 없는 한적한 지방도로 한가운데, 한 발을 앞으로 뻗고 다른 발로 땅을 밀어내는 mid-stride 자세로 묵묵히 나아가는 비장한 기세의 찰리의 넓은 전신 구도.\n\nLOCATION (lock): In the middle of an empty rural road in daylight, away from the auto-repair shop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track ahead of 찰리 and diagonally to one side from slightly above his head, tilting gently downward while retaining his complete body and generous empty road around him. Place him just left of center, one large boot reaching forward as the other pushes off, with the oversized straw hat and colorful raincoat preserving the comic contrast to his solemn expression. His attention stays on the road ahead beyond the frame, not the lens; maintain this forward three-quarter axis before the subsequent gradual approach and lowering.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Quiet provincial road (Empty around 찰리) — The road extends behind him and continues toward his off-screen destination; used as Give the solitary full-body stride spatial weight without adding roadside structures or weather.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves the raincoat's scripted colors within a restrained neutral grade and keeps the robot's visible hard surfaces precise.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): 찰리: His shoulder is repaired, and he now wears an oversized straw hat, oversized boots and a colorful raincoat while walking along the rural road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S52sh1__bgfirst_bg.png",
     "asset_id": "7dae5a11-d0cd-4944-9e7a-d897543bb761",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S52sh1.png",
     "asset_id": "f12b2d5c-4c71-4c55-a156-7891d8ed85b3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_solo_country_road_d24ba1.png",
     "asset_id": "70f6fa08-8e0e-4c75-99d5-4f9275a5d0bd",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
    "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
    "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 단, 레퍼런스의 고릴라 비율이 아닌 긴 다리의 인간형 체형이며, 마스크 형태가 다르고 가슴 원자로가 보이지 않음.",
    "hard_violations": [],
    "physics": "오른발을 앞으로 뻗고 왼발로 딛는 공중 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
    "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
    "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 짧은 다리와 육중한 팔, 마스크의 입 선, 가슴의 푸른 원자로 등 캐릭터 레퍼런스를 정확히 반영함.",
    "hard_violations": [],
    "physics": "오른발을 앞으로 뻗고 왼발로 밀어내는 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 고유한 고릴라형 신체 비율(짧은 다리와 육중한 팔)과 가슴의 원자로를 정확히 구현했으며, 지정된 걷는 자세와 넓은 구도를 잘 표현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "요청된 3/4 측면 구도와 걷는 자세는 잘 살렸으나, 찰리의 신체 비율을 무시하고 일반적인 인간형 체형으로 렌더링하여 캐릭터 정체성에서 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
        "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
        "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 단, 레퍼런스의 고릴라 비율이 아닌 긴 다리의 인간형 체형이며, 마스크 형태가 다르고 가슴 원자로가 보이지 않음.",
        "hard_violations": [],
        "physics": "오른발을 앞으로 뻗고 왼발로 딛는 공중 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
        "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
        "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 짧은 다리와 육중한 팔, 마스크의 입 선, 가슴의 푸른 원자로 등 캐릭터 레퍼런스를 정확히 반영함.",
        "hard_violations": [],
        "physics": "오른발을 앞으로 뻗고 왼발로 밀어내는 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 고유한 고릴라형 신체 비율(짧은 다리와 육중한 팔)과 가슴의 원자로를 정확히 구현했으며, 지정된 걷는 자세와 넓은 구도를 잘 표현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "요청된 3/4 측면 구도와 걷는 자세는 잘 살렸으나, 찰리의 신체 비율을 무시하고 일반적인 인간형 체형으로 렌더링하여 캐릭터 정체성에서 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
        "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
        "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 단, 레퍼런스의 고릴라 비율이 아닌 긴 다리의 인간형 체형이며, 마스크 형태가 다르고 가슴 원자로가 보이지 않음.",
        "hard_violations": [],
        "physics": "오른발을 앞으로 뻗고 왼발로 딛는 공중 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "캐릭터의 시선과 몸체가 프레임 밖 도로 정면을 향해 나아감.",
        "built_space": "중앙선과 우측 전신주가 있는 한적한 시골 도로. 배경 레퍼런스의 공간 구조와 일치함.",
        "entities": "밀짚모자, 화려한 우비, 장화를 착용한 로봇. 짧은 다리와 육중한 팔, 마스크의 입 선, 가슴의 푸른 원자로 등 캐릭터 레퍼런스를 정확히 반영함.",
        "hard_violations": [],
        "physics": "오른발을 앞으로 뻗고 왼발로 밀어내는 보행 자세. 지면의 그림자와 체중 지지 상태가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "A는 머리보다 높은 전방 사선 시점, 왼쪽에 놓인 전신과 도로 여백, 짧고 육중한 기계 체형을 더 충실히 구현하지만, 조명은 요구보다 강하고 뒷발의 밀어내기는 다소 약하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "B는 큰 보폭과 렌즈 밖을 향한 주시는 맞지만, 낮아진 카메라와 커진 인물 비중, 길고 사람 같은 다리 비율이 지정된 와이드 하향 시점과 찰리의 체형에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 화면 아래쪽의 도로 전방으로 걸어오며, 얼굴은 화면 오른쪽으로 조금 돌아 렌즈 오른편의 화면 밖을 향한다. 렌즈를 똑바로 응시하지 않는다. 화면 왼쪽의 큰 장화가 전방으로 뻗고 오른쪽 장화가 뒤에 남아 있어 진행 방향이 읽힌다. 겨누는 도구나 손에 든 물건은 없다.",
        "built_space": "노란 점선 중앙선 하나와 양쪽 흰 가장자리선 두 줄이 있는 아스팔트 도로다. 왼쪽에는 굽은 농로 접속부 하나와 숲, 양옆에는 논, 오른쪽에는 최소 여섯 개가 식별되는 전신주 열과 전선이 있다. 원경의 산과 작은 시설물도 위치 참조와 대체로 맞는다. 찰리는 도로 왼쪽 차로 안쪽이자 화면 중앙보다 왼쪽에 있고, 주변에 사람이나 차량은 없다. 모자 윗면과 노면이 많이 보이는 완만한 하향 시점으로 전신과 앞뒤 도로 여백을 확보했다.",
        "entities": "등장 개체는 기계 찰리 한 명뿐이다. 흰 마스크형 얼굴, 주황색 원형 눈 두 개, 검은 입 선 하나, 샌드 베이지 장갑판과 가슴의 푸른 원형 동력부가 보인다. 사람의 피부나 치아는 없다. 넓은 몸통, 긴 팔과 비교적 짧은 다리는 참조의 고릴라형 비율에 가깝다. 큰 밀짚모자, 노랑·파랑·적갈색 우비, 큰 검은 장화 두 짝이 있다. 어깨는 우비에 가려 수리 상태를 직접 확인하기 어렵지만 분리되거나 망가진 흔적은 보이지 않는다. 직사광과 진한 그림자는 차분한 낮 조명 요구보다 강하다.",
        "hard_violations": [],
        "physics": "뒤쪽 장화의 앞부분이 노면에 닿아 몸을 지지하고, 앞으로 뻗은 장화는 뒤꿈치가 내려오는 보행 단계로 읽힌다. 뒷발의 뒤꿈치 들림은 크지 않아 강한 밀어내기보다는 평범한 걸음에 가깝지만, 몸이 지지 없이 떠 있지는 않다. 팔은 어깨와 팔꿈치 관절에 연결되어 자연스럽게 내려오며, 모자는 머리에 얹히고 턱끈이 내려온다. 우비도 어깨와 몸통에 걸쳐 주름지고 처진다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 아래쪽 전방으로 크게 걸어오고, 얼굴과 눈은 렌즈보다 화면 오른쪽의 바깥 도로 방향을 향한다. 앞 장화는 화면 아래 왼쪽으로 뻗고 반대쪽 다리는 몸 뒤에 남아 있다. 팔도 보행에 맞춰 앞뒤로 엇갈린다. 손에 든 물건이나 특정 대상을 겨누는 도구는 없다.",
        "built_space": "노란 점선 중앙선 하나, 흰 가장자리선 두 줄, 왼쪽 농로 접속부 하나, 양옆 논과 왼쪽 숲, 오른쪽 전신주 열과 뒤쪽 산이 보여 장소 참조를 대체로 유지한다. 전신주는 최소 여섯 개가 식별되며 원경에서 더 겹쳐진다. 찰리는 화면 중앙보다 왼쪽의 차로 안에 있고 다른 사람이나 차량은 없다. 다만 수평선이 몸 뒤로 높게 펼쳐지고 모자 윗면의 노출이 적어, 요구된 머리 위 하향 시점보다 낮은 카메라로 보인다. 전신이 화면 높이 대부분을 차지해 A보다 위아래 도로 여백이 좁다.",
        "entities": "기계 찰리 한 명만 보이며 흰 얼굴판, 주황색 눈 두 개와 검은 입 선, 베이지 장갑판, 기계 손은 참조의 특징을 유지한다. 큰 밀짚모자와 노랑·파랑·붉은색 우비, 큰 장화 두 짝도 있다. 가슴 동력부와 어깨는 우비에 가려져 있어 그 부위의 일치나 수리 상태는 판단하지 않는다. 그러나 드러난 다리가 길고 몸이 세로로 늘어나, 짧고 육중한 고릴라형 기계보다는 보통 사람의 보행 비율에 가까워졌다. 햇빛과 그림자 역시 요구보다 강하다.",
        "hard_violations": [],
        "physics": "앞 장화는 발끝이 들리고 뒤꿈치가 노면에 닿는 단계로 보이며, 뒤쪽 장화는 앞다리 뒤로 일부 가려진 채 발끝을 남기는 보행 자세다. 지지 없이 공중에 매달린 몸으로 보이지는 않는다. 큰 보폭과 반대 방향의 팔 움직임은 걷기로 설명할 수 있다. 모자는 머리에 놓여 있고 우비는 어깨에서 매달려 몸과 다리 움직임에 따라 접힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "A는 머리보다 높은 전방 사선 시점, 왼쪽에 놓인 전신과 도로 여백, 짧고 육중한 기계 체형을 더 충실히 구현하지만, 조명은 요구보다 강하고 뒷발의 밀어내기는 다소 약하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "B는 큰 보폭과 렌즈 밖을 향한 주시는 맞지만, 낮아진 카메라와 커진 인물 비중, 길고 사람 같은 다리 비율이 지정된 와이드 하향 시점과 찰리의 체형에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 아래쪽의 도로 전방으로 걸어오며, 얼굴은 화면 오른쪽으로 조금 돌아 렌즈 오른편의 화면 밖을 향한다. 렌즈를 똑바로 응시하지 않는다. 화면 왼쪽의 큰 장화가 전방으로 뻗고 오른쪽 장화가 뒤에 남아 있어 진행 방향이 읽힌다. 겨누는 도구나 손에 든 물건은 없다.",
        "built_space": "노란 점선 중앙선 하나와 양쪽 흰 가장자리선 두 줄이 있는 아스팔트 도로다. 왼쪽에는 굽은 농로 접속부 하나와 숲, 양옆에는 논, 오른쪽에는 최소 여섯 개가 식별되는 전신주 열과 전선이 있다. 원경의 산과 작은 시설물도 위치 참조와 대체로 맞는다. 찰리는 도로 왼쪽 차로 안쪽이자 화면 중앙보다 왼쪽에 있고, 주변에 사람이나 차량은 없다. 모자 윗면과 노면이 많이 보이는 완만한 하향 시점으로 전신과 앞뒤 도로 여백을 확보했다.",
        "entities": "등장 개체는 기계 찰리 한 명뿐이다. 흰 마스크형 얼굴, 주황색 원형 눈 두 개, 검은 입 선 하나, 샌드 베이지 장갑판과 가슴의 푸른 원형 동력부가 보인다. 사람의 피부나 치아는 없다. 넓은 몸통, 긴 팔과 비교적 짧은 다리는 참조의 고릴라형 비율에 가깝다. 큰 밀짚모자, 노랑·파랑·적갈색 우비, 큰 검은 장화 두 짝이 있다. 어깨는 우비에 가려 수리 상태를 직접 확인하기 어렵지만 분리되거나 망가진 흔적은 보이지 않는다. 직사광과 진한 그림자는 차분한 낮 조명 요구보다 강하다.",
        "hard_violations": [],
        "physics": "뒤쪽 장화의 앞부분이 노면에 닿아 몸을 지지하고, 앞으로 뻗은 장화는 뒤꿈치가 내려오는 보행 단계로 읽힌다. 뒷발의 뒤꿈치 들림은 크지 않아 강한 밀어내기보다는 평범한 걸음에 가깝지만, 몸이 지지 없이 떠 있지는 않다. 팔은 어깨와 팔꿈치 관절에 연결되어 자연스럽게 내려오며, 모자는 머리에 얹히고 턱끈이 내려온다. 우비도 어깨와 몸통에 걸쳐 주름지고 처진다."
       },
       {
        "label": "A",
        "direction": "찰리는 화면 아래쪽 전방으로 크게 걸어오고, 얼굴과 눈은 렌즈보다 화면 오른쪽의 바깥 도로 방향을 향한다. 앞 장화는 화면 아래 왼쪽으로 뻗고 반대쪽 다리는 몸 뒤에 남아 있다. 팔도 보행에 맞춰 앞뒤로 엇갈린다. 손에 든 물건이나 특정 대상을 겨누는 도구는 없다.",
        "built_space": "노란 점선 중앙선 하나, 흰 가장자리선 두 줄, 왼쪽 농로 접속부 하나, 양옆 논과 왼쪽 숲, 오른쪽 전신주 열과 뒤쪽 산이 보여 장소 참조를 대체로 유지한다. 전신주는 최소 여섯 개가 식별되며 원경에서 더 겹쳐진다. 찰리는 화면 중앙보다 왼쪽의 차로 안에 있고 다른 사람이나 차량은 없다. 다만 수평선이 몸 뒤로 높게 펼쳐지고 모자 윗면의 노출이 적어, 요구된 머리 위 하향 시점보다 낮은 카메라로 보인다. 전신이 화면 높이 대부분을 차지해 A보다 위아래 도로 여백이 좁다.",
        "entities": "기계 찰리 한 명만 보이며 흰 얼굴판, 주황색 눈 두 개와 검은 입 선, 베이지 장갑판, 기계 손은 참조의 특징을 유지한다. 큰 밀짚모자와 노랑·파랑·붉은색 우비, 큰 장화 두 짝도 있다. 가슴 동력부와 어깨는 우비에 가려져 있어 그 부위의 일치나 수리 상태는 판단하지 않는다. 그러나 드러난 다리가 길고 몸이 세로로 늘어나, 짧고 육중한 고릴라형 기계보다는 보통 사람의 보행 비율에 가까워졌다. 햇빛과 그림자 역시 요구보다 강하다.",
        "hard_violations": [],
        "physics": "앞 장화는 발끝이 들리고 뒤꿈치가 노면에 닿는 단계로 보이며, 뒤쪽 장화는 앞다리 뒤로 일부 가려진 채 발끝을 남기는 보행 자세다. 지지 없이 공중에 매달린 몸으로 보이지는 않는다. 큰 보폭과 반대 방향의 팔 움직임은 걷기로 설명할 수 있다. 모자는 머리에 놓여 있고 우비는 어깨에서 매달려 몸과 다리 움직임에 따라 접힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "찰리의 고유한 고릴라형 신체 비율(짧은 다리와 육중한 팔)과 가슴의 원자로를 정확히 구현했으며, 지정된 걷는 자세와 넓은 구도를 잘 표현했습니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "요청된 3/4 측면 구도와 걷는 자세는 잘 살렸으나, 찰리의 신체 비율을 무시하고 일반적인 인간형 체형으로 렌더링하여 캐릭터 정체성에서 크게 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_solo_country_road_d24ba1.png",
    "asset_id": "70f6fa08-8e0e-4c75-99d5-4f9275a5d0bd",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0b99-918b-78e0-b961-c193ddd74ac9",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S52sh1__bgfirst_bg.png",
   "bg_asset_id": "7dae5a11-d0cd-4944-9e7a-d897543bb761",
   "bg_record_key": "S52sh1::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "solo_country_road",
   "groupbg_asset_id": "70f6fa08-8e0e-4c75-99d5-4f9275a5d0bd"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S53sh3::signage": {
  "fp": "281bfee6435307c5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh3": {
  "input_fingerprint": "1652d4c1b7ccb0aa",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 분노에 찬 일그러진 얼굴로 다트 핀을 책상 위 지도에 막 꽂아 넣은, 쥔 주먹에 강하게 힘이 들어간 순간.\n\nLOCATION (lock): At the map-covered desk inside the militia commander's city office, under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the desk corner, begin the tracking stage in a settled three-quarter view slightly above 박철진's eyes, tilting down enough to hold his distorted face in the upper left and his clenched hand in the lower center. His shoulders pitch toward the desk and his attention remains on the dart he has just planted in the map; keep the face and fist legible together, with the map occupying less than a third of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 (Supports the spread map and the freshly planted dart) — The corner recedes diagonally beneath his forearm; used as Connects the foreground hand to the face without exaggerated foreground scale; 한반도 지도 (Spread across the desktop with a dart just planted in it) — The printed map face is visible obliquely from above; used as Provides the immediate target of his anger below the fist; 다트 핀 (Just planted in the map beside his tightly clenched fingers) — The shaft projects from the map toward his hand; used as Makes the completed action readable at natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the tension in his expression and knuckles without introducing a conspicuous light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula is spread across the office desk. 박철진: He has darts at hand while considering directions.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 분노에 찬 일그러진 얼굴로 다트 핀을 책상 위 지도에 막 꽂아 넣은, 쥔 주먹에 강하게 힘이 들어간 순간.\n\nLOCATION (lock): At the map-covered desk inside the militia commander's city office, under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the desk corner, begin the tracking stage in a settled three-quarter view slightly above 박철진's eyes, tilting down enough to hold his distorted face in the upper left and his clenched hand in the lower center. His shoulders pitch toward the desk and his attention remains on the dart he has just planted in the map; keep the face and fist legible together, with the map occupying less than a third of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 (Supports the spread map and the freshly planted dart) — The corner recedes diagonally beneath his forearm; used as Connects the foreground hand to the face without exaggerated foreground scale; 한반도 지도 (Spread across the desktop with a dart just planted in it) — The printed map face is visible obliquely from above; used as Provides the immediate target of his anger below the fist; 다트 핀 (Just planted in the map beside his tightly clenched fingers) — The shaft projects from the map toward his hand; used as Makes the completed action readable at natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the tension in his expression and knuckles without introducing a conspicuous light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula is spread across the office desk. 박철진: He has darts at hand while considering directions.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박철진이 분노에 찬 일그러진 얼굴로 다트 핀을 책상 위 지도에 막 꽂아 넣은, 쥔 주먹에 강하게 힘이 들어간 순간.\n\nLOCATION (lock): At the map-covered desk inside the militia commander's city office, under ordinary office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the desk corner, begin the tracking stage in a settled three-quarter view slightly above 박철진's eyes, tilting down enough to hold his distorted face in the upper left and his clenched hand in the lower center. His shoulders pitch toward the desk and his attention remains on the dart he has just planted in the map; keep the face and fist legible together, with the map occupying less than a third of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 (Supports the spread map and the freshly planted dart) — The corner recedes diagonally beneath his forearm; used as Connects the foreground hand to the face without exaggerated foreground scale; 한반도 지도 (Spread across the desktop with a dart just planted in it) — The printed map face is visible obliquely from above; used as Provides the immediate target of his anger below the fist; 다트 핀 (Just planted in the map beside his tightly clenched fingers) — The shaft projects from the map toward his hand; used as Makes the completed action readable at natural scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the tension in his expression and knuckles without introducing a conspicuous light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A map of the Korean Peninsula is spread across the office desk. 박철진: He has darts at hand while considering directions.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 박철진 right now, so 박철진's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 박철진: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 다트가 아닌 앞쪽 아래 허공을 향하고 있으며, 다트는 책상 위 지도를 향해 꽂혀 있음.",
    "built_space": "사무실 내부. 창밖으로 야간의 도시 불빛이 보이며, 배경에 데스크 램프와 서류철이 배치됨.",
    "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 지도.",
    "hard_violations": [],
    "physics": "오른손이 다트 핀을 쥐고 지도에 누르고 있으며, 왼손은 책상을 짚어 상체를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "시선은 책상 위 지도에 꽂힌 다트를 정확히 향하고 있으며, 다트는 지도를 수직으로 찌르고 있음.",
    "built_space": "사무실 내부. 창밖으로 지시된 주간(낮)의 풍경이 보이며, 깃발과 벽면 지도가 제 위치에 있음.",
    "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 한반도 지도.",
    "hard_violations": [],
    "physics": "오른손이 다트를 단단히 쥐어 지도에 꽂고 있으며, 양팔과 몸의 무게 중심이 책상을 향해 자연스럽게 쏠려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 주간 시간대, 상단 좌측의 얼굴과 하단 중앙의 주먹을 배치하는 구도, 그리고 다트 핀을 향한 시선 처리를 모두 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "시간대가 야간으로 잘못 묘사되었고 프레이밍 지시(얼굴과 손의 위치)를 따르지 않았으며, 시선이 다트를 향하지 않고 앞을 향하고 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 다트가 아닌 앞쪽 아래 허공을 향하고 있으며, 다트는 책상 위 지도를 향해 꽂혀 있음.",
        "built_space": "사무실 내부. 창밖으로 야간의 도시 불빛이 보이며, 배경에 데스크 램프와 서류철이 배치됨.",
        "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 지도.",
        "hard_violations": [],
        "physics": "오른손이 다트 핀을 쥐고 지도에 누르고 있으며, 왼손은 책상을 짚어 상체를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 책상 위 지도에 꽂힌 다트를 정확히 향하고 있으며, 다트는 지도를 수직으로 찌르고 있음.",
        "built_space": "사무실 내부. 창밖으로 지시된 주간(낮)의 풍경이 보이며, 깃발과 벽면 지도가 제 위치에 있음.",
        "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 한반도 지도.",
        "hard_violations": [],
        "physics": "오른손이 다트를 단단히 쥐어 지도에 꽂고 있으며, 양팔과 몸의 무게 중심이 책상을 향해 자연스럽게 쏠려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 주간 시간대, 상단 좌측의 얼굴과 하단 중앙의 주먹을 배치하는 구도, 그리고 다트 핀을 향한 시선 처리를 모두 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "시간대가 야간으로 잘못 묘사되었고 프레이밍 지시(얼굴과 손의 위치)를 따르지 않았으며, 시선이 다트를 향하지 않고 앞을 향하고 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 다트가 아닌 앞쪽 아래 허공을 향하고 있으며, 다트는 책상 위 지도를 향해 꽂혀 있음.",
        "built_space": "사무실 내부. 창밖으로 야간의 도시 불빛이 보이며, 배경에 데스크 램프와 서류철이 배치됨.",
        "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 지도.",
        "hard_violations": [],
        "physics": "오른손이 다트 핀을 쥐고 지도에 누르고 있으며, 왼손은 책상을 짚어 상체를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "시선은 책상 위 지도에 꽂힌 다트를 정확히 향하고 있으며, 다트는 지도를 수직으로 찌르고 있음.",
        "built_space": "사무실 내부. 창밖으로 지시된 주간(낮)의 풍경이 보이며, 깃발과 벽면 지도가 제 위치에 있음.",
        "entities": "박철진(얼굴, 모자, 복장, 완장 일치), 다트 핀(빨간색 날개), 한반도 지도.",
        "hard_violations": [],
        "physics": "오른손이 다트를 단단히 쥐어 지도에 꽂고 있으며, 양팔과 몸의 무게 중심이 책상을 향해 자연스럽게 쏠려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "다트를 내려다보는 시선과 좌상단 얼굴·하단 중앙 주먹의 연결, 지도에 막 꽂은 동작이 정확하며, 다만 요구한 클로즈업보다 상체와 주변 사무실이 조금 넓게 보인다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 사무실의 연속성은 좋지만 시선이 다트가 아닌 화면 왼쪽 밖을 향하고, 주먹도 하단 왼쪽에 놓여 지정된 시선·배치·하향 카메라 구도를 놓친다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴을 아래로 숙이고 눈동자도 오른쪽 아래의 주먹과 다트가 있는 지도 영역을 향한다. 다트의 붉은 깃은 위에 있고 금속 촉은 아래의 지도에 닿아 있어 목표와 사용 방향이 맞는다.",
        "built_space": "전경에 나무 책상 하나, 그 위에 펼친 지도 한 장이 보인다. 책상 가장자리는 팔 아래로 비스듬히 이어진다. 배경에는 왼쪽 창 구역과 깃발 하나, 중앙 수납 선반, 오른쪽 벽걸이 지도 하나가 있어 참고 장소의 주요 요소를 유지한다. 오른쪽 책상에는 서류철 더미와 필기구 통 하나가 보인다. 참고의 탁상등은 이 구도에서 확인되지 않으며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 얼굴은 좌상단, 주먹은 하단 중앙 부근이고 지도는 화면의 약 4분의 1을 차지한다. 위에서 내려다보는 각도는 맞지만 상체 노출은 다소 넓다.",
        "entities": "중년 한국인 남성으로 묘사된 인물 한 명이며 얼굴과 체격은 참고의 박철진과 대체로 일치한다. 어두운 남색 전투복, 낡은 챙모자, 흰 문양이 있는 붉은 완장이 유지된다. 머리카락은 모자로 대부분 가려져 있다. 한반도 인쇄 지도와 붉은 깃의 다트 한 개가 식별된다. 주먹과 팔은 같은 인물의 피부·체격·소매로 자연스럽게 연결된다. 찌푸린 미간과 드러낸 이가 분노를 표현하며 눈의 해부학은 정상이다. 창밖은 낮이다.",
        "hard_violations": [],
        "physics": "오른손이 다트 몸통을 감싸 쥐고 있으며 아래로 나온 촉이 지도에 꽂히는 접점이 보인다. 지도는 책상에 평평하게 놓여 있다. 주먹과 전완은 책상에 닿고 반대 손바닥도 책상을 짚어 앞으로 숙인 상체를 지탱한다. 다트와 몸 어느 쪽에도 지지 없는 부유는 없으며, 강하게 내리꽂은 직후의 자세로 성립한다."
       },
       {
        "label": "B",
        "direction": "눈동자는 화면 왼쪽의 프레임 밖을 향하며, 바로 아래 주먹 옆에 꽂힌 다트를 보지 않는다. 다트 자체는 깃이 위, 촉이 아래로 향해 지도에 꽂혀 있으므로 소품 방향은 맞지만 명시된 주의 대상은 어긋난다.",
        "built_space": "전경에 나무 책상 하나와 펼친 지도 한 장이 있다. 배경에는 왼쪽 창 구역, 깃발 하나, 켜진 탁상등 하나, 중앙 수납 선반과 오른쪽 벽걸이 지도 하나가 보여 참고 사무실과 잘 이어진다. 인물은 책상 뒤에서 몸을 숙이고 양손을 책상에 둔다. 중복 설비나 불가능한 반사는 보이지 않는다. 지도 면적은 3분의 1 미만이지만, 주먹이 하단 중앙이 아닌 왼쪽에 있으며 얼굴과 어깨를 보는 각도도 요구한 약간 높은 하향 시점보다 정면 눈높이에 가깝다.",
        "entities": "인물은 한 명이며 참고의 박철진과 닮은 중년 한국인 남성의 얼굴, 어두운 전투복과 챙모자, 붉은 완장의 흰 문양을 유지한다. 짧은 검은 머리는 귀 주변에서 일부만 보인다. 책상 위 한반도 지도와 붉은 깃의 다트 한 개가 확인된다. 주먹의 피부와 소매는 해당 인물과 일치한다. 벌린 입과 찌푸린 얼굴은 분노를 나타내고 눈은 정상적인 인간의 눈이다. 창밖은 낮이지만 켜진 탁상등이 비교적 눈에 띈다.",
        "hard_violations": [],
        "physics": "오른손이 다트의 축을 쥐고 있으며 손 아래 금속 촉이 지도에 닿아 박혀 있다. 주먹은 지도 위에 놓이고 반대 손바닥은 책상 오른쪽을 짚어 상체를 지지한다. 지도와 주변 물품은 책상 위에 놓여 있다. 공중에 지지 없이 떠 있는 신체나 소품은 없으며, 다트를 꽂은 뒤 힘을 유지하는 자세 자체는 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "다트를 내려다보는 시선과 좌상단 얼굴·하단 중앙 주먹의 연결, 지도에 막 꽂은 동작이 정확하며, 다만 요구한 클로즈업보다 상체와 주변 사무실이 조금 넓게 보인다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 사무실의 연속성은 좋지만 시선이 다트가 아닌 화면 왼쪽 밖을 향하고, 주먹도 하단 왼쪽에 놓여 지정된 시선·배치·하향 카메라 구도를 놓친다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴을 아래로 숙이고 눈동자도 오른쪽 아래의 주먹과 다트가 있는 지도 영역을 향한다. 다트의 붉은 깃은 위에 있고 금속 촉은 아래의 지도에 닿아 있어 목표와 사용 방향이 맞는다.",
        "built_space": "전경에 나무 책상 하나, 그 위에 펼친 지도 한 장이 보인다. 책상 가장자리는 팔 아래로 비스듬히 이어진다. 배경에는 왼쪽 창 구역과 깃발 하나, 중앙 수납 선반, 오른쪽 벽걸이 지도 하나가 있어 참고 장소의 주요 요소를 유지한다. 오른쪽 책상에는 서류철 더미와 필기구 통 하나가 보인다. 참고의 탁상등은 이 구도에서 확인되지 않으며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 얼굴은 좌상단, 주먹은 하단 중앙 부근이고 지도는 화면의 약 4분의 1을 차지한다. 위에서 내려다보는 각도는 맞지만 상체 노출은 다소 넓다.",
        "entities": "중년 한국인 남성으로 묘사된 인물 한 명이며 얼굴과 체격은 참고의 박철진과 대체로 일치한다. 어두운 남색 전투복, 낡은 챙모자, 흰 문양이 있는 붉은 완장이 유지된다. 머리카락은 모자로 대부분 가려져 있다. 한반도 인쇄 지도와 붉은 깃의 다트 한 개가 식별된다. 주먹과 팔은 같은 인물의 피부·체격·소매로 자연스럽게 연결된다. 찌푸린 미간과 드러낸 이가 분노를 표현하며 눈의 해부학은 정상이다. 창밖은 낮이다.",
        "hard_violations": [],
        "physics": "오른손이 다트 몸통을 감싸 쥐고 있으며 아래로 나온 촉이 지도에 꽂히는 접점이 보인다. 지도는 책상에 평평하게 놓여 있다. 주먹과 전완은 책상에 닿고 반대 손바닥도 책상을 짚어 앞으로 숙인 상체를 지탱한다. 다트와 몸 어느 쪽에도 지지 없는 부유는 없으며, 강하게 내리꽂은 직후의 자세로 성립한다."
       },
       {
        "label": "A",
        "direction": "눈동자는 화면 왼쪽의 프레임 밖을 향하며, 바로 아래 주먹 옆에 꽂힌 다트를 보지 않는다. 다트 자체는 깃이 위, 촉이 아래로 향해 지도에 꽂혀 있으므로 소품 방향은 맞지만 명시된 주의 대상은 어긋난다.",
        "built_space": "전경에 나무 책상 하나와 펼친 지도 한 장이 있다. 배경에는 왼쪽 창 구역, 깃발 하나, 켜진 탁상등 하나, 중앙 수납 선반과 오른쪽 벽걸이 지도 하나가 보여 참고 사무실과 잘 이어진다. 인물은 책상 뒤에서 몸을 숙이고 양손을 책상에 둔다. 중복 설비나 불가능한 반사는 보이지 않는다. 지도 면적은 3분의 1 미만이지만, 주먹이 하단 중앙이 아닌 왼쪽에 있으며 얼굴과 어깨를 보는 각도도 요구한 약간 높은 하향 시점보다 정면 눈높이에 가깝다.",
        "entities": "인물은 한 명이며 참고의 박철진과 닮은 중년 한국인 남성의 얼굴, 어두운 전투복과 챙모자, 붉은 완장의 흰 문양을 유지한다. 짧은 검은 머리는 귀 주변에서 일부만 보인다. 책상 위 한반도 지도와 붉은 깃의 다트 한 개가 확인된다. 주먹의 피부와 소매는 해당 인물과 일치한다. 벌린 입과 찌푸린 얼굴은 분노를 나타내고 눈은 정상적인 인간의 눈이다. 창밖은 낮이지만 켜진 탁상등이 비교적 눈에 띈다.",
        "hard_violations": [],
        "physics": "오른손이 다트의 축을 쥐고 있으며 손 아래 금속 촉이 지도에 닿아 박혀 있다. 주먹은 지도 위에 놓이고 반대 손바닥은 책상 오른쪽을 짚어 상체를 지지한다. 지도와 주변 물품은 책상 위에 놓여 있다. 공중에 지지 없이 떠 있는 신체나 소품은 없으며, 다트를 꽂은 뒤 힘을 유지하는 자세 자체는 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.196,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.196,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1196
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트에 명시된 주간 시간대, 상단 좌측의 얼굴과 하단 중앙의 주먹을 배치하는 구도, 그리고 다트 핀을 향한 시선 처리를 모두 정확하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1196,
    "verdict_ko": "시간대가 야간으로 잘못 묘사되었고 프레이밍 지시(얼굴과 손의 위치)를 따르지 않았으며, 시선이 다트를 향하지 않고 앞을 향하고 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh14_sel.png",
    "asset_id": "ffb71e32-1c56-444e-a940-7dc06aa1db84",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ba4-447b-70e5-8ac1-d5d02db9854d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh14"
  }
 },
 "S53sh6::signage": {
  "fp": "4cd43ca38d225b70",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh6": {
  "input_fingerprint": "9251ff58eb72e21d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 특공대와 유빅 쪽에 감시를 붙이라며 허공을 향해 매섭게 삿대질하듯 손가락을 뻗고 있는 박철진의 역동적인 상반신.\n\nLOCATION (lock): In the desk area of the militia commander's city office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established approach from beside the off-screen subordinate, at 박철진's upper-chest height with a slight upward tilt and an oblique view of his upper body. Place 박철진 left of center, his extended finger cutting into the open right side without pointing into the lens, while his eyes stay on the subordinate just outside the frame. Let the closing camera distance carry the emphasis, preserving the conversation axis and enough space around the hand to retain the entire gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Remains between 박철진 and the established reporting position) — A narrow oblique portion of the desktop crosses the lower edge; used as Retains spatial continuity without competing with the pointing hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's restrained ambient illumination and controlled contrast so the command gains force through gesture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The map of the Korean Peninsula remains spread across the office desk. 박철진: He still has the darts used while considering directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 특공대와 유빅 쪽에 감시를 붙이라며 허공을 향해 매섭게 삿대질하듯 손가락을 뻗고 있는 박철진의 역동적인 상반신.\n\nLOCATION (lock): In the desk area of the militia commander's city office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established approach from beside the off-screen subordinate, at 박철진's upper-chest height with a slight upward tilt and an oblique view of his upper body. Place 박철진 left of center, his extended finger cutting into the open right side without pointing into the lens, while his eyes stay on the subordinate just outside the frame. Let the closing camera distance carry the emphasis, preserving the conversation axis and enough space around the hand to retain the entire gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Remains between 박철진 and the established reporting position) — A narrow oblique portion of the desktop crosses the lower edge; used as Retains spatial continuity without competing with the pointing hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's restrained ambient illumination and controlled contrast so the command gains force through gesture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The map of the Korean Peninsula remains spread across the office desk. 박철진: He still has the darts used while considering directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 특공대와 유빅 쪽에 감시를 붙이라며 허공을 향해 매섭게 삿대질하듯 손가락을 뻗고 있는 박철진의 역동적인 상반신.\n\nLOCATION (lock): In the desk area of the militia commander's city office, facing his subordinate under office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established approach from beside the off-screen subordinate, at 박철진's upper-chest height with a slight upward tilt and an oblique view of his upper body. Place 박철진 left of center, his extended finger cutting into the open right side without pointing into the lens, while his eyes stay on the subordinate just outside the frame. Let the closing camera distance carry the emphasis, preserving the conversation axis and enough space around the hand to retain the entire gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Remains between 박철진 and the established reporting position) — A narrow oblique portion of the desktop crosses the lower edge; used as Retains spatial continuity without competing with the pointing hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's restrained ambient illumination and controlled contrast so the command gains force through gesture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The map of the Korean Peninsula remains spread across the office desk. 박철진: He still has the darts used while considering directions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선과 뻗은 왼팔의 손가락이 화면 우측 바깥에 있는 대상을 향해 정확히 조준됨.",
    "built_space": "사무실 내부. 하단에 책상과 지도가 비스듬히 걸쳐 있고, 뒤편의 창문, 깃발, 수납장이 이전 샷의 배치와 일치함.",
    "entities": "박철진. 남색 전투복, 붉은 완장, 모자의 형태가 참조 이미지와 일치하며 이전 샷처럼 오른손에 붉은 다트를 쥐고 있음.",
    "hard_violations": [],
    "physics": "오른손으로 책상을 짚어 상체를 지탱하고 있으며, 삿대질을 위해 왼팔을 힘차게 뻗은 근육과 자세가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선과 뻗은 왼팔이 화면 우측 바깥을 향함.",
    "built_space": "사무실 배경이나 화면 우측 하단 전경에 검은 실루엣이 공간을 침범함.",
    "entities": "박철진의 복장과 외양은 일치하나, 이전 샷과 달리 뻗은 왼손에 다트를 쥐고 있음. 샷 텍스트에 없는 인물의 어깨 일부가 우측 하단에 노출됨.",
    "hard_violations": [
     "[gemini-pro] 지시문에 명시되지 않은 인물의 신체 일부(우측 하단 어깨 실루엣) 추가",
     "[gpt-high] 오른쪽 아래 전경에 부하의 어깨·상체 일부로 읽히는 형체가 등장한다. 박철진 외 인물의 신체를 넣지 말고 부하는 화면 밖에 두라는 명시적 조건을 위반한다."
    ],
    "physics": "왼팔을 뻗고 있으나 상체를 지탱하던 오른팔이 프레임에서 사라져 자세의 안정감과 이전 샷과의 물리적 연속성이 결여됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레임 밖 부하를 향한 시선과 역동적인 삿대질을 정확한 미디엄 샷으로 구현했으며, 이전 샷의 다트를 쥔 오른손 위치와 공간의 연속성을 완벽히 유지했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면 밖에 있어야 할 부하의 어깨가 프레임에 포함되어 지시를 위반했으며, 다트를 쥔 손이 왼손으로 바뀌어 이전 샷과의 연속성이 깨졌습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 뻗은 왼팔의 손가락이 화면 우측 바깥에 있는 대상을 향해 정확히 조준됨.",
        "built_space": "사무실 내부. 하단에 책상과 지도가 비스듬히 걸쳐 있고, 뒤편의 창문, 깃발, 수납장이 이전 샷의 배치와 일치함.",
        "entities": "박철진. 남색 전투복, 붉은 완장, 모자의 형태가 참조 이미지와 일치하며 이전 샷처럼 오른손에 붉은 다트를 쥐고 있음.",
        "hard_violations": [],
        "physics": "오른손으로 책상을 짚어 상체를 지탱하고 있으며, 삿대질을 위해 왼팔을 힘차게 뻗은 근육과 자세가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 뻗은 왼팔이 화면 우측 바깥을 향함.",
        "built_space": "사무실 배경이나 화면 우측 하단 전경에 검은 실루엣이 공간을 침범함.",
        "entities": "박철진의 복장과 외양은 일치하나, 이전 샷과 달리 뻗은 왼손에 다트를 쥐고 있음. 샷 텍스트에 없는 인물의 어깨 일부가 우측 하단에 노출됨.",
        "hard_violations": [
         "지시문에 명시되지 않은 인물의 신체 일부(우측 하단 어깨 실루엣) 추가"
        ],
        "physics": "왼팔을 뻗고 있으나 상체를 지탱하던 오른팔이 프레임에서 사라져 자세의 안정감과 이전 샷과의 물리적 연속성이 결여됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프레임 밖 부하를 향한 시선과 역동적인 삿대질을 정확한 미디엄 샷으로 구현했으며, 이전 샷의 다트를 쥔 오른손 위치와 공간의 연속성을 완벽히 유지했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "화면 밖에 있어야 할 부하의 어깨가 프레임에 포함되어 지시를 위반했으며, 다트를 쥔 손이 왼손으로 바뀌어 이전 샷과의 연속성이 깨졌습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 뻗은 왼팔의 손가락이 화면 우측 바깥에 있는 대상을 향해 정확히 조준됨.",
        "built_space": "사무실 내부. 하단에 책상과 지도가 비스듬히 걸쳐 있고, 뒤편의 창문, 깃발, 수납장이 이전 샷의 배치와 일치함.",
        "entities": "박철진. 남색 전투복, 붉은 완장, 모자의 형태가 참조 이미지와 일치하며 이전 샷처럼 오른손에 붉은 다트를 쥐고 있음.",
        "hard_violations": [],
        "physics": "오른손으로 책상을 짚어 상체를 지탱하고 있으며, 삿대질을 위해 왼팔을 힘차게 뻗은 근육과 자세가 물리적으로 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 뻗은 왼팔이 화면 우측 바깥을 향함.",
        "built_space": "사무실 배경이나 화면 우측 하단 전경에 검은 실루엣이 공간을 침범함.",
        "entities": "박철진의 복장과 외양은 일치하나, 이전 샷과 달리 뻗은 왼손에 다트를 쥐고 있음. 샷 텍스트에 없는 인물의 어깨 일부가 우측 하단에 노출됨.",
        "hard_violations": [
         "지시문에 명시되지 않은 인물의 신체 일부(우측 하단 어깨 실루엣) 추가"
        ],
        "physics": "왼팔을 뻗고 있으나 상체를 지탱하던 오른팔이 프레임에서 사라져 자세의 안정감과 이전 샷과의 물리적 연속성이 결여됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "삿대질 방향과 인물의 외형은 맞지만, 오른쪽 전경에 부하의 어깨로 읽히는 신체 일부를 추가해 박철진만 보여야 한다는 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박철진만 등장하며 화면 오른쪽을 향한 삿대질·화면 밖 상대를 보는 시선·다트 소지를 충족하지만, 책상 노출이 넓고 요구한 약한 올려다보기보다 이전 숏의 내려다보는 구도에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼팔의 검지는 렌즈가 아니라 화면 오른쪽 바깥의 허공을 향한다. 눈도 오른쪽 화면 밖 상대를 향해 있어 명령하는 대화 방향은 맞는다. 접힌 손가락 안에는 붉은 다트 깃이 보이지만 다트 촉의 방향은 가려져 확인할 수 없다.",
        "built_space": "왼쪽 창문 한 곳과 깃발 한 개, 뒤쪽 선반 한 개와 낮은 수납장 열, 벽걸이 한반도 지도 한 개가 보인다. 앞에는 지도 한 장이 펼쳐진 책상 한 개가 있고 오른쪽에 서류철 더미와 필기구 통 한 개가 있다. 재질과 배치는 이전 사무실과 대체로 이어지며 책상은 인물과 카메라 사이에 놓인다. 다만 오른쪽 아래에 흐릿한 어깨 형태가 들어와 부하가 완전히 화면 밖이어야 하는 배치와 어긋난다. 반사면에 의한 공간 오류는 보이지 않는다.",
        "entities": "박철진은 참조와 유사한 중년 한국인 남성 외형이며 얼굴, 짙은 남색 전투복, 낡은 챙모자, 흰 표식이 있는 붉은 완장이 유지된다. 머리카락은 모자에 가려 짧은 검은 머리를 직접 확인하기 어렵다. 한반도 책상 지도와 붉은 깃의 다트가 보인다. 이전 숏의 오른손에 있던 다트는 삿대질하는 왼손으로 옮겨져 있다. 오른쪽 전경의 남색 어깨 형태는 허용되지 않은 추가 인물의 일부로 읽힌다.",
        "hard_violations": [
         "오른쪽 아래 전경에 부하의 어깨·상체 일부로 읽히는 형체가 등장한다. 박철진 외 인물의 신체를 넣지 말고 부하는 화면 밖에 두라는 명시적 조건을 위반한다."
        ],
        "physics": "앞으로 기울인 몸통에서 왼팔을 뻗고 있으며 어깨·팔꿈치·손목의 연결은 자연스럽다. 다트는 접힌 손가락으로 잡혀 있어 떠 있지 않다. 오른팔은 책상 가장자리 아래로 이어지지만 손의 접촉점과 발은 프레임 밖이라 확인할 수 없다. 이것만으로 몸이 공중에 떠 있다고 볼 근거는 없다. 지도와 서류류는 책상에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "왼손 검지가 화면 오른쪽의 빈 공간을 거의 수평으로 가리키며 렌즈를 겨누지 않는다. 눈은 오른쪽 화면 밖 부하가 있을 방향을 향한다. 오른손에 쥔 다트는 촉이 아래로 향해 책상 지도에 닿아 있어 이전 행동의 방향도 이어진다.",
        "built_space": "왼쪽 창문 한 곳과 깃발 한 개, 뒤쪽 선반 한 개와 낮은 수납장 열, 벽걸이 지도 한 개, 전경 책상 한 개가 보인다. 책상 오른쪽에는 서류철 더미와 필기구 통 한 개가 있으며 참조 사무실의 배치와 마모된 재질을 유지한다. 박철진은 책상 뒤에 있고 보고 위치는 책상 반대편 화면 밖에 남는다. 다만 지도와 책상 윗면이 하단에서 상당한 면적을 차지해 '좁고 비스듬한 책상 일부'라는 요구보다 넓다. 불가능한 반사나 명백한 중복 설비는 없다.",
        "entities": "등장 인물은 박철진 한 명뿐이다. 참조와 유사한 중년 한국인 남성의 얼굴과 체격, 남색 전투복, 낡은 남색 챙모자, 흰 표식의 붉은 완장을 유지한다. 머리카락은 모자 아래 가려져 있다. 책상에는 한반도 지도가 펼쳐져 있고, 오른손에는 이전 숏과 같은 붉은 깃의 다트를 쥐고 있다. 추가 인물이나 사진 위에 덧씌운 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "오른쪽 전완과 주먹이 책상 위에 놓여 앞으로 숙인 상체를 지지한다. 오른손은 다트 몸통을 감싸고 촉은 지도 표면에 닿아 있다. 뻗은 왼팔과 검지는 어깨에서 자연스럽게 이어지며 실제 삿대질로 가능한 자세다. 하체는 프레임 밖이지만 보이는 상체에는 분명한 책상 접촉점이 있다. 지도와 서류류도 책상 표면에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "삿대질 방향과 인물의 외형은 맞지만, 오른쪽 전경에 부하의 어깨로 읽히는 신체 일부를 추가해 박철진만 보여야 한다는 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박철진만 등장하며 화면 오른쪽을 향한 삿대질·화면 밖 상대를 보는 시선·다트 소지를 충족하지만, 책상 노출이 넓고 요구한 약한 올려다보기보다 이전 숏의 내려다보는 구도에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼팔의 검지는 렌즈가 아니라 화면 오른쪽 바깥의 허공을 향한다. 눈도 오른쪽 화면 밖 상대를 향해 있어 명령하는 대화 방향은 맞는다. 접힌 손가락 안에는 붉은 다트 깃이 보이지만 다트 촉의 방향은 가려져 확인할 수 없다.",
        "built_space": "왼쪽 창문 한 곳과 깃발 한 개, 뒤쪽 선반 한 개와 낮은 수납장 열, 벽걸이 한반도 지도 한 개가 보인다. 앞에는 지도 한 장이 펼쳐진 책상 한 개가 있고 오른쪽에 서류철 더미와 필기구 통 한 개가 있다. 재질과 배치는 이전 사무실과 대체로 이어지며 책상은 인물과 카메라 사이에 놓인다. 다만 오른쪽 아래에 흐릿한 어깨 형태가 들어와 부하가 완전히 화면 밖이어야 하는 배치와 어긋난다. 반사면에 의한 공간 오류는 보이지 않는다.",
        "entities": "박철진은 참조와 유사한 중년 한국인 남성 외형이며 얼굴, 짙은 남색 전투복, 낡은 챙모자, 흰 표식이 있는 붉은 완장이 유지된다. 머리카락은 모자에 가려 짧은 검은 머리를 직접 확인하기 어렵다. 한반도 책상 지도와 붉은 깃의 다트가 보인다. 이전 숏의 오른손에 있던 다트는 삿대질하는 왼손으로 옮겨져 있다. 오른쪽 전경의 남색 어깨 형태는 허용되지 않은 추가 인물의 일부로 읽힌다.",
        "hard_violations": [
         "오른쪽 아래 전경에 부하의 어깨·상체 일부로 읽히는 형체가 등장한다. 박철진 외 인물의 신체를 넣지 말고 부하는 화면 밖에 두라는 명시적 조건을 위반한다."
        ],
        "physics": "앞으로 기울인 몸통에서 왼팔을 뻗고 있으며 어깨·팔꿈치·손목의 연결은 자연스럽다. 다트는 접힌 손가락으로 잡혀 있어 떠 있지 않다. 오른팔은 책상 가장자리 아래로 이어지지만 손의 접촉점과 발은 프레임 밖이라 확인할 수 없다. 이것만으로 몸이 공중에 떠 있다고 볼 근거는 없다. 지도와 서류류는 책상에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "왼손 검지가 화면 오른쪽의 빈 공간을 거의 수평으로 가리키며 렌즈를 겨누지 않는다. 눈은 오른쪽 화면 밖 부하가 있을 방향을 향한다. 오른손에 쥔 다트는 촉이 아래로 향해 책상 지도에 닿아 있어 이전 행동의 방향도 이어진다.",
        "built_space": "왼쪽 창문 한 곳과 깃발 한 개, 뒤쪽 선반 한 개와 낮은 수납장 열, 벽걸이 지도 한 개, 전경 책상 한 개가 보인다. 책상 오른쪽에는 서류철 더미와 필기구 통 한 개가 있으며 참조 사무실의 배치와 마모된 재질을 유지한다. 박철진은 책상 뒤에 있고 보고 위치는 책상 반대편 화면 밖에 남는다. 다만 지도와 책상 윗면이 하단에서 상당한 면적을 차지해 '좁고 비스듬한 책상 일부'라는 요구보다 넓다. 불가능한 반사나 명백한 중복 설비는 없다.",
        "entities": "등장 인물은 박철진 한 명뿐이다. 참조와 유사한 중년 한국인 남성의 얼굴과 체격, 남색 전투복, 낡은 남색 챙모자, 흰 표식의 붉은 완장을 유지한다. 머리카락은 모자 아래 가려져 있다. 책상에는 한반도 지도가 펼쳐져 있고, 오른손에는 이전 숏과 같은 붉은 깃의 다트를 쥐고 있다. 추가 인물이나 사진 위에 덧씌운 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "오른쪽 전완과 주먹이 책상 위에 놓여 앞으로 숙인 상체를 지지한다. 오른손은 다트 몸통을 감싸고 촉은 지도 표면에 닿아 있다. 뻗은 왼팔과 검지는 어깨에서 자연스럽게 이어지며 실제 삿대질로 가능한 자세다. 하체는 프레임 밖이지만 보이는 상체에는 분명한 책상 접촉점이 있다. 지도와 서류류도 책상 표면에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시문에 명시되지 않은 인물의 신체 일부(우측 하단 어깨 실루엣) 추가",
     "[gpt-high] 오른쪽 아래 전경에 부하의 어깨·상체 일부로 읽히는 형체가 등장한다. 박철진 외 인물의 신체를 넣지 말고 부하는 화면 밖에 두라는 명시적 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프레임 밖 부하를 향한 시선과 역동적인 삿대질을 정확한 미디엄 샷으로 구현했으며, 이전 샷의 다트를 쥔 오른손 위치와 공간의 연속성을 완벽히 유지했습니다."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "화면 밖에 있어야 할 부하의 어깨가 프레임에 포함되어 지시를 위반했으며, 다트를 쥔 손이 왼손으로 바뀌어 이전 샷과의 연속성이 깨졌습니다.  ★위반: [gemini-pro] 지시문에 명시되지 않은 인물의 신체 일부(우측 하단 어깨 실루엣) 추가 / [gpt-high] 오른쪽 아래 전경에 부하의 어깨·상체 일부로 읽히는 형체가 등장한다. 박철진 외 인물의 신체를 넣지 말고 부하는 화면 밖에 두라는 명시적 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S53sh3_sel.png",
    "asset_id": "7ab52611-548b-49eb-9f5c-9020b1af0c5e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ba9-4b78-7327-95e6-0ad512aa772e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh3"
  }
 },
 "S54sh4::signage": {
  "fp": "cdbc39b767c5d366",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh4": {
  "input_fingerprint": "12f0c1115d0d1a4e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 별도의 병력을 지원해달라는 듯 앞으로 손을 내민 채 멈춘 윤성찬의 상반신.\n\nLOCATION (lock): In the visitor conversation area inside the defense minister's private office, illuminated for daytime use. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside and behind 국방장관's shoulder position, keep the minister beyond the left crop and approach 윤성찬 at a three-quarter angle with a slight downward pitch from above his eye line. 윤성찬 occupies the right-center, leaning toward the minister with his offered hand suspended in the lower left and his gaze fixed on the minister off-screen. Slow the forward movement as the hand stops, keeping hand and upper body together rather than isolating the request as an insert.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination and restrained tonal separation support the transactional intimacy without specifying an unestablished fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 별도의 병력을 지원해달라는 듯 앞으로 손을 내민 채 멈춘 윤성찬의 상반신.\n\nLOCATION (lock): In the visitor conversation area inside the defense minister's private office, illuminated for daytime use. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside and behind 국방장관's shoulder position, keep the minister beyond the left crop and approach 윤성찬 at a three-quarter angle with a slight downward pitch from above his eye line. 윤성찬 occupies the right-center, leaning toward the minister with his offered hand suspended in the lower left and his gaze fixed on the minister off-screen. Slow the forward movement as the hand stops, keeping hand and upper body together rather than isolating the request as an insert.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination and restrained tonal separation support the transactional intimacy without specifying an unestablished fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신에게 별도의 병력을 지원해달라는 듯 앞으로 손을 내민 채 멈춘 윤성찬의 상반신.\n\nLOCATION (lock): In the visitor conversation area inside the defense minister's private office, illuminated for daytime use. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside and behind 국방장관's shoulder position, keep the minister beyond the left crop and approach 윤성찬 at a three-quarter angle with a slight downward pitch from above his eye line. 윤성찬 occupies the right-center, leaning toward the minister with his offered hand suspended in the lower left and his gaze fixed on the minister off-screen. Slow the forward movement as the hand stops, keeping hand and upper body together rather than isolating the request as an insert.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime ambient illumination and restrained tonal separation support the transactional intimacy without specifying an unestablished fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬은 화면 왼쪽 프레임 밖의 장관을 향해 시선을 고정하고 있으며, 오른손을 앞으로 내밀고 있음.",
    "built_space": "인물 앞쪽에 책상이 있고 뒤편으로는 이전 샷에서 장관의 뒤에 있던 창문, 태극기, 그림, 국방부 깃발이 배치되어 있음. 윤성찬과 장관의 위치가 완전히 뒤바뀐 상태임.",
    "entities": "윤성찬은 레퍼런스와 일치하는 노인 남성이며 지시된 차콜 그레이 정장과 넥타이를 착용함.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에서 확립된 공간 구조를 위반하여 윤성찬을 방문객 자리가 아닌 장관의 자리에 배치함",
     "[gemini-pro] 국방장관을 왼쪽 크롭 밖으로 두어 화면에서 제외하라는 지시를 어기고 어깨와 머리 일부를 프레임 안에 포함함",
     "[gpt-high] 장관을 왼쪽 크롭 밖에 두라는 명시적 지시를 어기고, 왼쪽 전경에 다른 남자의 머리와 정장 상체를 노출했다."
    ],
    "physics": "앞으로 내민 오른손은 허공에 떠 있으며 몸통과 팔에 의해 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "윤성찬이 화면 왼쪽 오프스크린 방향으로 시선을 향하고 있으며 오른손을 내밀고 있음.",
    "built_space": "앞쪽에 책상이 놓여 있고, 윤성찬의 뒤쪽으로 이전 샷의 장관 측 배경(창문, 태극기, 액자, 부대기)이 그대로 배치되어 공간적 위치가 심각하게 어긋남.",
    "entities": "윤성찬은 주어진 레퍼런스와 일치하는 인상착의(머리, 얼굴, 정장)를 갖추고 있음.",
    "hard_violations": [
     "[gemini-pro] 공간의 물리적 일관성을 무시하고 윤성찬이 장관의 위치와 배경을 차지하도록 잘못 배치함",
     "[gemini-pro] 장관을 프레임 왼쪽 밖으로 제외하라는 지시와 달리 어깨가 명확히 화면에 포함됨",
     "[gpt-high] 장관을 왼쪽 크롭 밖에 두고 윤성찬만 보여야 하는데, 왼쪽 전경에 다른 남자의 머리와 상체를 크게 넣었다."
    ],
    "physics": "내민 오른손이 허공에 떠 있으며 팔뚝과 어깨를 통해 정상적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "카메라의 약간 하향된 피치(downward pitch)는 어느 정도 구현했으나, 윤성찬이 장관의 자리에 앉아 이전 샷의 배경이 뒤바뀌는 치명적인 공간 오류를 범했고 장관을 크롭 바깥으로 빼내지 못했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이전 샷의 공간적 논리를 무시하고 윤성찬을 장관의 배경 쪽에 배치하는 하드 위반을 범했으며, 하향 카메라 앵글 지시도 따르지 않았고 장관의 어깨가 화면에 노출되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 화면 왼쪽 프레임 밖의 장관을 향해 시선을 고정하고 있으며, 오른손을 앞으로 내밀고 있음.",
        "built_space": "인물 앞쪽에 책상이 있고 뒤편으로는 이전 샷에서 장관의 뒤에 있던 창문, 태극기, 그림, 국방부 깃발이 배치되어 있음. 윤성찬과 장관의 위치가 완전히 뒤바뀐 상태임.",
        "entities": "윤성찬은 레퍼런스와 일치하는 노인 남성이며 지시된 차콜 그레이 정장과 넥타이를 착용함.",
        "hard_violations": [
         "이전 샷에서 확립된 공간 구조를 위반하여 윤성찬을 방문객 자리가 아닌 장관의 자리에 배치함",
         "국방장관을 왼쪽 크롭 밖으로 두어 화면에서 제외하라는 지시를 어기고 어깨와 머리 일부를 프레임 안에 포함함"
        ],
        "physics": "앞으로 내민 오른손은 허공에 떠 있으며 몸통과 팔에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "윤성찬이 화면 왼쪽 오프스크린 방향으로 시선을 향하고 있으며 오른손을 내밀고 있음.",
        "built_space": "앞쪽에 책상이 놓여 있고, 윤성찬의 뒤쪽으로 이전 샷의 장관 측 배경(창문, 태극기, 액자, 부대기)이 그대로 배치되어 공간적 위치가 심각하게 어긋남.",
        "entities": "윤성찬은 주어진 레퍼런스와 일치하는 인상착의(머리, 얼굴, 정장)를 갖추고 있음.",
        "hard_violations": [
         "공간의 물리적 일관성을 무시하고 윤성찬이 장관의 위치와 배경을 차지하도록 잘못 배치함",
         "장관을 프레임 왼쪽 밖으로 제외하라는 지시와 달리 어깨가 명확히 화면에 포함됨"
        ],
        "physics": "내민 오른손이 허공에 떠 있으며 팔뚝과 어깨를 통해 정상적으로 지지됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "카메라의 약간 하향된 피치(downward pitch)는 어느 정도 구현했으나, 윤성찬이 장관의 자리에 앉아 이전 샷의 배경이 뒤바뀌는 치명적인 공간 오류를 범했고 장관을 크롭 바깥으로 빼내지 못했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이전 샷의 공간적 논리를 무시하고 윤성찬을 장관의 배경 쪽에 배치하는 하드 위반을 범했으며, 하향 카메라 앵글 지시도 따르지 않았고 장관의 어깨가 화면에 노출되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 화면 왼쪽 프레임 밖의 장관을 향해 시선을 고정하고 있으며, 오른손을 앞으로 내밀고 있음.",
        "built_space": "인물 앞쪽에 책상이 있고 뒤편으로는 이전 샷에서 장관의 뒤에 있던 창문, 태극기, 그림, 국방부 깃발이 배치되어 있음. 윤성찬과 장관의 위치가 완전히 뒤바뀐 상태임.",
        "entities": "윤성찬은 레퍼런스와 일치하는 노인 남성이며 지시된 차콜 그레이 정장과 넥타이를 착용함.",
        "hard_violations": [
         "이전 샷에서 확립된 공간 구조를 위반하여 윤성찬을 방문객 자리가 아닌 장관의 자리에 배치함",
         "국방장관을 왼쪽 크롭 밖으로 두어 화면에서 제외하라는 지시를 어기고 어깨와 머리 일부를 프레임 안에 포함함"
        ],
        "physics": "앞으로 내민 오른손은 허공에 떠 있으며 몸통과 팔에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "윤성찬이 화면 왼쪽 오프스크린 방향으로 시선을 향하고 있으며 오른손을 내밀고 있음.",
        "built_space": "앞쪽에 책상이 놓여 있고, 윤성찬의 뒤쪽으로 이전 샷의 장관 측 배경(창문, 태극기, 액자, 부대기)이 그대로 배치되어 공간적 위치가 심각하게 어긋남.",
        "entities": "윤성찬은 주어진 레퍼런스와 일치하는 인상착의(머리, 얼굴, 정장)를 갖추고 있음.",
        "hard_violations": [
         "공간의 물리적 일관성을 무시하고 윤성찬이 장관의 위치와 배경을 차지하도록 잘못 배치함",
         "장관을 프레임 왼쪽 밖으로 제외하라는 지시와 달리 어깨가 명확히 화면에 포함됨"
        ],
        "physics": "내민 오른손이 허공에 떠 있으며 팔뚝과 어깨를 통해 정상적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "손과 상반신을 함께 담았지만, 제외해야 할 장관의 머리와 몸을 크게 노출했고 윤성찬의 접근 각도도 지정된 사선보다 정면에 가깝다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장관을 화면에 넣은 실격 오류는 같지만, 오른쪽 중심의 사선 자세와 손바닥을 내민 요청 동작, 명확한 낮빛은 A보다 지시에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 왼쪽 전경 남자의 얼굴을 바라보며 오른손을 그쪽으로 내민다. 시선과 손의 대상은 일치하지만, 대상이 화면 밖 장관이어야 한다는 조건과 달리 남자가 크게 보인다. 몸통은 카메라에 비교적 정면이다.",
        "built_space": "왼쪽 창과 달력 하나, 태극기 하나, 오른쪽 기관기 하나, 뒤쪽 산수화 하나와 오른쪽 높은 목재 책장이 보인다. 주요 사무실 재료와 비품은 참조와 유사하다. 윤성찬이 앉은 전경 가죽 좌석 외에 뒤쪽 중앙 등받이와 왼쪽 안락의자가 보이며, 앞에는 목재 탁자와 검은 매트가 있다. 약간 내려다보는 중간 크기 구도이나 왼쪽을 장관의 머리와 어깨가 차지한다.",
        "entities": "윤성찬은 고령의 한국인 남성으로 읽히며 회백색 짧은 머리, 콧수염, 차콜 정장, 흰 셔츠와 짙은 넥타이가 참조에 대체로 부합한다. 머리는 참조보다 높게 쓸어 올린 형태다. 왼쪽에는 허용되지 않은 검은 머리 남자의 머리·귀·목·정장 상체가 추가되어 있다. 별도 자막이나 도식은 없다.",
        "hard_violations": [
         "장관을 왼쪽 크롭 밖에 두고 윤성찬만 보여야 하는데, 왼쪽 전경에 다른 남자의 머리와 상체를 크게 넣었다."
        ],
        "physics": "윤성찬의 하체는 탁자에 가려지지만 가죽 좌석에 앉은 자세로 자연스럽게 연결된다. 내민 손은 손목과 팔꿈치를 거쳐 어깨에 연결되어 팔의 힘으로 유지된다. 손을 내밀다 멈춘 자세가 가능하며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "윤성찬은 몸을 앞으로 기울여 왼쪽 전경 남자를 바라보고, 오른손 손바닥을 위로 열어 같은 상대에게 내민다. 지원을 요청하는 방향과 동작은 분명하다. 다만 시선의 도착점인 장관이 화면 밖이 아니라 화면 안에 있다.",
        "built_space": "왼쪽 창과 달력 하나, 태극기 하나, 오른쪽 가장자리 기관기 하나, 뒤쪽 산수화 하나가 보인다. 윤성찬의 가죽 의자 하나와 뒤쪽 왼편의 가죽 좌석 하나가 식별되며 앞에는 목재 탁자와 검은 매트가 놓여 있다. 참조의 높은 오른쪽 책장 대신 낮은 목재 수납장 쪽이 강조되어 공간 일치는 부분적이다. 창밖은 명확한 낮이다. 윤성찬은 오른쪽 중심에 사선으로 배치되지만 시점은 지시된 약한 하향각보다 눈높이에 가까워 보인다.",
        "entities": "윤성찬의 고령 남성 얼굴, 짧게 가른 회백색 머리, 콧수염과 마른 체격은 인물 참조에 가깝다. 깨끗한 차콜 맞춤 정장, 흰 셔츠, 짙은 넥타이도 부합한다. 왼쪽 전경의 검은 머리 남자는 이 장면에서 보이면 안 되는 추가 인물이다. 별도 자막이나 그래픽은 없다.",
        "hard_violations": [
         "장관을 왼쪽 크롭 밖에 두라는 명시적 지시를 어기고, 왼쪽 전경에 다른 남자의 머리와 정장 상체를 노출했다."
        ],
        "physics": "윤성찬은 등받이가 뒤에 있는 가죽 의자에 앉아 상체를 앞으로 기울인다. 좌석이 몸을 받치며, 내민 손은 자연스럽게 연결된 팔로 유지되고 팔꿈치는 탁자 가까이에 놓인다. 손바닥을 펼친 채 동작을 멈추는 자세가 물리적으로 가능하고, 무지지 부유나 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "손과 상반신을 함께 담았지만, 제외해야 할 장관의 머리와 몸을 크게 노출했고 윤성찬의 접근 각도도 지정된 사선보다 정면에 가깝다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장관을 화면에 넣은 실격 오류는 같지만, 오른쪽 중심의 사선 자세와 손바닥을 내민 요청 동작, 명확한 낮빛은 A보다 지시에 가깝다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬은 왼쪽 전경 남자의 얼굴을 바라보며 오른손을 그쪽으로 내민다. 시선과 손의 대상은 일치하지만, 대상이 화면 밖 장관이어야 한다는 조건과 달리 남자가 크게 보인다. 몸통은 카메라에 비교적 정면이다.",
        "built_space": "왼쪽 창과 달력 하나, 태극기 하나, 오른쪽 기관기 하나, 뒤쪽 산수화 하나와 오른쪽 높은 목재 책장이 보인다. 주요 사무실 재료와 비품은 참조와 유사하다. 윤성찬이 앉은 전경 가죽 좌석 외에 뒤쪽 중앙 등받이와 왼쪽 안락의자가 보이며, 앞에는 목재 탁자와 검은 매트가 있다. 약간 내려다보는 중간 크기 구도이나 왼쪽을 장관의 머리와 어깨가 차지한다.",
        "entities": "윤성찬은 고령의 한국인 남성으로 읽히며 회백색 짧은 머리, 콧수염, 차콜 정장, 흰 셔츠와 짙은 넥타이가 참조에 대체로 부합한다. 머리는 참조보다 높게 쓸어 올린 형태다. 왼쪽에는 허용되지 않은 검은 머리 남자의 머리·귀·목·정장 상체가 추가되어 있다. 별도 자막이나 도식은 없다.",
        "hard_violations": [
         "장관을 왼쪽 크롭 밖에 두고 윤성찬만 보여야 하는데, 왼쪽 전경에 다른 남자의 머리와 상체를 크게 넣었다."
        ],
        "physics": "윤성찬의 하체는 탁자에 가려지지만 가죽 좌석에 앉은 자세로 자연스럽게 연결된다. 내민 손은 손목과 팔꿈치를 거쳐 어깨에 연결되어 팔의 힘으로 유지된다. 손을 내밀다 멈춘 자세가 가능하며, 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "윤성찬은 몸을 앞으로 기울여 왼쪽 전경 남자를 바라보고, 오른손 손바닥을 위로 열어 같은 상대에게 내민다. 지원을 요청하는 방향과 동작은 분명하다. 다만 시선의 도착점인 장관이 화면 밖이 아니라 화면 안에 있다.",
        "built_space": "왼쪽 창과 달력 하나, 태극기 하나, 오른쪽 가장자리 기관기 하나, 뒤쪽 산수화 하나가 보인다. 윤성찬의 가죽 의자 하나와 뒤쪽 왼편의 가죽 좌석 하나가 식별되며 앞에는 목재 탁자와 검은 매트가 놓여 있다. 참조의 높은 오른쪽 책장 대신 낮은 목재 수납장 쪽이 강조되어 공간 일치는 부분적이다. 창밖은 명확한 낮이다. 윤성찬은 오른쪽 중심에 사선으로 배치되지만 시점은 지시된 약한 하향각보다 눈높이에 가까워 보인다.",
        "entities": "윤성찬의 고령 남성 얼굴, 짧게 가른 회백색 머리, 콧수염과 마른 체격은 인물 참조에 가깝다. 깨끗한 차콜 맞춤 정장, 흰 셔츠, 짙은 넥타이도 부합한다. 왼쪽 전경의 검은 머리 남자는 이 장면에서 보이면 안 되는 추가 인물이다. 별도 자막이나 그래픽은 없다.",
        "hard_violations": [
         "장관을 왼쪽 크롭 밖에 두라는 명시적 지시를 어기고, 왼쪽 전경에 다른 남자의 머리와 정장 상체를 노출했다."
        ],
        "physics": "윤성찬은 등받이가 뒤에 있는 가죽 의자에 앉아 상체를 앞으로 기울인다. 좌석이 몸을 받치며, 내민 손은 자연스럽게 연결된 팔로 유지되고 팔꿈치는 탁자 가까이에 놓인다. 손바닥을 펼친 채 동작을 멈추는 자세가 물리적으로 가능하고, 무지지 부유나 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷에서 확립된 공간 구조를 위반하여 윤성찬을 방문객 자리가 아닌 장관의 자리에 배치함",
     "[gemini-pro] 국방장관을 왼쪽 크롭 밖으로 두어 화면에서 제외하라는 지시를 어기고 어깨와 머리 일부를 프레임 안에 포함함",
     "[gpt-high] 장관을 왼쪽 크롭 밖에 두라는 명시적 지시를 어기고, 왼쪽 전경에 다른 남자의 머리와 정장 상체를 노출했다."
    ],
    "B": [
     "[gemini-pro] 공간의 물리적 일관성을 무시하고 윤성찬이 장관의 위치와 배경을 차지하도록 잘못 배치함",
     "[gemini-pro] 장관을 프레임 왼쪽 밖으로 제외하라는 지시와 달리 어깨가 명확히 화면에 포함됨",
     "[gpt-high] 장관을 왼쪽 크롭 밖에 두고 윤성찬만 보여야 하는데, 왼쪽 전경에 다른 남자의 머리와 상체를 크게 넣었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1083
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "카메라의 약간 하향된 피치(downward pitch)는 어느 정도 구현했으나, 윤성찬이 장관의 자리에 앉아 이전 샷의 배경이 뒤바뀌는 치명적인 공간 오류를 범했고 장관을 크롭 바깥으로 빼내지 못했습니다.  ★위반: [gemini-pro] 이전 샷에서 확립된 공간 구조를 위반하여 윤성찬을 방문객 자리가 아닌 장관의 자리에 배치함 / [gemini-pro] 국방장관을 왼쪽 크롭 밖으로 두어 화면에서 제외하라는 지시를 어기고 어깨와 머리 일부를 프레임 안에 포함함 / [gpt-high] 장관을 왼쪽 크롭 밖에 두라는 명시적 지시를 어기고, 왼쪽 전경에 다른 남자의 머리와 정장 상체를 노출했다."
   },
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "이전 샷의 공간적 논리를 무시하고 윤성찬을 장관의 배경 쪽에 배치하는 하드 위반을 범했으며, 하향 카메라 앵글 지시도 따르지 않았고 장관의 어깨가 화면에 노출되었습니다.  ★위반: [gemini-pro] 공간의 물리적 일관성을 무시하고 윤성찬이 장관의 위치와 배경을 차지하도록 잘못 배치함 / [gemini-pro] 장관을 프레임 왼쪽 밖으로 제외하라는 지시와 달리 어깨가 명확히 화면에 포함됨 / [gpt-high] 장관을 왼쪽 크롭 밖에 두고 윤성찬만 보여야 하는데, 왼쪽 전경에 다른 남자의 머리와 상체를 크게 넣었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S43sh9_sel.png",
    "asset_id": "de30b32b-ef0a-4e80-8a60-da33f5a7da34",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bae-2750-7aed-bf1d-d3b15d6834eb",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh9"
  }
 },
 "S54sh6::signage": {
  "fp": "8f250a9a821458ef",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh6": {
  "input_fingerprint": "5230262c8f49ce41",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬을 의심스러운 눈길로 쏘아보는 국방장관의 찌푸린 측면.\n\nLOCATION (lock): At the minister's position in the private office's conversation area, under daytime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle beside 국방장관 at his eye level, nearly perpendicular to the conversation line, holding a static close profile on the same established side of the axis. His furrowed face occupies the left half, his chin slightly drawn back as he scrutinizes 윤성찬 beyond the right edge; leave that side open without bringing 윤성찬 into view before the subsequent pullback. Make the minister's redirected scrutiny the expressive change, with no additional lighting shift.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daytime illumination, retaining readable detail around the brow and mouth without turning suspicion into theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the minister's office furnishings, interior finishes, and daylight appearance from the reference. Exclude militia-office objects such as the dart-covered map and the knife embedded in a wall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬을 의심스러운 눈길로 쏘아보는 국방장관의 찌푸린 측면.\n\nLOCATION (lock): At the minister's position in the private office's conversation area, under daytime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle beside 국방장관 at his eye level, nearly perpendicular to the conversation line, holding a static close profile on the same established side of the axis. His furrowed face occupies the left half, his chin slightly drawn back as he scrutinizes 윤성찬 beyond the right edge; leave that side open without bringing 윤성찬 into view before the subsequent pullback. Make the minister's redirected scrutiny the expressive change, with no additional lighting shift.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daytime illumination, retaining readable detail around the brow and mouth without turning suspicion into theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the minister's office furnishings, interior finishes, and daylight appearance from the reference. Exclude militia-office objects such as the dart-covered map and the knife embedded in a wall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 윤성찬을 의심스러운 눈길로 쏘아보는 국방장관의 찌푸린 측면.\n\nLOCATION (lock): At the minister's position in the private office's conversation area, under daytime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle beside 국방장관 at his eye level, nearly perpendicular to the conversation line, holding a static close profile on the same established side of the axis. His furrowed face occupies the left half, his chin slightly drawn back as he scrutinizes 윤성찬 beyond the right edge; leave that side open without bringing 윤성찬 into view before the subsequent pullback. Make the minister's redirected scrutiny the expressive change, with no additional lighting shift.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daytime illumination, retaining readable detail around the brow and mouth without turning suspicion into theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the minister's office furnishings, interior finishes, and daylight appearance from the reference. Exclude militia-office objects such as the dart-covered map and the knife embedded in a wall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "장관이 화면 우측 바깥에 있는 윤성찬을 향해 날카로운 시선을 던지고 있음.",
    "built_space": "좌측에 창문, 중앙에 태극기, 우측에 책장이 위치함. 지정된 카메라 시점에서는 창문이 등 뒤(좌측 배경)에 올 수 없어 공간 구조상 불가능한 배치임.",
    "entities": "지침(짧은 검은 머리의 남성) 및 레퍼런스와 달리 이전 샷의 백발 노인으로 완전히 잘못 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
     "[gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
    ],
    "physics": "인물이 의자에 자연스럽게 앉아 무게를 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "장관이 화면 우측 바깥쪽의 타겟을 향해 시선을 고정하고 있음.",
    "built_space": "A와 동일하게 불가능한 창문 배치를 보이며, 태극기 문양의 상하좌우가 잘못된 방향으로 그려짐.",
    "entities": "지정된 흑발 인물이 아닌 이전 샷의 백발 노인으로 묘사되어 신원 지침을 실패함.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
     "[gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
    ],
    "physics": "신체가 의자에 올바르게 지지되어 안착된 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "명시된 캐릭터 설정을 무시하고 금지된 이전 샷의 노인을 그대로 렌더링하여 지침을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일한 캐릭터 설정 위반에 더해 배경의 태극기 문양 방향까지 크게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장관이 화면 우측 바깥에 있는 윤성찬을 향해 날카로운 시선을 던지고 있음.",
        "built_space": "좌측에 창문, 중앙에 태극기, 우측에 책장이 위치함. 지정된 카메라 시점에서는 창문이 등 뒤(좌측 배경)에 올 수 없어 공간 구조상 불가능한 배치임.",
        "entities": "지침(짧은 검은 머리의 남성) 및 레퍼런스와 달리 이전 샷의 백발 노인으로 완전히 잘못 묘사됨.",
        "hard_violations": [
         "지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
         "카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
        ],
        "physics": "인물이 의자에 자연스럽게 앉아 무게를 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "장관이 화면 우측 바깥쪽의 타겟을 향해 시선을 고정하고 있음.",
        "built_space": "A와 동일하게 불가능한 창문 배치를 보이며, 태극기 문양의 상하좌우가 잘못된 방향으로 그려짐.",
        "entities": "지정된 흑발 인물이 아닌 이전 샷의 백발 노인으로 묘사되어 신원 지침을 실패함.",
        "hard_violations": [
         "지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
         "카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
        ],
        "physics": "신체가 의자에 올바르게 지지되어 안착된 상태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "명시된 캐릭터 설정을 무시하고 금지된 이전 샷의 노인을 그대로 렌더링하여 지침을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일한 캐릭터 설정 위반에 더해 배경의 태극기 문양 방향까지 크게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "장관이 화면 우측 바깥에 있는 윤성찬을 향해 날카로운 시선을 던지고 있음.",
        "built_space": "좌측에 창문, 중앙에 태극기, 우측에 책장이 위치함. 지정된 카메라 시점에서는 창문이 등 뒤(좌측 배경)에 올 수 없어 공간 구조상 불가능한 배치임.",
        "entities": "지침(짧은 검은 머리의 남성) 및 레퍼런스와 달리 이전 샷의 백발 노인으로 완전히 잘못 묘사됨.",
        "hard_violations": [
         "지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
         "카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
        ],
        "physics": "인물이 의자에 자연스럽게 앉아 무게를 지지하고 있음."
       },
       {
        "label": "B",
        "direction": "장관이 화면 우측 바깥쪽의 타겟을 향해 시선을 고정하고 있음.",
        "built_space": "A와 동일하게 불가능한 창문 배치를 보이며, 태극기 문양의 상하좌우가 잘못된 방향으로 그려짐.",
        "entities": "지정된 흑발 인물이 아닌 이전 샷의 백발 노인으로 묘사되어 신원 지침을 실패함.",
        "hard_violations": [
         "지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
         "카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
        ],
        "physics": "신체가 의자에 올바르게 지지되어 안착된 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 얼굴 배치와 오른쪽 여백, 다문 입의 의심 표현은 더 가깝지만, 지정된 인물 대신 이전 장면의 노인을 재현했고 요구된 정측면도 아니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "오른쪽 밖을 의심하는 시선은 맞지만 정측면이 아닌 사선 얼굴이며, 인물 참조와 전혀 다른 백발·콧수염의 노인을 재현하고 상체가 여백을 더 차지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 화면 오른쪽 위 바깥을 향한다. 화면 밖 윤성찬을 살핀다는 설정과 양립하며, 윤성찬이나 다른 사람은 보이지 않는다. 다만 양쪽 눈과 얼굴 전면이 상당히 드러나는 사선 각도로, 대화선에 거의 수직인 정측면 촬영은 아니다.",
        "built_space": "왼쪽에 분할 창 하나, 그 옆 벽걸이 달력 하나, 중앙 뒤 태극기 하나, 오른쪽 위 풍경 액자 하나, 오른쪽 뒤 목재 수납장 하나와 책 묶음, 아래쪽 검은 가죽 의자 하나가 보인다. 식물도 창가와 인물 뒤에 보인다. 베이지 벽과 목재·가죽 마감, 낮빛은 장소 참조와 대체로 일치한다. 인물이 앉은 좌면은 잘려 있어 정확한 착석 위치는 확인할 수 없다. 중복 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 동아시아계로 보이는 노년 남성 한 명이다. 백발, 흰 콧수염, 깊은 주름은 이전 장면 인물과 닮았고, 지정 인물 참조의 짧은 검은 머리와 수염 없는 성숙한 남성 얼굴에는 맞지 않는다. 흰 셔츠는 맞지만 정장은 지정된 짙은 네이비보다 회색에 가깝다. 찌푸린 미간과 다문 입은 의심의 연기로 읽힌다. 지도, 벽에 꽂힌 칼, 추가 인물, 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고 상체는 약간 앞으로 기울어 있다. 하체와 좌면은 프레임 밖이므로 지지 접촉을 직접 확인할 수 없지만, 떠 있는 몸으로 보이지는 않는다. 책과 소품은 수납장 위에 놓여 있고 깃발은 깃대에 걸려 있다. 공중 동작이나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "눈동자는 화면 오른쪽 위 바깥을 향하며 카메라를 보지 않는다. 윤성찬은 보이지 않아 화면 밖 상대를 응시한다는 조건은 지킨다. 그러나 양쪽 눈과 코 양옆이 보이는 사선 얼굴이어서 요구된 가까운 정측면과 다르다.",
        "built_space": "왼쪽 분할 창 하나와 식물, 머리 뒤 일부 가려진 달력 하나, 중앙 뒤 태극기 하나, 오른쪽 위 풍경 액자 하나, 뒤쪽 목재 수납장 하나와 책 묶음이 보인다. 오른쪽 아래에는 인물 등 뒤의 가죽 의자 등받이 하나가 있다. 참조의 벽·창·목재·가죽과 낮 조명은 대체로 유지한다. 오른쪽 수납장 위에는 참조에서 확인되지 않는 작은 지구본 모양 장식이 보인다. 중복 설비나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "동아시아계로 보이는 노년 남성 한 명만 등장한다. 백발과 흰 콧수염, 노년의 얼굴은 지정된 검은 머리의 무수염 인물이 아니라 이전 장면 인물을 재현한 모습이다. 흰 셔츠는 맞지만 정장은 네이비보다 짙은 회색으로 보인다. 미간을 찌푸리고 입술을 조금 벌린 표정은 의심으로 읽힌다. 금지된 지도·칼이나 추가 인물, 자막은 없다.",
        "hard_violations": [],
        "physics": "상체 뒤로 의자 등받이가 보여 착석 자세와 양립한다. 엉덩이와 발은 잘려 있지만 머리·목·어깨의 연결과 기울기는 자연스럽다. 책과 장식은 수납장 상판에 지지되고 깃발은 깃대에 매달려 있다. 부유하는 신체나 물체, 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 얼굴 배치와 오른쪽 여백, 다문 입의 의심 표현은 더 가깝지만, 지정된 인물 대신 이전 장면의 노인을 재현했고 요구된 정측면도 아니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽 밖을 의심하는 시선은 맞지만 정측면이 아닌 사선 얼굴이며, 인물 참조와 전혀 다른 백발·콧수염의 노인을 재현하고 상체가 여백을 더 차지한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 화면 오른쪽 위 바깥을 향한다. 화면 밖 윤성찬을 살핀다는 설정과 양립하며, 윤성찬이나 다른 사람은 보이지 않는다. 다만 양쪽 눈과 얼굴 전면이 상당히 드러나는 사선 각도로, 대화선에 거의 수직인 정측면 촬영은 아니다.",
        "built_space": "왼쪽에 분할 창 하나, 그 옆 벽걸이 달력 하나, 중앙 뒤 태극기 하나, 오른쪽 위 풍경 액자 하나, 오른쪽 뒤 목재 수납장 하나와 책 묶음, 아래쪽 검은 가죽 의자 하나가 보인다. 식물도 창가와 인물 뒤에 보인다. 베이지 벽과 목재·가죽 마감, 낮빛은 장소 참조와 대체로 일치한다. 인물이 앉은 좌면은 잘려 있어 정확한 착석 위치는 확인할 수 없다. 중복 고정 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "보이는 사람은 동아시아계로 보이는 노년 남성 한 명이다. 백발, 흰 콧수염, 깊은 주름은 이전 장면 인물과 닮았고, 지정 인물 참조의 짧은 검은 머리와 수염 없는 성숙한 남성 얼굴에는 맞지 않는다. 흰 셔츠는 맞지만 정장은 지정된 짙은 네이비보다 회색에 가깝다. 찌푸린 미간과 다문 입은 의심의 연기로 읽힌다. 지도, 벽에 꽂힌 칼, 추가 인물, 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고 상체는 약간 앞으로 기울어 있다. 하체와 좌면은 프레임 밖이므로 지지 접촉을 직접 확인할 수 없지만, 떠 있는 몸으로 보이지는 않는다. 책과 소품은 수납장 위에 놓여 있고 깃발은 깃대에 걸려 있다. 공중 동작이나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "눈동자는 화면 오른쪽 위 바깥을 향하며 카메라를 보지 않는다. 윤성찬은 보이지 않아 화면 밖 상대를 응시한다는 조건은 지킨다. 그러나 양쪽 눈과 코 양옆이 보이는 사선 얼굴이어서 요구된 가까운 정측면과 다르다.",
        "built_space": "왼쪽 분할 창 하나와 식물, 머리 뒤 일부 가려진 달력 하나, 중앙 뒤 태극기 하나, 오른쪽 위 풍경 액자 하나, 뒤쪽 목재 수납장 하나와 책 묶음이 보인다. 오른쪽 아래에는 인물 등 뒤의 가죽 의자 등받이 하나가 있다. 참조의 벽·창·목재·가죽과 낮 조명은 대체로 유지한다. 오른쪽 수납장 위에는 참조에서 확인되지 않는 작은 지구본 모양 장식이 보인다. 중복 설비나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "동아시아계로 보이는 노년 남성 한 명만 등장한다. 백발과 흰 콧수염, 노년의 얼굴은 지정된 검은 머리의 무수염 인물이 아니라 이전 장면 인물을 재현한 모습이다. 흰 셔츠는 맞지만 정장은 네이비보다 짙은 회색으로 보인다. 미간을 찌푸리고 입술을 조금 벌린 표정은 의심으로 읽힌다. 금지된 지도·칼이나 추가 인물, 자막은 없다.",
        "hard_violations": [],
        "physics": "상체 뒤로 의자 등받이가 보여 착석 자세와 양립한다. 엉덩이와 발은 잘려 있지만 머리·목·어깨의 연결과 기울기는 자연스럽다. 책과 장식은 수납장 상판에 지지되고 깃발은 깃대에 매달려 있다. 부유하는 신체나 물체, 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
     "[gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
    ],
    "B": [
     "[gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people)",
     "[gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "명시된 캐릭터 설정을 무시하고 금지된 이전 샷의 노인을 그대로 렌더링하여 지침을 심각하게 위반했습니다.  ★위반: [gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people) / [gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "A와 동일한 캐릭터 설정 위반에 더해 배경의 태극기 문양 방향까지 크게 왜곡되었습니다.  ★위반: [gemini-pro] 지정된 인물이 아닌 금지된 이전 샷 인물 묘사 (Invented people) / [gemini-pro] 카메라 시점상 불가능한 창문 및 배경 구조 배치 (Physically impossible staging)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S54sh4_sel.png",
    "asset_id": "f529566f-2dfa-47be-9998-e0a55429f33d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:805558>",
    "asset_id": "2ee9d902-fac0-42dc-986b-7bd16edba439",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bb4-20f2-79c3-b4ea-b4688111986c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh4"
  }
 },
 "S55sh5::signage": {
  "fp": "22874e208f1371d3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh5::bgfirst_bg": {
  "input_fingerprint": "307c91492e2b29f7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 뒷좌석의 라울이 창밖 숲 쪽을 향해 손가락을 뻗은 측면 구도.\n\nLOCATION (lock): Inside the camper's rear passenger area beside a side window, looking toward the forest bordering the rural road. Daylight enters after a rain shower.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-seat tracking stage beside 라울 at his seated shoulder height, looking across his profile toward the side window. Place his tired upper body on the left and his pointing hand below the window on the right, with his gaze following the same outward line toward the forest rather than toward the camera. Preserve a comfortable distance between his face and the window, holding the gesture's destination before the camera advances toward the driver.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: forest visible through the side window, aligned with 라울's pointing finger in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 캠핑카 측면 창문 (The forest outside is visible through it) — Viewed obliquely across the interior, with the forest beyond 라울's pointing hand; used as Links his gesture and gaze to the proposed search direction; 창밖 숲 (Visible outside the camper as the subject of 라울's suggestion); used as Provides the destination in the right side of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime light through the window gives the rain-soaked, exhausted 라울 gentle definition without adding ongoing rain or atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 뒷좌석의 라울이 창밖 숲 쪽을 향해 손가락을 뻗은 측면 구도.\n\nLOCATION (lock): Inside the camper's rear passenger area beside a side window, looking toward the forest bordering the rural road. Daylight enters after a rain shower.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-seat tracking stage beside 라울 at his seated shoulder height, looking across his profile toward the side window. Place his tired upper body on the left and his pointing hand below the window on the right, with his gaze following the same outward line toward the forest rather than toward the camera. Preserve a comfortable distance between his face and the window, holding the gesture's destination before the camera advances toward the driver.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: forest visible through the side window, aligned with 라울's pointing finger in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 캠핑카 측면 창문 (The forest outside is visible through it) — Viewed obliquely across the interior, with the forest beyond 라울's pointing hand; used as Links his gesture and gaze to the proposed search direction; 창밖 숲 (Visible outside the camper as the subject of 라울's suggestion); used as Provides the destination in the right side of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime light through the window gives the rain-soaked, exhausted 라울 gentle definition without adding ongoing rain or atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh5__bgfirst_bg.png",
  "asset_id": "7a0d9017-f905-4faf-a276-564f50e89ffd",
  "input_asset_ids": [
   "14d22d90-78d6-4643-a901-3050ae29203b",
   "52016a0f-334c-4dbd-addb-6c0ad953ba30"
  ]
 },
 "S55sh5": {
  "input_fingerprint": "cab1951129a4962e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뒷좌석의 라울이 창밖 숲 쪽을 향해 손가락을 뻗은 측면 구도.\n\nLOCATION (lock): Inside the camper's rear passenger area beside a side window, looking toward the forest bordering the rural road. Daylight enters after a rain shower. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-seat tracking stage beside 라울 at his seated shoulder height, looking across his profile toward the side window. Place his tired upper body on the left and his pointing hand below the window on the right, with his gaze following the same outward line toward the forest rather than toward the camera. Preserve a comfortable distance between his face and the window, holding the gesture's destination before the camera advances toward the driver.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: forest visible through the side window, aligned with 라울's pointing finger in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 캠핑카 측면 창문 (The forest outside is visible through it) — Viewed obliquely across the interior, with the forest beyond 라울's pointing hand; used as Links his gesture and gaze to the proposed search direction; 창밖 숲 (Visible outside the camper as the subject of 라울's suggestion); used as Provides the destination in the right side of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime light through the window gives the rain-soaked, exhausted 라울 gentle definition without adding ongoing rain or atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A rain shower has passed over the rural road, and the camper still carries the packed luggage, added food and medicine. 라울: He is soaked and tired inside the camper, pointing toward the forest outside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뒷좌석의 라울이 창밖 숲 쪽을 향해 손가락을 뻗은 측면 구도.\n\nLOCATION (lock): Inside the camper's rear passenger area beside a side window, looking toward the forest bordering the rural road. Daylight enters after a rain shower. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-seat tracking stage beside 라울 at his seated shoulder height, looking across his profile toward the side window. Place his tired upper body on the left and his pointing hand below the window on the right, with his gaze following the same outward line toward the forest rather than toward the camera. Preserve a comfortable distance between his face and the window, holding the gesture's destination before the camera advances toward the driver.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: forest visible through the side window, aligned with 라울's pointing finger in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 캠핑카 측면 창문 (The forest outside is visible through it) — Viewed obliquely across the interior, with the forest beyond 라울's pointing hand; used as Links his gesture and gaze to the proposed search direction; 창밖 숲 (Visible outside the camper as the subject of 라울's suggestion); used as Provides the destination in the right side of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime light through the window gives the rain-soaked, exhausted 라울 gentle definition without adding ongoing rain or atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A rain shower has passed over the rural road, and the camper still carries the packed luggage, added food and medicine. 라울: He is soaked and tired inside the camper, pointing toward the forest outside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뒷좌석의 라울이 창밖 숲 쪽을 향해 손가락을 뻗은 측면 구도.\n\nLOCATION (lock): Inside the camper's rear passenger area beside a side window, looking toward the forest bordering the rural road. Daylight enters after a rain shower. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-seat tracking stage beside 라울 at his seated shoulder height, looking across his profile toward the side window. Place his tired upper body on the left and his pointing hand below the window on the right, with his gaze following the same outward line toward the forest rather than toward the camera. Preserve a comfortable distance between his face and the window, holding the gesture's destination before the camera advances toward the driver.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: forest visible through the side window, aligned with 라울's pointing finger in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 캠핑카 측면 창문 (The forest outside is visible through it) — Viewed obliquely across the interior, with the forest beyond 라울's pointing hand; used as Links his gesture and gaze to the proposed search direction; 창밖 숲 (Visible outside the camper as the subject of 라울's suggestion); used as Provides the destination in the right side of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime light through the window gives the rain-soaked, exhausted 라울 gentle definition without adding ongoing rain or atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A rain shower has passed over the rural road, and the camper still carries the packed luggage, added food and medicine. 라울: He is soaked and tired inside the camper, pointing toward the forest outside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh5__bgfirst_bg.png",
     "asset_id": "7a0d9017-f905-4faf-a276-564f50e89ffd",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S55sh5.png",
     "asset_id": "14d22d90-78d6-4643-a901-3050ae29203b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camper_living_space_8fc641.png",
     "asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "라울이 창문 밖 숲을 바라보며 오른손 검지로 숲을 가리키고 있음.",
    "built_space": "캠핑카 내부이나, 원본 공간에 없는 여러 줄의 앞을 향한 좌석들이 버스처럼 배치되어 있음. 창문은 우측에 위치함.",
    "entities": "라울의 얼굴, 꽁지머리, 낡고 오염된 티셔츠와 반바지 모두 참조 이미지와 잘 일치함.",
    "hard_violations": [
     "[gemini-pro] 캠핑카 내부 구조가 버스처럼 여러 줄의 정면 방향 좌석으로 왜곡되어 공간 레퍼런스(LOCATION lock)와 불일치함."
    ],
    "physics": "좌석에 앉아 안전벨트를 맨 상태로 오른팔을 들어 손가락을 창문 유리에 대고 있으며, 체중은 좌석이 지탱함."
   },
   {
    "label": "B",
    "direction": "라울이 우측 창밖의 숲을 응시하며 오른손을 뻗어 그 방향을 가리키고 있음.",
    "built_space": "캠핑카 내부. 원본과 일치하는 커튼과 창문 프레임이 있으며, 라울은 측면 벤치에 앉아 있음. 우측 하단에 내부 구조물 일부가 보임.",
    "entities": "라울의 인종, 연령, 꽁지머리, 오염된 의상 등 모든 특징이 참조 이미지와 완벽히 일치함.",
    "hard_violations": [],
    "physics": "벤치에 안정적으로 앉아 있으며, 허공으로 뻗은 오른팔은 어깨와 몸통 근육으로 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터의 외형을 완벽히 재현했고, 창문과 얼굴 사이의 거리를 유지하라는 연출 지시와 측면 구도를 매우 정확히 따름."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물 재현은 훌륭하나 캠핑카 내부가 버스 좌석처럼 왜곡되었고, 얼굴이 창문에 너무 가까워 지시사항을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울이 창문 밖 숲을 바라보며 오른손 검지로 숲을 가리키고 있음.",
        "built_space": "캠핑카 내부이나, 원본 공간에 없는 여러 줄의 앞을 향한 좌석들이 버스처럼 배치되어 있음. 창문은 우측에 위치함.",
        "entities": "라울의 얼굴, 꽁지머리, 낡고 오염된 티셔츠와 반바지 모두 참조 이미지와 잘 일치함.",
        "hard_violations": [
         "캠핑카 내부 구조가 버스처럼 여러 줄의 정면 방향 좌석으로 왜곡되어 공간 레퍼런스(LOCATION lock)와 불일치함."
        ],
        "physics": "좌석에 앉아 안전벨트를 맨 상태로 오른팔을 들어 손가락을 창문 유리에 대고 있으며, 체중은 좌석이 지탱함."
       },
       {
        "label": "B",
        "direction": "라울이 우측 창밖의 숲을 응시하며 오른손을 뻗어 그 방향을 가리키고 있음.",
        "built_space": "캠핑카 내부. 원본과 일치하는 커튼과 창문 프레임이 있으며, 라울은 측면 벤치에 앉아 있음. 우측 하단에 내부 구조물 일부가 보임.",
        "entities": "라울의 인종, 연령, 꽁지머리, 오염된 의상 등 모든 특징이 참조 이미지와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "벤치에 안정적으로 앉아 있으며, 허공으로 뻗은 오른팔은 어깨와 몸통 근육으로 자연스럽게 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "캐릭터의 외형을 완벽히 재현했고, 창문과 얼굴 사이의 거리를 유지하라는 연출 지시와 측면 구도를 매우 정확히 따름."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물 재현은 훌륭하나 캠핑카 내부가 버스 좌석처럼 왜곡되었고, 얼굴이 창문에 너무 가까워 지시사항을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "라울이 창문 밖 숲을 바라보며 오른손 검지로 숲을 가리키고 있음.",
        "built_space": "캠핑카 내부이나, 원본 공간에 없는 여러 줄의 앞을 향한 좌석들이 버스처럼 배치되어 있음. 창문은 우측에 위치함.",
        "entities": "라울의 얼굴, 꽁지머리, 낡고 오염된 티셔츠와 반바지 모두 참조 이미지와 잘 일치함.",
        "hard_violations": [
         "캠핑카 내부 구조가 버스처럼 여러 줄의 정면 방향 좌석으로 왜곡되어 공간 레퍼런스(LOCATION lock)와 불일치함."
        ],
        "physics": "좌석에 앉아 안전벨트를 맨 상태로 오른팔을 들어 손가락을 창문 유리에 대고 있으며, 체중은 좌석이 지탱함."
       },
       {
        "label": "B",
        "direction": "라울이 우측 창밖의 숲을 응시하며 오른손을 뻗어 그 방향을 가리키고 있음.",
        "built_space": "캠핑카 내부. 원본과 일치하는 커튼과 창문 프레임이 있으며, 라울은 측면 벤치에 앉아 있음. 우측 하단에 내부 구조물 일부가 보임.",
        "entities": "라울의 인종, 연령, 꽁지머리, 오염된 의상 등 모든 특징이 참조 이미지와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "벤치에 안정적으로 앉아 있으며, 허공으로 뻗은 오른팔은 어깨와 몸통 근육으로 자연스럽게 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 상체와 오른쪽 숲을 잇는 손짓, 어깨 높이에 가까운 측면 미디엄 구도와 낡은 실내 재질이 더 충실하지만, 손이 창 아래보다 높고 흠뻑 젖은 상태는 약하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "옆얼굴과 숲을 향한 시선·손짓은 정확하지만, 높은 시점과 넓어진 통로 노출, 헤드레스트 좌석과 성형 벽 패널이 지정 구도와 장소에서 더 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "라울은 화면 오른쪽을 바라보고 검지를 같은 방향으로 뻗는다. 손끝의 연장선은 창밖 도로 너머의 수풀과 나무에 닿으며, 시선도 그 숲을 향한다. 카메라를 보지 않는다. 손은 오른쪽 중앙에 있지만 창 아래가 아니라 유리 면 앞의 중간 높이에 놓인다.",
        "built_space": "오른쪽 큰 측면 창 하나, 인물 뒤의 창 하나와 왼쪽 끝 창 일부가 보인다. 큰 창 왼쪽에는 묶인 커튼 한 폭이 있고, 검은 창틀과 하단 잠금장치가 보인다. 왼쪽에는 천으로 덮인 좌석과 등받이, 위에는 목재 수납장 열, 오른쪽 아래에는 젖은 상판 일부가 있다. 낡은 갈색 내장과 직물은 장소 사진에 가깝다. 인물과 창 사이에도 충분한 거리가 있다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "보이는 사람은 어린 남자아이 한 명뿐이다. 갈색 피부, 어린 얼굴, 곱슬머리를 뒤로 묶은 형태와 마른 체격은 라울 참조에 가깝다. 지정된 혼혈 배경 자체는 외모만으로 확정할 수 없다. 빛바랜 얼룩진 올리브색 티셔츠와 베이지색 반바지가 일치한다. 창밖에는 숲과 젖은 도로가 있고 유리에 물방울이 남아 있으며, 진행 중인 비는 뚜렷하지 않다. 피로는 약간 보이지만 머리와 옷이 흠뻑 젖었다는 표현은 약하다. 짐·식품·약품은 이 크롭에서 확인되지 않는다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "골반과 허벅지는 좌석 방석에 지지되고 등 뒤에는 등받이가 있다. 왼팔은 허벅지 쪽으로 내려가며, 오른팔은 어깨와 팔꿈치에서 이어져 검지를 뻗는 자연스러운 자세다. 손이나 몸이 지지 없이 떠 있는 부분은 없다. 창문을 관통하지 않고 실내에서 가리키는 동작으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "라울의 얼굴은 오른쪽을 향한 선명한 측면이며 눈도 창밖을 본다. 뻗은 검지는 오른쪽 도로 가장자리 너머의 낮은 수풀을 가리킨다. 시선과 손짓 모두 숲을 목적지로 삼고 있다. 손은 창의 아래쪽 유리 면 앞에 있지만 하단 창틀 아래까지 내려가지는 않는다.",
        "built_space": "오른쪽 큰 측면 창 하나와 그 뒤의 측면 창 두 개, 멀리 끝쪽 유리창이 보인다. 큰 창 왼쪽에는 커튼 한 폭과 검은 창틀 손잡이가 있다. 천장등 하나와 양쪽 상부 수납장 열도 보인다. 라울은 창가 등받이 좌석에 앉아 있고 왼쪽에는 빈 좌석과 헤드레스트, 별도 안전벨트가 보인다. 창 아래에는 큰 성형 패널과 팔걸이가 있다. 이 좌석·패널 구성은 참조의 낡은 벤치형 좌석과 목재 중심 실내보다 승합차에 가깝다. 카메라도 앉은 어깨보다 높아서 통로와 천장이 많이 드러난다. 명백히 불가능한 반사는 없다.",
        "entities": "어린 남자아이 한 명만 보이며, 갈색 피부와 뒤로 묶은 곱슬머리, 어린 옆얼굴과 체격은 라울 참조에 가깝다. 혼혈 배경은 영상만으로 확정할 수 없다. 얼룩지고 해진 올리브색 티셔츠와 화면 아래 일부 보이는 베이지색 반바지는 맞는다. 몸을 가로지르는 안전벨트가 추가되어 있다. 뒤에는 여행가방 두 개와 큰 천가방 하나가 보이지만 식품과 약품 여부는 식별할 수 없다. 숲, 젖은 도로, 유리의 잔류 물방울과 낮빛은 맞는다. 흠뻑 젖고 지친 상태는 충분히 강하지 않다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "하체는 방석에 놓이고 등 뒤에는 좌석 등받이가 있으며 안전벨트는 몸통을 가로질러 내려온다. 오른팔은 어깨에서 창 쪽으로 자연스럽게 뻗어 있고 손목과 검지의 연결도 타당하다. 왼팔은 무릎 쪽으로 내려간다. 뒤의 가방들은 실내 좌석이나 적재면 위에 놓여 있으며 떠 있지 않다. 몸과 손의 지지 및 가리키는 동작에 명백한 물리적 모순은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 상체와 오른쪽 숲을 잇는 손짓, 어깨 높이에 가까운 측면 미디엄 구도와 낡은 실내 재질이 더 충실하지만, 손이 창 아래보다 높고 흠뻑 젖은 상태는 약하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "옆얼굴과 숲을 향한 시선·손짓은 정확하지만, 높은 시점과 넓어진 통로 노출, 헤드레스트 좌석과 성형 벽 패널이 지정 구도와 장소에서 더 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "라울은 화면 오른쪽을 바라보고 검지를 같은 방향으로 뻗는다. 손끝의 연장선은 창밖 도로 너머의 수풀과 나무에 닿으며, 시선도 그 숲을 향한다. 카메라를 보지 않는다. 손은 오른쪽 중앙에 있지만 창 아래가 아니라 유리 면 앞의 중간 높이에 놓인다.",
        "built_space": "오른쪽 큰 측면 창 하나, 인물 뒤의 창 하나와 왼쪽 끝 창 일부가 보인다. 큰 창 왼쪽에는 묶인 커튼 한 폭이 있고, 검은 창틀과 하단 잠금장치가 보인다. 왼쪽에는 천으로 덮인 좌석과 등받이, 위에는 목재 수납장 열, 오른쪽 아래에는 젖은 상판 일부가 있다. 낡은 갈색 내장과 직물은 장소 사진에 가깝다. 인물과 창 사이에도 충분한 거리가 있다. 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "보이는 사람은 어린 남자아이 한 명뿐이다. 갈색 피부, 어린 얼굴, 곱슬머리를 뒤로 묶은 형태와 마른 체격은 라울 참조에 가깝다. 지정된 혼혈 배경 자체는 외모만으로 확정할 수 없다. 빛바랜 얼룩진 올리브색 티셔츠와 베이지색 반바지가 일치한다. 창밖에는 숲과 젖은 도로가 있고 유리에 물방울이 남아 있으며, 진행 중인 비는 뚜렷하지 않다. 피로는 약간 보이지만 머리와 옷이 흠뻑 젖었다는 표현은 약하다. 짐·식품·약품은 이 크롭에서 확인되지 않는다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "골반과 허벅지는 좌석 방석에 지지되고 등 뒤에는 등받이가 있다. 왼팔은 허벅지 쪽으로 내려가며, 오른팔은 어깨와 팔꿈치에서 이어져 검지를 뻗는 자연스러운 자세다. 손이나 몸이 지지 없이 떠 있는 부분은 없다. 창문을 관통하지 않고 실내에서 가리키는 동작으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "라울의 얼굴은 오른쪽을 향한 선명한 측면이며 눈도 창밖을 본다. 뻗은 검지는 오른쪽 도로 가장자리 너머의 낮은 수풀을 가리킨다. 시선과 손짓 모두 숲을 목적지로 삼고 있다. 손은 창의 아래쪽 유리 면 앞에 있지만 하단 창틀 아래까지 내려가지는 않는다.",
        "built_space": "오른쪽 큰 측면 창 하나와 그 뒤의 측면 창 두 개, 멀리 끝쪽 유리창이 보인다. 큰 창 왼쪽에는 커튼 한 폭과 검은 창틀 손잡이가 있다. 천장등 하나와 양쪽 상부 수납장 열도 보인다. 라울은 창가 등받이 좌석에 앉아 있고 왼쪽에는 빈 좌석과 헤드레스트, 별도 안전벨트가 보인다. 창 아래에는 큰 성형 패널과 팔걸이가 있다. 이 좌석·패널 구성은 참조의 낡은 벤치형 좌석과 목재 중심 실내보다 승합차에 가깝다. 카메라도 앉은 어깨보다 높아서 통로와 천장이 많이 드러난다. 명백히 불가능한 반사는 없다.",
        "entities": "어린 남자아이 한 명만 보이며, 갈색 피부와 뒤로 묶은 곱슬머리, 어린 옆얼굴과 체격은 라울 참조에 가깝다. 혼혈 배경은 영상만으로 확정할 수 없다. 얼룩지고 해진 올리브색 티셔츠와 화면 아래 일부 보이는 베이지색 반바지는 맞는다. 몸을 가로지르는 안전벨트가 추가되어 있다. 뒤에는 여행가방 두 개와 큰 천가방 하나가 보이지만 식품과 약품 여부는 식별할 수 없다. 숲, 젖은 도로, 유리의 잔류 물방울과 낮빛은 맞는다. 흠뻑 젖고 지친 상태는 충분히 강하지 않다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "하체는 방석에 놓이고 등 뒤에는 좌석 등받이가 있으며 안전벨트는 몸통을 가로질러 내려온다. 오른팔은 어깨에서 창 쪽으로 자연스럽게 뻗어 있고 손목과 검지의 연결도 타당하다. 왼팔은 무릎 쪽으로 내려간다. 뒤의 가방들은 실내 좌석이나 적재면 위에 놓여 있으며 떠 있지 않다. 몸과 손의 지지 및 가리키는 동작에 명백한 물리적 모순은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.446,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.196,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 캠핑카 내부 구조가 버스처럼 여러 줄의 정면 방향 좌석으로 왜곡되어 공간 레퍼런스(LOCATION lock)와 불일치함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1196
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "캐릭터의 외형을 완벽히 재현했고, 창문과 얼굴 사이의 거리를 유지하라는 연출 지시와 측면 구도를 매우 정확히 따름."
   },
   {
    "label": "A",
    "score": 1196,
    "verdict_ko": "인물 재현은 훌륭하나 캠핑카 내부가 버스 좌석처럼 왜곡되었고, 얼굴이 창문에 너무 가까워 지시사항을 위반함.  ★위반: [gemini-pro] 캠핑카 내부 구조가 버스처럼 여러 줄의 정면 방향 좌석으로 왜곡되어 공간 레퍼런스(LOCATION lock)와 불일치함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_camper_living_space_8fc641.png",
    "asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bb9-f311-7d68-bc35-3e98c0a55e02",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh5__bgfirst_bg.png",
   "bg_asset_id": "7a0d9017-f905-4faf-a276-564f50e89ffd",
   "bg_record_key": "S55sh5::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "camper_living_space",
   "groupbg_asset_id": "52016a0f-334c-4dbd-addb-6c0ad953ba30"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh6::confined_fp_apt": {
  "applies": true,
  "reason_ko": "캠핑카 운전석 내부에서 운전대를 급격히 조작하는 숏으로, 운전석의 위치와 스티어링 휠 등 차량 제어 장치의 공간적 배치, 그리고 창밖으로 향하는 숲의 방향이 시각적으로 정확하게 일치해야 하므로 평면도 보조가 필요합니다.",
  "input_fingerprint": "99891626734d0faa"
 },
 "S55sh6::signage": {
  "fp": "fe4cf444c579cee6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::085f743cd061": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_085f743cd061.png",
  "place_text": "At the driver's seat inside the camper's front cab, turning off the rural road toward the forest. Daylight enters through the windshield.",
  "input_fingerprint": "2fe3665fa93d74a6"
 },
 "S55sh6::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located at the front-left Driver station.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "Positioned in the center aisle between the Driver and Passenger seats, just behind the seat backs, pointing diagonally forward and left toward the steering wheel and Driver.",
   "occupants": "이현우 occupies the Driver seat on the left."
  },
  "mismatches": [],
  "scene_description_en": "The camera looks forward and slightly left from the space between the front seats. On the left side of the frame in the near-to-middle ground, 이현우 sits in the driver's seat, facing away toward the front of the vehicle. In front of him, occupying the center-left middle ground, is the steering wheel. The wide windshield stretches across the background from left to right. The empty passenger seat is positioned out of frame to the camera's immediate right, while the camper living area sits entirely behind the camera. No mirrors are present in this layout.",
  "fixed": false,
  "input_fingerprint": "fc3280f60bb4435f"
 },
 "S55sh6": {
  "input_fingerprint": "ba83f43f460668da",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대를 숲 쪽으로 확 꺾은 채 팔에 힘이 들어간 이현우의 역동적인 찰나.\n\nLOCATION (lock): At the driver's seat inside the camper's front cab, turning off the rural road toward the forest. Daylight enters through the windshield. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the interior approach on the passenger side, slightly behind 이현우's shoulder line at his lower-chest height, looking obliquely upward across his arms and steering wheel. His torso occupies the left-center, his loaded forearms run toward the wheel in the lower right, and his partly visible profile remains directed along the vehicle's new heading beyond the frame. Emphasize the closer camera distance while retaining the driver's forward axis, with the wheel below a third of the image and the torso providing a natural scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 운전대 (Turned sharply as 이현우 redirects the camper toward the forest) — Seen obliquely from the passenger side beneath his hands; used as Shows the directional decision through the relationship between hands, arms, and wheel.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime illumination across his rain-soaked appearance, keeping arm tension readable without changing the established interior light character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper turns toward the forest after the passing shower, with the packed supplies still aboard. 이현우: He remains at the wheel, soaked and exhausted, with his treated leg still bandaged.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera looks forward and slightly left from the space between the front seats. On the left side of the frame in the near-to-middle ground, 이현우 sits in the driver's seat, facing away toward the front of the vehicle. In front of him, occupying the center-left middle ground, is the steering wheel. The wide windshield stretches across the background from left to right. The empty passenger seat is positioned out of frame to the camera's immediate right, while the camper living area sits entirely behind the camera. No mirrors are present in this layout.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대를 숲 쪽으로 확 꺾은 채 팔에 힘이 들어간 이현우의 역동적인 찰나.\n\nLOCATION (lock): At the driver's seat inside the camper's front cab, turning off the rural road toward the forest. Daylight enters through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime illumination across his rain-soaked appearance, keeping arm tension readable without changing the established interior light character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper turns toward the forest after the passing shower, with the packed supplies still aboard. 이현우: He remains at the wheel, soaked and exhausted, with his treated leg still bandaged.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera looks forward and slightly left from the space between the front seats. On the left side of the frame in the near-to-middle ground, 이현우 sits in the driver's seat, facing away toward the front of the vehicle. In front of him, occupying the center-left middle ground, is the steering wheel. The wide windshield stretches across the background from left to right. The empty passenger seat is positioned out of frame to the camera's immediate right, while the camper living area sits entirely behind the camera. No mirrors are present in this layout.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전대를 숲 쪽으로 확 꺾은 채 팔에 힘이 들어간 이현우의 역동적인 찰나.\n\nLOCATION (lock): At the driver's seat inside the camper's front cab, turning off the rural road toward the forest. Daylight enters through the windshield. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime illumination across his rain-soaked appearance, keeping arm tension readable without changing the established interior light character.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper turns toward the forest after the passing shower, with the packed supplies still aboard. 이현우: He remains at the wheel, soaked and exhausted, with his treated leg still bandaged.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh6_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh6_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "차량과 인물의 시선이 숲 방향의 흙길을 향함.",
    "built_space": "캠핑카 운전석 내부. 조수석 뒤편 시점이며 운전대와 대시보드가 배치됨.",
    "entities": "이현우 (짧은 검은 머리, 흙 묻은 셔츠, 인이어 무전기 일치).",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 운전대를 잡은 손이 모두 오른손이며, 가슴 부분에서 정체불명의 팔이 추가로 돌출됨"
    ],
    "physics": "몸은 좌석에 지지되어 있으나, 핸들을 잡은 두 손의 구조와 팔의 연결이 물리적으로 성립하지 않음."
   },
   {
    "label": "B",
    "direction": "차량의 주행 방향과 인물의 시선이 숲을 향함.",
    "built_space": "캠핑카 내부 운전석. 조수석 시점의 앵글이며 운전대와 센터 콘솔이 정상 위치함.",
    "entities": "이현우 (짧은 검은 머리, 낡고 오염된 셔츠, 인이어 무전기 일치).",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 운전대의 위와 아래를 잡은 두 손이 모두 오른손임"
    ],
    "physics": "좌석에 앉아 하중을 지탱하나, 왼쪽에서 뻗어나온 팔에 오른손이 달려 있어 동작과 구조가 불가능함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "차량 내부 묘사와 구도는 양호하나, 운전대를 잡은 두 손이 모두 오른손으로 생성되는 치명적 해부학 오류로 인해 사용 불가."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물 외형은 일치하나, 양손이 모두 오른손이고 가슴에서 추가적인 팔이 나오는 등 심각한 신체 구조 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "차량과 인물의 시선이 숲 방향의 흙길을 향함.",
        "built_space": "캠핑카 운전석 내부. 조수석 뒤편 시점이며 운전대와 대시보드가 배치됨.",
        "entities": "이현우 (짧은 검은 머리, 흙 묻은 셔츠, 인이어 무전기 일치).",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 운전대를 잡은 손이 모두 오른손이며, 가슴 부분에서 정체불명의 팔이 추가로 돌출됨"
        ],
        "physics": "몸은 좌석에 지지되어 있으나, 핸들을 잡은 두 손의 구조와 팔의 연결이 물리적으로 성립하지 않음."
       },
       {
        "label": "B",
        "direction": "차량의 주행 방향과 인물의 시선이 숲을 향함.",
        "built_space": "캠핑카 내부 운전석. 조수석 시점의 앵글이며 운전대와 센터 콘솔이 정상 위치함.",
        "entities": "이현우 (짧은 검은 머리, 낡고 오염된 셔츠, 인이어 무전기 일치).",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 운전대의 위와 아래를 잡은 두 손이 모두 오른손임"
        ],
        "physics": "좌석에 앉아 하중을 지탱하나, 왼쪽에서 뻗어나온 팔에 오른손이 달려 있어 동작과 구조가 불가능함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "차량 내부 묘사와 구도는 양호하나, 운전대를 잡은 두 손이 모두 오른손으로 생성되는 치명적 해부학 오류로 인해 사용 불가."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물 외형은 일치하나, 양손이 모두 오른손이고 가슴에서 추가적인 팔이 나오는 등 심각한 신체 구조 오류가 발생함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "차량과 인물의 시선이 숲 방향의 흙길을 향함.",
        "built_space": "캠핑카 운전석 내부. 조수석 뒤편 시점이며 운전대와 대시보드가 배치됨.",
        "entities": "이현우 (짧은 검은 머리, 흙 묻은 셔츠, 인이어 무전기 일치).",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 운전대를 잡은 손이 모두 오른손이며, 가슴 부분에서 정체불명의 팔이 추가로 돌출됨"
        ],
        "physics": "몸은 좌석에 지지되어 있으나, 핸들을 잡은 두 손의 구조와 팔의 연결이 물리적으로 성립하지 않음."
       },
       {
        "label": "B",
        "direction": "차량의 주행 방향과 인물의 시선이 숲을 향함.",
        "built_space": "캠핑카 내부 운전석. 조수석 시점의 앵글이며 운전대와 센터 콘솔이 정상 위치함.",
        "entities": "이현우 (짧은 검은 머리, 낡고 오염된 셔츠, 인이어 무전기 일치).",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 운전대의 위와 아래를 잡은 두 손이 모두 오른손임"
        ],
        "physics": "좌석에 앉아 하중을 지탱하나, 왼쪽에서 뻗어나온 팔에 오른손이 달려 있어 동작과 구조가 불가능함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "젖은 운전자와 숲을 향한 전방 관계는 맞지만, 손 배치가 통상적인 조향처럼 보여 급격히 꺾는 순간이 약하고 카메라도 지정된 낮은 상향 시점보다 높다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "교차한 팔과 돌아간 운전대가 힘을 실어 급조향하는 순간을 더 명확히 구현하지만, 낮은 가슴 높이에서 비스듬히 올려다보는 지정 시점은 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 일부 보이는 옆얼굴은 오른쪽 전방의 앞유리와 숲을 향한다. 눈이 가려져 정확한 주시점은 확인할 수 없다. 양손은 운전대 위쪽과 오른쪽 아래를 잡고 있으나, 손과 팔의 관계만으로는 숲 쪽으로 확 꺾은 상태가 뚜렷하지 않다.",
        "built_space": "왼쪽에 운전석 등받이 하나, 운전자 앞에 운전대 하나와 계기판 한 묶음, 전면에 앞유리 하나가 있다. 오른쪽 대시보드에는 화면 하나, 보이는 송풍구 두 곳과 변속 레버 하나가 있다. 운전자는 등받이 앞에서 계기판과 운전대를 마주하며 좌측 운전석 배치에 부합한다. 다만 카메라는 조수석 쪽 뒤편의 어깨 부근 높이에서 내려다보는 인상으로, 낮은 가슴 높이의 상향 구도와 다르다. 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 한 명이며, 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보인다. 얼굴 대부분이 가려져 참조 인물의 세부 얼굴과 정확한 나이는 확인하기 어렵다. 마른 체격, 소형 인이어 장치, 젖고 진흙 묻은 어두운 셔츠는 맞는다. 핏자국은 진흙과 분명히 구별되지 않는다. 바지는 일부만 보이고 붕대와 적재 물자는 프레임 밖이다. 앞유리 밖 숲과 흐린 낮빛은 요청에 부합한다.",
        "hard_violations": [],
        "physics": "몸은 운전석에 앉아 있으며 등 뒤의 등받이와 아래쪽 좌석 영역이 지지한다. 두 손은 운전대 테두리를 실제로 감싸 쥐고, 운전대는 조향축과 대시보드에 연결된다. 팔과 손목 자세는 가능한 조작 자세이나 급격한 회전에 필요한 비틀림과 힘의 표현은 비교적 약하다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "고개는 약간 숙인 채 앞유리 너머 숲길 쪽을 향하며 카메라를 보지 않는다. 눈은 보이지 않아 시선이 먼 진행 방향인지 가까운 노면인지는 확정할 수 없다. 양손이 운전대 오른쪽 위와 아래에 모이고 팔이 교차하여 큰 조향 동작이 읽힌다. 바깥에는 숲속으로 굽는 길이 보이지만, 농로에서 막 이탈하는 경계 자체는 명확하지 않다.",
        "built_space": "왼쪽 운전석 등받이 하나 앞에 인물이 앉고, 운전대 하나와 계기판 한 묶음이 정면에 있다. 앞유리 하나와 왼쪽 측면 창, 오른쪽 대시보드의 화면 하나, 뚜렷한 세로 송풍구 하나와 상단 환기구, 하단 레버 하나가 보인다. 운전대와 계기판은 운전자를 향해 정상 배치되어 있다. 조수석 쪽 후방 시점은 맞지만 계기판 윗면과 팔 윗면을 내려다보므로 지정된 낮은 상향 시점은 아니다. 중복 좌석이나 불가능한 반사는 없다.",
        "entities": "한 명의 젊은 동아시아계 남성이 보이며 검은 헝클어진 머리, 마른 체격, 귀의 소형 인이어 장치가 요청과 맞는다. 참조처럼 어두운 낡은 셔츠를 입었고 어깨의 적갈색 얼룩과 진흙, 젖은 머리와 피부가 보인다. 가려진 얼굴 때문에 동일 인물의 얼굴 특징과 10대 후반 여부는 제한적으로만 판단할 수 있다. 바지는 조금 보이며 붕대와 물자는 구도 밖이라 판단 대상이 아니다. 숲과 비에 젖은 앞유리, 절제된 낮빛이 유지된다.",
        "hard_violations": [],
        "physics": "운전석이 앉은 몸을 지지하고, 앞으로 기울인 상체에서 두 팔이 운전대로 이어진다. 양손 모두 테두리에 접촉해 잡고 있으며 교차한 팔은 운전대를 크게 돌리는 과정에서 가능한 자세다. 전완의 긴장과 기울어진 운전대 중심부가 급조향 동작을 뒷받침한다. 운전대와 레버는 차량 구조에 고정되어 있고 지지 없이 떠 있는 요소는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "젖은 운전자와 숲을 향한 전방 관계는 맞지만, 손 배치가 통상적인 조향처럼 보여 급격히 꺾는 순간이 약하고 카메라도 지정된 낮은 상향 시점보다 높다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "교차한 팔과 돌아간 운전대가 힘을 실어 급조향하는 순간을 더 명확히 구현하지만, 낮은 가슴 높이에서 비스듬히 올려다보는 지정 시점은 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개와 일부 보이는 옆얼굴은 오른쪽 전방의 앞유리와 숲을 향한다. 눈이 가려져 정확한 주시점은 확인할 수 없다. 양손은 운전대 위쪽과 오른쪽 아래를 잡고 있으나, 손과 팔의 관계만으로는 숲 쪽으로 확 꺾은 상태가 뚜렷하지 않다.",
        "built_space": "왼쪽에 운전석 등받이 하나, 운전자 앞에 운전대 하나와 계기판 한 묶음, 전면에 앞유리 하나가 있다. 오른쪽 대시보드에는 화면 하나, 보이는 송풍구 두 곳과 변속 레버 하나가 있다. 운전자는 등받이 앞에서 계기판과 운전대를 마주하며 좌측 운전석 배치에 부합한다. 다만 카메라는 조수석 쪽 뒤편의 어깨 부근 높이에서 내려다보는 인상으로, 낮은 가슴 높이의 상향 구도와 다르다. 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 한 명이며, 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보인다. 얼굴 대부분이 가려져 참조 인물의 세부 얼굴과 정확한 나이는 확인하기 어렵다. 마른 체격, 소형 인이어 장치, 젖고 진흙 묻은 어두운 셔츠는 맞는다. 핏자국은 진흙과 분명히 구별되지 않는다. 바지는 일부만 보이고 붕대와 적재 물자는 프레임 밖이다. 앞유리 밖 숲과 흐린 낮빛은 요청에 부합한다.",
        "hard_violations": [],
        "physics": "몸은 운전석에 앉아 있으며 등 뒤의 등받이와 아래쪽 좌석 영역이 지지한다. 두 손은 운전대 테두리를 실제로 감싸 쥐고, 운전대는 조향축과 대시보드에 연결된다. 팔과 손목 자세는 가능한 조작 자세이나 급격한 회전에 필요한 비틀림과 힘의 표현은 비교적 약하다. 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "고개는 약간 숙인 채 앞유리 너머 숲길 쪽을 향하며 카메라를 보지 않는다. 눈은 보이지 않아 시선이 먼 진행 방향인지 가까운 노면인지는 확정할 수 없다. 양손이 운전대 오른쪽 위와 아래에 모이고 팔이 교차하여 큰 조향 동작이 읽힌다. 바깥에는 숲속으로 굽는 길이 보이지만, 농로에서 막 이탈하는 경계 자체는 명확하지 않다.",
        "built_space": "왼쪽 운전석 등받이 하나 앞에 인물이 앉고, 운전대 하나와 계기판 한 묶음이 정면에 있다. 앞유리 하나와 왼쪽 측면 창, 오른쪽 대시보드의 화면 하나, 뚜렷한 세로 송풍구 하나와 상단 환기구, 하단 레버 하나가 보인다. 운전대와 계기판은 운전자를 향해 정상 배치되어 있다. 조수석 쪽 후방 시점은 맞지만 계기판 윗면과 팔 윗면을 내려다보므로 지정된 낮은 상향 시점은 아니다. 중복 좌석이나 불가능한 반사는 없다.",
        "entities": "한 명의 젊은 동아시아계 남성이 보이며 검은 헝클어진 머리, 마른 체격, 귀의 소형 인이어 장치가 요청과 맞는다. 참조처럼 어두운 낡은 셔츠를 입었고 어깨의 적갈색 얼룩과 진흙, 젖은 머리와 피부가 보인다. 가려진 얼굴 때문에 동일 인물의 얼굴 특징과 10대 후반 여부는 제한적으로만 판단할 수 있다. 바지는 조금 보이며 붕대와 물자는 구도 밖이라 판단 대상이 아니다. 숲과 비에 젖은 앞유리, 절제된 낮빛이 유지된다.",
        "hard_violations": [],
        "physics": "운전석이 앉은 몸을 지지하고, 앞으로 기울인 상체에서 두 팔이 운전대로 이어진다. 양손 모두 테두리에 접촉해 잡고 있으며 교차한 팔은 운전대를 크게 돌리는 과정에서 가능한 자세다. 전완의 긴장과 기울어진 운전대 중심부가 급조향 동작을 뒷받침한다. 운전대와 레버는 차량 구조에 고정되어 있고 지지 없이 떠 있는 요소는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.607
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 운전대를 잡은 손이 모두 오른손이며, 가슴 부분에서 정체불명의 팔이 추가로 돌출됨"
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 운전대의 위와 아래를 잡은 두 손이 모두 오른손임"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1607,
   "A": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1607,
    "verdict_ko": "차량 내부 묘사와 구도는 양호하나, 운전대를 잡은 두 손이 모두 오른손으로 생성되는 치명적 해부학 오류로 인해 사용 불가.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 운전대의 위와 아래를 잡은 두 손이 모두 오른손임"
   },
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "인물 외형은 일치하나, 양손이 모두 오른손이고 가슴에서 추가적인 팔이 나오는 등 심각한 신체 구조 오류가 발생함.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 운전대를 잡은 손이 모두 오른손이며, 가슴 부분에서 정체불명의 팔이 추가로 돌출됨"
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S55sh6_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bc1-5363-7cfb-bffb-b38305d5d3ee",
  "confined_fp": {
   "base_key": "confinedfp::085f743cd061",
   "apt_reason": "캠핑카 운전석 내부에서 운전대를 급격히 조작하는 숏으로, 운전석의 위치와 스티어링 휠 등 차량 제어 장치의 공간적 배치, 그리고 창밖으로 향하는 숲의 방향이 시각적으로 정확하게 일치해야 하므로 평면도 보조가 필요합니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh8::signage": {
  "fp": "f879c47c84464f98",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh8": {
  "input_fingerprint": "f0a42dee2dbc0587",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙바닥의 숲길 위에 멈춰 서 있는 캠핑카의 외부 구도.\n\nLOCATION (lock): On a rough dirt track within the forest, at the camper's stopping point near the marsh approach. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use an explicit spatial and temporal cut after the mountain-road drive, arriving at a static elevated rear three-quarter view from behind and beside the stopped camper. Place the whole vehicle in the lower-left middle distance, occupying less than a third of the image, with the rough earth road continuing toward the upper right and forest providing its spatial context. The increased camera distance supplies the principal visual release, and no occupants are discernible.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 캠핑카 (Stopped on the rough forest road) — Rear and one side are visible, with its front directed along the road's continuation; used as Provides the stationary narrative anchor at environmental scale; 숲길 (Rough earth road extending ahead of the stopped camper) — Runs diagonally from the vehicle toward the upper right; used as Makes the camper's heading and halted progress readable; 숲 (Surrounds the road); used as Supplies scale and encloses the route without obscuring its continuation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and controlled contrast hold the stopped vehicle and forest road together without adding rain, mist, or a new weather condition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): After the shower, the camper has reached the muddy end of the forest approach at the swamp; it is not on a safely cleared stopping area. The packed luggage, food and medicine remain aboard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙바닥의 숲길 위에 멈춰 서 있는 캠핑카의 외부 구도.\n\nLOCATION (lock): On a rough dirt track within the forest, at the camper's stopping point near the marsh approach. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use an explicit spatial and temporal cut after the mountain-road drive, arriving at a static elevated rear three-quarter view from behind and beside the stopped camper. Place the whole vehicle in the lower-left middle distance, occupying less than a third of the image, with the rough earth road continuing toward the upper right and forest providing its spatial context. The increased camera distance supplies the principal visual release, and no occupants are discernible.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 캠핑카 (Stopped on the rough forest road) — Rear and one side are visible, with its front directed along the road's continuation; used as Provides the stationary narrative anchor at environmental scale; 숲길 (Rough earth road extending ahead of the stopped camper) — Runs diagonally from the vehicle toward the upper right; used as Makes the camper's heading and halted progress readable; 숲 (Surrounds the road); used as Supplies scale and encloses the route without obscuring its continuation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and controlled contrast hold the stopped vehicle and forest road together without adding rain, mist, or a new weather condition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): After the shower, the camper has reached the muddy end of the forest approach at the swamp; it is not on a safely cleared stopping area. The packed luggage, food and medicine remain aboard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 흙바닥의 숲길 위에 멈춰 서 있는 캠핑카의 외부 구도.\n\nLOCATION (lock): On a rough dirt track within the forest, at the camper's stopping point near the marsh approach. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Use an explicit spatial and temporal cut after the mountain-road drive, arriving at a static elevated rear three-quarter view from behind and beside the stopped camper. Place the whole vehicle in the lower-left middle distance, occupying less than a third of the image, with the rough earth road continuing toward the upper right and forest providing its spatial context. The increased camera distance supplies the principal visual release, and no occupants are discernible.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 캠핑카 (Stopped on the rough forest road) — Rear and one side are visible, with its front directed along the road's continuation; used as Provides the stationary narrative anchor at environmental scale; 숲길 (Rough earth road extending ahead of the stopped camper) — Runs diagonally from the vehicle toward the upper right; used as Makes the camper's heading and halted progress readable; 숲 (Surrounds the road); used as Supplies scale and encloses the route without obscuring its continuation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient light and controlled contrast hold the stopped vehicle and forest road together without adding rain, mist, or a new weather condition.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): After the shower, the camper has reached the muddy end of the forest approach at the swamp; it is not on a safely cleared stopping area. The packed luggage, food and medicine remain aboard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캠핑카의 전면부가 화면 우측 상단으로 이어지는 길의 방향을 향해 정렬되어 있음.",
    "built_space": "숲으로 둘러싸인 거친 흙길 위에 캠핑카가 정차해 있으며, 카메라는 캠핑카의 좌측 후방의 약간 높은 곳에 위치함.",
    "entities": "지시된 캠핑카, 숲길, 숲이 모두 존재함. 요구사항대로 프레임 내에 인물은 전혀 보이지 않음. 다만 흙길이 비 온 뒤의 상태로 보이지 않음.",
    "hard_violations": [],
    "physics": "캠핑카가 흙길 지면 위에 물리적으로 자연스럽게 안착해 있음."
   },
   {
    "label": "B",
    "direction": "캠핑카의 머리 부분이 우측 상단으로 뻗은 길을 따라 향하고 있음.",
    "built_space": "수목이 우거진 숲길에 캠핑카가 놓여 있으며, 카메라는 캠핑카 뒤쪽 좌측의 높은 시점을 확보함.",
    "entities": "지붕과 후면에 짐이 실린 캠핑카, 물웅덩이가 고인 진흙길, 숲이 정확히 표현됨. 화면 내에 인물은 존재하지 않음.",
    "hard_violations": [],
    "physics": "캠핑카 바퀴가 지면에 안정적으로 닿아 있으며, 바닥의 물웅덩이가 빛과 주변 환경을 자연스럽게 반사하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 카메라 구도와 화면 배치는 훌륭하게 구현했으나, '소나기가 내린 후 진흙길'이라는 핵심 상태를 누락하고 건조한 흙길로 묘사하여 다소 아쉽습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 앵글과 피사체의 화면 비율을 정확히 지켰으며, 소나기가 지나간 후의 물웅덩이와 진흙길의 질감까지 완벽하게 표현하여 프롬프트의 상황을 가장 충실히 재현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 전면부가 화면 우측 상단으로 이어지는 길의 방향을 향해 정렬되어 있음.",
        "built_space": "숲으로 둘러싸인 거친 흙길 위에 캠핑카가 정차해 있으며, 카메라는 캠핑카의 좌측 후방의 약간 높은 곳에 위치함.",
        "entities": "지시된 캠핑카, 숲길, 숲이 모두 존재함. 요구사항대로 프레임 내에 인물은 전혀 보이지 않음. 다만 흙길이 비 온 뒤의 상태로 보이지 않음.",
        "hard_violations": [],
        "physics": "캠핑카가 흙길 지면 위에 물리적으로 자연스럽게 안착해 있음."
       },
       {
        "label": "B",
        "direction": "캠핑카의 머리 부분이 우측 상단으로 뻗은 길을 따라 향하고 있음.",
        "built_space": "수목이 우거진 숲길에 캠핑카가 놓여 있으며, 카메라는 캠핑카 뒤쪽 좌측의 높은 시점을 확보함.",
        "entities": "지붕과 후면에 짐이 실린 캠핑카, 물웅덩이가 고인 진흙길, 숲이 정확히 표현됨. 화면 내에 인물은 존재하지 않음.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 지면에 안정적으로 닿아 있으며, 바닥의 물웅덩이가 빛과 주변 환경을 자연스럽게 반사하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지정된 카메라 구도와 화면 배치는 훌륭하게 구현했으나, '소나기가 내린 후 진흙길'이라는 핵심 상태를 누락하고 건조한 흙길로 묘사하여 다소 아쉽습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 앵글과 피사체의 화면 비율을 정확히 지켰으며, 소나기가 지나간 후의 물웅덩이와 진흙길의 질감까지 완벽하게 표현하여 프롬프트의 상황을 가장 충실히 재현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 전면부가 화면 우측 상단으로 이어지는 길의 방향을 향해 정렬되어 있음.",
        "built_space": "숲으로 둘러싸인 거친 흙길 위에 캠핑카가 정차해 있으며, 카메라는 캠핑카의 좌측 후방의 약간 높은 곳에 위치함.",
        "entities": "지시된 캠핑카, 숲길, 숲이 모두 존재함. 요구사항대로 프레임 내에 인물은 전혀 보이지 않음. 다만 흙길이 비 온 뒤의 상태로 보이지 않음.",
        "hard_violations": [],
        "physics": "캠핑카가 흙길 지면 위에 물리적으로 자연스럽게 안착해 있음."
       },
       {
        "label": "B",
        "direction": "캠핑카의 머리 부분이 우측 상단으로 뻗은 길을 따라 향하고 있음.",
        "built_space": "수목이 우거진 숲길에 캠핑카가 놓여 있으며, 카메라는 캠핑카 뒤쪽 좌측의 높은 시점을 확보함.",
        "entities": "지붕과 후면에 짐이 실린 캠핑카, 물웅덩이가 고인 진흙길, 숲이 정확히 표현됨. 화면 내에 인물은 존재하지 않음.",
        "hard_violations": [],
        "physics": "캠핑카 바퀴가 지면에 안정적으로 닿아 있으며, 바닥의 물웅덩이가 빛과 주변 환경을 자연스럽게 반사하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "높은 후측면 와이드 구도와 우상단으로 이어지는 길을 충족하며, 진흙과 물웅덩이가 소나기 뒤 습지 접근로의 정차 상태를 더 충실하게 보여준다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "차량의 좌하단 배치와 작은 화면 점유율은 정확하지만, 길이 상대적으로 마르고 돌이 많아 진흙투성이 습지 접근 지점이라는 상태가 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캠핑카의 뒤와 오른쪽 측면이 보이고, 앞머리는 화면 우상단으로 이어지는 흙길을 향한다. 차량의 진행 방향과 길의 연장이 일치한다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "건축물 없이 숲으로 둘러싸인 흙길 하나와 캠핑카 한 대가 있다. 카메라는 차량 뒤쪽 옆의 높은 위치에서 내려다보며, 차량 전체는 좌하단 중경에 화면 면적의 3분의 1보다 작게 들어온다. 후면 사다리 하나와 후면 적재대 하나가 보이며, 길의 연장은 나무에 완전히 가려지지 않는다.",
        "entities": "흰색 캠핑카 한 대, 거친 흙길, 주변 숲이 모두 식별된다. 바퀴 자국과 여러 물웅덩이가 있어 젖고 정비되지 않은 접근로로 읽힌다. 후면에는 덮개를 씌운 적재물이 있지만 실내의 짐·식량·약품은 보이지 않아 확인할 수 없다. 사람이나 식별 가능한 탑승자는 없고, 비나 안개 없는 낮이다. 습지 자체는 화면에 드러나지 않는다.",
        "hard_violations": [],
        "physics": "보이는 앞뒤 바퀴가 흙길에 닿아 차체를 지지한다. 후면 적재물은 적재대와 결박끈으로 고정되어 있고 지붕 장치는 차체에 부착되어 있다. 물은 낮은 바퀴 자국에 고여 있으며, 떠 있는 물체나 불가능한 자세는 없다. 움직임을 나타내는 먼지나 흐림 없이 정차 상태로 읽힌다."
       },
       {
        "label": "B",
        "direction": "캠핑카 후면과 오른쪽 측면이 보이며, 앞머리는 숲길이 이어지는 우상단을 향한다. 길은 차량 앞에서 같은 방향으로 계속된다. 사람의 시선이나 다른 지향성 행동은 없다.",
        "built_space": "숲속 흙길 하나에 캠핑카 한 대가 있고 별도 건축물이나 정차장은 없다. 높은 후측면 시점에서 차량 전체가 좌하단 중경에 작게 배치된다. 후면 창 하나, 사다리 하나, 적재대 하나가 보인다. 숲은 길 양옆을 둘러싸지만 우상단의 연속된 노면을 가리지 않는다.",
        "entities": "흰색 캠핑카, 돌과 패인 자국이 있는 흙길, 울창한 숲이 보인다. 후면 적재대에는 상자형 적재물 두 개가 있다. 실내의 짐·식량·약품은 확인할 수 없으며 사람이나 얼굴도 보이지 않는다. 낮의 자연광이고 비나 안개는 없다. 일부 젖은 자국은 있지만 전체 노면은 비교적 건조하고, 습지 접근로의 진흙 상태는 약하게 표현된다.",
        "hard_violations": [],
        "physics": "차량은 보이는 바퀴로 노면에 지지되어 있다. 후면 상자들은 적재대 위에 놓여 고정되어 있고 지붕 설비도 차체에 붙어 있다. 공중에 뜬 물체나 불가능한 지지 관계는 없다. 바퀴와 차체에 이동 흔적이 없어 정차한 차량으로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "높은 후측면 와이드 구도와 우상단으로 이어지는 길을 충족하며, 진흙과 물웅덩이가 소나기 뒤 습지 접근로의 정차 상태를 더 충실하게 보여준다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "차량의 좌하단 배치와 작은 화면 점유율은 정확하지만, 길이 상대적으로 마르고 돌이 많아 진흙투성이 습지 접근 지점이라는 상태가 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캠핑카의 뒤와 오른쪽 측면이 보이고, 앞머리는 화면 우상단으로 이어지는 흙길을 향한다. 차량의 진행 방향과 길의 연장이 일치한다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "건축물 없이 숲으로 둘러싸인 흙길 하나와 캠핑카 한 대가 있다. 카메라는 차량 뒤쪽 옆의 높은 위치에서 내려다보며, 차량 전체는 좌하단 중경에 화면 면적의 3분의 1보다 작게 들어온다. 후면 사다리 하나와 후면 적재대 하나가 보이며, 길의 연장은 나무에 완전히 가려지지 않는다.",
        "entities": "흰색 캠핑카 한 대, 거친 흙길, 주변 숲이 모두 식별된다. 바퀴 자국과 여러 물웅덩이가 있어 젖고 정비되지 않은 접근로로 읽힌다. 후면에는 덮개를 씌운 적재물이 있지만 실내의 짐·식량·약품은 보이지 않아 확인할 수 없다. 사람이나 식별 가능한 탑승자는 없고, 비나 안개 없는 낮이다. 습지 자체는 화면에 드러나지 않는다.",
        "hard_violations": [],
        "physics": "보이는 앞뒤 바퀴가 흙길에 닿아 차체를 지지한다. 후면 적재물은 적재대와 결박끈으로 고정되어 있고 지붕 장치는 차체에 부착되어 있다. 물은 낮은 바퀴 자국에 고여 있으며, 떠 있는 물체나 불가능한 자세는 없다. 움직임을 나타내는 먼지나 흐림 없이 정차 상태로 읽힌다."
       },
       {
        "label": "A",
        "direction": "캠핑카 후면과 오른쪽 측면이 보이며, 앞머리는 숲길이 이어지는 우상단을 향한다. 길은 차량 앞에서 같은 방향으로 계속된다. 사람의 시선이나 다른 지향성 행동은 없다.",
        "built_space": "숲속 흙길 하나에 캠핑카 한 대가 있고 별도 건축물이나 정차장은 없다. 높은 후측면 시점에서 차량 전체가 좌하단 중경에 작게 배치된다. 후면 창 하나, 사다리 하나, 적재대 하나가 보인다. 숲은 길 양옆을 둘러싸지만 우상단의 연속된 노면을 가리지 않는다.",
        "entities": "흰색 캠핑카, 돌과 패인 자국이 있는 흙길, 울창한 숲이 보인다. 후면 적재대에는 상자형 적재물 두 개가 있다. 실내의 짐·식량·약품은 확인할 수 없으며 사람이나 얼굴도 보이지 않는다. 낮의 자연광이고 비나 안개는 없다. 일부 젖은 자국은 있지만 전체 노면은 비교적 건조하고, 습지 접근로의 진흙 상태는 약하게 표현된다.",
        "hard_violations": [],
        "physics": "차량은 보이는 바퀴로 노면에 지지되어 있다. 후면 상자들은 적재대 위에 놓여 고정되어 있고 지붕 설비도 차체에 붙어 있다. 공중에 뜬 물체나 불가능한 지지 관계는 없다. 바퀴와 차체에 이동 흔적이 없어 정차한 차량으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.603,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.603,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1603,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1603,
    "verdict_ko": "지정된 카메라 구도와 화면 배치는 훌륭하게 구현했으나, '소나기가 내린 후 진흙길'이라는 핵심 상태를 누락하고 건조한 흙길로 묘사하여 다소 아쉽습니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 앵글과 피사체의 화면 비율을 정확히 지켰으며, 소나기가 지나간 후의 물웅덩이와 진흙길의 질감까지 완벽하게 표현하여 프롬프트의 상황을 가장 충실히 재현했습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bcb-def3-7e68-b4b8-b02926a58a31",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S56sh14::signage": {
  "fp": "831349f5beacf4d7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::0b09e901a4504185": {
  "subjects": [],
  "subject_text": "익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이\n진흙과 물이 고인 넓은 늪지대. 부식된 출입 금지·방사능 표지판이 서 있고, 인접한 숲에는 나무와 꽃, 땅에 파인 함정 구덩이가 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L179",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::forest_capture_site": {
  "input_fingerprint": "99340b53ed7fbb00",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "forest_capture_site",
    "tags": [
     "S56sh14"
    ]
   },
   "context_sig": "1de0938ec48dd702"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /숲 길 -D\n- 그것을 올려다보는 순간 화면을 덮치는 그물망!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /숲 길 -D\n- 그것을 올려다보는 순간 화면을 덮치는 그물망!\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_capture_site_4efc3c.png",
  "asset_id": "50187464-125b-40d8-9db9-f8b1b64944a6",
  "input_asset_ids": [
   "1bbaa650-796c-4df1-9085-4a0a8af10bd3"
  ],
  "origin_tag": "S56sh14",
  "place_text": "On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.",
  "origin_inputs": {
   "place_text": "On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.",
   "time_of_day_en": "day",
   "conti_asset_id": "1bbaa650-796c-4df1-9085-4a0a8af10bd3"
  }
 },
 "S56sh14::bgfirst_bg": {
  "input_fingerprint": "3cd245a90757eafd",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하늘에서 떨어진 거대한 그물망이 찰리의 금속 몸체를 완전히 덮기 직전 공중에 넓게 펼쳐져 있는 역동적인 찰나.\n\nLOCATION (lock): On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the upward tilt from 찰리's waist-height front-quarter position, remaining outside his forward axis and holding his head, raised hand, and upper torso beneath the spreading net. Place 찰리 in the lower-left half with his face lifted after the ascending butterfly, while the descending net stretches across the upper third without yet obscuring him. Keep the forest visible around both forms so his lingering, gentle reach and the sudden threat read in the same direct observational image.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: descending net spread above 찰리, not yet covering his face in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 그물망 (Spread in midair immediately before descending over 찰리) — Its underside is visible above 찰리, with the spread crossing the upper third; used as Creates the impending enclosure while leaving his face and gesture exposed; 나비 (Ascending after leaving the flower); used as Remains a small visual endpoint above 찰리's raised gaze; 숲의 나무 (Surround 찰리's location on the forest path); used as Maintain environmental scale around the robot and airborne net.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves precise metal-body articulation and readable net strands, with controlled contrast rather than a dramatic lighting change announcing the trap.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하늘에서 떨어진 거대한 그물망이 찰리의 금속 몸체를 완전히 덮기 직전 공중에 넓게 펼쳐져 있는 역동적인 찰나.\n\nLOCATION (lock): On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the upward tilt from 찰리's waist-height front-quarter position, remaining outside his forward axis and holding his head, raised hand, and upper torso beneath the spreading net. Place 찰리 in the lower-left half with his face lifted after the ascending butterfly, while the descending net stretches across the upper third without yet obscuring him. Keep the forest visible around both forms so his lingering, gentle reach and the sudden threat read in the same direct observational image.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: descending net spread above 찰리, not yet covering his face in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 그물망 (Spread in midair immediately before descending over 찰리) — Its underside is visible above 찰리, with the spread crossing the upper third; used as Creates the impending enclosure while leaving his face and gesture exposed; 나비 (Ascending after leaving the flower); used as Remains a small visual endpoint above 찰리's raised gaze; 숲의 나무 (Surround 찰리's location on the forest path); used as Maintain environmental scale around the robot and airborne net.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves precise metal-body articulation and readable net strands, with controlled contrast rather than a dramatic lighting change announcing the trap.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh14__bgfirst_bg.png",
  "asset_id": "07be09ef-d192-4906-bbda-7d5864eb163f",
  "input_asset_ids": [
   "1bbaa650-796c-4df1-9085-4a0a8af10bd3",
   "50187464-125b-40d8-9db9-f8b1b64944a6"
  ]
 },
 "S56sh14": {
  "input_fingerprint": "c82296bcad49baa5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하늘에서 떨어진 거대한 그물망이 찰리의 금속 몸체를 완전히 덮기 직전 공중에 넓게 펼쳐져 있는 역동적인 찰나.\n\nLOCATION (lock): On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the upward tilt from 찰리's waist-height front-quarter position, remaining outside his forward axis and holding his head, raised hand, and upper torso beneath the spreading net. Place 찰리 in the lower-left half with his face lifted after the ascending butterfly, while the descending net stretches across the upper third without yet obscuring him. Keep the forest visible around both forms so his lingering, gentle reach and the sudden threat read in the same direct observational image.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: descending net spread above 찰리, not yet covering his face in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 그물망 (Spread in midair immediately before descending over 찰리) — Its underside is visible above 찰리, with the spread crossing the upper third; used as Creates the impending enclosure while leaving his face and gesture exposed; 나비 (Ascending after leaving the flower); used as Remains a small visual endpoint above 찰리's raised gaze; 숲의 나무 (Surround 찰리's location on the forest path); used as Maintain environmental scale around the robot and airborne net.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves precise metal-body articulation and readable net strands, with controlled contrast rather than a dramatic lighting change announcing the trap.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is stuck in the swamp near oversized bootprints, a no-entry sign and a corroded radiation-zone marker. In the forest, a butterfly has just lifted from a flower as a capture net descends. 찰리: He wears the oversized straw hat, boots and colorful raincoat, with his shoulder repaired. He is looking upward after reaching toward a butterfly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하늘에서 떨어진 거대한 그물망이 찰리의 금속 몸체를 완전히 덮기 직전 공중에 넓게 펼쳐져 있는 역동적인 찰나.\n\nLOCATION (lock): On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the upward tilt from 찰리's waist-height front-quarter position, remaining outside his forward axis and holding his head, raised hand, and upper torso beneath the spreading net. Place 찰리 in the lower-left half with his face lifted after the ascending butterfly, while the descending net stretches across the upper third without yet obscuring him. Keep the forest visible around both forms so his lingering, gentle reach and the sudden threat read in the same direct observational image.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: descending net spread above 찰리, not yet covering his face in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 그물망 (Spread in midair immediately before descending over 찰리) — Its underside is visible above 찰리, with the spread crossing the upper third; used as Creates the impending enclosure while leaving his face and gesture exposed; 나비 (Ascending after leaving the flower); used as Remains a small visual endpoint above 찰리's raised gaze; 숲의 나무 (Surround 찰리's location on the forest path); used as Maintain environmental scale around the robot and airborne net.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves precise metal-body articulation and readable net strands, with controlled contrast rather than a dramatic lighting change announcing the trap.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is stuck in the swamp near oversized bootprints, a no-entry sign and a corroded radiation-zone marker. In the forest, a butterfly has just lifted from a flower as a capture net descends. 찰리: He wears the oversized straw hat, boots and colorful raincoat, with his shoulder repaired. He is looking upward after reaching toward a butterfly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하늘에서 떨어진 거대한 그물망이 찰리의 금속 몸체를 완전히 덮기 직전 공중에 넓게 펼쳐져 있는 역동적인 찰나.\n\nLOCATION (lock): On a woodland path near the contaminated marsh, beside the flowers where the robot stopped to watch a butterfly. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the upward tilt from 찰리's waist-height front-quarter position, remaining outside his forward axis and holding his head, raised hand, and upper torso beneath the spreading net. Place 찰리 in the lower-left half with his face lifted after the ascending butterfly, while the descending net stretches across the upper third without yet obscuring him. Keep the forest visible around both forms so his lingering, gentle reach and the sudden threat read in the same direct observational image.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: descending net spread above 찰리, not yet covering his face in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 그물망 (Spread in midair immediately before descending over 찰리) — Its underside is visible above 찰리, with the spread crossing the upper third; used as Creates the impending enclosure while leaving his face and gesture exposed; 나비 (Ascending after leaving the flower); used as Remains a small visual endpoint above 찰리's raised gaze; 숲의 나무 (Surround 찰리's location on the forest path); used as Maintain environmental scale around the robot and airborne net.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime illumination preserves precise metal-body articulation and readable net strands, with controlled contrast rather than a dramatic lighting change announcing the trap.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper is stuck in the swamp near oversized bootprints, a no-entry sign and a corroded radiation-zone marker. In the forest, a butterfly has just lifted from a flower as a capture net descends. 찰리: He wears the oversized straw hat, boots and colorful raincoat, with his shoulder repaired. He is looking upward after reaching toward a butterfly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh14__bgfirst_bg.png",
     "asset_id": "07be09ef-d192-4906-bbda-7d5864eb163f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S56sh14.png",
     "asset_id": "1bbaa650-796c-4df1-9085-4a0a8af10bd3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_capture_site_4efc3c.png",
     "asset_id": "50187464-125b-40d8-9db9-f8b1b64944a6",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 얼굴과 뻗은 왼손이 하늘에 있는 나비를 향해 있습니다.",
    "built_space": "레퍼런스 사진과 동일한 늪지대 숲길로, 방사능 표지판, 진흙 바닥, 상단의 그물망 등 환경적 요소들이 지정된 앵글(로우 앵글)에 맞게 잘 배치되었습니다.",
    "entities": "찰리는 짚모자와 우비를 착용하고 있으나, 하얀 얼굴 마스크의 형태가 일그러져 눈이 하나만 보이고 가슴의 푸른 심볼 위치가 측면으로 치우쳐 있어 캐릭터 레퍼런스와의 일치도가 떨어집니다. 나비와 그물망은 존재합니다.",
    "hard_violations": [],
    "physics": "찰리가 바닥을 딛고 서 있는 자세와 팔을 뻗은 동작, 공중을 나는 나비와 매달린 그물망의 상태 모두 중력 및 물리적 조건에 부합합니다."
   },
   {
    "label": "B",
    "direction": "찰리의 고개와 시선, 그리고 위로 뻗은 왼손이 정확히 공중에 떠 있는 나비를 향하고 있습니다.",
    "built_space": "진흙투성이의 숲길 늪지대 배경으로, 왼쪽 나무에 매달린 방사능 경고판과 철조망이 레퍼런스 사진과 완벽히 일치합니다. 화면 상단 3분의 1 지점에 그물망이 넓게 펼쳐져 있으며, 카메라 앵글은 찰리의 허리 높이에서 위로 올려다보는 구도를 정확히 따릅니다.",
    "entities": "찰리는 짚모자와 화려한 다채색 우비를 입고 있으며, 캐릭터 레퍼런스와 동일한 귀여운 하얀 마스크(주황색 두 눈과 입 선)와 가슴 중앙의 푸른 원자로를 정확히 갖추고 있습니다. 위쪽에는 나비와 그물망이 명확히 존재합니다.",
    "hard_violations": [],
    "physics": "찰리는 진흙 바닥에 안정적으로 두 발을 지지하고 서 있으며, 위로 뻗은 팔과 날아가는 나비, 공중에 펼쳐진 그물망 모두 물리적으로 자연스럽게 표현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "후보 B는 찰리의 하얀 마스크 얼굴(두 눈과 입 선)과 가슴의 원자로 심볼을 캐릭터 레퍼런스와 정확하게 일치시켰으며, 요청된 로우 앵글과 역동적인 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 배경은 잘 반영되었으나, 찰리의 얼굴 마스크가 심하게 왜곡되어 캐릭터 레퍼런스의 특징(두 눈과 입 선)을 잃었으며 가슴의 원자로 위치도 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 고개와 시선, 그리고 위로 뻗은 왼손이 정확히 공중에 떠 있는 나비를 향하고 있습니다.",
        "built_space": "진흙투성이의 숲길 늪지대 배경으로, 왼쪽 나무에 매달린 방사능 경고판과 철조망이 레퍼런스 사진과 완벽히 일치합니다. 화면 상단 3분의 1 지점에 그물망이 넓게 펼쳐져 있으며, 카메라 앵글은 찰리의 허리 높이에서 위로 올려다보는 구도를 정확히 따릅니다.",
        "entities": "찰리는 짚모자와 화려한 다채색 우비를 입고 있으며, 캐릭터 레퍼런스와 동일한 귀여운 하얀 마스크(주황색 두 눈과 입 선)와 가슴 중앙의 푸른 원자로를 정확히 갖추고 있습니다. 위쪽에는 나비와 그물망이 명확히 존재합니다.",
        "hard_violations": [],
        "physics": "찰리는 진흙 바닥에 안정적으로 두 발을 지지하고 서 있으며, 위로 뻗은 팔과 날아가는 나비, 공중에 펼쳐진 그물망 모두 물리적으로 자연스럽게 표현되었습니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 뻗은 왼손이 하늘에 있는 나비를 향해 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 늪지대 숲길로, 방사능 표지판, 진흙 바닥, 상단의 그물망 등 환경적 요소들이 지정된 앵글(로우 앵글)에 맞게 잘 배치되었습니다.",
        "entities": "찰리는 짚모자와 우비를 착용하고 있으나, 하얀 얼굴 마스크의 형태가 일그러져 눈이 하나만 보이고 가슴의 푸른 심볼 위치가 측면으로 치우쳐 있어 캐릭터 레퍼런스와의 일치도가 떨어집니다. 나비와 그물망은 존재합니다.",
        "hard_violations": [],
        "physics": "찰리가 바닥을 딛고 서 있는 자세와 팔을 뻗은 동작, 공중을 나는 나비와 매달린 그물망의 상태 모두 중력 및 물리적 조건에 부합합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "후보 B는 찰리의 하얀 마스크 얼굴(두 눈과 입 선)과 가슴의 원자로 심볼을 캐릭터 레퍼런스와 정확하게 일치시켰으며, 요청된 로우 앵글과 역동적인 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 배경은 잘 반영되었으나, 찰리의 얼굴 마스크가 심하게 왜곡되어 캐릭터 레퍼런스의 특징(두 눈과 입 선)을 잃었으며 가슴의 원자로 위치도 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 고개와 시선, 그리고 위로 뻗은 왼손이 정확히 공중에 떠 있는 나비를 향하고 있습니다.",
        "built_space": "진흙투성이의 숲길 늪지대 배경으로, 왼쪽 나무에 매달린 방사능 경고판과 철조망이 레퍼런스 사진과 완벽히 일치합니다. 화면 상단 3분의 1 지점에 그물망이 넓게 펼쳐져 있으며, 카메라 앵글은 찰리의 허리 높이에서 위로 올려다보는 구도를 정확히 따릅니다.",
        "entities": "찰리는 짚모자와 화려한 다채색 우비를 입고 있으며, 캐릭터 레퍼런스와 동일한 귀여운 하얀 마스크(주황색 두 눈과 입 선)와 가슴 중앙의 푸른 원자로를 정확히 갖추고 있습니다. 위쪽에는 나비와 그물망이 명확히 존재합니다.",
        "hard_violations": [],
        "physics": "찰리는 진흙 바닥에 안정적으로 두 발을 지지하고 서 있으며, 위로 뻗은 팔과 날아가는 나비, 공중에 펼쳐진 그물망 모두 물리적으로 자연스럽게 표현되었습니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 뻗은 왼손이 하늘에 있는 나비를 향해 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 늪지대 숲길로, 방사능 표지판, 진흙 바닥, 상단의 그물망 등 환경적 요소들이 지정된 앵글(로우 앵글)에 맞게 잘 배치되었습니다.",
        "entities": "찰리는 짚모자와 우비를 착용하고 있으나, 하얀 얼굴 마스크의 형태가 일그러져 눈이 하나만 보이고 가슴의 푸른 심볼 위치가 측면으로 치우쳐 있어 캐릭터 레퍼런스와의 일치도가 떨어집니다. 나비와 그물망은 존재합니다.",
        "hard_violations": [],
        "physics": "찰리가 바닥을 딛고 서 있는 자세와 팔을 뻗은 동작, 공중을 나는 나비와 매달린 그물망의 상태 모두 중력 및 물리적 조건에 부합합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 얼굴과 뻗은 손을 더 크게 담은 전방 사선 구도가 우세하지만, 무릎까지 보이는 촬영 범위와 강한 직사광은 지정된 미디엄 숏·차분한 낮 조명에서 벗어난다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "나비를 향한 시선과 머리 위 그물의 배치는 맞지만, 장화까지 드러내는 넓은 구도와 측면에 가까운 얼굴 방향이 지정된 상반신 중심 전방 사선 미디엄 숏보다 멀다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 얼굴을 위쪽 오른편으로 들고 있으며, 시선과 펼친 손끝의 연장선 위에 작은 나비가 있다. 그물은 찰리 머리 위 화면 상단을 가로지르고 얼굴이나 손을 가리지 않는다. 나비의 비행 자세는 보이지만 상승 방향과 그물의 하강 속도는 정지 화면만으로 확정하기 어렵다.",
        "built_space": "왼쪽과 오른쪽의 굵은 나무 사이로 진흙길이 습지까지 이어진다. 왼쪽 나무 옆 방사능 표지 하나, 좌우 철망 울타리 구간, 왼쪽 전경의 흰색·보라색 꽃이 보여 장소 참조의 주요 배치와 맞는다. 찰리는 꽃 옆 길의 왼쪽에 있다. 상단에는 그물 하나가 있으며 중복된 시설이나 반사는 없다. 다만 무릎 부근까지 포함해 지정된 상반신 중심 미디엄 숏보다 넓다.",
        "entities": "찰리 한 대, 나비 한 마리, 큰 밧줄 그물 하나가 보이며 추가 인물은 없다. 찰리의 샌드 베이지 금속 장갑, 육중한 팔, 흰 마스크, 주황색 점 형태의 두 눈과 입 선, 푸른 원형 가슴 동력부는 참조와 부합한다. 밀짚모자와 여러 색의 낡은 우비도 있다. 어깨 수리 흔적은 명확하게 식별되지 않으며 장화는 화면에서 확인하기 어렵지만, 잘린 의상을 결함으로 보지는 않는다. 꽃과 습지의 수목도 표현되어 있다.",
        "hard_violations": [],
        "physics": "찰리의 하체는 화면 아래 지면 쪽으로 이어지고 몸이 공중에 떠 있는 모습은 아니다. 들어 올린 손은 손목·팔꿈치·어깨 관절로 몸체에 연결되어 있으며 나비를 향한 뻗기 자세가 가능하다. 모자는 머리에, 우비는 어깨와 몸통에 걸쳐 있다. 나비는 펼친 날개로 비행한다. 그물은 매듭진 테두리와 처진 밧줄 곡선을 갖추고 있으며, 위에서 투하되어 중력으로 내려오는 순간으로 해석 가능한 형태다. 다만 낙하보다는 매달린 그물처럼 정적으로 읽힐 여지도 있다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴은 오른쪽 위를 향하고, 높이 든 손보다 조금 위쪽 오른편에 나비가 있어 시선과 손짓의 대상이 일치한다. 그물은 상단 전체에 펼쳐져 찰리 위에 놓이고 얼굴을 가리지 않는다. 나비의 상승과 그물의 하강은 배치로 암시되지만 움직임 자체는 선명하지 않다.",
        "built_space": "좌우의 큰 나무, 왼쪽 방사능 표지 하나, 양옆 철망 울타리, 중앙의 습지와 진흙길, 왼쪽 전경의 꽃들이 참조 장소와 대응한다. 찰리는 꽃 옆 길의 왼쪽에 서 있다. 그물은 하나이며 시설 중복이나 불가능한 반사는 없다. 카메라는 낮은 위치지만 찰리의 장화와 주변 길까지 넓게 보여, 요청된 머리·손·상반신 중심 미디엄 숏보다 전신 구도에 가깝다.",
        "entities": "찰리 한 대와 작은 나비 한 마리, 밧줄 그물 하나가 있다. 베이지색 기계 장갑, 긴 팔과 짧은 다리, 흰 마스크와 선 형태의 입, 푸른 가슴 동력부가 보인다. 얼굴이 측면에 가까워 한쪽 발광 눈이 주로 드러나며 인간 피부나 인간 눈은 추가되지 않았다. 큰 밀짚모자, 노랑·주황 계열 우비와 장화가 보인다. 어깨가 수리되었다는 구체적 흔적은 확인하기 어렵다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "찰리의 장화는 진흙길과 전경 식물 뒤 지면에 닿아 있고, 다리와 몸통의 기울기로 서서 팔을 뻗는 자세가 성립한다. 손과 팔은 관절로 연결되어 있으며 모자와 우비도 몸에 지지된다. 나비는 날개를 편 비행 상태다. 그물 양쪽 가장자리의 추 모양 매듭과 아래로 처진 테두리는 중력에 따른 낙하와 양립한다. 공중 투하된 그물에 손 지지가 없는 것은 불가능한 부유가 아니지만, 화면상으로는 정지해 걸린 그물과 낙하 중인 그물의 차이가 뚜렷하지 않다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 얼굴과 뻗은 손을 더 크게 담은 전방 사선 구도가 우세하지만, 무릎까지 보이는 촬영 범위와 강한 직사광은 지정된 미디엄 숏·차분한 낮 조명에서 벗어난다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "나비를 향한 시선과 머리 위 그물의 배치는 맞지만, 장화까지 드러내는 넓은 구도와 측면에 가까운 얼굴 방향이 지정된 상반신 중심 전방 사선 미디엄 숏보다 멀다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 얼굴을 위쪽 오른편으로 들고 있으며, 시선과 펼친 손끝의 연장선 위에 작은 나비가 있다. 그물은 찰리 머리 위 화면 상단을 가로지르고 얼굴이나 손을 가리지 않는다. 나비의 비행 자세는 보이지만 상승 방향과 그물의 하강 속도는 정지 화면만으로 확정하기 어렵다.",
        "built_space": "왼쪽과 오른쪽의 굵은 나무 사이로 진흙길이 습지까지 이어진다. 왼쪽 나무 옆 방사능 표지 하나, 좌우 철망 울타리 구간, 왼쪽 전경의 흰색·보라색 꽃이 보여 장소 참조의 주요 배치와 맞는다. 찰리는 꽃 옆 길의 왼쪽에 있다. 상단에는 그물 하나가 있으며 중복된 시설이나 반사는 없다. 다만 무릎 부근까지 포함해 지정된 상반신 중심 미디엄 숏보다 넓다.",
        "entities": "찰리 한 대, 나비 한 마리, 큰 밧줄 그물 하나가 보이며 추가 인물은 없다. 찰리의 샌드 베이지 금속 장갑, 육중한 팔, 흰 마스크, 주황색 점 형태의 두 눈과 입 선, 푸른 원형 가슴 동력부는 참조와 부합한다. 밀짚모자와 여러 색의 낡은 우비도 있다. 어깨 수리 흔적은 명확하게 식별되지 않으며 장화는 화면에서 확인하기 어렵지만, 잘린 의상을 결함으로 보지는 않는다. 꽃과 습지의 수목도 표현되어 있다.",
        "hard_violations": [],
        "physics": "찰리의 하체는 화면 아래 지면 쪽으로 이어지고 몸이 공중에 떠 있는 모습은 아니다. 들어 올린 손은 손목·팔꿈치·어깨 관절로 몸체에 연결되어 있으며 나비를 향한 뻗기 자세가 가능하다. 모자는 머리에, 우비는 어깨와 몸통에 걸쳐 있다. 나비는 펼친 날개로 비행한다. 그물은 매듭진 테두리와 처진 밧줄 곡선을 갖추고 있으며, 위에서 투하되어 중력으로 내려오는 순간으로 해석 가능한 형태다. 다만 낙하보다는 매달린 그물처럼 정적으로 읽힐 여지도 있다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴은 오른쪽 위를 향하고, 높이 든 손보다 조금 위쪽 오른편에 나비가 있어 시선과 손짓의 대상이 일치한다. 그물은 상단 전체에 펼쳐져 찰리 위에 놓이고 얼굴을 가리지 않는다. 나비의 상승과 그물의 하강은 배치로 암시되지만 움직임 자체는 선명하지 않다.",
        "built_space": "좌우의 큰 나무, 왼쪽 방사능 표지 하나, 양옆 철망 울타리, 중앙의 습지와 진흙길, 왼쪽 전경의 꽃들이 참조 장소와 대응한다. 찰리는 꽃 옆 길의 왼쪽에 서 있다. 그물은 하나이며 시설 중복이나 불가능한 반사는 없다. 카메라는 낮은 위치지만 찰리의 장화와 주변 길까지 넓게 보여, 요청된 머리·손·상반신 중심 미디엄 숏보다 전신 구도에 가깝다.",
        "entities": "찰리 한 대와 작은 나비 한 마리, 밧줄 그물 하나가 있다. 베이지색 기계 장갑, 긴 팔과 짧은 다리, 흰 마스크와 선 형태의 입, 푸른 가슴 동력부가 보인다. 얼굴이 측면에 가까워 한쪽 발광 눈이 주로 드러나며 인간 피부나 인간 눈은 추가되지 않았다. 큰 밀짚모자, 노랑·주황 계열 우비와 장화가 보인다. 어깨가 수리되었다는 구체적 흔적은 확인하기 어렵다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "찰리의 장화는 진흙길과 전경 식물 뒤 지면에 닿아 있고, 다리와 몸통의 기울기로 서서 팔을 뻗는 자세가 성립한다. 손과 팔은 관절로 연결되어 있으며 모자와 우비도 몸에 지지된다. 나비는 날개를 편 비행 상태다. 그물 양쪽 가장자리의 추 모양 매듭과 아래로 처진 테두리는 중력에 따른 낙하와 양립한다. 공중 투하된 그물에 손 지지가 없는 것은 불가능한 부유가 아니지만, 화면상으로는 정지해 걸린 그물과 낙하 중인 그물의 차이가 뚜렷하지 않다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.524,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.524,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1524
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "후보 B는 찰리의 하얀 마스크 얼굴(두 눈과 입 선)과 가슴의 원자로 심볼을 캐릭터 레퍼런스와 정확하게 일치시켰으며, 요청된 로우 앵글과 역동적인 구도를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1524,
    "verdict_ko": "구도와 배경은 잘 반영되었으나, 찰리의 얼굴 마스크가 심하게 왜곡되어 캐릭터 레퍼런스의 특징(두 눈과 입 선)을 잃었으며 가슴의 원자로 위치도 어색합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_capture_site_4efc3c.png",
    "asset_id": "50187464-125b-40d8-9db9-f8b1b64944a6",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bd0-569f-74e4-a4b6-12580eb186f5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh14__bgfirst_bg.png",
   "bg_asset_id": "07be09ef-d192-4906-bbda-7d5864eb163f",
   "bg_record_key": "S56sh14::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "forest_capture_site",
   "groupbg_asset_id": "50187464-125b-40d8-9db9-f8b1b64944a6"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S56sh23::signage": {
  "fp": "3c92b1962c85a283",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::forest_trap_pit": {
  "input_fingerprint": "49c7d3830de1961c",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "forest_trap_pit",
    "tags": [
     "S56sh23",
     "S56sh25"
    ]
   },
   "context_sig": "5db564e2085babcb"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 가면(하회탈 모양의 가죽 마스크)을 형형색색을 쓴 남자들이 구덩이에 빠진 현우와 앰버를 내려다본다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 가면(하회탈 모양의 가죽 마스크)을 형형색색을 쓴 남자들이 구덩이에 빠진 현우와 앰버를 내려다본다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
  "asset_id": "ea39b7e5-54ed-4fa5-83ac-c978c0f43983",
  "input_asset_ids": [
   "1d96b2cd-1952-4cbd-9db0-5c648a0e829f"
  ],
  "origin_tag": "S56sh23",
  "place_text": "At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.",
  "origin_inputs": {
   "place_text": "At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.",
   "time_of_day_en": "day",
   "conti_asset_id": "1d96b2cd-1952-4cbd-9db0-5c648a0e829f"
  }
 },
 "S56sh23::bgfirst_bg": {
  "input_fingerprint": "f19a036b9450ae01",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하회탈 모양의 마스크를 쓴 무장한 남자들이 함정 구덩이 위에서 이현우와 앰버를 굽어보는 위압적인 구도.\n\nLOCATION (lock): At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the threat-reveal stage inside the pit, low behind and beside 이현우 and 앰버, with a steep upward angle across their near shoulders toward the rim. Keep 이현우 at the lower left and 앰버 at the lower right, recoiling and looking up from clear of the wall, while the armed men look down from the upper band of the frame. Before the pan isolates the shooter, distinguish the men through uneven spacing, different downward head tilts, and varied weapon angles, preserving their common threatening attention without creating a synchronized lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: pit rim occupied by the armed men above the trapped pair in the upper-center of the frame, background; inside of pit containing 이현우 and 앰버 below the rim in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 함정 구덩이 (Contains 이현우 and 앰버 below the armed men at its rim) — Interior walls rise from the near shoulders toward the visible opening; used as Establishes the vertical separation and prevents the men from reading as standing beside the trapped pair; 하회탈 모양의 가죽 마스크 (Worn by the armed men in varied colors) — Mask fronts angle downward toward the trapped pair rather than directly toward the lens; used as Make the watching group threatening without erasing individual head-angle differences; 긴 창, 도끼, 활 (Held by the men around the pit opening) — Their silhouettes occupy differing angles beside their holders rather than repeating one identical pose; used as Identify the group's armed control while keeping faces and the opening readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime illumination from the pit opening separates the figures at the rim from the nearer shoulders while retaining readable detail below and the masks' established varied colors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하회탈 모양의 마스크를 쓴 무장한 남자들이 함정 구덩이 위에서 이현우와 앰버를 굽어보는 위압적인 구도.\n\nLOCATION (lock): At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the threat-reveal stage inside the pit, low behind and beside 이현우 and 앰버, with a steep upward angle across their near shoulders toward the rim. Keep 이현우 at the lower left and 앰버 at the lower right, recoiling and looking up from clear of the wall, while the armed men look down from the upper band of the frame. Before the pan isolates the shooter, distinguish the men through uneven spacing, different downward head tilts, and varied weapon angles, preserving their common threatening attention without creating a synchronized lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: pit rim occupied by the armed men above the trapped pair in the upper-center of the frame, background; inside of pit containing 이현우 and 앰버 below the rim in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 함정 구덩이 (Contains 이현우 and 앰버 below the armed men at its rim) — Interior walls rise from the near shoulders toward the visible opening; used as Establishes the vertical separation and prevents the men from reading as standing beside the trapped pair; 하회탈 모양의 가죽 마스크 (Worn by the armed men in varied colors) — Mask fronts angle downward toward the trapped pair rather than directly toward the lens; used as Make the watching group threatening without erasing individual head-angle differences; 긴 창, 도끼, 활 (Held by the men around the pit opening) — Their silhouettes occupy differing angles beside their holders rather than repeating one identical pose; used as Identify the group's armed control while keeping faces and the opening readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime illumination from the pit opening separates the figures at the rim from the nearer shoulders while retaining readable detail below and the masks' established varied colors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh23__bgfirst_bg.png",
  "asset_id": "47a6acbb-5c66-4c93-a4f5-042d6a3bbc9c",
  "input_asset_ids": [
   "1d96b2cd-1952-4cbd-9db0-5c648a0e829f",
   "ea39b7e5-54ed-4fa5-83ac-c978c0f43983"
  ]
 },
 "S56sh23": {
  "input_fingerprint": "154d864c5647bad6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 모양의 마스크를 쓴 무장한 남자들이 함정 구덩이 위에서 이현우와 앰버를 굽어보는 위압적인 구도.\n\nLOCATION (lock): At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the threat-reveal stage inside the pit, low behind and beside 이현우 and 앰버, with a steep upward angle across their near shoulders toward the rim. Keep 이현우 at the lower left and 앰버 at the lower right, recoiling and looking up from clear of the wall, while the armed men look down from the upper band of the frame. Before the pan isolates the shooter, distinguish the men through uneven spacing, different downward head tilts, and varied weapon angles, preserving their common threatening attention without creating a synchronized lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: pit rim occupied by the armed men above the trapped pair in the upper-center of the frame, background; inside of pit containing 이현우 and 앰버 below the rim in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 함정 구덩이 (Contains 이현우 and 앰버 below the armed men at its rim) — Interior walls rise from the near shoulders toward the visible opening; used as Establishes the vertical separation and prevents the men from reading as standing beside the trapped pair; 하회탈 모양의 가죽 마스크 (Worn by the armed men in varied colors) — Mask fronts angle downward toward the trapped pair rather than directly toward the lens; used as Make the watching group threatening without erasing individual head-angle differences; 긴 창, 도끼, 활 (Held by the men around the pit opening) — Their silhouettes occupy differing angles beside their holders rather than repeating one identical pose; used as Identify the group's armed control while keeping faces and the opening readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime illumination from the pit opening separates the figures at the rim from the nearer shoulders while retaining readable detail below and the masks' established varied colors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An arrow remains embedded in a tree beside the approach to the trap pit. The camper remains stuck in the nearby swamp, where the large bootprints and weathered warning signs persist. 이현우: He is trapped at the bottom of the pit, with his treated leg still bandaged. 앰버: She is trapped at the bottom of the pit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 모양의 마스크를 쓴 무장한 남자들이 함정 구덩이 위에서 이현우와 앰버를 굽어보는 위압적인 구도.\n\nLOCATION (lock): At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the threat-reveal stage inside the pit, low behind and beside 이현우 and 앰버, with a steep upward angle across their near shoulders toward the rim. Keep 이현우 at the lower left and 앰버 at the lower right, recoiling and looking up from clear of the wall, while the armed men look down from the upper band of the frame. Before the pan isolates the shooter, distinguish the men through uneven spacing, different downward head tilts, and varied weapon angles, preserving their common threatening attention without creating a synchronized lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: pit rim occupied by the armed men above the trapped pair in the upper-center of the frame, background; inside of pit containing 이현우 and 앰버 below the rim in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 함정 구덩이 (Contains 이현우 and 앰버 below the armed men at its rim) — Interior walls rise from the near shoulders toward the visible opening; used as Establishes the vertical separation and prevents the men from reading as standing beside the trapped pair; 하회탈 모양의 가죽 마스크 (Worn by the armed men in varied colors) — Mask fronts angle downward toward the trapped pair rather than directly toward the lens; used as Make the watching group threatening without erasing individual head-angle differences; 긴 창, 도끼, 활 (Held by the men around the pit opening) — Their silhouettes occupy differing angles beside their holders rather than repeating one identical pose; used as Identify the group's armed control while keeping faces and the opening readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime illumination from the pit opening separates the figures at the rim from the nearer shoulders while retaining readable detail below and the masks' established varied colors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An arrow remains embedded in a tree beside the approach to the trap pit. The camper remains stuck in the nearby swamp, where the large bootprints and weathered warning signs persist. 이현우: He is trapped at the bottom of the pit, with his treated leg still bandaged. 앰버: She is trapped at the bottom of the pit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 모양의 마스크를 쓴 무장한 남자들이 함정 구덩이 위에서 이현우와 앰버를 굽어보는 위압적인 구도.\n\nLOCATION (lock): At the rim of a concealed pit trap in the forest near the marsh, overlooking the prisoners below. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the threat-reveal stage inside the pit, low behind and beside 이현우 and 앰버, with a steep upward angle across their near shoulders toward the rim. Keep 이현우 at the lower left and 앰버 at the lower right, recoiling and looking up from clear of the wall, while the armed men look down from the upper band of the frame. Before the pan isolates the shooter, distinguish the men through uneven spacing, different downward head tilts, and varied weapon angles, preserving their common threatening attention without creating a synchronized lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: pit rim occupied by the armed men above the trapped pair in the upper-center of the frame, background; inside of pit containing 이현우 and 앰버 below the rim in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 함정 구덩이 (Contains 이현우 and 앰버 below the armed men at its rim) — Interior walls rise from the near shoulders toward the visible opening; used as Establishes the vertical separation and prevents the men from reading as standing beside the trapped pair; 하회탈 모양의 가죽 마스크 (Worn by the armed men in varied colors) — Mask fronts angle downward toward the trapped pair rather than directly toward the lens; used as Make the watching group threatening without erasing individual head-angle differences; 긴 창, 도끼, 활 (Held by the men around the pit opening) — Their silhouettes occupy differing angles beside their holders rather than repeating one identical pose; used as Identify the group's armed control while keeping faces and the opening readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime illumination from the pit opening separates the figures at the rim from the nearer shoulders while retaining readable detail below and the masks' established varied colors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): An arrow remains embedded in a tree beside the approach to the trap pit. The camper remains stuck in the nearby swamp, where the large bootprints and weathered warning signs persist. 이현우: He is trapped at the bottom of the pit, with his treated leg still bandaged. 앰버: She is trapped at the bottom of the pit.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh23__bgfirst_bg.png",
     "asset_id": "47a6acbb-5c66-4c93-a4f5-042d6a3bbc9c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S56sh23.png",
     "asset_id": "1d96b2cd-1952-4cbd-9db0-5c648a0e829f",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
     "asset_id": "ea39b7e5-54ed-4fa5-83ac-c978c0f43983",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "구덩이 위 남자들은 아래를, 이현우와 앰버는 위를 쳐다보고 있으며, 무기들은 다양한 각도를 향함.",
    "built_space": "레퍼런스와 일치하는 흙벽 구덩이 내부. 카메라는 인물들 뒤에서 위를 향하는 로우 앵글.",
    "entities": "이현우(검은 머리, 붕대 감은 다리)와 앰버(금발)의 뒷모습. 남자들은 가죽 마스크를 쓰고 무기를 들고 있음.",
    "hard_violations": [
     "[gemini-pro] 허공에 떠 있는 화살들 (unsupported objects)",
     "[gpt-high] 가면 쓴 남자들 옆 하늘에 방향을 설명하는 도식 화살표 다섯 개가 노출되어, 그래픽·마커 오버레이 금지 조건을 위반한다."
    ],
    "physics": "인물들은 바닥에 지지되어 있으나, 상단 남자들 주위에 발사 원점이나 지지체 없이 허공에 뜬 화살들이 있음."
   },
   {
    "label": "B",
    "direction": "위쪽 남자들은 아래를 내려다보고, 이현우와 앰버는 정면 위를 쳐다봄.",
    "built_space": "흙 구덩이 우측에 거대한 나무 계단이 임의로 추가됨. 카메라는 인물들 정면에 위치함.",
    "entities": "이현우와 앰버가 정면으로 서 있음. 남자들은 도깨비/동물 형태의 마스크를 착용함.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 나무 계단 구조물 추가 (invented objects)",
     "[gpt-high] 장소 참조와 지시에 없는 대형 목제 계단을 구덩이 안에 추가하여, 갇힌 포로와 함정의 공간 구조를 바꾸었다."
    ],
    "physics": "모든 인물과 사물이 바닥에 안정적으로 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "명시된 후면 로우 앵글과 이현우의 붕대 감은 다리 등 디테일과 구도를 훌륭하게 구현했으나, 상단 남자들 주위에 허공에 뜬 화살들이 존재하여 물리 법칙을 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 요구한 후면 어깨 너머 앵글을 완전히 무시하고 정면에서 촬영했으며, 지정된 장소에 없는 거대한 나무 계단을 임의로 생성하여 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "구덩이 위 남자들은 아래를, 이현우와 앰버는 위를 쳐다보고 있으며, 무기들은 다양한 각도를 향함.",
        "built_space": "레퍼런스와 일치하는 흙벽 구덩이 내부. 카메라는 인물들 뒤에서 위를 향하는 로우 앵글.",
        "entities": "이현우(검은 머리, 붕대 감은 다리)와 앰버(금발)의 뒷모습. 남자들은 가죽 마스크를 쓰고 무기를 들고 있음.",
        "hard_violations": [
         "허공에 떠 있는 화살들 (unsupported objects)"
        ],
        "physics": "인물들은 바닥에 지지되어 있으나, 상단 남자들 주위에 발사 원점이나 지지체 없이 허공에 뜬 화살들이 있음."
       },
       {
        "label": "B",
        "direction": "위쪽 남자들은 아래를 내려다보고, 이현우와 앰버는 정면 위를 쳐다봄.",
        "built_space": "흙 구덩이 우측에 거대한 나무 계단이 임의로 추가됨. 카메라는 인물들 정면에 위치함.",
        "entities": "이현우와 앰버가 정면으로 서 있음. 남자들은 도깨비/동물 형태의 마스크를 착용함.",
        "hard_violations": [
         "레퍼런스에 없는 나무 계단 구조물 추가 (invented objects)"
        ],
        "physics": "모든 인물과 사물이 바닥에 안정적으로 지지되어 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "명시된 후면 로우 앵글과 이현우의 붕대 감은 다리 등 디테일과 구도를 훌륭하게 구현했으나, 상단 남자들 주위에 허공에 뜬 화살들이 존재하여 물리 법칙을 위반했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트가 요구한 후면 어깨 너머 앵글을 완전히 무시하고 정면에서 촬영했으며, 지정된 장소에 없는 거대한 나무 계단을 임의로 생성하여 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "구덩이 위 남자들은 아래를, 이현우와 앰버는 위를 쳐다보고 있으며, 무기들은 다양한 각도를 향함.",
        "built_space": "레퍼런스와 일치하는 흙벽 구덩이 내부. 카메라는 인물들 뒤에서 위를 향하는 로우 앵글.",
        "entities": "이현우(검은 머리, 붕대 감은 다리)와 앰버(금발)의 뒷모습. 남자들은 가죽 마스크를 쓰고 무기를 들고 있음.",
        "hard_violations": [
         "허공에 떠 있는 화살들 (unsupported objects)"
        ],
        "physics": "인물들은 바닥에 지지되어 있으나, 상단 남자들 주위에 발사 원점이나 지지체 없이 허공에 뜬 화살들이 있음."
       },
       {
        "label": "B",
        "direction": "위쪽 남자들은 아래를 내려다보고, 이현우와 앰버는 정면 위를 쳐다봄.",
        "built_space": "흙 구덩이 우측에 거대한 나무 계단이 임의로 추가됨. 카메라는 인물들 정면에 위치함.",
        "entities": "이현우와 앰버가 정면으로 서 있음. 남자들은 도깨비/동물 형태의 마스크를 착용함.",
        "hard_violations": [
         "레퍼런스에 없는 나무 계단 구조물 추가 (invented objects)"
        ],
        "physics": "모든 인물과 사물이 바닥에 안정적으로 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 사람의 어깨 너머 급격한 올려다보기와 위협적인 상하 배치는 정확하지만, 남자들 옆에 노출된 도식 화살표 때문에 최종 프레임으로는 실격이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "참조에 없는 탈출용 목제 계단을 추가했고, 두 사람의 뒤가 아닌 앞에서 촬영하여 핵심 카메라 구도와 움츠러드는 행동을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 화면 왼쪽 위의 가장자리 쪽을 올려다보지만, 앰버는 가장자리의 남자들보다 이현우의 얼굴을 보는 것으로 읽힌다. 위쪽 남자 여섯 명은 대체로 구덩이 안을 향하지만 가면의 정면 노출이 많고 고개 기울기 차이는 작다. 긴 자루들은 대부분 수직에 가까우며, 포로를 겨누는 방향이나 창·도끼·활의 서로 다른 실루엣이 명확하지 않다.",
        "built_space": "흙벽과 노출된 뿌리로 둘러싸인 구덩이 하나, 가장자리 뒤 철망 울타리, 오른쪽의 목제 계단 하나가 보인다. 남자 여섯 명은 가장자리에, 두 포로는 아래쪽에 있어 높이 차이는 성립한다. 그러나 참조에 없는 계단이 바닥과 지상을 연결한다. 카메라는 두 포로의 뒤·옆이 아니라 정면에 있고, 어깨 너머 시점 대신 몸통과 다리를 크게 보여준다. 상단의 두꺼운 흙 돌출부도 참조의 열린 구덩이와 다르게 동굴 입구처럼 보인다.",
        "entities": "검은 머리의 젊은 동아시아계 남성과 금발 여자아이, 가면을 쓴 성인 남자 여섯 명이 보인다. 이현우의 얼굴과 머리는 참조에 대체로 가깝지만 남색 티셔츠 대신 녹색 외투를 입었다. 앰버는 어린 금발 여자아이로 읽히지만 참조의 반소매 대신 긴소매 상의를 입었다. 가면은 여러 색이지만 일부는 뿔 달린 도깨비 형상에 가까워 하회탈 형태가 약하다. 긴 자루는 보이나 도끼와 활은 확실히 식별되지 않는다. 붕대는 확인되지 않으며, 나무에 박힌 화살·늪의 캠퍼·발자국은 이 화면에서 확인할 수 없다.",
        "hard_violations": [
         "장소 참조와 지시에 없는 대형 목제 계단을 구덩이 안에 추가하여, 갇힌 포로와 함정의 공간 구조를 바꾸었다."
        ],
        "physics": "가장자리 남자들은 흙 지면에 발을 딛고 있다. 두 포로는 다리가 화면 아래로 이어지는 직립 자세여서 공중에 떠 있는 증거는 없지만, 움츠러들거나 뒤로 물러나는 동작 없이 팔을 내린 채 서 있다. 보이는 긴 자루들은 손 또는 가장자리 지면과 접촉한다. 계단도 지면과 가장자리에 걸쳐 물리적으로 지지되지만, 그 존재 자체가 지시 위반이다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽 아래에서, 앰버는 오른쪽 아래에서 고개를 들어 가장자리 남자들을 본다. 남자 다섯 명의 가면은 서로 다른 기울기로 구덩이 아래를 향한다. 중앙 창은 위로 곧게, 오른쪽 창은 왼쪽 위로 비스듬히, 왼쪽 도끼는 위쪽으로 향하고 양끝의 활은 낮게 들려 있어 무기 각도가 다양하다. 남자들 옆의 가느다란 화살표 다섯 개는 아래 방향을 표시하는 도식으로 보인다.",
        "built_space": "하나의 열린 구덩이와 높은 흙벽, 노출된 뿌리, 가장자리의 숲이 보인다. 포로 둘은 바닥 전경에 있고 남자 다섯 명은 위쪽 가장자리에 불균등하게 배치되어 수직 분리가 명확하다. 낮은 카메라가 두 사람의 뒤·옆에서 어깨를 넘어 위를 보는 와이드 구도를 구현한다. 다만 참조 장소의 철망 울타리와 방사능 경고 표지판은 보이지 않아 장소 고유 요소의 일치는 부족하다.",
        "entities": "짧은 검은 머리의 젊은 남성과 금발 여자아이, 가면 쓴 성인 남자 다섯 명이 보인다. 두 포로의 남색 반소매는 참조와 맞고, 얼굴이 돌아가 있어 정확한 얼굴 일치와 앰버의 혼혈 외모는 검증하기 어렵다. 앰버의 머리는 참조와 달리 묶여 있다. 이현우의 아래쪽 다리에는 붕대가 보인다. 창 두 자루, 도끼 하나, 양끝의 활 두 개와 화살통이 식별된다. 황갈색·녹색·어두운색 가면은 있으나 하회탈 특유의 표정보다 매끈한 일반 인면 가면에 가깝다. 나무에 박힌 화살과 늪의 캠퍼·발자국은 프레임 밖이므로 불일치로 판단하지 않는다.",
        "hard_violations": [
         "가면 쓴 남자들 옆 하늘에 방향을 설명하는 도식 화살표 다섯 개가 노출되어, 그래픽·마커 오버레이 금지 조건을 위반한다."
        ],
        "physics": "두 포로는 엉덩이와 굽힌 다리를 구덩이 바닥에 둔 채 뒤로 기울어져 있으며, 이현우는 뒤로 뻗은 팔로 몸을 받치는 자세다. 앰버의 들어 올린 손은 팔에 자연스럽게 연결되어 방어적인 반응으로 읽힌다. 가장자리 남자들의 하체는 흙 가장자리로 이어지고, 창·도끼·활은 손으로 쥐고 있어 실제 인물이나 무기가 지지 없이 떠 있지는 않다. 하늘의 화살표는 물리적 화살이 아니라 별도로 삽입된 도식으로 보인다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 사람의 어깨 너머 급격한 올려다보기와 위협적인 상하 배치는 정확하지만, 남자들 옆에 노출된 도식 화살표 때문에 최종 프레임으로는 실격이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "참조에 없는 탈출용 목제 계단을 추가했고, 두 사람의 뒤가 아닌 앞에서 촬영하여 핵심 카메라 구도와 움츠러드는 행동을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 화면 왼쪽 위의 가장자리 쪽을 올려다보지만, 앰버는 가장자리의 남자들보다 이현우의 얼굴을 보는 것으로 읽힌다. 위쪽 남자 여섯 명은 대체로 구덩이 안을 향하지만 가면의 정면 노출이 많고 고개 기울기 차이는 작다. 긴 자루들은 대부분 수직에 가까우며, 포로를 겨누는 방향이나 창·도끼·활의 서로 다른 실루엣이 명확하지 않다.",
        "built_space": "흙벽과 노출된 뿌리로 둘러싸인 구덩이 하나, 가장자리 뒤 철망 울타리, 오른쪽의 목제 계단 하나가 보인다. 남자 여섯 명은 가장자리에, 두 포로는 아래쪽에 있어 높이 차이는 성립한다. 그러나 참조에 없는 계단이 바닥과 지상을 연결한다. 카메라는 두 포로의 뒤·옆이 아니라 정면에 있고, 어깨 너머 시점 대신 몸통과 다리를 크게 보여준다. 상단의 두꺼운 흙 돌출부도 참조의 열린 구덩이와 다르게 동굴 입구처럼 보인다.",
        "entities": "검은 머리의 젊은 동아시아계 남성과 금발 여자아이, 가면을 쓴 성인 남자 여섯 명이 보인다. 이현우의 얼굴과 머리는 참조에 대체로 가깝지만 남색 티셔츠 대신 녹색 외투를 입었다. 앰버는 어린 금발 여자아이로 읽히지만 참조의 반소매 대신 긴소매 상의를 입었다. 가면은 여러 색이지만 일부는 뿔 달린 도깨비 형상에 가까워 하회탈 형태가 약하다. 긴 자루는 보이나 도끼와 활은 확실히 식별되지 않는다. 붕대는 확인되지 않으며, 나무에 박힌 화살·늪의 캠퍼·발자국은 이 화면에서 확인할 수 없다.",
        "hard_violations": [
         "장소 참조와 지시에 없는 대형 목제 계단을 구덩이 안에 추가하여, 갇힌 포로와 함정의 공간 구조를 바꾸었다."
        ],
        "physics": "가장자리 남자들은 흙 지면에 발을 딛고 있다. 두 포로는 다리가 화면 아래로 이어지는 직립 자세여서 공중에 떠 있는 증거는 없지만, 움츠러들거나 뒤로 물러나는 동작 없이 팔을 내린 채 서 있다. 보이는 긴 자루들은 손 또는 가장자리 지면과 접촉한다. 계단도 지면과 가장자리에 걸쳐 물리적으로 지지되지만, 그 존재 자체가 지시 위반이다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽 아래에서, 앰버는 오른쪽 아래에서 고개를 들어 가장자리 남자들을 본다. 남자 다섯 명의 가면은 서로 다른 기울기로 구덩이 아래를 향한다. 중앙 창은 위로 곧게, 오른쪽 창은 왼쪽 위로 비스듬히, 왼쪽 도끼는 위쪽으로 향하고 양끝의 활은 낮게 들려 있어 무기 각도가 다양하다. 남자들 옆의 가느다란 화살표 다섯 개는 아래 방향을 표시하는 도식으로 보인다.",
        "built_space": "하나의 열린 구덩이와 높은 흙벽, 노출된 뿌리, 가장자리의 숲이 보인다. 포로 둘은 바닥 전경에 있고 남자 다섯 명은 위쪽 가장자리에 불균등하게 배치되어 수직 분리가 명확하다. 낮은 카메라가 두 사람의 뒤·옆에서 어깨를 넘어 위를 보는 와이드 구도를 구현한다. 다만 참조 장소의 철망 울타리와 방사능 경고 표지판은 보이지 않아 장소 고유 요소의 일치는 부족하다.",
        "entities": "짧은 검은 머리의 젊은 남성과 금발 여자아이, 가면 쓴 성인 남자 다섯 명이 보인다. 두 포로의 남색 반소매는 참조와 맞고, 얼굴이 돌아가 있어 정확한 얼굴 일치와 앰버의 혼혈 외모는 검증하기 어렵다. 앰버의 머리는 참조와 달리 묶여 있다. 이현우의 아래쪽 다리에는 붕대가 보인다. 창 두 자루, 도끼 하나, 양끝의 활 두 개와 화살통이 식별된다. 황갈색·녹색·어두운색 가면은 있으나 하회탈 특유의 표정보다 매끈한 일반 인면 가면에 가깝다. 나무에 박힌 화살과 늪의 캠퍼·발자국은 프레임 밖이므로 불일치로 판단하지 않는다.",
        "hard_violations": [
         "가면 쓴 남자들 옆 하늘에 방향을 설명하는 도식 화살표 다섯 개가 노출되어, 그래픽·마커 오버레이 금지 조건을 위반한다."
        ],
        "physics": "두 포로는 엉덩이와 굽힌 다리를 구덩이 바닥에 둔 채 뒤로 기울어져 있으며, 이현우는 뒤로 뻗은 팔로 몸을 받치는 자세다. 앰버의 들어 올린 손은 팔에 자연스럽게 연결되어 방어적인 반응으로 읽힌다. 가장자리 남자들의 하체는 흙 가장자리로 이어지고, 창·도끼·활은 손으로 쥐고 있어 실제 인물이나 무기가 지지 없이 떠 있지는 않다. 하늘의 화살표는 물리적 화살이 아니라 별도로 삽입된 도식으로 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.267
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.017
   },
   "violations": {
    "A": [
     "[gemini-pro] 허공에 떠 있는 화살들 (unsupported objects)",
     "[gpt-high] 가면 쓴 남자들 옆 하늘에 방향을 설명하는 도식 화살표 다섯 개가 노출되어, 그래픽·마커 오버레이 금지 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] 레퍼런스에 없는 나무 계단 구조물 추가 (invented objects)",
     "[gpt-high] 장소 참조와 지시에 없는 대형 목제 계단을 구덩이 안에 추가하여, 갇힌 포로와 함정의 공간 구조를 바꾸었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1017
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "명시된 후면 로우 앵글과 이현우의 붕대 감은 다리 등 디테일과 구도를 훌륭하게 구현했으나, 상단 남자들 주위에 허공에 뜬 화살들이 존재하여 물리 법칙을 위반했습니다.  ★위반: [gemini-pro] 허공에 떠 있는 화살들 (unsupported objects) / [gpt-high] 가면 쓴 남자들 옆 하늘에 방향을 설명하는 도식 화살표 다섯 개가 노출되어, 그래픽·마커 오버레이 금지 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1017,
    "verdict_ko": "프롬프트가 요구한 후면 어깨 너머 앵글을 완전히 무시하고 정면에서 촬영했으며, 지정된 장소에 없는 거대한 나무 계단을 임의로 생성하여 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 레퍼런스에 없는 나무 계단 구조물 추가 (invented objects) / [gpt-high] 장소 참조와 지시에 없는 대형 목제 계단을 구덩이 안에 추가하여, 갇힌 포로와 함정의 공간 구조를 바꾸었다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_forest_trap_pit_2a4ec2.png",
    "asset_id": "ea39b7e5-54ed-4fa5-83ac-c978c0f43983",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bda-70e0-79bb-8785-1954afb5001f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh23__bgfirst_bg.png",
   "bg_asset_id": "47a6acbb-5c66-4c93-a4f5-042d6a3bbc9c",
   "bg_record_key": "S56sh23::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "forest_trap_pit",
   "groupbg_asset_id": "ea39b7e5-54ed-4fa5-83ac-c978c0f43983"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C03",
   "C05"
  ]
 },
 "S56sh25::signage": {
  "fp": "8f44e1c18c647bfa",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh25": {
  "input_fingerprint": "8e9945b1b477b60f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마취 화살에 맞아 의식을 잃은 이현우의 몸이 구덩이 바닥을 향해 중심을 잃고 기울어진 순간.\n\nLOCATION (lock): At the bottom of an open-topped pit trap in the woodland near the contaminated marsh. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the downward rotation and descending move beside and slightly behind 이현우, holding an oblique side view with a gentle downward pitch toward the pit bottom. His unconscious head and torso tip diagonally from the upper left toward the lower center, knees giving way and eyes closed without purposeful visual attention; retain visible ground beneath him before contact. Let his falling position carry the change, keeping the established shooter-to-victim side and leaving 앰버 and the men outside this tighter frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 구덩이 바닥 (Visible beneath 이현우 before his falling body reaches it) — Extends below his descending head and torso; used as Preserves the remaining gap to the ground and makes the collapse physically legible; 구덩이 안쪽 벽 (Encloses the collapse location) — An oblique interior face remains behind his torso; used as Maintains continuity with the preceding view from inside the pit.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the pit's subdued daytime illumination and readable ground detail, treating the loss of consciousness as an observed physical event rather than a subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness after being struck by a tranquilizer dart inside the pit. The scene text does not establish a settled posture, a facing direction, or the positions and contact points of his head, torso, arms, and legs during the collapse.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The trap pit remains open beneath the masked captors, and the earlier arrow remains lodged in the nearby tree. The camper is still immobilized in the swamp. 이현우: He is losing consciousness inside the pit after being struck by a tranquilizer dart. His previously treated leg remains bandaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마취 화살에 맞아 의식을 잃은 이현우의 몸이 구덩이 바닥을 향해 중심을 잃고 기울어진 순간.\n\nLOCATION (lock): At the bottom of an open-topped pit trap in the woodland near the contaminated marsh. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the downward rotation and descending move beside and slightly behind 이현우, holding an oblique side view with a gentle downward pitch toward the pit bottom. His unconscious head and torso tip diagonally from the upper left toward the lower center, knees giving way and eyes closed without purposeful visual attention; retain visible ground beneath him before contact. Let his falling position carry the change, keeping the established shooter-to-victim side and leaving 앰버 and the men outside this tighter frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 구덩이 바닥 (Visible beneath 이현우 before his falling body reaches it) — Extends below his descending head and torso; used as Preserves the remaining gap to the ground and makes the collapse physically legible; 구덩이 안쪽 벽 (Encloses the collapse location) — An oblique interior face remains behind his torso; used as Maintains continuity with the preceding view from inside the pit.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the pit's subdued daytime illumination and readable ground detail, treating the loss of consciousness as an observed physical event rather than a subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness after being struck by a tranquilizer dart inside the pit. The scene text does not establish a settled posture, a facing direction, or the positions and contact points of his head, torso, arms, and legs during the collapse.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The trap pit remains open beneath the masked captors, and the earlier arrow remains lodged in the nearby tree. The camper is still immobilized in the swamp. 이현우: He is losing consciousness inside the pit after being struck by a tranquilizer dart. His previously treated leg remains bandaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 마취 화살에 맞아 의식을 잃은 이현우의 몸이 구덩이 바닥을 향해 중심을 잃고 기울어진 순간.\n\nLOCATION (lock): At the bottom of an open-topped pit trap in the woodland near the contaminated marsh. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the downward rotation and descending move beside and slightly behind 이현우, holding an oblique side view with a gentle downward pitch toward the pit bottom. His unconscious head and torso tip diagonally from the upper left toward the lower center, knees giving way and eyes closed without purposeful visual attention; retain visible ground beneath him before contact. Let his falling position carry the change, keeping the established shooter-to-victim side and leaving 앰버 and the men outside this tighter frame.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 구덩이 바닥 (Visible beneath 이현우 before his falling body reaches it) — Extends below his descending head and torso; used as Preserves the remaining gap to the ground and makes the collapse physically legible; 구덩이 안쪽 벽 (Encloses the collapse location) — An oblique interior face remains behind his torso; used as Maintains continuity with the preceding view from inside the pit.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the pit's subdued daytime illumination and readable ground detail, treating the loss of consciousness as an observed physical event rather than a subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is losing consciousness after being struck by a tranquilizer dart inside the pit. The scene text does not establish a settled posture, a facing direction, or the positions and contact points of his head, torso, arms, and legs during the collapse.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The trap pit remains open beneath the masked captors, and the earlier arrow remains lodged in the nearby tree. The camper is still immobilized in the swamp. 이현우: He is losing consciousness inside the pit after being struck by a tranquilizer dart. His previously treated leg remains bandaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "고개와 상체가 화면 우측 아래를 향해 기울어지고 있으며, 눈은 감겨 있음.",
    "built_space": "나무 뿌리가 드러난 흙벽과 진흙 바닥이 있는 구덩이 내부. 카메라가 아래를 향하며 피사체 밑의 바닥 공간을 보여줌.",
    "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
    "hard_violations": [],
    "physics": "우측 무릎이 지면에 닿으려 하거나 구부러진 상태로, 중력에 의해 몸이 앞으로 쏠리며 쓰러지는 자연스러운 동작."
   },
   {
    "label": "B",
    "direction": "고개와 상체가 화면 우측 아래를 향해 비스듬히 기울어짐. 눈은 감겨 있음.",
    "built_space": "나무 뿌리가 드러난 흙벽과 젖은 진흙 바닥이 있는 구덩이 내부.",
    "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
    "hard_violations": [
     "[gemini-pro] 지지대 없이 허공에 떠 있는 인물 (발이 지면에 닿지 않은 채 레퍼런스 자세를 회전시킴)"
    ],
    "physics": "지탱하는 곳이 전혀 없음. 발끝이 진흙 바닥 위 공중에 명백히 떠 있으며, 낙하가 아닌 뻣뻣한 동상이 기울어진 듯한 비현실적 상태."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 구도와 쓰러지는 동작을 물리적으로 자연스럽게 구현하여 우수한 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "레퍼런스 이미지를 그대로 회전시켜 인물의 발이 허공에 떠 있는 치명적인 물리적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 상체가 화면 우측 아래를 향해 기울어지고 있으며, 눈은 감겨 있음.",
        "built_space": "나무 뿌리가 드러난 흙벽과 진흙 바닥이 있는 구덩이 내부. 카메라가 아래를 향하며 피사체 밑의 바닥 공간을 보여줌.",
        "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
        "hard_violations": [],
        "physics": "우측 무릎이 지면에 닿으려 하거나 구부러진 상태로, 중력에 의해 몸이 앞으로 쏠리며 쓰러지는 자연스러운 동작."
       },
       {
        "label": "B",
        "direction": "고개와 상체가 화면 우측 아래를 향해 비스듬히 기울어짐. 눈은 감겨 있음.",
        "built_space": "나무 뿌리가 드러난 흙벽과 젖은 진흙 바닥이 있는 구덩이 내부.",
        "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
        "hard_violations": [
         "지지대 없이 허공에 떠 있는 인물 (발이 지면에 닿지 않은 채 레퍼런스 자세를 회전시킴)"
        ],
        "physics": "지탱하는 곳이 전혀 없음. 발끝이 진흙 바닥 위 공중에 명백히 떠 있으며, 낙하가 아닌 뻣뻣한 동상이 기울어진 듯한 비현실적 상태."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트에 명시된 구도와 쓰러지는 동작을 물리적으로 자연스럽게 구현하여 우수한 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "레퍼런스 이미지를 그대로 회전시켜 인물의 발이 허공에 떠 있는 치명적인 물리적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개와 상체가 화면 우측 아래를 향해 기울어지고 있으며, 눈은 감겨 있음.",
        "built_space": "나무 뿌리가 드러난 흙벽과 진흙 바닥이 있는 구덩이 내부. 카메라가 아래를 향하며 피사체 밑의 바닥 공간을 보여줌.",
        "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
        "hard_violations": [],
        "physics": "우측 무릎이 지면에 닿으려 하거나 구부러진 상태로, 중력에 의해 몸이 앞으로 쏠리며 쓰러지는 자연스러운 동작."
       },
       {
        "label": "B",
        "direction": "고개와 상체가 화면 우측 아래를 향해 비스듬히 기울어짐. 눈은 감겨 있음.",
        "built_space": "나무 뿌리가 드러난 흙벽과 젖은 진흙 바닥이 있는 구덩이 내부.",
        "entities": "이현우 (레퍼런스와 일치하는 얼굴, 검은 머리, 흙이 묻은 셔츠와 바지 착용).",
        "hard_violations": [
         "지지대 없이 허공에 떠 있는 인물 (발이 지면에 닿지 않은 채 레퍼런스 자세를 회전시킴)"
        ],
        "physics": "지탱하는 곳이 전혀 없음. 발끝이 진흙 바닥 위 공중에 명백히 떠 있으며, 낙하가 아닌 뻣뻣한 동상이 기울어진 듯한 비현실적 상태."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "옆뒤에서 내려다보는 카메라와 접촉 전 바닥 간격은 더 충실하지만, 구도가 다소 넓고 머리의 하강이 약하며 두 후보 모두 이전 장면의 반팔·반바지 복장을 유지하지 못했다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "미디엄 크기와 옆으로 무너지는 동작은 읽히지만, 옆뒤가 아닌 앞쪽에서 가슴을 보여 주어 지정 카메라 축을 어겼고 이전 장면의 복장도 바뀌었다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "눈은 감겨 있어 의도적인 시선 대상이 없다. 얼굴은 오른쪽 아래의 진흙 바닥을 향하지만, 머리는 화면 상단에 남아 있고 몸통은 왼쪽 뒤로 젖혀져 있다. 따라서 지정된 상단 왼쪽에서 하단 중앙으로 머리와 몸통이 내려가는 방향은 충분히 드러나지 않는다. 겨누는 무기나 화살은 보이지 않는다.",
        "built_space": "뒤쪽과 왼쪽에 뿌리가 드러난 흙벽이 이어지고, 오른쪽과 몸 아래에는 돌과 얕은 물이 있는 진흙 바닥이 보인다. 별도 인공 설비는 없다. 인물은 구덩이 내부에 있으며 카메라는 등과 옆얼굴을 함께 보는 높은 옆뒤 위치다. 이전 장면의 흙벽·뿌리·젖은 바닥 재질은 이어진다. 발까지 들어와 지정 미디엄보다 다소 넓다.",
        "entities": "젊은 동아시아계 남성 한 명이며 짧고 헝클어진 검은 머리와 옆얼굴은 인물 참조에 대체로 부합한다. 앰버나 남자 포획자들은 없다. 다만 회갈색 긴팔 셔츠와 긴바지는 인물 참조 의상을 따랐을 뿐, 연속성이 고정된 이전 장면의 남색 반팔과 반바지에 맞지 않는다. 마취 화살은 식별되지 않고, 긴바지 때문에 다리 붕대도 확인할 수 없다. 프레임 밖 나무의 화살과 늪의 인물은 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "오른쪽 아래 부츠의 뒤꿈치 쪽이 바닥에 닿아 있고 무릎이 굽어 있어, 발을 축으로 지지를 잃고 주저앉는 순간으로 읽힌다. 엉덩이와 몸통 아래에는 아직 바닥과의 간격이 있다. 머리는 목에서 꺾이고 손은 아래로 처져 있다. 완전히 지지 없이 떠 있는 몸은 아니지만, 상체가 비교적 높고 뒤로 젖혀져 머리부터 아래로 무너지는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "눈은 감겨 있고 얼굴은 오른쪽 아래 바닥을 향한다. 몸통은 왼쪽에서 오른쪽으로 기울어 옆으로 쓰러지는 방향이 읽히지만, 머리는 여전히 화면 위쪽 중앙에 있다. 카메라가 가슴과 얼굴 앞면을 보여 주므로 지정된 옆뒤 관찰 방향과 다르다. 무기나 마취 화살의 방향은 확인할 수 없다.",
        "built_space": "뒤쪽 흙벽과 오른쪽으로 돌아가는 벽면, 노출된 뿌리, 진흙과 돌이 있는 바닥이 보인다. 별도 인공 설비나 추가 인물은 없다. 몸통 아래와 오른쪽에 접촉 전 바닥 공간이 남아 있으며 장소 재질도 이전 장면과 대체로 이어진다. 허벅지 부근까지 크게 잡은 크기는 미디엄에 가깝지만 카메라는 인물 앞쪽에 놓여 있다.",
        "entities": "짧고 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명으로, 얼굴과 체격은 인물 참조와 대체로 맞는다. 회갈색 긴팔 셔츠와 긴바지는 이전 장면의 남색 반팔·반바지와 불일치한다. 마취 화살은 식별되지 않는다. 붕대가 있어야 할 하퇴는 프레임 밖이므로 붕대 누락으로 판단하지 않는다. 다른 인물이나 문자 삽입은 없다.",
        "hard_violations": [],
        "physics": "화면 아래의 굽힌 무릎 부근이 바닥에 내려앉는 지지점으로 읽히며, 하퇴와 발의 정확한 접촉은 잘려 있다. 상체는 그 지지점에서 옆으로 넘어가고 오른쪽 아래 손은 바닥에 닿기 전 축 늘어져 있다. 왼쪽 팔은 몸에서 벌어져 있어 완전히 이완된 인상은 약하지만, 붕괴 중 관성으로 가능한 범위이며 몸 전체가 근거 없이 공중에 떠 있다고 볼 수는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "옆뒤에서 내려다보는 카메라와 접촉 전 바닥 간격은 더 충실하지만, 구도가 다소 넓고 머리의 하강이 약하며 두 후보 모두 이전 장면의 반팔·반바지 복장을 유지하지 못했다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "미디엄 크기와 옆으로 무너지는 동작은 읽히지만, 옆뒤가 아닌 앞쪽에서 가슴을 보여 주어 지정 카메라 축을 어겼고 이전 장면의 복장도 바뀌었다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "눈은 감겨 있어 의도적인 시선 대상이 없다. 얼굴은 오른쪽 아래의 진흙 바닥을 향하지만, 머리는 화면 상단에 남아 있고 몸통은 왼쪽 뒤로 젖혀져 있다. 따라서 지정된 상단 왼쪽에서 하단 중앙으로 머리와 몸통이 내려가는 방향은 충분히 드러나지 않는다. 겨누는 무기나 화살은 보이지 않는다.",
        "built_space": "뒤쪽과 왼쪽에 뿌리가 드러난 흙벽이 이어지고, 오른쪽과 몸 아래에는 돌과 얕은 물이 있는 진흙 바닥이 보인다. 별도 인공 설비는 없다. 인물은 구덩이 내부에 있으며 카메라는 등과 옆얼굴을 함께 보는 높은 옆뒤 위치다. 이전 장면의 흙벽·뿌리·젖은 바닥 재질은 이어진다. 발까지 들어와 지정 미디엄보다 다소 넓다.",
        "entities": "젊은 동아시아계 남성 한 명이며 짧고 헝클어진 검은 머리와 옆얼굴은 인물 참조에 대체로 부합한다. 앰버나 남자 포획자들은 없다. 다만 회갈색 긴팔 셔츠와 긴바지는 인물 참조 의상을 따랐을 뿐, 연속성이 고정된 이전 장면의 남색 반팔과 반바지에 맞지 않는다. 마취 화살은 식별되지 않고, 긴바지 때문에 다리 붕대도 확인할 수 없다. 프레임 밖 나무의 화살과 늪의 인물은 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "오른쪽 아래 부츠의 뒤꿈치 쪽이 바닥에 닿아 있고 무릎이 굽어 있어, 발을 축으로 지지를 잃고 주저앉는 순간으로 읽힌다. 엉덩이와 몸통 아래에는 아직 바닥과의 간격이 있다. 머리는 목에서 꺾이고 손은 아래로 처져 있다. 완전히 지지 없이 떠 있는 몸은 아니지만, 상체가 비교적 높고 뒤로 젖혀져 머리부터 아래로 무너지는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "눈은 감겨 있고 얼굴은 오른쪽 아래 바닥을 향한다. 몸통은 왼쪽에서 오른쪽으로 기울어 옆으로 쓰러지는 방향이 읽히지만, 머리는 여전히 화면 위쪽 중앙에 있다. 카메라가 가슴과 얼굴 앞면을 보여 주므로 지정된 옆뒤 관찰 방향과 다르다. 무기나 마취 화살의 방향은 확인할 수 없다.",
        "built_space": "뒤쪽 흙벽과 오른쪽으로 돌아가는 벽면, 노출된 뿌리, 진흙과 돌이 있는 바닥이 보인다. 별도 인공 설비나 추가 인물은 없다. 몸통 아래와 오른쪽에 접촉 전 바닥 공간이 남아 있으며 장소 재질도 이전 장면과 대체로 이어진다. 허벅지 부근까지 크게 잡은 크기는 미디엄에 가깝지만 카메라는 인물 앞쪽에 놓여 있다.",
        "entities": "짧고 헝클어진 검은 머리의 젊은 동아시아계 남성 한 명으로, 얼굴과 체격은 인물 참조와 대체로 맞는다. 회갈색 긴팔 셔츠와 긴바지는 이전 장면의 남색 반팔·반바지와 불일치한다. 마취 화살은 식별되지 않는다. 붕대가 있어야 할 하퇴는 프레임 밖이므로 붕대 누락으로 판단하지 않는다. 다른 인물이나 문자 삽입은 없다.",
        "hard_violations": [],
        "physics": "화면 아래의 굽힌 무릎 부근이 바닥에 내려앉는 지지점으로 읽히며, 하퇴와 발의 정확한 접촉은 잘려 있다. 상체는 그 지지점에서 옆으로 넘어가고 오른쪽 아래 손은 바닥에 닿기 전 축 늘어져 있다. 왼쪽 팔은 몸에서 벌어져 있어 완전히 이완된 인상은 약하지만, 붕괴 중 관성으로 가능한 범위이며 몸 전체가 근거 없이 공중에 떠 있다고 볼 수는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 지지대 없이 허공에 떠 있는 인물 (발이 지면에 닿지 않은 채 레퍼런스 자세를 회전시킴)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "프롬프트에 명시된 구도와 쓰러지는 동작을 물리적으로 자연스럽게 구현하여 우수한 결과를 보여줍니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "레퍼런스 이미지를 그대로 회전시켜 인물의 발이 허공에 떠 있는 치명적인 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 지지대 없이 허공에 떠 있는 인물 (발이 지면에 닿지 않은 채 레퍼런스 자세를 회전시킴)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S56sh23_sel.png",
    "asset_id": "5be1ff18-dcdd-489c-aaec-bc0c0ad0b11b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1164888>",
    "asset_id": "a6131853-595d-4497-ab83-bbe9e4e4fa3b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bf3-8506-7759-b603-2f7913054cb1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh23"
  }
 },
 "S57sh3::signage": {
  "fp": "c4defff8617f5279",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4817d7f679f8d484": {
  "subjects": [],
  "subject_text": "익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞\n낡은 한옥들이 좁은 골목과 중앙 공터를 둘러싼 마을. 높은 물탱크 탑, 성당 앞 노천 격투장과 관중석·단상, 외곽도로와 폐창고가 이어진다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L158",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::village_procession_street": {
  "input_fingerprint": "1b1923459642a3a4",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "village_procession_street",
    "tags": [
     "S57sh3",
     "S57sh5"
    ]
   },
   "context_sig": "14a8d30b26782f2e"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a barred cage carried on the open load bed of a truck moving through the village streets.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우와 앰버가 쇠창살(케이지) 안에 갇힌 채 트럭 위에 실려서 끌려가고 있는 상황이다.\n- 한쪽에는 사제복을 입은 하회탈 남자가 주민들에게 물병을 나눠주고 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a barred cage carried on the open load bed of a truck moving through the village streets.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 현우와 앰버가 쇠창살(케이지) 안에 갇힌 채 트럭 위에 실려서 끌려가고 있는 상황이다.\n- 한쪽에는 사제복을 입은 하회탈 남자가 주민들에게 물병을 나눠주고 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_procession_street_9fa083.png",
  "asset_id": "02d6c885-c945-43eb-893f-b67c4e352ff2",
  "input_asset_ids": [
   "81a9e533-7359-4c08-908d-202e415b09ae"
  ],
  "origin_tag": "S57sh3",
  "place_text": "Inside a barred cage carried on the open load bed of a truck moving through the village streets.",
  "origin_inputs": {
   "place_text": "Inside a barred cage carried on the open load bed of a truck moving through the village streets.",
   "time_of_day_en": "day",
   "conti_asset_id": "81a9e533-7359-4c08-908d-202e415b09ae"
  }
 },
 "S57sh3::bgfirst_bg": {
  "input_fingerprint": "ad345a07a4f39d23",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 주행 중인 트럭 적재함 위의 쇠창살 케이지 안에 갇혀 있는 이현우와 앰버의 전경.\n\nLOCATION (lock): Inside a barred cage carried on the open load bed of a truck moving through the village streets.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track steadily outside the moving truck's rear quarter, above the captives, looking diagonally downward through the cage bars in a direct observational wide shot. Place 이현우 left of center, leaning toward 앰버 as she finishes shaking her head on the right; both turn their attention toward the shouting people beyond the off-screen side of the cage. Keep their complete figures and a strip of truck bed visible, with separated foreground bars framing rather than crossing their faces.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 쇠창살 케이지 (Contains both captives on the moving truck) — The rear corner and adjoining side bars are seen obliquely; used as Sparse foreground divisions establish imprisonment without concealing faces; 트럭 적재함 (Carrying the cage while the truck moves) — A strip of the upper bed surface remains visible below the captives; used as Provides spatial support and a scale reference beneath the cage.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve facial readability without softening the confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 주행 중인 트럭 적재함 위의 쇠창살 케이지 안에 갇혀 있는 이현우와 앰버의 전경.\n\nLOCATION (lock): Inside a barred cage carried on the open load bed of a truck moving through the village streets.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track steadily outside the moving truck's rear quarter, above the captives, looking diagonally downward through the cage bars in a direct observational wide shot. Place 이현우 left of center, leaning toward 앰버 as she finishes shaking her head on the right; both turn their attention toward the shouting people beyond the off-screen side of the cage. Keep their complete figures and a strip of truck bed visible, with separated foreground bars framing rather than crossing their faces.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 쇠창살 케이지 (Contains both captives on the moving truck) — The rear corner and adjoining side bars are seen obliquely; used as Sparse foreground divisions establish imprisonment without concealing faces; 트럭 적재함 (Carrying the cage while the truck moves) — A strip of the upper bed surface remains visible below the captives; used as Provides spatial support and a scale reference beneath the cage.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve facial readability without softening the confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh3__bgfirst_bg.png",
  "asset_id": "99754353-0cd3-47d2-a088-6abf902c57fc",
  "input_asset_ids": [
   "81a9e533-7359-4c08-908d-202e415b09ae",
   "02d6c885-c945-43eb-893f-b67c4e352ff2"
  ]
 },
 "S57sh3": {
  "input_fingerprint": "37321ef45688625e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 주행 중인 트럭 적재함 위의 쇠창살 케이지 안에 갇혀 있는 이현우와 앰버의 전경.\n\nLOCATION (lock): Inside a barred cage carried on the open load bed of a truck moving through the village streets. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track steadily outside the moving truck's rear quarter, above the captives, looking diagonally downward through the cage bars in a direct observational wide shot. Place 이현우 left of center, leaning toward 앰버 as she finishes shaking her head on the right; both turn their attention toward the shouting people beyond the off-screen side of the cage. Keep their complete figures and a strip of truck bed visible, with separated foreground bars framing rather than crossing their faces.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 쇠창살 케이지 (Contains both captives on the moving truck) — The rear corner and adjoining side bars are seen obliquely; used as Sparse foreground divisions establish imprisonment without concealing faces; 트럭 적재함 (Carrying the cage while the truck moves) — A strip of the upper bed surface remains visible below the captives; used as Provides spatial support and a scale reference beneath the cage.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve facial readability without softening the confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage rides on the truck bed through the village; a tall water-tank tower stands nearby. 이현우: He is bound inside the cage, newly awake after the tranquilizer dart. His previously treated leg remains bandaged. 앰버: She is bound inside the cage and has no new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 주행 중인 트럭 적재함 위의 쇠창살 케이지 안에 갇혀 있는 이현우와 앰버의 전경.\n\nLOCATION (lock): Inside a barred cage carried on the open load bed of a truck moving through the village streets. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track steadily outside the moving truck's rear quarter, above the captives, looking diagonally downward through the cage bars in a direct observational wide shot. Place 이현우 left of center, leaning toward 앰버 as she finishes shaking her head on the right; both turn their attention toward the shouting people beyond the off-screen side of the cage. Keep their complete figures and a strip of truck bed visible, with separated foreground bars framing rather than crossing their faces.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 쇠창살 케이지 (Contains both captives on the moving truck) — The rear corner and adjoining side bars are seen obliquely; used as Sparse foreground divisions establish imprisonment without concealing faces; 트럭 적재함 (Carrying the cage while the truck moves) — A strip of the upper bed surface remains visible below the captives; used as Provides spatial support and a scale reference beneath the cage.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve facial readability without softening the confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage rides on the truck bed through the village; a tall water-tank tower stands nearby. 이현우: He is bound inside the cage, newly awake after the tranquilizer dart. His previously treated leg remains bandaged. 앰버: She is bound inside the cage and has no new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 주행 중인 트럭 적재함 위의 쇠창살 케이지 안에 갇혀 있는 이현우와 앰버의 전경.\n\nLOCATION (lock): Inside a barred cage carried on the open load bed of a truck moving through the village streets. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track steadily outside the moving truck's rear quarter, above the captives, looking diagonally downward through the cage bars in a direct observational wide shot. Place 이현우 left of center, leaning toward 앰버 as she finishes shaking her head on the right; both turn their attention toward the shouting people beyond the off-screen side of the cage. Keep their complete figures and a strip of truck bed visible, with separated foreground bars framing rather than crossing their faces.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 쇠창살 케이지 (Contains both captives on the moving truck) — The rear corner and adjoining side bars are seen obliquely; used as Sparse foreground divisions establish imprisonment without concealing faces; 트럭 적재함 (Carrying the cage while the truck moves) — A strip of the upper bed surface remains visible below the captives; used as Provides spatial support and a scale reference beneath the cage.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and controlled contrast preserve facial readability without softening the confinement.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A barred cage rides on the truck bed through the village; a tall water-tank tower stands nearby. 이현우: He is bound inside the cage, newly awake after the tranquilizer dart. His previously treated leg remains bandaged. 앰버: She is bound inside the cage and has no new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh3__bgfirst_bg.png",
     "asset_id": "99754353-0cd3-47d2-a088-6abf902c57fc",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S57sh3.png",
     "asset_id": "81a9e533-7359-4c08-908d-202e415b09ae",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_procession_street_9fa083.png",
     "asset_id": "02d6c885-c945-43eb-893f-b67c4e352ff2",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "현우는 왼쪽에서 앰버 쪽으로 상체를 기울이고 얼굴과 시선을 화면 오른쪽으로 돌린다. 앰버 역시 오른쪽 케이지 밖을 본다. 두 사람 모두 화면 밖에서 소리치는 사람들에게 주의를 돌리는 설정과 맞는 방향이며, 실제 대상 인물은 보이지 않는다. 앰버가 고개 흔들기를 막 끝냈다는 순간까지는 명확하지 않다. 도로와 담장의 흐림은 트럭의 주행을 뒷받침한다.",
    "built_space": "트럭 적재함 하나 위에 녹슨 철제 케이지 하나가 있고, 후방 가장자리와 인접 측면을 비스듬히 내려다본다. 오른쪽에 운전석 일부와 사이드미러 하나, 배경에 물탱크 탑 하나와 교회 첨탑 하나가 보인다. 기와지붕, 돌담, 전신주가 장소 참조와 대응한다. 두 사람 아래에 적재함 바닥이 충분히 보이고 얼굴을 가로지르는 전경 창살은 없다. 다만 가까운 오른쪽 상부 난간이 앰버의 한쪽 다리와 발을 가린다.",
    "entities": "보이는 인물은 현우와 앰버 두 명뿐이다. 현우는 헝클어진 짧은 검은 머리의 동아시아계 십 대 후반 남성으로 보이며, 마른 체형과 오염된 어두운 셔츠·바지, 인이어 장치, 피가 묻은 다리 붕대가 확인된다. 앰버는 금발의 밝은 피부를 가진 어린 여자아이로, 참조의 대략적인 연령과 외형에 맞고 미래형 방진 마스크, 카키 작업복, 가죽 공구 벨트를 착용했다. 새로운 부상은 보이지 않는다. 두 사람 모두 팔을 뒤로 두었지만 결박 끈은 확인되지 않는다.",
    "hard_violations": [],
    "physics": "현우는 벌린 두 발을 적재함 바닥에 디디고 무릎과 허리를 굽혀 균형을 잡는다. 앰버도 보이는 부츠를 바닥에 디디며 서 있고 다른 발은 난간에 가려져 있다. 두 사람의 기울어진 자세는 움직이는 차량 위에서 가능한 자세다. 케이지는 적재함에 얹혀 있고 마스크와 벨트도 신체에 착용되어 있어 떠 있는 물체나 지지 없는 몸은 보이지 않는다."
   },
   {
    "label": "A",
    "direction": "현우는 왼쪽에서 오른쪽의 앰버 쪽으로 몸과 고개를 돌리고, 앰버는 더 오른쪽의 케이지 밖을 바라본다. 둘의 관심 방향은 화면 밖 사람들을 향한다는 지시와 양립한다. 앰버의 고개 흔들기 종료 동작은 뚜렷하지 않다. 도로와 배경이 선명하여 트럭이 주행 중이라는 시각적 단서는 약하다.",
    "built_space": "적재함 하나와 철제 케이지 하나 안에 두 사람이 나란히 바닥에 앉아 있다. 케이지의 후면과 오른쪽 측면이 사선으로 보이며, 오른쪽 운전석 일부와 사이드미러 하나, 배경의 물탱크 탑 하나와 교회 첨탑 하나가 참조 장소와 대응한다. 두 사람의 전신과 바닥이 보이지만, 가까운 수직 창살 여러 개가 다리와 몸통을 분할하고 하나는 앰버의 머리카락 앞을 지난다. 눈과 마스크는 대체로 가리지 않는다. 카메라는 인물보다 높지만 A보다 내려다보는 각도가 약하다.",
    "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 짧은 검은 머리, 젊은 동아시아계 남성 외형, 오염된 어두운 옷, 인이어 장치와 다리 붕대가 보인다. 앰버는 금발의 어린 여자아이로 표현되며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 두 사람의 상체와 팔을 감싼 밧줄이 명확해 결박 상태가 A보다 분명하다. 앰버에게 새 부상은 보이지 않는다.",
    "hard_violations": [],
    "physics": "두 사람 모두 엉덩이와 다리를 적재함 바닥에 두고 앉아 있어 체중 지지가 명확하다. 현우는 한쪽 무릎을 굽히고 붕대를 감은 다리를 앞으로 뻗으며, 앰버도 다리를 앞으로 뻗는다. 팔이 뒤로 묶인 채 가능한 앉은 자세이고, 약간 들린 신발 끝은 바닥에 지지된 다리에 연결되어 있다. 케이지와 인물 모두 적재함에 지지되며 물리적으로 불가능한 부유는 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "후방 사선 하이앵글과 얼굴을 가리지 않는 전경 창살, 앰버 쪽으로 기운 현우와 두 사람의 화면 밖 시선이 더 정확하지만, 앰버의 한쪽 하체가 난간에 가려지고 결박 자체는 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 사람의 결박과 전신, 장소 구조는 명확하지만, 촘촘한 전경 창살과 상대적으로 약한 내려다보기 구도 및 주행 표현이 지정된 관찰 와이드숏에 덜 부합한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 왼쪽에서 앰버 쪽으로 상체를 기울이고 얼굴과 시선을 화면 오른쪽으로 돌린다. 앰버 역시 오른쪽 케이지 밖을 본다. 두 사람 모두 화면 밖에서 소리치는 사람들에게 주의를 돌리는 설정과 맞는 방향이며, 실제 대상 인물은 보이지 않는다. 앰버가 고개 흔들기를 막 끝냈다는 순간까지는 명확하지 않다. 도로와 담장의 흐림은 트럭의 주행을 뒷받침한다.",
        "built_space": "트럭 적재함 하나 위에 녹슨 철제 케이지 하나가 있고, 후방 가장자리와 인접 측면을 비스듬히 내려다본다. 오른쪽에 운전석 일부와 사이드미러 하나, 배경에 물탱크 탑 하나와 교회 첨탑 하나가 보인다. 기와지붕, 돌담, 전신주가 장소 참조와 대응한다. 두 사람 아래에 적재함 바닥이 충분히 보이고 얼굴을 가로지르는 전경 창살은 없다. 다만 가까운 오른쪽 상부 난간이 앰버의 한쪽 다리와 발을 가린다.",
        "entities": "보이는 인물은 현우와 앰버 두 명뿐이다. 현우는 헝클어진 짧은 검은 머리의 동아시아계 십 대 후반 남성으로 보이며, 마른 체형과 오염된 어두운 셔츠·바지, 인이어 장치, 피가 묻은 다리 붕대가 확인된다. 앰버는 금발의 밝은 피부를 가진 어린 여자아이로, 참조의 대략적인 연령과 외형에 맞고 미래형 방진 마스크, 카키 작업복, 가죽 공구 벨트를 착용했다. 새로운 부상은 보이지 않는다. 두 사람 모두 팔을 뒤로 두었지만 결박 끈은 확인되지 않는다.",
        "hard_violations": [],
        "physics": "현우는 벌린 두 발을 적재함 바닥에 디디고 무릎과 허리를 굽혀 균형을 잡는다. 앰버도 보이는 부츠를 바닥에 디디며 서 있고 다른 발은 난간에 가려져 있다. 두 사람의 기울어진 자세는 움직이는 차량 위에서 가능한 자세다. 케이지는 적재함에 얹혀 있고 마스크와 벨트도 신체에 착용되어 있어 떠 있는 물체나 지지 없는 몸은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "현우는 왼쪽에서 오른쪽의 앰버 쪽으로 몸과 고개를 돌리고, 앰버는 더 오른쪽의 케이지 밖을 바라본다. 둘의 관심 방향은 화면 밖 사람들을 향한다는 지시와 양립한다. 앰버의 고개 흔들기 종료 동작은 뚜렷하지 않다. 도로와 배경이 선명하여 트럭이 주행 중이라는 시각적 단서는 약하다.",
        "built_space": "적재함 하나와 철제 케이지 하나 안에 두 사람이 나란히 바닥에 앉아 있다. 케이지의 후면과 오른쪽 측면이 사선으로 보이며, 오른쪽 운전석 일부와 사이드미러 하나, 배경의 물탱크 탑 하나와 교회 첨탑 하나가 참조 장소와 대응한다. 두 사람의 전신과 바닥이 보이지만, 가까운 수직 창살 여러 개가 다리와 몸통을 분할하고 하나는 앰버의 머리카락 앞을 지난다. 눈과 마스크는 대체로 가리지 않는다. 카메라는 인물보다 높지만 A보다 내려다보는 각도가 약하다.",
        "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 짧은 검은 머리, 젊은 동아시아계 남성 외형, 오염된 어두운 옷, 인이어 장치와 다리 붕대가 보인다. 앰버는 금발의 어린 여자아이로 표현되며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 두 사람의 상체와 팔을 감싼 밧줄이 명확해 결박 상태가 A보다 분명하다. 앰버에게 새 부상은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 엉덩이와 다리를 적재함 바닥에 두고 앉아 있어 체중 지지가 명확하다. 현우는 한쪽 무릎을 굽히고 붕대를 감은 다리를 앞으로 뻗으며, 앰버도 다리를 앞으로 뻗는다. 팔이 뒤로 묶인 채 가능한 앉은 자세이고, 약간 들린 신발 끝은 바닥에 지지된 다리에 연결되어 있다. 케이지와 인물 모두 적재함에 지지되며 물리적으로 불가능한 부유는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "후방 사선 하이앵글과 얼굴을 가리지 않는 전경 창살, 앰버 쪽으로 기운 현우와 두 사람의 화면 밖 시선이 더 정확하지만, 앰버의 한쪽 하체가 난간에 가려지고 결박 자체는 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 사람의 결박과 전신, 장소 구조는 명확하지만, 촘촘한 전경 창살과 상대적으로 약한 내려다보기 구도 및 주행 표현이 지정된 관찰 와이드숏에 덜 부합한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 왼쪽에서 앰버 쪽으로 상체를 기울이고 얼굴과 시선을 화면 오른쪽으로 돌린다. 앰버 역시 오른쪽 케이지 밖을 본다. 두 사람 모두 화면 밖에서 소리치는 사람들에게 주의를 돌리는 설정과 맞는 방향이며, 실제 대상 인물은 보이지 않는다. 앰버가 고개 흔들기를 막 끝냈다는 순간까지는 명확하지 않다. 도로와 담장의 흐림은 트럭의 주행을 뒷받침한다.",
        "built_space": "트럭 적재함 하나 위에 녹슨 철제 케이지 하나가 있고, 후방 가장자리와 인접 측면을 비스듬히 내려다본다. 오른쪽에 운전석 일부와 사이드미러 하나, 배경에 물탱크 탑 하나와 교회 첨탑 하나가 보인다. 기와지붕, 돌담, 전신주가 장소 참조와 대응한다. 두 사람 아래에 적재함 바닥이 충분히 보이고 얼굴을 가로지르는 전경 창살은 없다. 다만 가까운 오른쪽 상부 난간이 앰버의 한쪽 다리와 발을 가린다.",
        "entities": "보이는 인물은 현우와 앰버 두 명뿐이다. 현우는 헝클어진 짧은 검은 머리의 동아시아계 십 대 후반 남성으로 보이며, 마른 체형과 오염된 어두운 셔츠·바지, 인이어 장치, 피가 묻은 다리 붕대가 확인된다. 앰버는 금발의 밝은 피부를 가진 어린 여자아이로, 참조의 대략적인 연령과 외형에 맞고 미래형 방진 마스크, 카키 작업복, 가죽 공구 벨트를 착용했다. 새로운 부상은 보이지 않는다. 두 사람 모두 팔을 뒤로 두었지만 결박 끈은 확인되지 않는다.",
        "hard_violations": [],
        "physics": "현우는 벌린 두 발을 적재함 바닥에 디디고 무릎과 허리를 굽혀 균형을 잡는다. 앰버도 보이는 부츠를 바닥에 디디며 서 있고 다른 발은 난간에 가려져 있다. 두 사람의 기울어진 자세는 움직이는 차량 위에서 가능한 자세다. 케이지는 적재함에 얹혀 있고 마스크와 벨트도 신체에 착용되어 있어 떠 있는 물체나 지지 없는 몸은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "현우는 왼쪽에서 오른쪽의 앰버 쪽으로 몸과 고개를 돌리고, 앰버는 더 오른쪽의 케이지 밖을 바라본다. 둘의 관심 방향은 화면 밖 사람들을 향한다는 지시와 양립한다. 앰버의 고개 흔들기 종료 동작은 뚜렷하지 않다. 도로와 배경이 선명하여 트럭이 주행 중이라는 시각적 단서는 약하다.",
        "built_space": "적재함 하나와 철제 케이지 하나 안에 두 사람이 나란히 바닥에 앉아 있다. 케이지의 후면과 오른쪽 측면이 사선으로 보이며, 오른쪽 운전석 일부와 사이드미러 하나, 배경의 물탱크 탑 하나와 교회 첨탑 하나가 참조 장소와 대응한다. 두 사람의 전신과 바닥이 보이지만, 가까운 수직 창살 여러 개가 다리와 몸통을 분할하고 하나는 앰버의 머리카락 앞을 지난다. 눈과 마스크는 대체로 가리지 않는다. 카메라는 인물보다 높지만 A보다 내려다보는 각도가 약하다.",
        "entities": "인물은 현우와 앰버 두 명뿐이다. 현우의 짧은 검은 머리, 젊은 동아시아계 남성 외형, 오염된 어두운 옷, 인이어 장치와 다리 붕대가 보인다. 앰버는 금발의 어린 여자아이로 표현되며 방진 마스크, 카키 작업복, 가죽 공구 벨트를 갖췄다. 두 사람의 상체와 팔을 감싼 밧줄이 명확해 결박 상태가 A보다 분명하다. 앰버에게 새 부상은 보이지 않는다.",
        "hard_violations": [],
        "physics": "두 사람 모두 엉덩이와 다리를 적재함 바닥에 두고 앉아 있어 체중 지지가 명확하다. 현우는 한쪽 무릎을 굽히고 붕대를 감은 다리를 앞으로 뻗으며, 앰버도 다리를 앞으로 뻗는다. 팔이 뒤로 묶인 채 가능한 앉은 자세이고, 약간 들린 신발 끝은 바닥에 지지된 다리에 연결되어 있다. 케이지와 인물 모두 적재함에 지지되며 물리적으로 불가능한 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 8,
   "A": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "후방 사선 하이앵글과 얼굴을 가리지 않는 전경 창살, 앰버 쪽으로 기운 현우와 두 사람의 화면 밖 시선이 더 정확하지만, 앰버의 한쪽 하체가 난간에 가려지고 결박 자체는 확인되지 않는다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "두 사람의 결박과 전신, 장소 구조는 명확하지만, 촘촘한 전경 창살과 상대적으로 약한 내려다보기 구도 및 주행 표현이 지정된 관찰 와이드숏에 덜 부합한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_procession_street_9fa083.png",
    "asset_id": "02d6c885-c945-43eb-893f-b67c4e352ff2",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0bf8-e2d8-7090-8e19-1724011334cc",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh3__bgfirst_bg.png",
   "bg_asset_id": "99754353-0cd3-47d2-a088-6abf902c57fc",
   "bg_record_key": "S57sh3::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "village_procession_street",
   "groupbg_asset_id": "02d6c885-c945-43eb-893f-b67c4e352ff2"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S57sh5::signage": {
  "fp": "a2ff9ddb9ccbbbbc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh5": {
  "input_fingerprint": "ed1e42e00ba3788b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물을 나눠주는 사제복 차림의 남자가 건네는 물병을 향해 짓무른 얼굴의 사람들이 미친 듯이 손을 뻗는 찰나.\n\nLOCATION (lock): At an outdoor water-distribution point beside the village street traveled by the prisoner truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at adult shoulder height along the route-facing crowd edge, oblique to the exchange, with a slight downward tilt toward the bottle. Put the masked distributor at the right edge extending the small bottle into the central gap, while reaching hands cross the lower foreground and visibly disfigured recipients remain readable beyond them. Their attention converges on the handoff below their eyelines; stagger elbows, fingers, shoulder angles and forward weight shifts so the desperation is collective without duplicating poses, emphasizing proximity rather than a lighting change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 물병 (Being extended toward the reaching recipients) — Seen obliquely between the distributor's hand and the approaching hands; used as A small central focal point linking the foreground hands and middle-distance faces; 분배자의 하회탈 (Worn by the man distributing water) — Its sculpted face is seen in partial profile toward the recipients; used as Contrasts the distributor's concealed expression with the exposed recipients.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure and restrained contrast, keeping the damaged faces legible without theatrical accent lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cage remains on the truck bed, and the tall water-tank tower remains a village landmark.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사제복 차림의 남자 right now, so 사제복 차림의 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사제복 차림의 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물을 나눠주는 사제복 차림의 남자가 건네는 물병을 향해 짓무른 얼굴의 사람들이 미친 듯이 손을 뻗는 찰나.\n\nLOCATION (lock): At an outdoor water-distribution point beside the village street traveled by the prisoner truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at adult shoulder height along the route-facing crowd edge, oblique to the exchange, with a slight downward tilt toward the bottle. Put the masked distributor at the right edge extending the small bottle into the central gap, while reaching hands cross the lower foreground and visibly disfigured recipients remain readable beyond them. Their attention converges on the handoff below their eyelines; stagger elbows, fingers, shoulder angles and forward weight shifts so the desperation is collective without duplicating poses, emphasizing proximity rather than a lighting change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 물병 (Being extended toward the reaching recipients) — Seen obliquely between the distributor's hand and the approaching hands; used as A small central focal point linking the foreground hands and middle-distance faces; 분배자의 하회탈 (Worn by the man distributing water) — Its sculpted face is seen in partial profile toward the recipients; used as Contrasts the distributor's concealed expression with the exposed recipients.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure and restrained contrast, keeping the damaged faces legible without theatrical accent lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cage remains on the truck bed, and the tall water-tank tower remains a village landmark.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사제복 차림의 남자 right now, so 사제복 차림의 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사제복 차림의 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물을 나눠주는 사제복 차림의 남자가 건네는 물병을 향해 짓무른 얼굴의 사람들이 미친 듯이 손을 뻗는 찰나.\n\nLOCATION (lock): At an outdoor water-distribution point beside the village street traveled by the prisoner truck. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at adult shoulder height along the route-facing crowd edge, oblique to the exchange, with a slight downward tilt toward the bottle. Put the masked distributor at the right edge extending the small bottle into the central gap, while reaching hands cross the lower foreground and visibly disfigured recipients remain readable beyond them. Their attention converges on the handoff below their eyelines; stagger elbows, fingers, shoulder angles and forward weight shifts so the desperation is collective without duplicating poses, emphasizing proximity rather than a lighting change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 물병 (Being extended toward the reaching recipients) — Seen obliquely between the distributor's hand and the approaching hands; used as A small central focal point linking the foreground hands and middle-distance faces; 분배자의 하회탈 (Worn by the man distributing water) — Its sculpted face is seen in partial profile toward the recipients; used as Contrasts the distributor's concealed expression with the exposed recipients.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight exposure and restrained contrast, keeping the damaged faces legible without theatrical accent lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cage remains on the truck bed, and the tall water-tank tower remains a village landmark.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 사제복 차림의 남자 right now, so 사제복 차림의 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 사제복 차림의 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "사제가 화면 우측에서 중앙으로 물병을 내밀고 있으며, 군중의 시선과 뻗은 손들이 정확히 물병에 집중됨.",
    "built_space": "레퍼런스와 동일한 거리 배경(돌담, 급수탑, 교회 첨탑)이 보임. 군중과 사제 사이에 금속 철책이 있으나 레퍼런스의 트럭 철창 형태와는 약간 다름. 카메라는 지시된 대로 어깨 높이에 위치함.",
    "entities": "사제는 검은 사제복을 입고 전통 하회탈을 착용함. 군중은 프롬프트의 요구대로 상처 입고 짓무른 얼굴을 매우 사실적으로 보여줌.",
    "hard_violations": [],
    "physics": "사제의 장갑 낀 손이 물병을 자연스럽게 쥐고 있으며, 군중이 뻗은 여러 갈래의 손과 팔, 체중이 실린 자세가 철책에 기대어 물리적으로 타당하게 지탱됨."
   },
   {
    "label": "B",
    "direction": "사제가 물병을 건네고 군중이 이를 향해 시선과 손을 뻗고 있음.",
    "built_space": "배경에 돌담, 급수탑, 교회가 올바르게 위치함. 사제 뒤쪽으로 레퍼런스에서 요구된 트럭의 금속 철창 구조물이 보임.",
    "entities": "사제복과 하회탈은 묘사되었으나, 군중의 얼굴은 흙먼지가 묻은 정도로 연출되어 '짓무른' 상태의 묘사가 부족함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (물병 아래로 손을 뻗은 중앙 여성의 손가락이 엄지를 포함해 총 6개임)"
    ],
    "physics": "사제가 물병 뚜껑 부근을 손가락으로 잡고 있음. 중앙 군중의 뻗은 손에서 심각한 손가락 개수 오류가 발생해 인체 구조에 어긋남."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 명시한 전경의 손 교차 구도와 군중의 짓무른 얼굴을 훌륭하게 구현했으며, 지정된 배경과 사제의 복장 및 하회탈을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "레퍼런스의 철창 구조는 잘 반영했으나, 물병을 향해 뻗은 중앙 여성의 손가락이 6개로 렌더링된 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사제가 화면 우측에서 중앙으로 물병을 내밀고 있으며, 군중의 시선과 뻗은 손들이 정확히 물병에 집중됨.",
        "built_space": "레퍼런스와 동일한 거리 배경(돌담, 급수탑, 교회 첨탑)이 보임. 군중과 사제 사이에 금속 철책이 있으나 레퍼런스의 트럭 철창 형태와는 약간 다름. 카메라는 지시된 대로 어깨 높이에 위치함.",
        "entities": "사제는 검은 사제복을 입고 전통 하회탈을 착용함. 군중은 프롬프트의 요구대로 상처 입고 짓무른 얼굴을 매우 사실적으로 보여줌.",
        "hard_violations": [],
        "physics": "사제의 장갑 낀 손이 물병을 자연스럽게 쥐고 있으며, 군중이 뻗은 여러 갈래의 손과 팔, 체중이 실린 자세가 철책에 기대어 물리적으로 타당하게 지탱됨."
       },
       {
        "label": "B",
        "direction": "사제가 물병을 건네고 군중이 이를 향해 시선과 손을 뻗고 있음.",
        "built_space": "배경에 돌담, 급수탑, 교회가 올바르게 위치함. 사제 뒤쪽으로 레퍼런스에서 요구된 트럭의 금속 철창 구조물이 보임.",
        "entities": "사제복과 하회탈은 묘사되었으나, 군중의 얼굴은 흙먼지가 묻은 정도로 연출되어 '짓무른' 상태의 묘사가 부족함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (물병 아래로 손을 뻗은 중앙 여성의 손가락이 엄지를 포함해 총 6개임)"
        ],
        "physics": "사제가 물병 뚜껑 부근을 손가락으로 잡고 있음. 중앙 군중의 뻗은 손에서 심각한 손가락 개수 오류가 발생해 인체 구조에 어긋남."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 명시한 전경의 손 교차 구도와 군중의 짓무른 얼굴을 훌륭하게 구현했으며, 지정된 배경과 사제의 복장 및 하회탈을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "레퍼런스의 철창 구조는 잘 반영했으나, 물병을 향해 뻗은 중앙 여성의 손가락이 6개로 렌더링된 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "사제가 화면 우측에서 중앙으로 물병을 내밀고 있으며, 군중의 시선과 뻗은 손들이 정확히 물병에 집중됨.",
        "built_space": "레퍼런스와 동일한 거리 배경(돌담, 급수탑, 교회 첨탑)이 보임. 군중과 사제 사이에 금속 철책이 있으나 레퍼런스의 트럭 철창 형태와는 약간 다름. 카메라는 지시된 대로 어깨 높이에 위치함.",
        "entities": "사제는 검은 사제복을 입고 전통 하회탈을 착용함. 군중은 프롬프트의 요구대로 상처 입고 짓무른 얼굴을 매우 사실적으로 보여줌.",
        "hard_violations": [],
        "physics": "사제의 장갑 낀 손이 물병을 자연스럽게 쥐고 있으며, 군중이 뻗은 여러 갈래의 손과 팔, 체중이 실린 자세가 철책에 기대어 물리적으로 타당하게 지탱됨."
       },
       {
        "label": "B",
        "direction": "사제가 물병을 건네고 군중이 이를 향해 시선과 손을 뻗고 있음.",
        "built_space": "배경에 돌담, 급수탑, 교회가 올바르게 위치함. 사제 뒤쪽으로 레퍼런스에서 요구된 트럭의 금속 철창 구조물이 보임.",
        "entities": "사제복과 하회탈은 묘사되었으나, 군중의 얼굴은 흙먼지가 묻은 정도로 연출되어 '짓무른' 상태의 묘사가 부족함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (물병 아래로 손을 뻗은 중앙 여성의 손가락이 엄지를 포함해 총 6개임)"
        ],
        "physics": "사제가 물병 뚜껑 부근을 손가락으로 잡고 있음. 중앙 군중의 뻗은 손에서 심각한 손가락 개수 오류가 발생해 인체 구조에 어긋남."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "오른쪽 분배자와 중앙의 작은 물병은 맞지만, 높은 철창 뒤에서 내려주는 배치와 위를 보는 수령자들 때문에 어깨 높이의 근접 전달 장면 및 물병으로 모이는 주의가 약해진다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 가장자리의 하회탈 사제, 중앙 물병, 하단을 가로지르는 손과 그 너머 짓무른 얼굴을 중간 거리 구도로 충실히 연결하지만 일부 시선은 병보다 사제 얼굴을 향한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사제는 왼쪽 아래 수령자 쪽으로 탈을 돌리고 작은 병을 중앙으로 내민다. 중앙 남성과 후드를 쓴 사람의 손, 왼쪽 전경의 손은 병 쪽으로 향한다. 반면 오른쪽 아래 머릿수건을 쓴 사람의 손은 병보다 상당히 아래·오른쪽으로 뻗는다. 중앙 남성과 여러 수령자의 눈은 병보다 높은 사제 얼굴 쪽을 향해, 전달 지점 아래로 주의가 모이라는 지시는 부분적으로만 구현된다.",
        "built_space": "기와를 얹은 돌담, 전신주와 전선, 먼 급수탑 한 기와 교회 첨탑 한 개가 보여 이전 장면의 마을 재료와 낮 풍경은 이어진다. 오른쪽에는 높은 녹슨 철창 구획과 물병이 든 금속 상자 한 개가 뚜렷하다. 사제는 철창 안쪽의 높은 위치, 수령자들은 바깥의 낮은 위치에 있어 길가 배급소보다 철창 차량에서 내려주는 장면처럼 보인다. 차량 적재함 자체는 충분히 보이지 않아 기존 수감용 철창인지 확정할 수 없다. 반사면이나 중복 급수탑은 보이지 않는다.",
        "entities": "성인 남성 분배자는 검은 사제복, 흰 성직자 칼라, 보라색 영대와 십자가를 착용한다. 탈은 조각된 웃는 얼굴이지만 금속성 테두리와 장식이 두드러져 명시된 하회탈 형태와는 차이가 있다. 맨손으로 잡은 투명한 소형 물병에는 파란 뚜껑이 있다. 수령자들은 주로 동아시아인으로 보이는 성인 남녀이며 얼굴에 염증성 상처와 딱지가 보인다. 별도의 인물 신원이나 정확한 연령은 참고 자료에서 고정되지 않았다. 이전 장면의 두 수감자나 그 복장을 그대로 옮긴 인물은 식별되지 않는다.",
        "hard_violations": [],
        "physics": "물병은 사제의 손가락과 엄지 사이에 실제로 잡혀 있고 손목은 검은 소매의 팔로 연결된다. 수령자들의 뻗은 팔도 보이는 어깨와 몸통으로 이어지며, 일부는 허리를 숙이거나 몸을 낮춘 상태다. 발과 바닥 접점은 화면 밖이지만 공중에 떠 있는 몸으로 보이지는 않는다. 여분 물병은 금속 상자 안에 담겨 있다. 명백히 지지 없이 떠 있는 물체나 불가능한 관절은 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "오른쪽 사제의 탈은 왼쪽 아래의 전달 공간으로 향하고, 장갑 낀 손이 물병을 중앙에 내민다. 왼쪽 남성의 펼친 손, 중앙 뒤 남성의 굽힌 손가락, 아래쪽에서 올라오는 손들이 병 주변으로 수렴한다. 전경의 가장 낮은 손은 병보다 아래를 지나가지만 다른 높이의 손들과 함께 접근 동작을 이룬다. 중앙 여성과 오른쪽 어린 수령자 등 일부는 병보다 사제 얼굴 쪽을 올려다보므로 시선의 완전한 수렴은 부족하다.",
        "built_space": "기와지붕과 돌담, 전신주, 천막이 이어진 도로, 급수탑 한 기와 교회 첨탑 한 개가 보인다. 참고 장면의 마을 구조와 햇빛을 유지하면서 거리 가장자리로 카메라를 옮긴 구도다. 수령자들은 왼쪽의 낮은 금속 경계 바깥, 사제는 오른쪽 배급 공간에 있으며, 오른쪽 앞뒤로 물병 상자가 최소 네 개 보인다. 상자들은 깊이에 따라 작아지고 중앙 전달 공간을 막지 않는다. 기존 수감용 철창과 트럭 적재함은 이 구도에서 식별되지 않으며, 이를 보여주려고 화면을 넓히지는 않았다.",
        "entities": "분배자는 성인 남성으로 보이며 검은 사제복, 흰 칼라, 십자가 목걸이와 검은 장갑을 착용한다. 끈으로 고정된 탈은 돌출된 코와 조각된 눈·입을 가진 하회탈로 읽히고 수령자 쪽의 부분 측면이 보인다. 중앙에는 파란 뚜껑의 투명한 소형 물병이 있다. 수령자는 주로 동아시아인으로 보이는 성인 남녀와 어린 인물들이며 얼굴의 짓무름과 상처가 손 너머로 명확하다. 눈은 정상적인 사람 눈으로 표현된다. 이전 참고 장면의 수감자 신원이나 의상을 그대로 복제한 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "사제의 장갑 낀 엄지와 손가락이 병 몸통을 감싸 지지하고, 손목과 전완은 같은 검은 소매로 자연스럽게 이어진다. 수령자들의 팔은 각기 다른 어깨 높이와 굽힘으로 뻗어 있으며 몸통은 앞으로 기울어 있다. 아래쪽 인물들은 낮게 웅크린 자세로 읽히고, 화면 밖의 발이 보이지 않는 것 외에 부유를 시사하는 요소는 없다. 여분 병들은 상자 바닥에 놓여 있다. 무지지 물체나 명백히 불가능한 신체 연결은 확인되지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "오른쪽 분배자와 중앙의 작은 물병은 맞지만, 높은 철창 뒤에서 내려주는 배치와 위를 보는 수령자들 때문에 어깨 높이의 근접 전달 장면 및 물병으로 모이는 주의가 약해진다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 가장자리의 하회탈 사제, 중앙 물병, 하단을 가로지르는 손과 그 너머 짓무른 얼굴을 중간 거리 구도로 충실히 연결하지만 일부 시선은 병보다 사제 얼굴을 향한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "사제는 왼쪽 아래 수령자 쪽으로 탈을 돌리고 작은 병을 중앙으로 내민다. 중앙 남성과 후드를 쓴 사람의 손, 왼쪽 전경의 손은 병 쪽으로 향한다. 반면 오른쪽 아래 머릿수건을 쓴 사람의 손은 병보다 상당히 아래·오른쪽으로 뻗는다. 중앙 남성과 여러 수령자의 눈은 병보다 높은 사제 얼굴 쪽을 향해, 전달 지점 아래로 주의가 모이라는 지시는 부분적으로만 구현된다.",
        "built_space": "기와를 얹은 돌담, 전신주와 전선, 먼 급수탑 한 기와 교회 첨탑 한 개가 보여 이전 장면의 마을 재료와 낮 풍경은 이어진다. 오른쪽에는 높은 녹슨 철창 구획과 물병이 든 금속 상자 한 개가 뚜렷하다. 사제는 철창 안쪽의 높은 위치, 수령자들은 바깥의 낮은 위치에 있어 길가 배급소보다 철창 차량에서 내려주는 장면처럼 보인다. 차량 적재함 자체는 충분히 보이지 않아 기존 수감용 철창인지 확정할 수 없다. 반사면이나 중복 급수탑은 보이지 않는다.",
        "entities": "성인 남성 분배자는 검은 사제복, 흰 성직자 칼라, 보라색 영대와 십자가를 착용한다. 탈은 조각된 웃는 얼굴이지만 금속성 테두리와 장식이 두드러져 명시된 하회탈 형태와는 차이가 있다. 맨손으로 잡은 투명한 소형 물병에는 파란 뚜껑이 있다. 수령자들은 주로 동아시아인으로 보이는 성인 남녀이며 얼굴에 염증성 상처와 딱지가 보인다. 별도의 인물 신원이나 정확한 연령은 참고 자료에서 고정되지 않았다. 이전 장면의 두 수감자나 그 복장을 그대로 옮긴 인물은 식별되지 않는다.",
        "hard_violations": [],
        "physics": "물병은 사제의 손가락과 엄지 사이에 실제로 잡혀 있고 손목은 검은 소매의 팔로 연결된다. 수령자들의 뻗은 팔도 보이는 어깨와 몸통으로 이어지며, 일부는 허리를 숙이거나 몸을 낮춘 상태다. 발과 바닥 접점은 화면 밖이지만 공중에 떠 있는 몸으로 보이지는 않는다. 여분 물병은 금속 상자 안에 담겨 있다. 명백히 지지 없이 떠 있는 물체나 불가능한 관절은 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "오른쪽 사제의 탈은 왼쪽 아래의 전달 공간으로 향하고, 장갑 낀 손이 물병을 중앙에 내민다. 왼쪽 남성의 펼친 손, 중앙 뒤 남성의 굽힌 손가락, 아래쪽에서 올라오는 손들이 병 주변으로 수렴한다. 전경의 가장 낮은 손은 병보다 아래를 지나가지만 다른 높이의 손들과 함께 접근 동작을 이룬다. 중앙 여성과 오른쪽 어린 수령자 등 일부는 병보다 사제 얼굴 쪽을 올려다보므로 시선의 완전한 수렴은 부족하다.",
        "built_space": "기와지붕과 돌담, 전신주, 천막이 이어진 도로, 급수탑 한 기와 교회 첨탑 한 개가 보인다. 참고 장면의 마을 구조와 햇빛을 유지하면서 거리 가장자리로 카메라를 옮긴 구도다. 수령자들은 왼쪽의 낮은 금속 경계 바깥, 사제는 오른쪽 배급 공간에 있으며, 오른쪽 앞뒤로 물병 상자가 최소 네 개 보인다. 상자들은 깊이에 따라 작아지고 중앙 전달 공간을 막지 않는다. 기존 수감용 철창과 트럭 적재함은 이 구도에서 식별되지 않으며, 이를 보여주려고 화면을 넓히지는 않았다.",
        "entities": "분배자는 성인 남성으로 보이며 검은 사제복, 흰 칼라, 십자가 목걸이와 검은 장갑을 착용한다. 끈으로 고정된 탈은 돌출된 코와 조각된 눈·입을 가진 하회탈로 읽히고 수령자 쪽의 부분 측면이 보인다. 중앙에는 파란 뚜껑의 투명한 소형 물병이 있다. 수령자는 주로 동아시아인으로 보이는 성인 남녀와 어린 인물들이며 얼굴의 짓무름과 상처가 손 너머로 명확하다. 눈은 정상적인 사람 눈으로 표현된다. 이전 참고 장면의 수감자 신원이나 의상을 그대로 복제한 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "사제의 장갑 낀 엄지와 손가락이 병 몸통을 감싸 지지하고, 손목과 전완은 같은 검은 소매로 자연스럽게 이어진다. 수령자들의 팔은 각기 다른 어깨 높이와 굽힘으로 뻗어 있으며 몸통은 앞으로 기울어 있다. 아래쪽 인물들은 낮게 웅크린 자세로 읽히고, 화면 밖의 발이 보이지 않는 것 외에 부유를 시사하는 요소는 없다. 여분 병들은 상자 바닥에 놓여 있다. 무지지 물체나 명백히 불가능한 신체 연결은 확인되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.125
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.875
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (물병 아래로 손을 뻗은 중앙 여성의 손가락이 엄지를 포함해 총 6개임)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 명시한 전경의 손 교차 구도와 군중의 짓무른 얼굴을 훌륭하게 구현했으며, 지정된 배경과 사제의 복장 및 하회탈을 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 875,
    "verdict_ko": "레퍼런스의 철창 구조는 잘 반영했으나, 물병을 향해 뻗은 중앙 여성의 손가락이 6개로 렌더링된 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (물병 아래로 손을 뻗은 중앙 여성의 손가락이 엄지를 포함해 총 6개임)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh3_sel.png",
    "asset_id": "af03b0a7-cd6f-497a-8503-672c3d6c1226",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c01-6cb8-7f8a-9a3d-4a68e4cf142f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S57sh3"
  }
 },
 "S57sh10::signage": {
  "fp": "e2b1043333e3b903",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::masked_church_front": {
  "input_fingerprint": "51be2c0c67b8a9c7",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "masked_church_front",
    "tags": [
     "S57sh10"
    ]
   },
   "context_sig": "1fb4e8feb7f2b0ae"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the approach to the village church entrance, beneath the masked religious statue outside the building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 성당인 듯 보이는 건물 앞에 트럭이 서자, 결박된 채로 현우와 앰버를 끌어내는 병사들.\n- 두 팔을 벌린 예수 동상에도 하회탈 가면이 씌워진.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the approach to the village church entrance, beneath the masked religious statue outside the building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 성당인 듯 보이는 건물 앞에 트럭이 서자, 결박된 채로 현우와 앰버를 끌어내는 병사들.\n- 두 팔을 벌린 예수 동상에도 하회탈 가면이 씌워진.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_masked_church_front_59ceeb.png",
  "asset_id": "7c288265-8830-4ead-b81d-f4281e48bd28",
  "input_asset_ids": [
   "65f1dd1f-3f20-4554-a6af-adaa53196391"
  ],
  "origin_tag": "S57sh10",
  "place_text": "On the approach to the village church entrance, beneath the masked religious statue outside the building.",
  "origin_inputs": {
   "place_text": "On the approach to the village church entrance, beneath the masked religious statue outside the building.",
   "time_of_day_en": "day",
   "conti_asset_id": "65f1dd1f-3f20-4554-a6af-adaa53196391"
  }
 },
 "S57sh10::bgfirst_bg": {
  "input_fingerprint": "5045b779e3628d44",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하회탈 예수 동상을 경악에 찬 눈빛으로 쳐다보는 이현우와 앰버의 압송 뒷모습.\n\nLOCATION (lock): On the approach to the village church entrance, beneath the masked religious statue outside the building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-following track from a low, laterally offset position behind 이현우 and 앰버, tilting upward past their backs toward the entrance statue in a direct wide view. Their bound figures occupy the lower left and lower center, caught at different phases of an unwilling step as escorting soldiers flank them; raised chins and narrow profile slivers convey their alarm at the statue ahead, without forcing their necks vertically upward. Keep the masked, outstretched figure in the upper-right background at less than a third of the image, making the captives' redirected gaze the principal emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: masked Jesus statue with outstretched arms in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 하회탈을 쓴 예수 동상 (Arms outstretched with a Hahoe mask covering the face) — Its front and one side are visible beyond the captives' raised heads; used as Upper-background destination of the captives' gaze; 성당처럼 보이는 건물 입구 (The captives are being escorted inside) — Seen obliquely along the escort's approach; used as Connects the rear-following figures and the statue within one entrance space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep daylight subdued and consistent with the exterior approach, retaining readable profiles and the statue's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 하회탈 예수 동상을 경악에 찬 눈빛으로 쳐다보는 이현우와 앰버의 압송 뒷모습.\n\nLOCATION (lock): On the approach to the village church entrance, beneath the masked religious statue outside the building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-following track from a low, laterally offset position behind 이현우 and 앰버, tilting upward past their backs toward the entrance statue in a direct wide view. Their bound figures occupy the lower left and lower center, caught at different phases of an unwilling step as escorting soldiers flank them; raised chins and narrow profile slivers convey their alarm at the statue ahead, without forcing their necks vertically upward. Keep the masked, outstretched figure in the upper-right background at less than a third of the image, making the captives' redirected gaze the principal emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: masked Jesus statue with outstretched arms in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 하회탈을 쓴 예수 동상 (Arms outstretched with a Hahoe mask covering the face) — Its front and one side are visible beyond the captives' raised heads; used as Upper-background destination of the captives' gaze; 성당처럼 보이는 건물 입구 (The captives are being escorted inside) — Seen obliquely along the escort's approach; used as Connects the rear-following figures and the statue within one entrance space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep daylight subdued and consistent with the exterior approach, retaining readable profiles and the statue's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh10__bgfirst_bg.png",
  "asset_id": "4bcb8810-8eb0-4689-a811-7177ab4c36b7",
  "input_asset_ids": [
   "65f1dd1f-3f20-4554-a6af-adaa53196391",
   "7c288265-8830-4ead-b81d-f4281e48bd28"
  ]
 },
 "S57sh10": {
  "input_fingerprint": "0eb8ca03e94d75f2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 예수 동상을 경악에 찬 눈빛으로 쳐다보는 이현우와 앰버의 압송 뒷모습.\n\nLOCATION (lock): On the approach to the village church entrance, beneath the masked religious statue outside the building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-following track from a low, laterally offset position behind 이현우 and 앰버, tilting upward past their backs toward the entrance statue in a direct wide view. Their bound figures occupy the lower left and lower center, caught at different phases of an unwilling step as escorting soldiers flank them; raised chins and narrow profile slivers convey their alarm at the statue ahead, without forcing their necks vertically upward. Keep the masked, outstretched figure in the upper-right background at less than a third of the image, making the captives' redirected gaze the principal emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: masked Jesus statue with outstretched arms in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 하회탈을 쓴 예수 동상 (Arms outstretched with a Hahoe mask covering the face) — Its front and one side are visible beyond the captives' raised heads; used as Upper-background destination of the captives' gaze; 성당처럼 보이는 건물 입구 (The captives are being escorted inside) — Seen obliquely along the escort's approach; used as Connects the rear-following figures and the statue within one entrance space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep daylight subdued and consistent with the exterior approach, retaining readable profiles and the statue's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped outside the church-like building with its cage still on the bed. The outstretched-armed Jesus statue wears a Hahoe mask. 이현우: He is bound and being escorted into the building after removal from the cage. His previously treated leg remains bandaged. 앰버: She is bound and being escorted into the building, still without a new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 예수 동상을 경악에 찬 눈빛으로 쳐다보는 이현우와 앰버의 압송 뒷모습.\n\nLOCATION (lock): On the approach to the village church entrance, beneath the masked religious statue outside the building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-following track from a low, laterally offset position behind 이현우 and 앰버, tilting upward past their backs toward the entrance statue in a direct wide view. Their bound figures occupy the lower left and lower center, caught at different phases of an unwilling step as escorting soldiers flank them; raised chins and narrow profile slivers convey their alarm at the statue ahead, without forcing their necks vertically upward. Keep the masked, outstretched figure in the upper-right background at less than a third of the image, making the captives' redirected gaze the principal emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: masked Jesus statue with outstretched arms in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 하회탈을 쓴 예수 동상 (Arms outstretched with a Hahoe mask covering the face) — Its front and one side are visible beyond the captives' raised heads; used as Upper-background destination of the captives' gaze; 성당처럼 보이는 건물 입구 (The captives are being escorted inside) — Seen obliquely along the escort's approach; used as Connects the rear-following figures and the statue within one entrance space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep daylight subdued and consistent with the exterior approach, retaining readable profiles and the statue's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped outside the church-like building with its cage still on the bed. The outstretched-armed Jesus statue wears a Hahoe mask. 이현우: He is bound and being escorted into the building after removal from the cage. His previously treated leg remains bandaged. 앰버: She is bound and being escorted into the building, still without a new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하회탈 예수 동상을 경악에 찬 눈빛으로 쳐다보는 이현우와 앰버의 압송 뒷모습.\n\nLOCATION (lock): On the approach to the village church entrance, beneath the masked religious statue outside the building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the rear-following track from a low, laterally offset position behind 이현우 and 앰버, tilting upward past their backs toward the entrance statue in a direct wide view. Their bound figures occupy the lower left and lower center, caught at different phases of an unwilling step as escorting soldiers flank them; raised chins and narrow profile slivers convey their alarm at the statue ahead, without forcing their necks vertically upward. Keep the masked, outstretched figure in the upper-right background at less than a third of the image, making the captives' redirected gaze the principal emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: masked Jesus statue with outstretched arms in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 하회탈을 쓴 예수 동상 (Arms outstretched with a Hahoe mask covering the face) — Its front and one side are visible beyond the captives' raised heads; used as Upper-background destination of the captives' gaze; 성당처럼 보이는 건물 입구 (The captives are being escorted inside) — Seen obliquely along the escort's approach; used as Connects the rear-following figures and the statue within one entrance space.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep daylight subdued and consistent with the exterior approach, retaining readable profiles and the statue's mask.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has stopped outside the church-like building with its cage still on the bed. The outstretched-armed Jesus statue wears a Hahoe mask. 이현우: He is bound and being escorted into the building after removal from the cage. His previously treated leg remains bandaged. 앰버: She is bound and being escorted into the building, still without a new injury from the capture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh10__bgfirst_bg.png",
     "asset_id": "4bcb8810-8eb0-4689-a811-7177ab4c36b7",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S57sh10.png",
     "asset_id": "65f1dd1f-3f20-4554-a6af-adaa53196391",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_masked_church_front_59ceeb.png",
     "asset_id": "7c288265-8830-4ead-b81d-f4281e48bd28",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "이현우와 앰버는 등을 보인 채 얼굴을 오른쪽 위로 돌려 입구 위 동상 쪽을 바라본다. 좁은 옆얼굴은 보이지만 눈의 경악은 뚜렷하지 않다. 병사들은 입구 쪽을 향하며, 오른쪽 병사의 총구는 아래를 향하고 포로를 겨누지 않는다. 동상은 양팔을 좌우로 펼치고 바깥 접근 공간을 향한다.",
    "built_space": "오른쪽에 첨두아치 출입구 하나와 열린 양쪽 문짝, 그 위 벽감의 동상 하나가 있다. 전면의 좁은 창 세 개와 측벽 창 하나, 입구로 올라가는 여러 단의 계단이 보인다. 왼쪽의 원통형 물탱크 하나, 십자가 첨탑 하나, 가장자리의 철창 적재함 일부가 참조 장소와 연결된다. 인물들은 계단 앞 접근 공간에 있으나 크게 잡혀 하체가 화면 아래로 잘리고, 요청한 직접적인 와이드 숏보다 가까운 구도다.",
    "entities": "포로 두 명과 구도 지시가 명시한 호송병 두 명이 있다. 이현우는 짧고 흐트러진 검은 머리의 젊은 동아시아계 남성으로, 오염된 어두운 긴소매 셔츠와 바지가 참조에 가깝다. 인이어는 식별하기 어렵다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니가 보이지만 머리가 참조보다 길다. 두 사람의 손목은 뒤에서 묶여 있다. 동상은 하회탈 형태의 금빛 가면을 쓴 장포 차림 예수상이다. 다리 붕대는 프레임 밖이라 평가할 수 없다.",
    "hard_violations": [],
    "physics": "병사들의 손이 각 포로의 위팔에 닿아 압송 접촉이 성립하고, 결박은 실제 손목을 감싼다. 총과 허리 주머니는 휴대 장구에 연결되어 있다. 발은 화면 밖이므로 지면 접촉과 보행 단계는 확인할 수 없지만, 공중에 떠 있다는 증거도 없다. 동상은 벽감 안 받침대에 지지되어 있다."
   },
   {
    "label": "A",
    "direction": "이현우는 오른쪽 위로 턱을 들고, 앰버도 마스크를 쓴 얼굴을 오른쪽 위 동상 쪽으로 돌린다. 두 사람의 시선 목적지가 입구 위 가면 동상으로 읽히며 목을 수직으로 꺾지는 않는다. 병사들은 계단과 출입구를 향해 이동한다. 양쪽 병사의 소총은 몸 옆에 세로로 휴대되어 총구가 지면을 향한다.",
    "built_space": "참조처럼 오른쪽에 첨두아치 출입구 하나, 열린 문짝 두 개, 상부 벽감의 동상 하나가 배치된다. 전면의 좁은 창 세 개가 보이고 측벽 창은 인물에게 상당 부분 가려진다. 넓은 계단은 지면에서 입구로 올라간다. 왼쪽에는 철창을 실은 트럭 한 대, 원통형 물탱크 하나, 십자가 첨탑 하나가 보인다. 카메라는 계단 아래 접근로의 낮은 후방에 있고, 포로들은 왼쪽과 중앙, 동상은 우측 상단 배경에 놓여 하나의 진입 동선을 이룬다.",
    "entities": "이현우와 앰버, 양옆의 호송병 두 명이 보이며 불필요한 추가 인물은 없다. 이현우의 젊은 동아시아계 외모, 짧은 검은 머리, 마른 체격과 오염된 어두운 바지는 맞지만 셔츠가 참조의 긴소매가 아닌 반소매다. 한쪽 종아리에는 붕대가 있고 인이어는 확인하기 어렵다. 앰버는 작은 체구의 금발 여자아이이며 방진 마스크, 오염된 카키 작업복, 가죽 공구 벨트를 착용한다. 뒷모습만으로 혼혈 외모나 정확한 얼굴 일치까지 판단하기는 어렵다. 두 포로의 뒤로 묶인 손목, 하회탈을 쓴 예수상, 트럭 적재함의 철창이 확인된다.",
    "hard_violations": [],
    "physics": "두 포로와 병사들은 각각 지면을 딛는 발과 들리거나 뒤로 빠지는 발을 보여 진행 중인 보행으로 성립한다. 앰버는 한쪽 신발 바닥이 보이도록 발을 들었고 반대 발이 몸을 지지한다. 병사들의 손은 포로들의 위팔을 붙잡고 있다. 결박은 손목에, 공구 주머니는 허리띠에, 소총은 휴대 끈과 장구에 지지된다. 동상은 받침대 위에 있고 트럭은 바퀴로 지면에 서 있어 지지 없는 물체나 인물은 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "낮은 후방 시점과 동상을 향한 고개 방향은 맞지만, 인물들이 크게 잘려 요청한 와이드 숏과 서로 다른 압송 걸음의 순간이 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "낮은 후방 와이드 구도에서 두 포로의 시선과 서로 다른 걸음, 양옆의 호송병, 우측 상단 동상을 가장 충실하게 연결하지만 이현우의 반소매 복장은 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 앰버는 등을 보인 채 얼굴을 오른쪽 위로 돌려 입구 위 동상 쪽을 바라본다. 좁은 옆얼굴은 보이지만 눈의 경악은 뚜렷하지 않다. 병사들은 입구 쪽을 향하며, 오른쪽 병사의 총구는 아래를 향하고 포로를 겨누지 않는다. 동상은 양팔을 좌우로 펼치고 바깥 접근 공간을 향한다.",
        "built_space": "오른쪽에 첨두아치 출입구 하나와 열린 양쪽 문짝, 그 위 벽감의 동상 하나가 있다. 전면의 좁은 창 세 개와 측벽 창 하나, 입구로 올라가는 여러 단의 계단이 보인다. 왼쪽의 원통형 물탱크 하나, 십자가 첨탑 하나, 가장자리의 철창 적재함 일부가 참조 장소와 연결된다. 인물들은 계단 앞 접근 공간에 있으나 크게 잡혀 하체가 화면 아래로 잘리고, 요청한 직접적인 와이드 숏보다 가까운 구도다.",
        "entities": "포로 두 명과 구도 지시가 명시한 호송병 두 명이 있다. 이현우는 짧고 흐트러진 검은 머리의 젊은 동아시아계 남성으로, 오염된 어두운 긴소매 셔츠와 바지가 참조에 가깝다. 인이어는 식별하기 어렵다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니가 보이지만 머리가 참조보다 길다. 두 사람의 손목은 뒤에서 묶여 있다. 동상은 하회탈 형태의 금빛 가면을 쓴 장포 차림 예수상이다. 다리 붕대는 프레임 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "병사들의 손이 각 포로의 위팔에 닿아 압송 접촉이 성립하고, 결박은 실제 손목을 감싼다. 총과 허리 주머니는 휴대 장구에 연결되어 있다. 발은 화면 밖이므로 지면 접촉과 보행 단계는 확인할 수 없지만, 공중에 떠 있다는 증거도 없다. 동상은 벽감 안 받침대에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 위로 턱을 들고, 앰버도 마스크를 쓴 얼굴을 오른쪽 위 동상 쪽으로 돌린다. 두 사람의 시선 목적지가 입구 위 가면 동상으로 읽히며 목을 수직으로 꺾지는 않는다. 병사들은 계단과 출입구를 향해 이동한다. 양쪽 병사의 소총은 몸 옆에 세로로 휴대되어 총구가 지면을 향한다.",
        "built_space": "참조처럼 오른쪽에 첨두아치 출입구 하나, 열린 문짝 두 개, 상부 벽감의 동상 하나가 배치된다. 전면의 좁은 창 세 개가 보이고 측벽 창은 인물에게 상당 부분 가려진다. 넓은 계단은 지면에서 입구로 올라간다. 왼쪽에는 철창을 실은 트럭 한 대, 원통형 물탱크 하나, 십자가 첨탑 하나가 보인다. 카메라는 계단 아래 접근로의 낮은 후방에 있고, 포로들은 왼쪽과 중앙, 동상은 우측 상단 배경에 놓여 하나의 진입 동선을 이룬다.",
        "entities": "이현우와 앰버, 양옆의 호송병 두 명이 보이며 불필요한 추가 인물은 없다. 이현우의 젊은 동아시아계 외모, 짧은 검은 머리, 마른 체격과 오염된 어두운 바지는 맞지만 셔츠가 참조의 긴소매가 아닌 반소매다. 한쪽 종아리에는 붕대가 있고 인이어는 확인하기 어렵다. 앰버는 작은 체구의 금발 여자아이이며 방진 마스크, 오염된 카키 작업복, 가죽 공구 벨트를 착용한다. 뒷모습만으로 혼혈 외모나 정확한 얼굴 일치까지 판단하기는 어렵다. 두 포로의 뒤로 묶인 손목, 하회탈을 쓴 예수상, 트럭 적재함의 철창이 확인된다.",
        "hard_violations": [],
        "physics": "두 포로와 병사들은 각각 지면을 딛는 발과 들리거나 뒤로 빠지는 발을 보여 진행 중인 보행으로 성립한다. 앰버는 한쪽 신발 바닥이 보이도록 발을 들었고 반대 발이 몸을 지지한다. 병사들의 손은 포로들의 위팔을 붙잡고 있다. 결박은 손목에, 공구 주머니는 허리띠에, 소총은 휴대 끈과 장구에 지지된다. 동상은 받침대 위에 있고 트럭은 바퀴로 지면에 서 있어 지지 없는 물체나 인물은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "낮은 후방 시점과 동상을 향한 고개 방향은 맞지만, 인물들이 크게 잘려 요청한 와이드 숏과 서로 다른 압송 걸음의 순간이 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "낮은 후방 와이드 구도에서 두 포로의 시선과 서로 다른 걸음, 양옆의 호송병, 우측 상단 동상을 가장 충실하게 연결하지만 이현우의 반소매 복장은 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우와 앰버는 등을 보인 채 얼굴을 오른쪽 위로 돌려 입구 위 동상 쪽을 바라본다. 좁은 옆얼굴은 보이지만 눈의 경악은 뚜렷하지 않다. 병사들은 입구 쪽을 향하며, 오른쪽 병사의 총구는 아래를 향하고 포로를 겨누지 않는다. 동상은 양팔을 좌우로 펼치고 바깥 접근 공간을 향한다.",
        "built_space": "오른쪽에 첨두아치 출입구 하나와 열린 양쪽 문짝, 그 위 벽감의 동상 하나가 있다. 전면의 좁은 창 세 개와 측벽 창 하나, 입구로 올라가는 여러 단의 계단이 보인다. 왼쪽의 원통형 물탱크 하나, 십자가 첨탑 하나, 가장자리의 철창 적재함 일부가 참조 장소와 연결된다. 인물들은 계단 앞 접근 공간에 있으나 크게 잡혀 하체가 화면 아래로 잘리고, 요청한 직접적인 와이드 숏보다 가까운 구도다.",
        "entities": "포로 두 명과 구도 지시가 명시한 호송병 두 명이 있다. 이현우는 짧고 흐트러진 검은 머리의 젊은 동아시아계 남성으로, 오염된 어두운 긴소매 셔츠와 바지가 참조에 가깝다. 인이어는 식별하기 어렵다. 앰버는 금발의 어린 여자아이로 방진 마스크, 카키 작업복, 허리 공구 주머니가 보이지만 머리가 참조보다 길다. 두 사람의 손목은 뒤에서 묶여 있다. 동상은 하회탈 형태의 금빛 가면을 쓴 장포 차림 예수상이다. 다리 붕대는 프레임 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "병사들의 손이 각 포로의 위팔에 닿아 압송 접촉이 성립하고, 결박은 실제 손목을 감싼다. 총과 허리 주머니는 휴대 장구에 연결되어 있다. 발은 화면 밖이므로 지면 접촉과 보행 단계는 확인할 수 없지만, 공중에 떠 있다는 증거도 없다. 동상은 벽감 안 받침대에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 위로 턱을 들고, 앰버도 마스크를 쓴 얼굴을 오른쪽 위 동상 쪽으로 돌린다. 두 사람의 시선 목적지가 입구 위 가면 동상으로 읽히며 목을 수직으로 꺾지는 않는다. 병사들은 계단과 출입구를 향해 이동한다. 양쪽 병사의 소총은 몸 옆에 세로로 휴대되어 총구가 지면을 향한다.",
        "built_space": "참조처럼 오른쪽에 첨두아치 출입구 하나, 열린 문짝 두 개, 상부 벽감의 동상 하나가 배치된다. 전면의 좁은 창 세 개가 보이고 측벽 창은 인물에게 상당 부분 가려진다. 넓은 계단은 지면에서 입구로 올라간다. 왼쪽에는 철창을 실은 트럭 한 대, 원통형 물탱크 하나, 십자가 첨탑 하나가 보인다. 카메라는 계단 아래 접근로의 낮은 후방에 있고, 포로들은 왼쪽과 중앙, 동상은 우측 상단 배경에 놓여 하나의 진입 동선을 이룬다.",
        "entities": "이현우와 앰버, 양옆의 호송병 두 명이 보이며 불필요한 추가 인물은 없다. 이현우의 젊은 동아시아계 외모, 짧은 검은 머리, 마른 체격과 오염된 어두운 바지는 맞지만 셔츠가 참조의 긴소매가 아닌 반소매다. 한쪽 종아리에는 붕대가 있고 인이어는 확인하기 어렵다. 앰버는 작은 체구의 금발 여자아이이며 방진 마스크, 오염된 카키 작업복, 가죽 공구 벨트를 착용한다. 뒷모습만으로 혼혈 외모나 정확한 얼굴 일치까지 판단하기는 어렵다. 두 포로의 뒤로 묶인 손목, 하회탈을 쓴 예수상, 트럭 적재함의 철창이 확인된다.",
        "hard_violations": [],
        "physics": "두 포로와 병사들은 각각 지면을 딛는 발과 들리거나 뒤로 빠지는 발을 보여 진행 중인 보행으로 성립한다. 앰버는 한쪽 신발 바닥이 보이도록 발을 들었고 반대 발이 몸을 지지한다. 병사들의 손은 포로들의 위팔을 붙잡고 있다. 결박은 손목에, 공구 주머니는 허리띠에, 소총은 휴대 끈과 장구에 지지된다. 동상은 받침대 위에 있고 트럭은 바퀴로 지면에 서 있어 지지 없는 물체나 인물은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 7,
   "A": 9
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "낮은 후방 시점과 동상을 향한 고개 방향은 맞지만, 인물들이 크게 잘려 요청한 와이드 숏과 서로 다른 압송 걸음의 순간이 약하다."
   },
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "낮은 후방 와이드 구도에서 두 포로의 시선과 서로 다른 걸음, 양옆의 호송병, 우측 상단 동상을 가장 충실하게 연결하지만 이현우의 반소매 복장은 참조와 다르다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_masked_church_front_59ceeb.png",
    "asset_id": "7c288265-8830-4ead-b81d-f4281e48bd28",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c08-b30c-7ec4-b348-ad147a0de738",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S57sh10__bgfirst_bg.png",
   "bg_asset_id": "4bcb8810-8eb0-4689-a811-7177ab4c36b7",
   "bg_record_key": "S57sh10::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "masked_church_front",
   "groupbg_asset_id": "7c288265-8830-4ead-b81d-f4281e48bd28"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S58sh11::signage": {
  "fp": "102705cac474735a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::9f6e4e0d38836f3c": {
  "subjects": [],
  "subject_text": "익산 마을 성당 내부·단상·발코니, 장백산의 단상·권좌 공간\n낡은 고딕 양식 성당 내부. 높은 기둥과 2층 발코니가 중앙 홀을 둘러싸며, 정면 단상에 큰 의자가 놓이고 스테인드글라스 빛이 들어온다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L177",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::ruined_church_hall": {
  "input_fingerprint": "6355320746e9cb50",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "ruined_church_hall",
    "tags": [
     "S58sh11",
     "S58sh28",
     "S58sh41"
    ]
   },
   "context_sig": "77e033581a75d30f"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 성당 내부·단상·발코니, 장백산의 단상·권좌 공간: 파괴된 성당 내부로 스테인드글라스 빛이 들어오며 권력자의 옥좌처럼 개조된 공간이다. (특징: 빛이 들어오는 깨진 스테인드글라스 창문; 하회탈 가면이 씌워진 거대한 예수 조각상; 활을 겨누고 있는 2층 발코니의 가죽 가면 궁사들; 황금빛 하회탈과 화려한 왕권풍 예복을 걸친 100kg 이상의 육중한 백산 체구; 전기 스파크가 튀는 몽둥이를 든 하회탈 병사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 성당 안으로 들어서자 2층 발코니 구간에는 하회탈을 쓴 궁사들이 일제히 활시위를 당기고 있다.\n\nTIME OF DAY (lock): day to dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 성당 내부·단상·발코니, 장백산의 단상·권좌 공간: 파괴된 성당 내부로 스테인드글라스 빛이 들어오며 권력자의 옥좌처럼 개조된 공간이다. (특징: 빛이 들어오는 깨진 스테인드글라스 창문; 하회탈 가면이 씌워진 거대한 예수 조각상; 활을 겨누고 있는 2층 발코니의 가죽 가면 궁사들; 황금빛 하회탈과 화려한 왕권풍 예복을 걸친 100kg 이상의 육중한 백산 체구; 전기 스파크가 튀는 몽둥이를 든 하회탈 병사)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 성당 안으로 들어서자 2층 발코니 구간에는 하회탈을 쓴 궁사들이 일제히 활시위를 당기고 있다.\n\nTIME OF DAY (lock): day to dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_ruined_church_hall_7bdda2.png",
  "asset_id": "514d1d17-44f3-40d7-948f-9f7085a81d43",
  "input_asset_ids": [
   "517cf288-e775-47d6-af70-4759d0214bb0"
  ],
  "origin_tag": "S58sh11",
  "place_text": "On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.",
  "origin_inputs": {
   "place_text": "On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.",
   "time_of_day_en": "day to dusk",
   "conti_asset_id": "517cf288-e775-47d6-af70-4759d0214bb0"
  }
 },
 "S58sh11::bgfirst_bg": {
  "input_fingerprint": "1cdb436e4a960349",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 황금빛 하회탈을 쓰고 화려한 옷을 입은 거구의 장백산이 위풍당당하게 단상 위로 한 발을 내디디며 걷고 있는 전신 구도.\n\nLOCATION (lock): On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.\n\nTIME OF DAY (lock): day to dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entry composition at the foot of the dais, low and oblique to 장백산's path, looking upward while retaining his entire body with clear space above his head and beneath his feet. Place him just right of center as his first foot takes weight on the dais, his masked face directed toward his seat beyond the frame rather than toward the lens. Bowed members of the crowd occupy the lower side edges with individually varied neck angles and shoulder heights, making his elevated position—not a change in exposure—the source of authority.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 단상 (Being stepped onto by 장백산) — The near edge and upper surface are visible from below and to one side; used as Establishes the physical elevation separating authority from the bowed crowd; 파괴된 성당 내부 (Ruined interior surrounding the dais) — The interior recedes behind the full-body figure; used as Provides architectural scale without competing with the entrance gesture; 황금빛 하회탈 (Worn with elaborate, kinglike clothing) — The mask's sculpted front is visible at a three-quarter angle; used as A small, precise identity and authority cue within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the ruined church's stained glass gives restrained sacred weight while preserving the gold-colored mask and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 황금빛 하회탈을 쓰고 화려한 옷을 입은 거구의 장백산이 위풍당당하게 단상 위로 한 발을 내디디며 걷고 있는 전신 구도.\n\nLOCATION (lock): On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows.\n\nTIME OF DAY (lock): day to dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entry composition at the foot of the dais, low and oblique to 장백산's path, looking upward while retaining his entire body with clear space above his head and beneath his feet. Place him just right of center as his first foot takes weight on the dais, his masked face directed toward his seat beyond the frame rather than toward the lens. Bowed members of the crowd occupy the lower side edges with individually varied neck angles and shoulder heights, making his elevated position—not a change in exposure—the source of authority.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 단상 (Being stepped onto by 장백산) — The near edge and upper surface are visible from below and to one side; used as Establishes the physical elevation separating authority from the bowed crowd; 파괴된 성당 내부 (Ruined interior surrounding the dais) — The interior recedes behind the full-body figure; used as Provides architectural scale without competing with the entrance gesture; 황금빛 하회탈 (Worn with elaborate, kinglike clothing) — The mask's sculpted front is visible at a three-quarter angle; used as A small, precise identity and authority cue within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the ruined church's stained glass gives restrained sacred weight while preserving the gold-colored mask and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S58sh11__bgfirst_bg.png",
  "asset_id": "e31cd8a9-82bf-4a35-a406-9002281a9496",
  "input_asset_ids": [
   "517cf288-e775-47d6-af70-4759d0214bb0",
   "514d1d17-44f3-40d7-948f-9f7085a81d43"
  ]
 },
 "S58sh11": {
  "input_fingerprint": "f507ba38128e52ba",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈을 쓰고 화려한 옷을 입은 거구의 장백산이 위풍당당하게 단상 위로 한 발을 내디디며 걷고 있는 전신 구도.\n\nLOCATION (lock): On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entry composition at the foot of the dais, low and oblique to 장백산's path, looking upward while retaining his entire body with clear space above his head and beneath his feet. Place him just right of center as his first foot takes weight on the dais, his masked face directed toward his seat beyond the frame rather than toward the lens. Bowed members of the crowd occupy the lower side edges with individually varied neck angles and shoulder heights, making his elevated position—not a change in exposure—the source of authority.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 단상 (Being stepped onto by 장백산) — The near edge and upper surface are visible from below and to one side; used as Establishes the physical elevation separating authority from the bowed crowd; 파괴된 성당 내부 (Ruined interior surrounding the dais) — The interior recedes behind the full-body figure; used as Provides architectural scale without competing with the entrance gesture; 황금빛 하회탈 (Worn with elaborate, kinglike clothing) — The mask's sculpted front is visible at a three-quarter angle; used as A small, precise identity and authority cue within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the ruined church's stained glass gives restrained sacred weight while preserving the gold-colored mask and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church interior is damaged, with daylight entering through surviving stained glass. Charlie is already confined in a steel cage before its reveal, still wearing the oversized straw hat, boots and colorful raincoat established on the road. 장백산: He has a massive build and wears a golden Hahoe mask and regal clothing. His face is already severely disfigured by radiation beneath the intact mask.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈을 쓰고 화려한 옷을 입은 거구의 장백산이 위풍당당하게 단상 위로 한 발을 내디디며 걷고 있는 전신 구도.\n\nLOCATION (lock): On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entry composition at the foot of the dais, low and oblique to 장백산's path, looking upward while retaining his entire body with clear space above his head and beneath his feet. Place him just right of center as his first foot takes weight on the dais, his masked face directed toward his seat beyond the frame rather than toward the lens. Bowed members of the crowd occupy the lower side edges with individually varied neck angles and shoulder heights, making his elevated position—not a change in exposure—the source of authority.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 단상 (Being stepped onto by 장백산) — The near edge and upper surface are visible from below and to one side; used as Establishes the physical elevation separating authority from the bowed crowd; 파괴된 성당 내부 (Ruined interior surrounding the dais) — The interior recedes behind the full-body figure; used as Provides architectural scale without competing with the entrance gesture; 황금빛 하회탈 (Worn with elaborate, kinglike clothing) — The mask's sculpted front is visible at a three-quarter angle; used as A small, precise identity and authority cue within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the ruined church's stained glass gives restrained sacred weight while preserving the gold-colored mask and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church interior is damaged, with daylight entering through surviving stained glass. Charlie is already confined in a steel cage before its reveal, still wearing the oversized straw hat, boots and colorful raincoat established on the road. 장백산: He has a massive build and wears a golden Hahoe mask and regal clothing. His face is already severely disfigured by radiation beneath the intact mask.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 황금빛 하회탈을 쓰고 화려한 옷을 입은 거구의 장백산이 위풍당당하게 단상 위로 한 발을 내디디며 걷고 있는 전신 구도.\n\nLOCATION (lock): On the raised platform inside the ruined village church. Daylight filters through the stained-glass windows. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entry composition at the foot of the dais, low and oblique to 장백산's path, looking upward while retaining his entire body with clear space above his head and beneath his feet. Place him just right of center as his first foot takes weight on the dais, his masked face directed toward his seat beyond the frame rather than toward the lens. Bowed members of the crowd occupy the lower side edges with individually varied neck angles and shoulder heights, making his elevated position—not a change in exposure—the source of authority.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 단상 (Being stepped onto by 장백산) — The near edge and upper surface are visible from below and to one side; used as Establishes the physical elevation separating authority from the bowed crowd; 파괴된 성당 내부 (Ruined interior surrounding the dais) — The interior recedes behind the full-body figure; used as Provides architectural scale without competing with the entrance gesture; 황금빛 하회탈 (Worn with elaborate, kinglike clothing) — The mask's sculpted front is visible at a three-quarter angle; used as A small, precise identity and authority cue within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Light entering through the ruined church's stained glass gives restrained sacred weight while preserving the gold-colored mask and controlled highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The church interior is damaged, with daylight entering through surviving stained glass. Charlie is already confined in a steel cage before its reveal, still wearing the oversized straw hat, boots and colorful raincoat established on the road. 장백산: He has a massive build and wears a golden Hahoe mask and regal clothing. His face is already severely disfigured by radiation beneath the intact mask.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S58sh11__bgfirst_bg.png",
     "asset_id": "e31cd8a9-82bf-4a35-a406-9002281a9496",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S58sh11.png",
     "asset_id": "517cf288-e775-47d6-af70-4759d0214bb0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:750246>",
     "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_ruined_church_hall_7bdda2.png",
     "asset_id": "514d1d17-44f3-40d7-948f-9f7085a81d43",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:750246>",
     "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "장백산의 얼굴이 프레임 우측 밖 의자를 향하고 있으며, 단상 위로 걷는 방향이 정확함.",
    "built_space": "성당 내부의 계단, 제단, 철창, 우측 단상이 정확히 배치되었으나 원본의 조각상이 제거됨.",
    "entities": "장백산(붉은 예복, 하회탈), 찰리(철창 속 밀짚모자와 우비), 엎드린 군중이 지시대로 잘 구현됨.",
    "hard_violations": [],
    "physics": "오른발이 단상 윗단을 단단히 딛고 체중을 싣고 있어 지지면과의 접촉이 확실함."
   },
   {
    "label": "B",
    "direction": "장백산의 시선이 정면/좌측을 향하며, 단상 위로 올라가는 대신 계단을 내려오는 방향임.",
    "built_space": "성당 구조물이 배치되었으나 배경의 거대 조각상 얼굴에 하회탈이 씌워진 구조적 오류가 있음.",
    "entities": "장백산과 엎드린 군중은 존재하나, 철창 안의 찰리(밀짚모자, 우비)가 묘사되지 않음.",
    "hard_violations": [
     "[gemini-pro] 배경 조각상에 하회탈이 중복 생성됨",
     "[gemini-pro] 단상을 올라가는 지시를 어기고 계단을 내려오는 방향으로 묘사됨"
    ],
    "physics": "왼발이 허공이나 계단 모서리에 닿을 듯 말 듯 떠 있어 지지가 불안정하고 자연스럽지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "로우 앵글, 상승 동작, 철창 안의 인물을 잘 묘사했으나 배경의 거대 조각상이 생략됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "조각상에 하회탈이 중복 생성되었고 단상을 내려오는 방향으로 묘사되어 치명적인 오류가 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장백산의 얼굴이 프레임 우측 밖 의자를 향하고 있으며, 단상 위로 걷는 방향이 정확함.",
        "built_space": "성당 내부의 계단, 제단, 철창, 우측 단상이 정확히 배치되었으나 원본의 조각상이 제거됨.",
        "entities": "장백산(붉은 예복, 하회탈), 찰리(철창 속 밀짚모자와 우비), 엎드린 군중이 지시대로 잘 구현됨.",
        "hard_violations": [],
        "physics": "오른발이 단상 윗단을 단단히 딛고 체중을 싣고 있어 지지면과의 접촉이 확실함."
       },
       {
        "label": "B",
        "direction": "장백산의 시선이 정면/좌측을 향하며, 단상 위로 올라가는 대신 계단을 내려오는 방향임.",
        "built_space": "성당 구조물이 배치되었으나 배경의 거대 조각상 얼굴에 하회탈이 씌워진 구조적 오류가 있음.",
        "entities": "장백산과 엎드린 군중은 존재하나, 철창 안의 찰리(밀짚모자, 우비)가 묘사되지 않음.",
        "hard_violations": [
         "배경 조각상에 하회탈이 중복 생성됨",
         "단상을 올라가는 지시를 어기고 계단을 내려오는 방향으로 묘사됨"
        ],
        "physics": "왼발이 허공이나 계단 모서리에 닿을 듯 말 듯 떠 있어 지지가 불안정하고 자연스럽지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "로우 앵글, 상승 동작, 철창 안의 인물을 잘 묘사했으나 배경의 거대 조각상이 생략됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "조각상에 하회탈이 중복 생성되었고 단상을 내려오는 방향으로 묘사되어 치명적인 오류가 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "장백산의 얼굴이 프레임 우측 밖 의자를 향하고 있으며, 단상 위로 걷는 방향이 정확함.",
        "built_space": "성당 내부의 계단, 제단, 철창, 우측 단상이 정확히 배치되었으나 원본의 조각상이 제거됨.",
        "entities": "장백산(붉은 예복, 하회탈), 찰리(철창 속 밀짚모자와 우비), 엎드린 군중이 지시대로 잘 구현됨.",
        "hard_violations": [],
        "physics": "오른발이 단상 윗단을 단단히 딛고 체중을 싣고 있어 지지면과의 접촉이 확실함."
       },
       {
        "label": "B",
        "direction": "장백산의 시선이 정면/좌측을 향하며, 단상 위로 올라가는 대신 계단을 내려오는 방향임.",
        "built_space": "성당 구조물이 배치되었으나 배경의 거대 조각상 얼굴에 하회탈이 씌워진 구조적 오류가 있음.",
        "entities": "장백산과 엎드린 군중은 존재하나, 철창 안의 찰리(밀짚모자, 우비)가 묘사되지 않음.",
        "hard_violations": [
         "배경 조각상에 하회탈이 중복 생성됨",
         "단상을 올라가는 지시를 어기고 계단을 내려오는 방향으로 묘사됨"
        ],
        "physics": "왼발이 허공이나 계단 모서리에 닿을 듯 말 듯 떠 있어 지지가 불안정하고 자연스럽지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 사선 전신 구도와 자연스러운 진입 동작, 기준 장소의 동상·단상 배치를 더 충실히 살렸지만 화면 밖이어야 할 왕좌까지 보인다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "발의 지지는 명확하지만 높은 계단을 크게 딛는 자세가 진입 보행보다 강조되고, 기준 장소의 대형 동상이 확인되지 않으며 왕좌도 화면 안에 드러난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장백산의 몸과 금빛 가면은 화면 오른쪽 위의 왕좌 방향으로 향하며 렌즈를 정면으로 보지 않는다. 가면 앞면은 삼사분면으로 보인다. 다만 목적지인 왕좌가 실제 화면 오른쪽에 들어와 있어 ‘화면 밖의 자리’를 향한다는 조건은 어긋난다. 양쪽 아래 군중은 서로 다른 깊이로 고개를 숙여 바닥 쪽을 본다.",
        "built_space": "오른쪽의 높은 단상과 왕좌 1개, 그 왼쪽의 대형 동상 1개, 뒤쪽 중앙의 제단 1개와 십자가 1개, 왼쪽 철창 1개가 보인다. 큰 색유리창 세 면과 오른쪽의 좁은 색유리창 한 면, 무너진 지붕과 석조 기둥도 기준 장소에 대응한다. 카메라는 단상 아래에서 비스듬히 올려다보며 가까운 단차의 전면과 윗면을 함께 보여준다. 장백산은 중앙보다 약간 오른쪽에 있고 군중은 그보다 낮은 양쪽 가장자리에 배치된다.",
        "entities": "주인공은 넓은 체격의 성인 남성으로 보이며 검은 머리 부분, 관, 붉은색 금문양 예복, 장식 허리띠와 붉은 신발이 인물 기준에 부합한다. 얼굴에는 웃는 조형의 금빛 하회탈이 착용되어 있다. 가면 아래 얼굴의 나이·민족적 특징과 방사선 손상은 확인할 수 없으며, 이를 드러내지 않은 것은 지시에 맞는다. 가장자리에는 카메라 지시가 요구한 실제 사람 형태의 군중이 있다. 철창 안에도 인물이 보이지만 작고 어두워 밀짚모자·우비·부츠의 일치를 모두 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "앞으로 내민 신발은 높은 단차의 모서리 부근에 닿고 뒤쪽 발은 더 낮은 단차에 놓여 있어 몸을 지지할 경로가 보인다. 앞발 전체에 체중이 완전히 실린 순간인지는 다소 모호하지만, 계단을 오르며 체중을 옮기는 동작으로 가능하다. 손은 몸 옆에서 자연스럽게 움직이고 무거운 옷자락은 아래로 처진다. 군중의 하체는 대부분 가려져 있으나 바닥 높이에서 몸을 숙인 자세이며 공중에 떠 있는 몸은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "장백산의 얼굴과 진행 방향은 오른쪽 왕좌를 향하고 렌즈를 피한다. 금빛 가면의 앞면도 비스듬히 보인다. 그러나 왕좌 자체가 화면 오른쪽에 크게 보여 목적지를 화면 밖에 두라는 조건을 충족하지 못한다. 양쪽 군중은 아래를 향해 고개를 숙이며 목과 어깨 높이에 차이가 있다.",
        "built_space": "오른쪽 단상 위 왕좌 1개, 중앙 뒤편 제단 1개와 십자가 1개, 왼쪽 철창 1개가 보인다. 큰 색유리창 세 면과 좁은 창 한 면, 파괴된 지붕과 석조 아치는 기준 장소와 유사하다. 다만 기준 사진과 A에서 왕좌 왼쪽에 보이는 대형 동상은 확인되지 않아 그 위치의 건축적 표지가 약해졌다. 단상 아래의 낮은 사선 시점, 단차의 전면과 윗면, 중앙 오른쪽 전신 배치는 대체로 맞는다.",
        "entities": "장백산은 육중한 성인 남성으로 표현되며 붉은 금문양 왕실 예복, 관, 허리띠, 붉은 신발과 착용한 금빛 하회탈이 요구에 부합한다. 가면 때문에 실제 얼굴의 동일성과 손상 상태는 판단할 수 없다. 철창 안에는 큰 밀짚모자를 쓰고 다채로운 겉옷과 부츠를 착용한 사람이 비교적 선명하게 보여 찰리의 소품 조건에 대응하지만, 이 진입 장면에서 불필요하게 먼저 드러난다. 양쪽 군중은 구체적인 사람의 머리와 몸으로 표현되어 있다.",
        "hard_violations": [],
        "physics": "뒤쪽 신발은 낮은 계단 윗면에 놓이고 앞으로 든 신발은 높은 계단에 접촉하므로 몸의 지지는 명확하다. 무릎을 높게 굽혀 여러 단차를 크게 딛는 자세는 신체적으로 가능하지만, 첫발에 체중을 싣고 위풍당당하게 걷는 순간보다는 큰 계단 오르기 동작으로 읽힌다. 철창 속 인물은 앉은 자세로 내부 받침에 지지되는 것으로 보이며, 군중도 낮은 바닥에 몸을 굽힌 상태다. 지지 없이 떠 있는 물체나 사람은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 사선 전신 구도와 자연스러운 진입 동작, 기준 장소의 동상·단상 배치를 더 충실히 살렸지만 화면 밖이어야 할 왕좌까지 보인다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "발의 지지는 명확하지만 높은 계단을 크게 딛는 자세가 진입 보행보다 강조되고, 기준 장소의 대형 동상이 확인되지 않으며 왕좌도 화면 안에 드러난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "장백산의 몸과 금빛 가면은 화면 오른쪽 위의 왕좌 방향으로 향하며 렌즈를 정면으로 보지 않는다. 가면 앞면은 삼사분면으로 보인다. 다만 목적지인 왕좌가 실제 화면 오른쪽에 들어와 있어 ‘화면 밖의 자리’를 향한다는 조건은 어긋난다. 양쪽 아래 군중은 서로 다른 깊이로 고개를 숙여 바닥 쪽을 본다.",
        "built_space": "오른쪽의 높은 단상과 왕좌 1개, 그 왼쪽의 대형 동상 1개, 뒤쪽 중앙의 제단 1개와 십자가 1개, 왼쪽 철창 1개가 보인다. 큰 색유리창 세 면과 오른쪽의 좁은 색유리창 한 면, 무너진 지붕과 석조 기둥도 기준 장소에 대응한다. 카메라는 단상 아래에서 비스듬히 올려다보며 가까운 단차의 전면과 윗면을 함께 보여준다. 장백산은 중앙보다 약간 오른쪽에 있고 군중은 그보다 낮은 양쪽 가장자리에 배치된다.",
        "entities": "주인공은 넓은 체격의 성인 남성으로 보이며 검은 머리 부분, 관, 붉은색 금문양 예복, 장식 허리띠와 붉은 신발이 인물 기준에 부합한다. 얼굴에는 웃는 조형의 금빛 하회탈이 착용되어 있다. 가면 아래 얼굴의 나이·민족적 특징과 방사선 손상은 확인할 수 없으며, 이를 드러내지 않은 것은 지시에 맞는다. 가장자리에는 카메라 지시가 요구한 실제 사람 형태의 군중이 있다. 철창 안에도 인물이 보이지만 작고 어두워 밀짚모자·우비·부츠의 일치를 모두 확정하기 어렵다.",
        "hard_violations": [],
        "physics": "앞으로 내민 신발은 높은 단차의 모서리 부근에 닿고 뒤쪽 발은 더 낮은 단차에 놓여 있어 몸을 지지할 경로가 보인다. 앞발 전체에 체중이 완전히 실린 순간인지는 다소 모호하지만, 계단을 오르며 체중을 옮기는 동작으로 가능하다. 손은 몸 옆에서 자연스럽게 움직이고 무거운 옷자락은 아래로 처진다. 군중의 하체는 대부분 가려져 있으나 바닥 높이에서 몸을 숙인 자세이며 공중에 떠 있는 몸은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "장백산의 얼굴과 진행 방향은 오른쪽 왕좌를 향하고 렌즈를 피한다. 금빛 가면의 앞면도 비스듬히 보인다. 그러나 왕좌 자체가 화면 오른쪽에 크게 보여 목적지를 화면 밖에 두라는 조건을 충족하지 못한다. 양쪽 군중은 아래를 향해 고개를 숙이며 목과 어깨 높이에 차이가 있다.",
        "built_space": "오른쪽 단상 위 왕좌 1개, 중앙 뒤편 제단 1개와 십자가 1개, 왼쪽 철창 1개가 보인다. 큰 색유리창 세 면과 좁은 창 한 면, 파괴된 지붕과 석조 아치는 기준 장소와 유사하다. 다만 기준 사진과 A에서 왕좌 왼쪽에 보이는 대형 동상은 확인되지 않아 그 위치의 건축적 표지가 약해졌다. 단상 아래의 낮은 사선 시점, 단차의 전면과 윗면, 중앙 오른쪽 전신 배치는 대체로 맞는다.",
        "entities": "장백산은 육중한 성인 남성으로 표현되며 붉은 금문양 왕실 예복, 관, 허리띠, 붉은 신발과 착용한 금빛 하회탈이 요구에 부합한다. 가면 때문에 실제 얼굴의 동일성과 손상 상태는 판단할 수 없다. 철창 안에는 큰 밀짚모자를 쓰고 다채로운 겉옷과 부츠를 착용한 사람이 비교적 선명하게 보여 찰리의 소품 조건에 대응하지만, 이 진입 장면에서 불필요하게 먼저 드러난다. 양쪽 군중은 구체적인 사람의 머리와 몸으로 표현되어 있다.",
        "hard_violations": [],
        "physics": "뒤쪽 신발은 낮은 계단 윗면에 놓이고 앞으로 든 신발은 높은 계단에 접촉하므로 몸의 지지는 명확하다. 무릎을 높게 굽혀 여러 단차를 크게 딛는 자세는 신체적으로 가능하지만, 첫발에 체중을 싣고 위풍당당하게 걷는 순간보다는 큰 계단 오르기 동작으로 읽힌다. 철창 속 인물은 앉은 자세로 내부 받침에 지지되는 것으로 보이며, 군중도 낮은 바닥에 몸을 굽힌 상태다. 지지 없이 떠 있는 물체나 사람은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 배경 조각상에 하회탈이 중복 생성됨",
     "[gemini-pro] 단상을 올라가는 지시를 어기고 계단을 내려오는 방향으로 묘사됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "로우 앵글, 상승 동작, 철창 안의 인물을 잘 묘사했으나 배경의 거대 조각상이 생략됨."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "조각상에 하회탈이 중복 생성되었고 단상을 내려오는 방향으로 묘사되어 치명적인 오류가 있음.  ★위반: [gemini-pro] 배경 조각상에 하회탈이 중복 생성됨 / [gemini-pro] 단상을 올라가는 지시를 어기고 계단을 내려오는 방향으로 묘사됨"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_ruined_church_hall_7bdda2.png",
    "asset_id": "514d1d17-44f3-40d7-948f-9f7085a81d43",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:750246>",
    "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c13-6607-70b8-b5f7-a2b27c3901a6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S58sh11__bgfirst_bg.png",
   "bg_asset_id": "e31cd8a9-82bf-4a35-a406-9002281a9496",
   "bg_record_key": "S58sh11::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "ruined_church_hall",
   "groupbg_asset_id": "514d1d17-44f3-40d7-948f-9f7085a81d43"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S58sh28::signage": {
  "fp": "2b14747e4b3bc869",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh28": {
  "input_fingerprint": "547ad9176607fab8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 강철 케이지 안에 갇힌 찰리의 거대한 기계 몸체가 시무룩하게 웅크리고 있는 전신 구도.\n\nLOCATION (lock): Inside the steel cage brought into the village church's main assembly space, near the raised platform. Stained-glass daylight reaches the interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly approach outside the cage's near corner, looking diagonally through its bars from a shallow high angle while retaining 찰리's complete crouched body. Place him slightly left of center with knees folded inward, shoulders lowered and his head inclined toward the floor; a minimal lateral correction leaves his head between bars rather than behind one. Preserve church space around the cage and natural mechanical proportions, making the stopped camera distance express his confinement without pushing into a face close-up.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Contains 찰리, whose complete crouched body remains visible) — The near corner reveals two adjoining barred sides; used as Surrounds the figure with visible divisions while keeping the head unobstructed; 성당 바닥과 내부 공간 (Part of the ruined church surrounding the cage) — Floor space extends beside and behind the cage; used as Maintains scale and leaves visual room for the impending reverse movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the church's stained-glass daylight with controlled highlights that distinguish 찰리's hard-surface construction while preserving the subdued emotional tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains confined in the steel cage, wearing the oversized straw hat, boots and colorful raincoat; the barred access door has opened to reveal the cage. Daylight continues to enter the ruined church through stained glass.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 강철 케이지 안에 갇힌 찰리의 거대한 기계 몸체가 시무룩하게 웅크리고 있는 전신 구도.\n\nLOCATION (lock): Inside the steel cage brought into the village church's main assembly space, near the raised platform. Stained-glass daylight reaches the interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly approach outside the cage's near corner, looking diagonally through its bars from a shallow high angle while retaining 찰리's complete crouched body. Place him slightly left of center with knees folded inward, shoulders lowered and his head inclined toward the floor; a minimal lateral correction leaves his head between bars rather than behind one. Preserve church space around the cage and natural mechanical proportions, making the stopped camera distance express his confinement without pushing into a face close-up.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Contains 찰리, whose complete crouched body remains visible) — The near corner reveals two adjoining barred sides; used as Surrounds the figure with visible divisions while keeping the head unobstructed; 성당 바닥과 내부 공간 (Part of the ruined church surrounding the cage) — Floor space extends beside and behind the cage; used as Maintains scale and leaves visual room for the impending reverse movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the church's stained-glass daylight with controlled highlights that distinguish 찰리's hard-surface construction while preserving the subdued emotional tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains confined in the steel cage, wearing the oversized straw hat, boots and colorful raincoat; the barred access door has opened to reveal the cage. Daylight continues to enter the ruined church through stained glass.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 강철 케이지 안에 갇힌 찰리의 거대한 기계 몸체가 시무룩하게 웅크리고 있는 전신 구도.\n\nLOCATION (lock): Inside the steel cage brought into the village church's main assembly space, near the raised platform. Stained-glass daylight reaches the interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly approach outside the cage's near corner, looking diagonally through its bars from a shallow high angle while retaining 찰리's complete crouched body. Place him slightly left of center with knees folded inward, shoulders lowered and his head inclined toward the floor; a minimal lateral correction leaves his head between bars rather than behind one. Preserve church space around the cage and natural mechanical proportions, making the stopped camera distance express his confinement without pushing into a face close-up.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 강철 케이지 (Contains 찰리, whose complete crouched body remains visible) — The near corner reveals two adjoining barred sides; used as Surrounds the figure with visible divisions while keeping the head unobstructed; 성당 바닥과 내부 공간 (Part of the ruined church surrounding the cage) — Floor space extends beside and behind the cage; used as Maintains scale and leaves visual room for the impending reverse movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the church's stained-glass daylight with controlled highlights that distinguish 찰리's hard-surface construction while preserving the subdued emotional tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains confined in the steel cage, wearing the oversized straw hat, boots and colorful raincoat; the barred access door has opened to reveal the cage. Daylight continues to enter the ruined church through stained glass.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "고개를 바닥으로 숙이고 시선은 아래를 향하고 있음.",
    "built_space": "폐허가 된 성당 내부의 우측 제단과 계단이 배경에 위치함. 전경에 철제 케이지의 모서리가 배치되어 대각선 구도를 형성하며, 캐릭터의 얼굴이 창살 사이에 정확히 위치함.",
    "entities": "찰리(흰색 마스크, 파란 원자로, 샌드 베이지 장갑)가 밀짚모자, 다채로운 색상의 우비, 그리고 뚜렷한 부츠를 착용하고 있음.",
    "hard_violations": [],
    "physics": "케이지 바닥에 두 발을 딛고 안정적으로 쭈그리고 앉아 체중을 지탱함."
   },
   {
    "label": "B",
    "direction": "고개를 바닥 쪽으로 숙이고 시선은 아래를 향함.",
    "built_space": "성당 내부와 우측 제단이 일치하게 렌더링됨. 대각선 방향에서 케이지를 바라보는 구도이며, 캐릭터의 머리가 창살 사이에 위치함.",
    "entities": "찰리의 기본 외형과 밀짚모자, 우비는 존재하나 부츠의 형태가 모호하며 기계 발이 그대로 드러나 보임. 얼굴의 입선 디테일이 생략됨.",
    "hard_violations": [],
    "physics": "케이지 바닥에 두 발을 지지하고 쭈그리고 앉은 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터의 외형(얼굴, 원자로) 및 지시된 소품(부츠, 우비, 모자)을 정확히 구현했으며, 요구된 카메라 구도와 프레이밍을 완벽하게 따름."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 구도와 배경은 잘 구현했으나, 부츠 대신 기계 맨발이 노출되었고 얼굴 디테일이 레퍼런스와 다소 차이가 있음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 바닥으로 숙이고 시선은 아래를 향하고 있음.",
        "built_space": "폐허가 된 성당 내부의 우측 제단과 계단이 배경에 위치함. 전경에 철제 케이지의 모서리가 배치되어 대각선 구도를 형성하며, 캐릭터의 얼굴이 창살 사이에 정확히 위치함.",
        "entities": "찰리(흰색 마스크, 파란 원자로, 샌드 베이지 장갑)가 밀짚모자, 다채로운 색상의 우비, 그리고 뚜렷한 부츠를 착용하고 있음.",
        "hard_violations": [],
        "physics": "케이지 바닥에 두 발을 딛고 안정적으로 쭈그리고 앉아 체중을 지탱함."
       },
       {
        "label": "B",
        "direction": "고개를 바닥 쪽으로 숙이고 시선은 아래를 향함.",
        "built_space": "성당 내부와 우측 제단이 일치하게 렌더링됨. 대각선 방향에서 케이지를 바라보는 구도이며, 캐릭터의 머리가 창살 사이에 위치함.",
        "entities": "찰리의 기본 외형과 밀짚모자, 우비는 존재하나 부츠의 형태가 모호하며 기계 발이 그대로 드러나 보임. 얼굴의 입선 디테일이 생략됨.",
        "hard_violations": [],
        "physics": "케이지 바닥에 두 발을 지지하고 쭈그리고 앉은 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터의 외형(얼굴, 원자로) 및 지시된 소품(부츠, 우비, 모자)을 정확히 구현했으며, 요구된 카메라 구도와 프레이밍을 완벽하게 따름."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 구도와 배경은 잘 구현했으나, 부츠 대신 기계 맨발이 노출되었고 얼굴 디테일이 레퍼런스와 다소 차이가 있음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개를 바닥으로 숙이고 시선은 아래를 향하고 있음.",
        "built_space": "폐허가 된 성당 내부의 우측 제단과 계단이 배경에 위치함. 전경에 철제 케이지의 모서리가 배치되어 대각선 구도를 형성하며, 캐릭터의 얼굴이 창살 사이에 정확히 위치함.",
        "entities": "찰리(흰색 마스크, 파란 원자로, 샌드 베이지 장갑)가 밀짚모자, 다채로운 색상의 우비, 그리고 뚜렷한 부츠를 착용하고 있음.",
        "hard_violations": [],
        "physics": "케이지 바닥에 두 발을 딛고 안정적으로 쭈그리고 앉아 체중을 지탱함."
       },
       {
        "label": "B",
        "direction": "고개를 바닥 쪽으로 숙이고 시선은 아래를 향함.",
        "built_space": "성당 내부와 우측 제단이 일치하게 렌더링됨. 대각선 방향에서 케이지를 바라보는 구도이며, 캐릭터의 머리가 창살 사이에 위치함.",
        "entities": "찰리의 기본 외형과 밀짚모자, 우비는 존재하나 부츠의 형태가 모호하며 기계 발이 그대로 드러나 보임. 얼굴의 입선 디테일이 생략됨.",
        "hard_violations": [],
        "physics": "케이지 바닥에 두 발을 지지하고 쭈그리고 앉은 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "전신과 주변 바닥을 담은 얕은 부감, 왼쪽 배치, 쇠창살 사이로 드러난 얼굴이 지시에 더 가깝지만 열린 출입문 상태는 재현하지 못했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "웅크린 전신과 의상은 맞지만 쇠창살이 얼굴 중앙을 가리고 부감이 약해, 명시된 카메라 위치와 머리 가림 방지 지시에서 A보다 뒤처진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 고개를 앞으로 낮추고 얼굴을 자기 앞쪽 케이지 바닥 방향으로 기울인다. 발광하는 눈은 고정된 기계 눈이라 정확한 응시점은 불명확하지만, 카메라를 정면으로 바라보는 자세는 아니다. 얼굴의 두 눈과 입은 세로 창살 사이에 드러나며, 머리 왼쪽 가장자리에는 창살이 겹친다.",
        "built_space": "케이지 하나의 앞면과 오른쪽 인접 면, 격자 바닥이 보이며 카메라는 가까운 모서리 밖에서 비스듬히 내려다본다. 찰리의 모자부터 두 부츠까지 들어오고 오른쪽과 뒤쪽에 교회 바닥이 남는다. 뒤에는 올라가는 계단열 하나, 제단 하나, 십자가 하나와 여러 촛대가 있으며, 파손된 석벽과 스테인드글라스가 이전 장소의 재료와 채광을 이어간다. 다만 이전 사진보다 케이지가 제단에서 떨어진 낮은 바닥에 놓인 듯 보이며, 열린 출입문은 보이지 않고 오른쪽 걸쇠 주변의 문은 닫힌 형태다.",
        "entities": "등장 개체는 찰리 하나뿐이다. 긴 육중한 팔, 짧게 접힌 다리, 샌드 베이지 장갑판, 흰 마스크, 주황색 점눈 두 개와 선형 입, 푸른 가슴 원자로가 참조와 부합한다. 큰 밀짚모자, 여러 색이 섞인 낡은 우비, 부츠도 보인다. 인간 피부나 치아, 다른 사람, 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 케이지 바닥에 닿고 접힌 다리가 낮춘 몸통을 지탱한다. 긴 팔은 어깨와 팔꿈치 관절에서 아래로 내려오며 손은 발 가까이에 놓인다. 모자는 머리에 얹혀 있고 우비는 어깨에서 중력 방향으로 늘어진다. 케이지 하부 틀도 교회 바닥에 놓여 있어 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "찰리는 고개와 어깨를 낮추고 자기 앞 바닥 쪽으로 얼굴을 기울인다. 다만 가까운 세로 창살 하나가 흰 얼굴 중앙과 입 부근을 직접 가르므로, 머리를 창살 사이에 두라는 지시를 충족하지 못한다.",
        "built_space": "케이지 하나의 앞면과 오른쪽 면, 바닥이 보이고 찰리의 전신이 중앙보다 약간 왼쪽에 들어온다. 오른쪽에는 바닥 여백과 올라가는 계단열 하나, 제단 하나, 십자가 하나 및 촛대들이 보인다. 파손된 석벽과 스테인드글라스의 낮빛은 장소 참조와 유사하다. A보다 바닥을 내려다보는 정도가 약해 요청된 얕은 부감이 덜 분명하다. 케이지는 계단 아래에 놓인 듯 보이며, 오른쪽 걸쇠가 있는 출입문도 열린 모습이 아니라 닫힌 형태다.",
        "entities": "찰리 하나만 등장하며 베이지색 기계 장갑, 긴 팔과 짧은 다리, 흰 마스크, 점눈 두 개, 선형 입, 푸른 원자로를 갖췄다. 밀짚모자와 다색 우비, 두 부츠도 확인된다. 얼굴 일부는 창살에 가려지지만 인간 얼굴로 바뀌지는 않았고 다른 사람이나 문구도 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 격자 바닥을 딛고 무릎을 깊게 굽혀 몸통을 받친다. 두 손은 연결된 팔 관절에서 무릎 앞쪽으로 자연스럽게 내려와 있어 손이 바닥에 닿지 않아도 지지가 성립한다. 모자와 우비는 각각 머리와 어깨에 지지되고 케이지는 바닥에 놓여 있다. 공중 부유나 명백히 불가능한 관절 배치는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전신과 주변 바닥을 담은 얕은 부감, 왼쪽 배치, 쇠창살 사이로 드러난 얼굴이 지시에 더 가깝지만 열린 출입문 상태는 재현하지 못했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "웅크린 전신과 의상은 맞지만 쇠창살이 얼굴 중앙을 가리고 부감이 약해, 명시된 카메라 위치와 머리 가림 방지 지시에서 A보다 뒤처진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 고개를 앞으로 낮추고 얼굴을 자기 앞쪽 케이지 바닥 방향으로 기울인다. 발광하는 눈은 고정된 기계 눈이라 정확한 응시점은 불명확하지만, 카메라를 정면으로 바라보는 자세는 아니다. 얼굴의 두 눈과 입은 세로 창살 사이에 드러나며, 머리 왼쪽 가장자리에는 창살이 겹친다.",
        "built_space": "케이지 하나의 앞면과 오른쪽 인접 면, 격자 바닥이 보이며 카메라는 가까운 모서리 밖에서 비스듬히 내려다본다. 찰리의 모자부터 두 부츠까지 들어오고 오른쪽과 뒤쪽에 교회 바닥이 남는다. 뒤에는 올라가는 계단열 하나, 제단 하나, 십자가 하나와 여러 촛대가 있으며, 파손된 석벽과 스테인드글라스가 이전 장소의 재료와 채광을 이어간다. 다만 이전 사진보다 케이지가 제단에서 떨어진 낮은 바닥에 놓인 듯 보이며, 열린 출입문은 보이지 않고 오른쪽 걸쇠 주변의 문은 닫힌 형태다.",
        "entities": "등장 개체는 찰리 하나뿐이다. 긴 육중한 팔, 짧게 접힌 다리, 샌드 베이지 장갑판, 흰 마스크, 주황색 점눈 두 개와 선형 입, 푸른 가슴 원자로가 참조와 부합한다. 큰 밀짚모자, 여러 색이 섞인 낡은 우비, 부츠도 보인다. 인간 피부나 치아, 다른 사람, 추가 문구는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 케이지 바닥에 닿고 접힌 다리가 낮춘 몸통을 지탱한다. 긴 팔은 어깨와 팔꿈치 관절에서 아래로 내려오며 손은 발 가까이에 놓인다. 모자는 머리에 얹혀 있고 우비는 어깨에서 중력 방향으로 늘어진다. 케이지 하부 틀도 교회 바닥에 놓여 있어 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "찰리는 고개와 어깨를 낮추고 자기 앞 바닥 쪽으로 얼굴을 기울인다. 다만 가까운 세로 창살 하나가 흰 얼굴 중앙과 입 부근을 직접 가르므로, 머리를 창살 사이에 두라는 지시를 충족하지 못한다.",
        "built_space": "케이지 하나의 앞면과 오른쪽 면, 바닥이 보이고 찰리의 전신이 중앙보다 약간 왼쪽에 들어온다. 오른쪽에는 바닥 여백과 올라가는 계단열 하나, 제단 하나, 십자가 하나 및 촛대들이 보인다. 파손된 석벽과 스테인드글라스의 낮빛은 장소 참조와 유사하다. A보다 바닥을 내려다보는 정도가 약해 요청된 얕은 부감이 덜 분명하다. 케이지는 계단 아래에 놓인 듯 보이며, 오른쪽 걸쇠가 있는 출입문도 열린 모습이 아니라 닫힌 형태다.",
        "entities": "찰리 하나만 등장하며 베이지색 기계 장갑, 긴 팔과 짧은 다리, 흰 마스크, 점눈 두 개, 선형 입, 푸른 원자로를 갖췄다. 밀짚모자와 다색 우비, 두 부츠도 확인된다. 얼굴 일부는 창살에 가려지지만 인간 얼굴로 바뀌지는 않았고 다른 사람이나 문구도 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 격자 바닥을 딛고 무릎을 깊게 굽혀 몸통을 받친다. 두 손은 연결된 팔 관절에서 무릎 앞쪽으로 자연스럽게 내려와 있어 손이 바닥에 닿지 않아도 지지가 성립한다. 모자와 우비는 각각 머리와 어깨에 지지되고 케이지는 바닥에 놓여 있다. 공중 부유나 명백히 불가능한 관절 배치는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "캐릭터의 외형(얼굴, 원자로) 및 지시된 소품(부츠, 우비, 모자)을 정확히 구현했으며, 요구된 카메라 구도와 프레이밍을 완벽하게 따름."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "지정된 구도와 배경은 잘 구현했으나, 부츠 대신 기계 맨발이 노출되었고 얼굴 디테일이 레퍼런스와 다소 차이가 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S58sh11_sel.png",
    "asset_id": "5da71df1-fc1f-4b45-8efe-69742855dfc4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c1d-90ab-7218-bef8-4610e0ca7e39",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh11"
  }
 },
 "S58sh41::signage": {
  "fp": "96b167e48687282f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh41": {
  "input_fingerprint": "8e4891774daf1bff",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우가 붙잡고 있던 병사의 가슴 정중앙에 화살촉이 퍽 꽂히는 잔혹한 근접 구도.\n\nLOCATION (lock): On the church floor immediately before the raised platform, below the archers' upper balcony. Daylight enters through the stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Lock the camera low beside the restrained soldier, oblique to his chest and tilted slightly downward, observing the arrow's impact directly rather than through a character's eyes. His upper torso occupies the central half of the frame, with the embedded arrow at its center and 이현우's restraining forearm entering from the left beside the soldier's neck; retain a narrow floor margin to establish their grounded position. Keep both faces outside the crop, with 이현우's attention remaining on the captive beyond the upper edge, and let the tightened distance alone emphasize the fatal interruption without added wound detail.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Striking and lodging in the center of the soldier's chest) — The visible shaft projects obliquely from the chest; the embedded tip is not exposed; used as Pinpoints the fatal interruption without obscuring the restraining arm; 성당 바닥 (Visible beside the restrained soldier); used as A narrow contextual margin establishes the low physical placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established stained-glass daylight and controlled contrast without a new impact spotlight or exaggerated blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The masked soldier is held on the floor by Hyunwoo at the neck, with an arrow embedded in his chest at the instant of his death. The turn of his head, his torso's facing direction, and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inside the steel cage in his oversized hat, boots and colorful raincoat; the access door has not yet been closed again. The church remains ruined and lit through stained glass. 이현우: He is bent over with his arms engaged in a choking hold after throwing a guard down. His leg remains previously bandaged, and he has just endured the electric-club assault.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우가 붙잡고 있던 병사의 가슴 정중앙에 화살촉이 퍽 꽂히는 잔혹한 근접 구도.\n\nLOCATION (lock): On the church floor immediately before the raised platform, below the archers' upper balcony. Daylight enters through the stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Lock the camera low beside the restrained soldier, oblique to his chest and tilted slightly downward, observing the arrow's impact directly rather than through a character's eyes. His upper torso occupies the central half of the frame, with the embedded arrow at its center and 이현우's restraining forearm entering from the left beside the soldier's neck; retain a narrow floor margin to establish their grounded position. Keep both faces outside the crop, with 이현우's attention remaining on the captive beyond the upper edge, and let the tightened distance alone emphasize the fatal interruption without added wound detail.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Striking and lodging in the center of the soldier's chest) — The visible shaft projects obliquely from the chest; the embedded tip is not exposed; used as Pinpoints the fatal interruption without obscuring the restraining arm; 성당 바닥 (Visible beside the restrained soldier); used as A narrow contextual margin establishes the low physical placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established stained-glass daylight and controlled contrast without a new impact spotlight or exaggerated blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The masked soldier is held on the floor by Hyunwoo at the neck, with an arrow embedded in his chest at the instant of his death. The turn of his head, his torso's facing direction, and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inside the steel cage in his oversized hat, boots and colorful raincoat; the access door has not yet been closed again. The church remains ruined and lit through stained glass. 이현우: He is bent over with his arms engaged in a choking hold after throwing a guard down. His leg remains previously bandaged, and he has just endured the electric-club assault.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우가 붙잡고 있던 병사의 가슴 정중앙에 화살촉이 퍽 꽂히는 잔혹한 근접 구도.\n\nLOCATION (lock): On the church floor immediately before the raised platform, below the archers' upper balcony. Daylight enters through the stained glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Lock the camera low beside the restrained soldier, oblique to his chest and tilted slightly downward, observing the arrow's impact directly rather than through a character's eyes. His upper torso occupies the central half of the frame, with the embedded arrow at its center and 이현우's restraining forearm entering from the left beside the soldier's neck; retain a narrow floor margin to establish their grounded position. Keep both faces outside the crop, with 이현우's attention remaining on the captive beyond the upper edge, and let the tightened distance alone emphasize the fatal interruption without added wound detail.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 화살 (Striking and lodging in the center of the soldier's chest) — The visible shaft projects obliquely from the chest; the embedded tip is not exposed; used as Pinpoints the fatal interruption without obscuring the restraining arm; 성당 바닥 (Visible beside the restrained soldier); used as A narrow contextual margin establishes the low physical placement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established stained-glass daylight and controlled contrast without a new impact spotlight or exaggerated blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The masked soldier is held on the floor by Hyunwoo at the neck, with an arrow embedded in his chest at the instant of his death. The turn of his head, his torso's facing direction, and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inside the steel cage in his oversized hat, boots and colorful raincoat; the access door has not yet been closed again. The church remains ruined and lit through stained glass. 이현우: He is bent over with his arms engaged in a choking hold after throwing a guard down. His leg remains previously bandaged, and he has just endured the electric-club assault.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화살이 병사의 가슴 중앙을 향해 사선으로 꽂혀 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 향해 뻗어 있습니다.",
    "built_space": "성당 바닥에 누워 있으며, 배경에 이전 샷에서 확인되는 제단으로 향하는 계단이 배치되어 위치가 일치합니다.",
    "entities": "병사의 붉은 금박 도포와 금색 가면 하단, 이현우의 어두운 회색 긴팔 소매, 가슴에 꽂힌 화살이 모두 식별됩니다.",
    "hard_violations": [],
    "physics": "병사의 몸은 바닥에 완전히 지지되어 누워 있고, 이현우의 손은 목을 단단히 쥐고 있으며, 화살은 가슴에 자연스럽게 고정되어 있습니다."
   },
   {
    "label": "B",
    "direction": "화살이 병사의 가슴 정중앙을 찌르고 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 강하게 틀어쥐고 있습니다.",
    "built_space": "잔해가 흩어진 성당 바닥과 스테인드글라스 빛이 보이나, 배경의 제단 계단 형태가 명확하지 않습니다.",
    "entities": "병사의 화려한 붉은 도포, 걷어 올린 소매 아래로 드러난 이현우의 맨팔, 가슴에 꽂힌 화살이 식별됩니다.",
    "hard_violations": [],
    "physics": "병사는 바닥에 밀착되어 누워 있고, 손의 압박과 꽂힌 화살의 물리적 상태가 안정적으로 묘사되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 근접 구도를 잘 구현했으며, 배경의 제단 계단과 이현우의 긴팔 의상 등 세부 설정이 레퍼런스와 정확히 일치합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "화살 피격과 구도는 요구사항을 따랐으나, 배경에서 제단 계단이 명확히 보이지 않고 이현우의 소매가 걷혀 있어 레퍼런스 일치도가 다소 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화살이 병사의 가슴 중앙을 향해 사선으로 꽂혀 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 향해 뻗어 있습니다.",
        "built_space": "성당 바닥에 누워 있으며, 배경에 이전 샷에서 확인되는 제단으로 향하는 계단이 배치되어 위치가 일치합니다.",
        "entities": "병사의 붉은 금박 도포와 금색 가면 하단, 이현우의 어두운 회색 긴팔 소매, 가슴에 꽂힌 화살이 모두 식별됩니다.",
        "hard_violations": [],
        "physics": "병사의 몸은 바닥에 완전히 지지되어 누워 있고, 이현우의 손은 목을 단단히 쥐고 있으며, 화살은 가슴에 자연스럽게 고정되어 있습니다."
       },
       {
        "label": "B",
        "direction": "화살이 병사의 가슴 정중앙을 찌르고 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 강하게 틀어쥐고 있습니다.",
        "built_space": "잔해가 흩어진 성당 바닥과 스테인드글라스 빛이 보이나, 배경의 제단 계단 형태가 명확하지 않습니다.",
        "entities": "병사의 화려한 붉은 도포, 걷어 올린 소매 아래로 드러난 이현우의 맨팔, 가슴에 꽂힌 화살이 식별됩니다.",
        "hard_violations": [],
        "physics": "병사는 바닥에 밀착되어 누워 있고, 손의 압박과 꽂힌 화살의 물리적 상태가 안정적으로 묘사되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 근접 구도를 잘 구현했으며, 배경의 제단 계단과 이현우의 긴팔 의상 등 세부 설정이 레퍼런스와 정확히 일치합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "화살 피격과 구도는 요구사항을 따랐으나, 배경에서 제단 계단이 명확히 보이지 않고 이현우의 소매가 걷혀 있어 레퍼런스 일치도가 다소 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화살이 병사의 가슴 중앙을 향해 사선으로 꽂혀 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 향해 뻗어 있습니다.",
        "built_space": "성당 바닥에 누워 있으며, 배경에 이전 샷에서 확인되는 제단으로 향하는 계단이 배치되어 위치가 일치합니다.",
        "entities": "병사의 붉은 금박 도포와 금색 가면 하단, 이현우의 어두운 회색 긴팔 소매, 가슴에 꽂힌 화살이 모두 식별됩니다.",
        "hard_violations": [],
        "physics": "병사의 몸은 바닥에 완전히 지지되어 누워 있고, 이현우의 손은 목을 단단히 쥐고 있으며, 화살은 가슴에 자연스럽게 고정되어 있습니다."
       },
       {
        "label": "B",
        "direction": "화살이 병사의 가슴 정중앙을 찌르고 있으며, 이현우의 팔이 왼쪽에서 들어와 목을 강하게 틀어쥐고 있습니다.",
        "built_space": "잔해가 흩어진 성당 바닥과 스테인드글라스 빛이 보이나, 배경의 제단 계단 형태가 명확하지 않습니다.",
        "entities": "병사의 화려한 붉은 도포, 걷어 올린 소매 아래로 드러난 이현우의 맨팔, 가슴에 꽂힌 화살이 식별됩니다.",
        "hard_violations": [],
        "physics": "병사는 바닥에 밀착되어 누워 있고, 손의 압박과 꽂힌 화살의 물리적 상태가 안정적으로 묘사되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "가슴 중심의 화살과 왼쪽에서 들어오는 제압 팔, 내려다보는 근접 구도는 더 충실하지만, 이전 인물의 붉은 예복을 병사에게 옮겼고 상처와 출혈을 과도하게 강조했다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "화살의 명중과 목 제압은 보이지만, 이전 인물의 의상을 그대로 사용하고 가면·복부·성당 배경까지 넓혀 지정된 하향 근접 구도에서 더 멀어졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화살 한 발의 축이 오른쪽 위에서 가슴 중앙의 관통 지점으로 이어지고 촉은 몸 안에 가려져 있다. 이현우의 팔은 왼쪽에서 들어와 손으로 병사의 목을 붙잡는다. 두 사람의 눈은 보이지 않아 시선의 목표는 확인할 수 없다.",
        "built_space": "몸 오른쪽과 상단에 잔해가 흩어진 석재 바닥과 유색광이 보인다. 단상, 계단, 발코니 같은 고정 구조물은 크롭 밖이어서 개수나 정확한 위치를 확인할 수 없다. 바닥 재질과 폐허의 채광은 참고 장소에 부합한다. 카메라는 몸 옆에서 아래로 비스듬히 내려다보지만 복부까지 포함하며, 상단에 가면 일부가 남아 있다.",
        "entities": "보이는 인체는 포로의 몸과 이현우의 팔·몸 일부로, 추가 인물은 없다. 이현우의 오염된 어두운 셔츠와 드러난 팔은 설정과 대체로 맞지만 얼굴이 없어 정확한 나이·민족적 외양·정체성은 확인할 수 없다. 포로는 금색 가면과 붉은 금문양 예복, 장식 허리띠를 착용하여 이전 장면의 중앙 인물 의상을 명백히 반복한다. 이는 이전 인물의 의상을 옮기지 말라는 지시와 다르다. 화살은 한 발이며 촉은 노출되지 않았다. 관통부의 벌어진 구멍과 흘러내리는 피는 상처 세부를 추가하지 말라는 요구보다 강하다.",
        "hard_violations": [],
        "physics": "병사의 몸은 바닥에 누워 있고 보이는 팔도 몸 옆으로 내려가 있다. 목은 이현우의 손에 붙잡혀 지지된다. 화살은 가슴에 박힌 부분을 지점으로 돌출되어 있어 떠 있는 물체가 아니다. 보이는 범위에서 지지 없는 사지나 불가능한 관절은 없다."
       },
       {
        "label": "B",
        "direction": "화살 한 발이 오른쪽 위에서 왼쪽 아래로 향하며 가슴 중앙 부근에 박혀 있다. 촉은 보이지 않고 깃과 축이 오른쪽 위로 돌출된다. 왼쪽에서 들어온 이현우의 손은 병사의 목을 조인다. 눈은 크롭 밖이므로 두 사람의 시선 방향은 확인할 수 없다.",
        "built_space": "오른쪽에 잔해가 있는 석재 바닥, 뒤에 여러 계단으로 이루어진 단상 접근부 한 곳, 기둥들과 스테인드글라스 창 일부가 보인다. 동일 시설의 명백한 중복은 없다. 폐허 성당이라는 장소는 맞지만 배경 구조물이 상당히 드러나고 가면 하부와 복부도 크게 포함된다. 시점은 몸 옆의 낮은 위치이나 지정된 약한 하향 관찰보다 몸통을 따라 뒤쪽을 바라보는 구도에 가깝다.",
        "entities": "포로와 이현우의 일부만 보이며 추가 인물은 없다. 이현우의 어둡고 흙 묻은 소매는 의상 설정과 맞지만 얼굴·머리·인이어는 확인할 수 없다. 포로의 금색 가면, 붉은 금문양 예복과 금속 장식 허리띠는 이전 장면의 중앙 인물에게서 가져온 것으로 보이며, 별도의 병사를 묘사해야 하는 지시와 어긋난다. 화살 한 발과 작은 어두운 혈흔이 있고 촉은 감춰져 있다. 출혈 표현은 A보다 절제되어 있다.",
        "hard_violations": [],
        "physics": "병사는 뒤로 기대어 있으며 목은 이현우의 손이 받치고, 몸 왼쪽 뒤에는 이현우의 검은 옷을 입은 다리 또는 몸 일부가 접해 있다. 하체와 내려간 팔은 바닥 쪽으로 이어진다. 등 전체의 접촉면은 가려져 있지만 몸이 무지지 상태로 공중에 떠 있다고 단정할 근거는 없다. 화살은 가슴에 박혀 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "가슴 중심의 화살과 왼쪽에서 들어오는 제압 팔, 내려다보는 근접 구도는 더 충실하지만, 이전 인물의 붉은 예복을 병사에게 옮겼고 상처와 출혈을 과도하게 강조했다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "화살의 명중과 목 제압은 보이지만, 이전 인물의 의상을 그대로 사용하고 가면·복부·성당 배경까지 넓혀 지정된 하향 근접 구도에서 더 멀어졌다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "화살 한 발의 축이 오른쪽 위에서 가슴 중앙의 관통 지점으로 이어지고 촉은 몸 안에 가려져 있다. 이현우의 팔은 왼쪽에서 들어와 손으로 병사의 목을 붙잡는다. 두 사람의 눈은 보이지 않아 시선의 목표는 확인할 수 없다.",
        "built_space": "몸 오른쪽과 상단에 잔해가 흩어진 석재 바닥과 유색광이 보인다. 단상, 계단, 발코니 같은 고정 구조물은 크롭 밖이어서 개수나 정확한 위치를 확인할 수 없다. 바닥 재질과 폐허의 채광은 참고 장소에 부합한다. 카메라는 몸 옆에서 아래로 비스듬히 내려다보지만 복부까지 포함하며, 상단에 가면 일부가 남아 있다.",
        "entities": "보이는 인체는 포로의 몸과 이현우의 팔·몸 일부로, 추가 인물은 없다. 이현우의 오염된 어두운 셔츠와 드러난 팔은 설정과 대체로 맞지만 얼굴이 없어 정확한 나이·민족적 외양·정체성은 확인할 수 없다. 포로는 금색 가면과 붉은 금문양 예복, 장식 허리띠를 착용하여 이전 장면의 중앙 인물 의상을 명백히 반복한다. 이는 이전 인물의 의상을 옮기지 말라는 지시와 다르다. 화살은 한 발이며 촉은 노출되지 않았다. 관통부의 벌어진 구멍과 흘러내리는 피는 상처 세부를 추가하지 말라는 요구보다 강하다.",
        "hard_violations": [],
        "physics": "병사의 몸은 바닥에 누워 있고 보이는 팔도 몸 옆으로 내려가 있다. 목은 이현우의 손에 붙잡혀 지지된다. 화살은 가슴에 박힌 부분을 지점으로 돌출되어 있어 떠 있는 물체가 아니다. 보이는 범위에서 지지 없는 사지나 불가능한 관절은 없다."
       },
       {
        "label": "A",
        "direction": "화살 한 발이 오른쪽 위에서 왼쪽 아래로 향하며 가슴 중앙 부근에 박혀 있다. 촉은 보이지 않고 깃과 축이 오른쪽 위로 돌출된다. 왼쪽에서 들어온 이현우의 손은 병사의 목을 조인다. 눈은 크롭 밖이므로 두 사람의 시선 방향은 확인할 수 없다.",
        "built_space": "오른쪽에 잔해가 있는 석재 바닥, 뒤에 여러 계단으로 이루어진 단상 접근부 한 곳, 기둥들과 스테인드글라스 창 일부가 보인다. 동일 시설의 명백한 중복은 없다. 폐허 성당이라는 장소는 맞지만 배경 구조물이 상당히 드러나고 가면 하부와 복부도 크게 포함된다. 시점은 몸 옆의 낮은 위치이나 지정된 약한 하향 관찰보다 몸통을 따라 뒤쪽을 바라보는 구도에 가깝다.",
        "entities": "포로와 이현우의 일부만 보이며 추가 인물은 없다. 이현우의 어둡고 흙 묻은 소매는 의상 설정과 맞지만 얼굴·머리·인이어는 확인할 수 없다. 포로의 금색 가면, 붉은 금문양 예복과 금속 장식 허리띠는 이전 장면의 중앙 인물에게서 가져온 것으로 보이며, 별도의 병사를 묘사해야 하는 지시와 어긋난다. 화살 한 발과 작은 어두운 혈흔이 있고 촉은 감춰져 있다. 출혈 표현은 A보다 절제되어 있다.",
        "hard_violations": [],
        "physics": "병사는 뒤로 기대어 있으며 목은 이현우의 손이 받치고, 몸 왼쪽 뒤에는 이현우의 검은 옷을 입은 다리 또는 몸 일부가 접해 있다. 하체와 내려간 팔은 바닥 쪽으로 이어진다. 등 전체의 접촉면은 가려져 있지만 몸이 무지지 상태로 공중에 떠 있다고 단정할 근거는 없다. 화살은 가슴에 박혀 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.8,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1800,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1800,
    "verdict_ko": "지정된 로우 앵글 근접 구도를 잘 구현했으며, 배경의 제단 계단과 이현우의 긴팔 의상 등 세부 설정이 레퍼런스와 정확히 일치합니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "화살 피격과 구도는 요구사항을 따랐으나, 배경에서 제단 계단이 명확히 보이지 않고 이현우의 소매가 걷혀 있어 레퍼런스 일치도가 다소 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S58sh11_sel.png",
    "asset_id": "5da71df1-fc1f-4b45-8efe-69742855dfc4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c22-aaef-71ba-aab9-a5caf2a41017",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh11"
  }
 },
 "S59sh9::signage": {
  "fp": "ee276885badbebbf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7113099bbf757fd9": {
  "subjects": [],
  "subject_text": "익산 마을 지하감옥 감방, 복도·계단, 찰리 수용 케이지실\n쇠창살 감방들이 복도를 따라 이어진 지하 공간. 바닥에 건초와 고인 물이 있고, 작은 지상창과 계단, 천장 개구부 아래 승강 케이지가 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L161",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::underground_cell": {
  "input_fingerprint": "8fbc2ca3416bb056",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "underground_cell",
    "tags": [
     "S59sh25",
     "S59sh44",
     "S59sh9"
    ]
   },
   "context_sig": "70c716b71b82014e"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 지하감옥 감방, 복도·계단, 찰리 수용 케이지실: 어둡고 습한 지하 쇠창살 감옥으로 바닥에 얕은 물이 고여 있다. (특징: 두꺼운 쇠창살과 지하로 내려가는 돌계단; 바닥에 고인 탁한 물과 마른 건초더미; 땟국물이 묻은 얼굴과 방사능 피폭 흉터가 배에 선명한 수빈; 열쇠로 열리는 강철 케이지 문)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 감옥에 끌려와 갇히는 현우와 앰버.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 마을 지하감옥 감방, 복도·계단, 찰리 수용 케이지실: 어둡고 습한 지하 쇠창살 감옥으로 바닥에 얕은 물이 고여 있다. (특징: 두꺼운 쇠창살과 지하로 내려가는 돌계단; 바닥에 고인 탁한 물과 마른 건초더미; 땟국물이 묻은 얼굴과 방사능 피폭 흉터가 배에 선명한 수빈; 열쇠로 열리는 강철 케이지 문)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 감옥에 끌려와 갇히는 현우와 앰버.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_underground_cell_e9765d.png",
  "asset_id": "5cb97c24-4a32-42d0-a1ec-e12ec940b442",
  "input_asset_ids": [
   "978b2d8e-a4da-4cf1-892f-be6f6f2781c6"
  ],
  "origin_tag": "S59sh9",
  "place_text": "Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.",
  "origin_inputs": {
   "place_text": "Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.",
   "time_of_day_en": "day",
   "conti_asset_id": "978b2d8e-a4da-4cf1-892f-be6f6f2781c6"
  }
 },
 "S59sh9::bgfirst_bg": {
  "input_fingerprint": "74f165b64dca70b7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 품에 안긴 앰버의 코 바로 밑으로 수빈의 손이 젖은 건초를 바짝 들이민 근접 구도.\n\nLOCATION (lock): Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 앰버's nose height, oblique to her face and beside 이현우's enclosing arm, in a direct observational close shot. Her face occupies the center-right, his arm curves along the left and lower edges, and 수빈's hand enters from the right with a naturally sized handful of wet hay immediately beneath the nose, leaving both nostrils unobstructed. 앰버 lowers her gaze toward the offered hay as her breathing settles, while 이현우 looks down protectively from the upper-left edge and 수빈, visible only by her hand and forearm, attends to 앰버 from outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held immediately beneath 앰버's nose after being soaked in pooled cell water); used as A small lower-center focal detail connecting care, breath and the offered hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the cell, preserving gentle facial detail and the wet hay without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 품에 안긴 앰버의 코 바로 밑으로 수빈의 손이 젖은 건초를 바짝 들이민 근접 구도.\n\nLOCATION (lock): Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 앰버's nose height, oblique to her face and beside 이현우's enclosing arm, in a direct observational close shot. Her face occupies the center-right, his arm curves along the left and lower edges, and 수빈's hand enters from the right with a naturally sized handful of wet hay immediately beneath the nose, leaving both nostrils unobstructed. 앰버 lowers her gaze toward the offered hay as her breathing settles, while 이현우 looks down protectively from the upper-left edge and 수빈, visible only by her hand and forearm, attends to 앰버 from outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held immediately beneath 앰버's nose after being soaked in pooled cell water); used as A small lower-center focal detail connecting care, breath and the offered hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the cell, preserving gentle facial detail and the wet hay without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S59sh9__bgfirst_bg.png",
  "asset_id": "a834cb97-0cc7-4e05-a921-57f22dd9c460",
  "input_asset_ids": [
   "978b2d8e-a4da-4cf1-892f-be6f6f2781c6",
   "5cb97c24-4a32-42d0-a1ec-e12ec940b442"
  ]
 },
 "S59sh9": {
  "input_fingerprint": "82707165990ee25b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 품에 안긴 앰버의 코 바로 밑으로 수빈의 손이 젖은 건초를 바짝 들이민 근접 구도.\n\nLOCATION (lock): Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 앰버's nose height, oblique to her face and beside 이현우's enclosing arm, in a direct observational close shot. Her face occupies the center-right, his arm curves along the left and lower edges, and 수빈's hand enters from the right with a naturally sized handful of wet hay immediately beneath the nose, leaving both nostrils unobstructed. 앰버 lowers her gaze toward the offered hay as her breathing settles, while 이현우 looks down protectively from the upper-left edge and 수빈, visible only by her hand and forearm, attends to 앰버 from outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held immediately beneath 앰버's nose after being soaked in pooled cell water); used as A small lower-center focal detail connecting care, breath and the offered hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the cell, preserving gentle facial detail and the wet hay without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell is locked, with hay on the floor and pooled water available inside. 이현우: He is confined in the cell with his arms held in a protective embrace. His previously bandaged leg and the effects of the recent assaults remain. 앰버: She is confined in the cell during a severe coughing episode; her breathing begins to settle at this point. 수빈: She has a dirty, wounded face and radiation lesions on her torso beneath her T-shirt. She holds a handful of hay soaked in the cell's pooled water.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 품에 안긴 앰버의 코 바로 밑으로 수빈의 손이 젖은 건초를 바짝 들이민 근접 구도.\n\nLOCATION (lock): Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 앰버's nose height, oblique to her face and beside 이현우's enclosing arm, in a direct observational close shot. Her face occupies the center-right, his arm curves along the left and lower edges, and 수빈's hand enters from the right with a naturally sized handful of wet hay immediately beneath the nose, leaving both nostrils unobstructed. 앰버 lowers her gaze toward the offered hay as her breathing settles, while 이현우 looks down protectively from the upper-left edge and 수빈, visible only by her hand and forearm, attends to 앰버 from outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held immediately beneath 앰버's nose after being soaked in pooled cell water); used as A small lower-center focal detail connecting care, breath and the offered hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the cell, preserving gentle facial detail and the wet hay without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell is locked, with hay on the floor and pooled water available inside. 이현우: He is confined in the cell with his arms held in a protective embrace. His previously bandaged leg and the effects of the recent assaults remain. 앰버: She is confined in the cell during a severe coughing episode; her breathing begins to settle at this point. 수빈: She has a dirty, wounded face and radiation lesions on her torso beneath her T-shirt. She holds a handful of hay soaked in the cell's pooled water.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이현우의 품에 안긴 앰버의 코 바로 밑으로 수빈의 손이 젖은 건초를 바짝 들이민 근접 구도.\n\nLOCATION (lock): Inside a barred underground prison cell, near the damp floor and loose hay. Light is limited, with no specific fixture established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach at 앰버's nose height, oblique to her face and beside 이현우's enclosing arm, in a direct observational close shot. Her face occupies the center-right, his arm curves along the left and lower edges, and 수빈's hand enters from the right with a naturally sized handful of wet hay immediately beneath the nose, leaving both nostrils unobstructed. 앰버 lowers her gaze toward the offered hay as her breathing settles, while 이현우 looks down protectively from the upper-left edge and 수빈, visible only by her hand and forearm, attends to 앰버 from outside the crop.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 물에 적신 건초 (Held immediately beneath 앰버's nose after being soaked in pooled cell water); used as A small lower-center focal detail connecting care, breath and the offered hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient illumination appropriate to the cell, preserving gentle facial detail and the wet hay without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The underground cell is locked, with hay on the floor and pooled water available inside. 이현우: He is confined in the cell with his arms held in a protective embrace. His previously bandaged leg and the effects of the recent assaults remain. 앰버: She is confined in the cell during a severe coughing episode; her breathing begins to settle at this point. 수빈: She has a dirty, wounded face and radiation lesions on her torso beneath her T-shirt. She holds a handful of hay soaked in the cell's pooled water.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S59sh9__bgfirst_bg.png",
     "asset_id": "a834cb97-0cc7-4e05-a921-57f22dd9c460",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S59sh9.png",
     "asset_id": "978b2d8e-a4da-4cf1-892f-be6f6f2781c6",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_underground_cell_e9765d.png",
     "asset_id": "5cb97c24-4a32-42d0-a1ec-e12ec940b442",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들의 시선은 건초를 향하지만, 건초가 마른 상태이며 앰버의 가슴 높이에 위치함.",
    "built_space": "제공된 레퍼런스(돌벽 구조)와 전혀 일치하지 않는 콘크리트 벽체로 이루어진 공간임.",
    "entities": "앰버의 필수 소품인 방진 마스크가 누락되었으며, 프레임 밖에서 손만 등장해야 할 수빈의 상반신과 얼굴이 노출됨.",
    "hard_violations": [],
    "physics": "손으로 건초를 쥐고 있는 상태와 인물들의 앉은 자세는 물리적으로 정상적임."
   },
   {
    "label": "B",
    "direction": "앰버와 이현우의 시선이 건초를 향하나, 건초 위치가 코 바로 아래가 아닌 턱 부근임.",
    "built_space": "레퍼런스의 돌벽, 아치형 입구, 철창, 계단 등 공간 구조를 완벽하게 재현함.",
    "entities": "앰버의 마스크가 목에 걸쳐 구현되었으나, 손만 등장해야 할 수빈의 머리와 어깨가 프레임 우측에 노출됨.",
    "hard_violations": [],
    "physics": "수빈의 손이 젖은 건초를 자연스럽게 받치고 있으며, 이현우의 포옹 자세에 무리가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 지하던전 배경을 정확히 반영하고 클로즈업 구도에 근접했으나, 수빈이 손만 보여야 한다는 제약을 위반하고 어깨와 머리가 노출된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 구성되었고, 배경이 레퍼런스와 완전히 다르며 앰버의 방진 마스크가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 이현우의 시선이 건초를 향하나, 건초 위치가 코 바로 아래가 아닌 턱 부근임.",
        "built_space": "레퍼런스의 돌벽, 아치형 입구, 철창, 계단 등 공간 구조를 완벽하게 재현함.",
        "entities": "앰버의 마스크가 목에 걸쳐 구현되었으나, 손만 등장해야 할 수빈의 머리와 어깨가 프레임 우측에 노출됨.",
        "hard_violations": [],
        "physics": "수빈의 손이 젖은 건초를 자연스럽게 받치고 있으며, 이현우의 포옹 자세에 무리가 없음."
       },
       {
        "label": "A",
        "direction": "인물들의 시선은 건초를 향하지만, 건초가 마른 상태이며 앰버의 가슴 높이에 위치함.",
        "built_space": "제공된 레퍼런스(돌벽 구조)와 전혀 일치하지 않는 콘크리트 벽체로 이루어진 공간임.",
        "entities": "앰버의 필수 소품인 방진 마스크가 누락되었으며, 프레임 밖에서 손만 등장해야 할 수빈의 상반신과 얼굴이 노출됨.",
        "hard_violations": [],
        "physics": "손으로 건초를 쥐고 있는 상태와 인물들의 앉은 자세는 물리적으로 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 지하던전 배경을 정확히 반영하고 클로즈업 구도에 근접했으나, 수빈이 손만 보여야 한다는 제약을 위반하고 어깨와 머리가 노출된 점이 아쉽습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 구성되었고, 배경이 레퍼런스와 완전히 다르며 앰버의 방진 마스크가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 이현우의 시선이 건초를 향하나, 건초 위치가 코 바로 아래가 아닌 턱 부근임.",
        "built_space": "레퍼런스의 돌벽, 아치형 입구, 철창, 계단 등 공간 구조를 완벽하게 재현함.",
        "entities": "앰버의 마스크가 목에 걸쳐 구현되었으나, 손만 등장해야 할 수빈의 머리와 어깨가 프레임 우측에 노출됨.",
        "hard_violations": [],
        "physics": "수빈의 손이 젖은 건초를 자연스럽게 받치고 있으며, 이현우의 포옹 자세에 무리가 없음."
       },
       {
        "label": "A",
        "direction": "인물들의 시선은 건초를 향하지만, 건초가 마른 상태이며 앰버의 가슴 높이에 위치함.",
        "built_space": "제공된 레퍼런스(돌벽 구조)와 전혀 일치하지 않는 콘크리트 벽체로 이루어진 공간임.",
        "entities": "앰버의 필수 소품인 방진 마스크가 누락되었으며, 프레임 밖에서 손만 등장해야 할 수빈의 상반신과 얼굴이 노출됨.",
        "hard_violations": [],
        "physics": "손으로 건초를 쥐고 있는 상태와 인물들의 앉은 자세는 물리적으로 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "젖은 건초, 보호하는 포옹, 석조 감방은 더 충실하지만, 건초가 코 바로 밑에 있지 않고 구도가 넓어 수빈의 머리와 상체까지 보인다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "건초를 가슴 앞에 내민 넓은 구도로 핵심 근접 동작을 놓쳤으며, 건조해 보이는 건초와 현대식 회색 감방도 기준 장소 및 소품 상태와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 오른쪽 아래에 내밀어진 건초 쪽으로 시선을 낮추고, 이현우는 왼쪽 위에서 앰버를 내려다본다. 수빈의 손은 오른쪽에서 들어오지만 건초는 코 바로 밑이 아니라 턱보다 아래인 가슴 높이에 머문다. 콧구멍은 가려지지 않는다. 수빈의 얼굴은 대부분 돌아서 있어 정확한 시선은 확인하기 어렵다.",
        "built_space": "뒤쪽에 자물쇠판이 있는 철창문 한 개, 그 너머 아치형 통로와 올라가는 계단, 거친 석조 벽, 젖은 바닥과 흩어진 건초가 보인다. 작은 벽등 하나도 보이며 장소 참조의 통로와 대체로 대응한다. 이현우와 앰버는 바닥 가까이 포개져 앉아 있다. 다만 무릎과 허리까지 포함하고 배경을 넓게 드러내어 요구된 코 높이의 밀착 클로즈업이 아니다. 앰버의 얼굴도 중앙 오른쪽보다는 중앙에 가깝다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 더러운 어두운 셔츠와 인이어 장치를 착용한다. 앰버는 금발의 어린 여자아이이며 참조의 얼굴 인상과 카키 작업복에 비교적 가깝다. 내려진 방진 마스크와 공구 벨트도 보인다. 수빈은 검은 단발머리, 올리브색 소매와 맨손으로 나타나지만, 손과 전완만 보여야 하는 지시와 달리 머리와 어깨까지 들어온다. 손에는 참조의 반장갑이 없다. 건초에는 물방울과 젖은 광택이 보인다.",
        "hard_violations": [],
        "physics": "앰버의 몸은 뒤에 있는 이현우의 몸통과 둘러싼 팔에 기대어 지지된다. 이현우의 굽힌 다리와 낮은 자세는 바닥에 앉은 상태와 양립한다. 수빈은 손바닥과 굽힌 손가락으로 건초를 받치고 있으며 물방울은 아래로 떨어진다. 방진 마스크는 목 쪽 끈에 매달려 몸에 닿아 있다. 지지 없이 떠 있는 물체나 명백히 불가능한 자세는 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 가슴 앞의 건초를 내려다보고, 이현우는 앰버 쪽으로 고개와 시선을 기울인다. 수빈 역시 손과 건초 쪽을 내려다본다. 오른쪽에서 들어온 손의 건초는 코 바로 밑이 아니라 가슴 앞에 있어 요구된 접근 지점에 도달하지 않는다. 코는 노출되어 있다.",
        "built_space": "왼쪽에 회색 철창 구획 한 면, 뒤쪽에 평평한 회색 벽과 수직 이음부, 바닥에 건초 더미가 보인다. 참조의 거친 석재, 낡은 철창문과 잠금판, 아치형 통로를 확인할 수 없어 같은 장소라는 일치도가 낮다. 이현우는 앰버 뒤왼쪽의 낮은 위치에서 허리를 감싸며, 수빈은 오른쪽에서 몸을 숙인다. 세 사람의 상체와 앰버의 허리까지 보여 지정된 클로즈업보다 상당히 넓다. 반사나 중복된 고정 설비의 문제는 보이지 않는다.",
        "entities": "이현우는 검은 머리의 젊은 동아시아계 남성으로 더러운 어두운 셔츠와 귀의 장치를 갖췄다. 앰버는 금발의 어린 여자아이지만 참조보다 얼굴 인상이 다르고, 카키 작업복 앞부분의 구조도 다르다. 허리 벨트는 보이나 참조의 정교한 방진 마스크는 식별되지 않는다. 수빈은 검은 단발머리와 올리브색 옷을 입었지만 얼굴과 상체까지 노출되어 손과 전완만 허용한 구도를 벗어난다. 맨손은 참조의 반장갑과 다르며, 얼굴의 상처도 뚜렷하지 않다. 건초는 길고 성긴 다발로, 물에 흠뻑 적신 광택이나 물방울이 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 팔과 손이 앰버의 허리를 감싸고 있어 접촉과 보호 동작은 성립한다. 하체와 바닥 접점은 잘렸지만 공중에 떠 있다는 징후는 없다. 수빈은 손가락으로 건초 줄기를 실제로 쥐고 있으며 전완은 화면 밖 몸통으로 연결된다. 건초 다발은 손에서 양쪽으로 뻗어 중력에 따라 일부 처진다. 명백한 무지지 물체나 불가능한 신체 구조는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "젖은 건초, 보호하는 포옹, 석조 감방은 더 충실하지만, 건초가 코 바로 밑에 있지 않고 구도가 넓어 수빈의 머리와 상체까지 보인다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "건초를 가슴 앞에 내민 넓은 구도로 핵심 근접 동작을 놓쳤으며, 건조해 보이는 건초와 현대식 회색 감방도 기준 장소 및 소품 상태와 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 오른쪽 아래에 내밀어진 건초 쪽으로 시선을 낮추고, 이현우는 왼쪽 위에서 앰버를 내려다본다. 수빈의 손은 오른쪽에서 들어오지만 건초는 코 바로 밑이 아니라 턱보다 아래인 가슴 높이에 머문다. 콧구멍은 가려지지 않는다. 수빈의 얼굴은 대부분 돌아서 있어 정확한 시선은 확인하기 어렵다.",
        "built_space": "뒤쪽에 자물쇠판이 있는 철창문 한 개, 그 너머 아치형 통로와 올라가는 계단, 거친 석조 벽, 젖은 바닥과 흩어진 건초가 보인다. 작은 벽등 하나도 보이며 장소 참조의 통로와 대체로 대응한다. 이현우와 앰버는 바닥 가까이 포개져 앉아 있다. 다만 무릎과 허리까지 포함하고 배경을 넓게 드러내어 요구된 코 높이의 밀착 클로즈업이 아니다. 앰버의 얼굴도 중앙 오른쪽보다는 중앙에 가깝다.",
        "entities": "이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 더러운 어두운 셔츠와 인이어 장치를 착용한다. 앰버는 금발의 어린 여자아이이며 참조의 얼굴 인상과 카키 작업복에 비교적 가깝다. 내려진 방진 마스크와 공구 벨트도 보인다. 수빈은 검은 단발머리, 올리브색 소매와 맨손으로 나타나지만, 손과 전완만 보여야 하는 지시와 달리 머리와 어깨까지 들어온다. 손에는 참조의 반장갑이 없다. 건초에는 물방울과 젖은 광택이 보인다.",
        "hard_violations": [],
        "physics": "앰버의 몸은 뒤에 있는 이현우의 몸통과 둘러싼 팔에 기대어 지지된다. 이현우의 굽힌 다리와 낮은 자세는 바닥에 앉은 상태와 양립한다. 수빈은 손바닥과 굽힌 손가락으로 건초를 받치고 있으며 물방울은 아래로 떨어진다. 방진 마스크는 목 쪽 끈에 매달려 몸에 닿아 있다. 지지 없이 떠 있는 물체나 명백히 불가능한 자세는 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 가슴 앞의 건초를 내려다보고, 이현우는 앰버 쪽으로 고개와 시선을 기울인다. 수빈 역시 손과 건초 쪽을 내려다본다. 오른쪽에서 들어온 손의 건초는 코 바로 밑이 아니라 가슴 앞에 있어 요구된 접근 지점에 도달하지 않는다. 코는 노출되어 있다.",
        "built_space": "왼쪽에 회색 철창 구획 한 면, 뒤쪽에 평평한 회색 벽과 수직 이음부, 바닥에 건초 더미가 보인다. 참조의 거친 석재, 낡은 철창문과 잠금판, 아치형 통로를 확인할 수 없어 같은 장소라는 일치도가 낮다. 이현우는 앰버 뒤왼쪽의 낮은 위치에서 허리를 감싸며, 수빈은 오른쪽에서 몸을 숙인다. 세 사람의 상체와 앰버의 허리까지 보여 지정된 클로즈업보다 상당히 넓다. 반사나 중복된 고정 설비의 문제는 보이지 않는다.",
        "entities": "이현우는 검은 머리의 젊은 동아시아계 남성으로 더러운 어두운 셔츠와 귀의 장치를 갖췄다. 앰버는 금발의 어린 여자아이지만 참조보다 얼굴 인상이 다르고, 카키 작업복 앞부분의 구조도 다르다. 허리 벨트는 보이나 참조의 정교한 방진 마스크는 식별되지 않는다. 수빈은 검은 단발머리와 올리브색 옷을 입었지만 얼굴과 상체까지 노출되어 손과 전완만 허용한 구도를 벗어난다. 맨손은 참조의 반장갑과 다르며, 얼굴의 상처도 뚜렷하지 않다. 건초는 길고 성긴 다발로, 물에 흠뻑 적신 광택이나 물방울이 보이지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 팔과 손이 앰버의 허리를 감싸고 있어 접촉과 보호 동작은 성립한다. 하체와 바닥 접점은 잘렸지만 공중에 떠 있다는 징후는 없다. 수빈은 손가락으로 건초 줄기를 실제로 쥐고 있으며 전완은 화면 밖 몸통으로 연결된다. 건초 다발은 손에서 양쪽으로 뻗어 중력에 따라 일부 처진다. 명백한 무지지 물체나 불가능한 신체 구조는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.1,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.1,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1100
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 지하던전 배경을 정확히 반영하고 클로즈업 구도에 근접했으나, 수빈이 손만 보여야 한다는 제약을 위반하고 어깨와 머리가 노출된 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 1100,
    "verdict_ko": "요구된 클로즈업 대신 미디엄 샷으로 구성되었고, 배경이 레퍼런스와 완전히 다르며 앰버의 방진 마스크가 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_underground_cell_e9765d.png",
    "asset_id": "5cb97c24-4a32-42d0-a1ec-e12ec940b442",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c27-f0dd-7228-9159-4f03fb074852",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S59sh9__bgfirst_bg.png",
   "bg_asset_id": "a834cb97-0cc7-4e05-a921-57f22dd9c460",
   "bg_record_key": "S59sh9::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "underground_cell",
   "groupbg_asset_id": "5cb97c24-4a32-42d0-a1ec-e12ec940b442"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S59sh25::signage": {
  "fp": "c1e97dc4f16f964c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S59sh25": {
  "input_fingerprint": "06a0bd5c67cb3932",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 들어 올려진 티셔츠 아래로 붉게 진물고 짓무른 수빈의 방사능 피폭 흉터가 적나라하게 드러난 복부 근접 구도.\n\nLOCATION (lock): Inside the shared underground prison cell, in the open standing space beside the prisoners. The cell remains dim, with no defined light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop the forward move at 수빈's abdominal height on 이현우's side of the dialogue axis, viewing her exposed torso obliquely rather than straight on. Frame the radiation-damaged abdomen across the central portion, with her hand holding the raised T-shirt at the upper edge and a small margin of cell space beside her, keeping the red, weeping damage factual rather than magnified into an extreme close-up. Her face remains above the crop and her attention stays on 이현우 off-screen as she deliberately holds the evidence for him to see.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 들어 올린 티셔츠 (Held above the exposed radiation-damaged abdomen) — The lifted front hem enters along the upper frame boundary; used as Makes the deliberate act of disclosure readable while limiting the crop to the abdomen.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cell's subdued ambient illumination with enough tonal separation to reveal the red, weeping skin damage without sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked, with hay and pooled water on the floor. 수빈: Her face remains dirty and wounded. Her raised T-shirt exposes the radiation lesions across her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 들어 올려진 티셔츠 아래로 붉게 진물고 짓무른 수빈의 방사능 피폭 흉터가 적나라하게 드러난 복부 근접 구도.\n\nLOCATION (lock): Inside the shared underground prison cell, in the open standing space beside the prisoners. The cell remains dim, with no defined light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop the forward move at 수빈's abdominal height on 이현우's side of the dialogue axis, viewing her exposed torso obliquely rather than straight on. Frame the radiation-damaged abdomen across the central portion, with her hand holding the raised T-shirt at the upper edge and a small margin of cell space beside her, keeping the red, weeping damage factual rather than magnified into an extreme close-up. Her face remains above the crop and her attention stays on 이현우 off-screen as she deliberately holds the evidence for him to see.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 들어 올린 티셔츠 (Held above the exposed radiation-damaged abdomen) — The lifted front hem enters along the upper frame boundary; used as Makes the deliberate act of disclosure readable while limiting the crop to the abdomen.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cell's subdued ambient illumination with enough tonal separation to reveal the red, weeping skin damage without sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked, with hay and pooled water on the floor. 수빈: Her face remains dirty and wounded. Her raised T-shirt exposes the radiation lesions across her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 들어 올려진 티셔츠 아래로 붉게 진물고 짓무른 수빈의 방사능 피폭 흉터가 적나라하게 드러난 복부 근접 구도.\n\nLOCATION (lock): Inside the shared underground prison cell, in the open standing space beside the prisoners. The cell remains dim, with no defined light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop the forward move at 수빈's abdominal height on 이현우's side of the dialogue axis, viewing her exposed torso obliquely rather than straight on. Frame the radiation-damaged abdomen across the central portion, with her hand holding the raised T-shirt at the upper edge and a small margin of cell space beside her, keeping the red, weeping damage factual rather than magnified into an extreme close-up. Her face remains above the crop and her attention stays on 이현우 off-screen as she deliberately holds the evidence for him to see.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 들어 올린 티셔츠 (Held above the exposed radiation-damaged abdomen) — The lifted front hem enters along the upper frame boundary; used as Makes the deliberate act of disclosure readable while limiting the crop to the abdomen.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cell's subdued ambient illumination with enough tonal separation to reveal the red, weeping skin damage without sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell remains locked, with hay and pooled water on the floor. 수빈: Her face remains dirty and wounded. Her raised T-shirt exposes the radiation lesions across her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 밖 왼쪽을 향함. 손으로 셔츠를 쥐고 올림.",
    "built_space": "돌벽, 배경의 쇠창살, 바닥의 짚 등 참조된 지하 감옥 세트와 일치함.",
    "entities": "수빈의 얼굴과 의상은 일치하나, 흉터 대신 흰색 복대가 있음. 좌측 가장자리에 프롬프트에 없는 인물 뒷모습이 보임.",
    "hard_violations": [
     "[gemini-pro] 화면 밖(off-screen)에 있어야 할 인물의 신체 일부가 프레임 왼쪽에 등장함 (추가 인물)",
     "[gemini-pro] 지정된 방사능 흉터 대신 프롬프트에 없는 흰색 복대가 그려짐 (발명된 사물)",
     "[gpt-high] 수빈만 보여야 하는 장면에 다른 인물의 머리와 어깨를 전경으로 추가했다.",
     "[gpt-high] 노출해야 할 피폭 복부를 덮는 복대를 새로 만들어 핵심 증거를 가렸다.",
     "[gpt-high] 지정된 열린 공간의 서 있는 배치를 앉은 자세로 바꾸었다."
    ],
    "physics": "손으로 셔츠 자락을 쥔 형태와 신체 자세는 정상적으로 지지됨."
   },
   {
    "label": "B",
    "direction": "몸통은 사선 방향. 손이 셔츠 자락을 위로 당기고 있음.",
    "built_space": "돌벽, 아치형 쇠창살 문, 물 고인 바닥의 짚 등 감옥 배경이 정확히 구현됨.",
    "entities": "수빈의 의상(재킷, 벨트)이 참조와 일치하며, 복부에 붉게 진물은 방사능 피폭 흉터가 명확하게 묘사됨.",
    "hard_violations": [],
    "physics": "셔츠를 당기는 손의 파지와 서 있는 몸통의 지지가 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴을 배제하는 복부 클로즈업 프레이밍 지시를 어겼으며, 핵심인 피폭 흉터 대신 복대를 그리고 화면 밖 인물까지 등장시켜 주요 조건을 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "얼굴을 프레임 밖으로 자른 정확한 복부 클로즈업 구도를 구현했으며, 요구된 붉은 피폭 흉터와 배경 세트를 훌륭하게 재현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 밖 왼쪽을 향함. 손으로 셔츠를 쥐고 올림.",
        "built_space": "돌벽, 배경의 쇠창살, 바닥의 짚 등 참조된 지하 감옥 세트와 일치함.",
        "entities": "수빈의 얼굴과 의상은 일치하나, 흉터 대신 흰색 복대가 있음. 좌측 가장자리에 프롬프트에 없는 인물 뒷모습이 보임.",
        "hard_violations": [
         "화면 밖(off-screen)에 있어야 할 인물의 신체 일부가 프레임 왼쪽에 등장함 (추가 인물)",
         "지정된 방사능 흉터 대신 프롬프트에 없는 흰색 복대가 그려짐 (발명된 사물)"
        ],
        "physics": "손으로 셔츠 자락을 쥔 형태와 신체 자세는 정상적으로 지지됨."
       },
       {
        "label": "B",
        "direction": "몸통은 사선 방향. 손이 셔츠 자락을 위로 당기고 있음.",
        "built_space": "돌벽, 아치형 쇠창살 문, 물 고인 바닥의 짚 등 감옥 배경이 정확히 구현됨.",
        "entities": "수빈의 의상(재킷, 벨트)이 참조와 일치하며, 복부에 붉게 진물은 방사능 피폭 흉터가 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "셔츠를 당기는 손의 파지와 서 있는 몸통의 지지가 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴을 배제하는 복부 클로즈업 프레이밍 지시를 어겼으며, 핵심인 피폭 흉터 대신 복대를 그리고 화면 밖 인물까지 등장시켜 주요 조건을 위반함."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "얼굴을 프레임 밖으로 자른 정확한 복부 클로즈업 구도를 구현했으며, 요구된 붉은 피폭 흉터와 배경 세트를 훌륭하게 재현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 밖 왼쪽을 향함. 손으로 셔츠를 쥐고 올림.",
        "built_space": "돌벽, 배경의 쇠창살, 바닥의 짚 등 참조된 지하 감옥 세트와 일치함.",
        "entities": "수빈의 얼굴과 의상은 일치하나, 흉터 대신 흰색 복대가 있음. 좌측 가장자리에 프롬프트에 없는 인물 뒷모습이 보임.",
        "hard_violations": [
         "화면 밖(off-screen)에 있어야 할 인물의 신체 일부가 프레임 왼쪽에 등장함 (추가 인물)",
         "지정된 방사능 흉터 대신 프롬프트에 없는 흰색 복대가 그려짐 (발명된 사물)"
        ],
        "physics": "손으로 셔츠 자락을 쥔 형태와 신체 자세는 정상적으로 지지됨."
       },
       {
        "label": "B",
        "direction": "몸통은 사선 방향. 손이 셔츠 자락을 위로 당기고 있음.",
        "built_space": "돌벽, 아치형 쇠창살 문, 물 고인 바닥의 짚 등 감옥 배경이 정확히 구현됨.",
        "entities": "수빈의 의상(재킷, 벨트)이 참조와 일치하며, 복부에 붉게 진물은 방사능 피폭 흉터가 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "셔츠를 당기는 손의 파지와 서 있는 몸통의 지지가 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 비스듬한 복부 근접 구도와 손으로 올린 티셔츠, 붉고 젖은 피폭 병변을 충실히 구현했으나 손과 옷단이 지정된 상단 경계보다 조금 아래에 있다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "얼굴까지 포함한 앉은 구도로 바뀌고 추가 인물이 들어왔으며, 핵심인 노출된 피폭 병변을 새로 만든 복대로 가려 장면의 요구를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복부를 약간 비스듬히 바라보며 배꼽과 옆구리가 함께 보인다. 손은 티셔츠를 위로 당겨 병변을 카메라 쪽에 노출한다. 얼굴과 이현우는 화면 밖이므로 시선의 실제 도착점이나 대화축의 정확한 쪽은 확인할 수 없다.",
        "built_space": "왼쪽에 거친 석벽, 오른쪽 뒤에 닫힌 철창문 한 개와 그 너머 계단이 보인다. 작은 따뜻한 발광부 두 곳, 짚과 물이 고인 바닥이 이전 장면의 공간을 이어 간다. 수빈은 벽과 철창 사이 열린 공간에 서 있는 몸통 배치이며 다른 수감자는 보이지 않는다. 젖은 바닥의 밝은 반사는 뒤쪽 밝은 공간과 양립한다.",
        "entities": "수빈 한 명의 마른 몸통과 손, 낡고 먼지 묻은 올리브 재킷, 회색 티셔츠, 카고 바지 허리 부분, 벨트와 허리 주머니가 보인다. 의상과 체격은 참조에 부합한다. 얼굴과 머리는 요구대로 잘려 나가 나이·민족적 외양·얼굴 동일성은 판정할 수 없다. 복부에는 붉은 불규칙 병변과 진물처럼 빛나는 표면이 있어 요구한 피폭 손상을 표현한다.",
        "hard_violations": [],
        "physics": "손가락이 티셔츠를 실제로 움켜쥐고 있으며 주름이 잡은 지점으로 모여 올라간 옷의 지지가 명확하다. 재킷은 몸에 걸쳐지고 주머니는 허리 장비에 연결되어 있다. 하체와 발은 프레임 밖이지만 몸통은 자연스러운 직립 자세이며 공중에 뜬 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "수빈은 화면 왼쪽 전경의 흐릿한 상대 인물 쪽을 바라본다. 손은 티셔츠를 위로 당기지만 상대에게 드러내는 것은 병변이 아니라 복대다. 상대를 보는 행동 자체는 읽히지만 이현우를 화면 밖에 두라는 조건과 맞지 않는다.",
        "built_space": "거친 석벽, 오른쪽 뒤 철창문 한 개와 계단, 벽등 한 개, 왼쪽 짚 덮인 낮은 침상 한 개가 보인다. 바닥의 짚과 물 반사는 이전 장소와 대체로 이어진다. 그러나 수빈은 열린 공간에 서 있는 대신 무릎을 앞으로 내민 앉은 자세로 보이고, 카메라는 복부 높이의 근접 구도보다 얼굴과 상체를 넓게 포함한다.",
        "entities": "검은 단발의 젊은 동아시아 여성과 낡은 올리브 재킷, 회색 티셔츠, 카고 바지와 벨트는 수빈의 참조 외양에 대체로 맞는다. 다만 얼굴의 상처는 뚜렷하지 않다. 복부는 흰 천과 넓은 베이지색 복대로 덮여 붉고 진물 나는 병변이 전혀 보이지 않는다. 왼쪽 전경에는 허용되지 않은 다른 인물의 머리와 어깨 일부가 들어온다.",
        "hard_violations": [
         "수빈만 보여야 하는 장면에 다른 인물의 머리와 어깨를 전경으로 추가했다.",
         "노출해야 할 피폭 복부를 덮는 복대를 새로 만들어 핵심 증거를 가렸다.",
         "지정된 열린 공간의 서 있는 배치를 앉은 자세로 바꾸었다."
        ],
        "physics": "손이 티셔츠를 잡고 있어 들어 올린 천의 지지는 자연스럽다. 복대는 몸통을 둘러 고정되어 있으며 떠 있지 않다. 굽힌 허벅지와 기울어진 몸통은 앉은 자세로 가능하지만 엉덩이의 정확한 좌면 접촉은 화면 밖이라 확인되지 않는다. 이를 지지 없는 공중 부양으로 볼 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴을 제외한 비스듬한 복부 근접 구도와 손으로 올린 티셔츠, 붉고 젖은 피폭 병변을 충실히 구현했으나 손과 옷단이 지정된 상단 경계보다 조금 아래에 있다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "얼굴까지 포함한 앉은 구도로 바뀌고 추가 인물이 들어왔으며, 핵심인 노출된 피폭 병변을 새로 만든 복대로 가려 장면의 요구를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "복부를 약간 비스듬히 바라보며 배꼽과 옆구리가 함께 보인다. 손은 티셔츠를 위로 당겨 병변을 카메라 쪽에 노출한다. 얼굴과 이현우는 화면 밖이므로 시선의 실제 도착점이나 대화축의 정확한 쪽은 확인할 수 없다.",
        "built_space": "왼쪽에 거친 석벽, 오른쪽 뒤에 닫힌 철창문 한 개와 그 너머 계단이 보인다. 작은 따뜻한 발광부 두 곳, 짚과 물이 고인 바닥이 이전 장면의 공간을 이어 간다. 수빈은 벽과 철창 사이 열린 공간에 서 있는 몸통 배치이며 다른 수감자는 보이지 않는다. 젖은 바닥의 밝은 반사는 뒤쪽 밝은 공간과 양립한다.",
        "entities": "수빈 한 명의 마른 몸통과 손, 낡고 먼지 묻은 올리브 재킷, 회색 티셔츠, 카고 바지 허리 부분, 벨트와 허리 주머니가 보인다. 의상과 체격은 참조에 부합한다. 얼굴과 머리는 요구대로 잘려 나가 나이·민족적 외양·얼굴 동일성은 판정할 수 없다. 복부에는 붉은 불규칙 병변과 진물처럼 빛나는 표면이 있어 요구한 피폭 손상을 표현한다.",
        "hard_violations": [],
        "physics": "손가락이 티셔츠를 실제로 움켜쥐고 있으며 주름이 잡은 지점으로 모여 올라간 옷의 지지가 명확하다. 재킷은 몸에 걸쳐지고 주머니는 허리 장비에 연결되어 있다. 하체와 발은 프레임 밖이지만 몸통은 자연스러운 직립 자세이며 공중에 뜬 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "수빈은 화면 왼쪽 전경의 흐릿한 상대 인물 쪽을 바라본다. 손은 티셔츠를 위로 당기지만 상대에게 드러내는 것은 병변이 아니라 복대다. 상대를 보는 행동 자체는 읽히지만 이현우를 화면 밖에 두라는 조건과 맞지 않는다.",
        "built_space": "거친 석벽, 오른쪽 뒤 철창문 한 개와 계단, 벽등 한 개, 왼쪽 짚 덮인 낮은 침상 한 개가 보인다. 바닥의 짚과 물 반사는 이전 장소와 대체로 이어진다. 그러나 수빈은 열린 공간에 서 있는 대신 무릎을 앞으로 내민 앉은 자세로 보이고, 카메라는 복부 높이의 근접 구도보다 얼굴과 상체를 넓게 포함한다.",
        "entities": "검은 단발의 젊은 동아시아 여성과 낡은 올리브 재킷, 회색 티셔츠, 카고 바지와 벨트는 수빈의 참조 외양에 대체로 맞는다. 다만 얼굴의 상처는 뚜렷하지 않다. 복부는 흰 천과 넓은 베이지색 복대로 덮여 붉고 진물 나는 병변이 전혀 보이지 않는다. 왼쪽 전경에는 허용되지 않은 다른 인물의 머리와 어깨 일부가 들어온다.",
        "hard_violations": [
         "수빈만 보여야 하는 장면에 다른 인물의 머리와 어깨를 전경으로 추가했다.",
         "노출해야 할 피폭 복부를 덮는 복대를 새로 만들어 핵심 증거를 가렸다.",
         "지정된 열린 공간의 서 있는 배치를 앉은 자세로 바꾸었다."
        ],
        "physics": "손이 티셔츠를 잡고 있어 들어 올린 천의 지지는 자연스럽다. 복대는 몸통을 둘러 고정되어 있으며 떠 있지 않다. 굽힌 허벅지와 기울어진 몸통은 앉은 자세로 가능하지만 엉덩이의 정확한 좌면 접촉은 화면 밖이라 확인되지 않는다. 이를 지지 없는 공중 부양으로 볼 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.54,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.29,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 화면 밖(off-screen)에 있어야 할 인물의 신체 일부가 프레임 왼쪽에 등장함 (추가 인물)",
     "[gemini-pro] 지정된 방사능 흉터 대신 프롬프트에 없는 흰색 복대가 그려짐 (발명된 사물)",
     "[gpt-high] 수빈만 보여야 하는 장면에 다른 인물의 머리와 어깨를 전경으로 추가했다.",
     "[gpt-high] 노출해야 할 피폭 복부를 덮는 복대를 새로 만들어 핵심 증거를 가렸다.",
     "[gpt-high] 지정된 열린 공간의 서 있는 배치를 앉은 자세로 바꾸었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 290,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 290,
    "verdict_ko": "얼굴을 배제하는 복부 클로즈업 프레이밍 지시를 어겼으며, 핵심인 피폭 흉터 대신 복대를 그리고 화면 밖 인물까지 등장시켜 주요 조건을 위반함.  ★위반: [gemini-pro] 화면 밖(off-screen)에 있어야 할 인물의 신체 일부가 프레임 왼쪽에 등장함 (추가 인물) / [gemini-pro] 지정된 방사능 흉터 대신 프롬프트에 없는 흰색 복대가 그려짐 (발명된 사물) / [gpt-high] 수빈만 보여야 하는 장면에 다른 인물의 머리와 어깨를 전경으로 추가했다. / [gpt-high] 노출해야 할 피폭 복부를 덮는 복대를 새로 만들어 핵심 증거를 가렸다. / [gpt-high] 지정된 열린 공간의 서 있는 배치를 앉은 자세로 바꾸었다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "얼굴을 프레임 밖으로 자른 정확한 복부 클로즈업 구도를 구현했으며, 요구된 붉은 피폭 흉터와 배경 세트를 훌륭하게 재현함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S59sh9_sel.png",
    "asset_id": "91817dd5-83bf-4e96-8d4b-b1167d1a8625",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c40-9e66-7890-9911-e1e3b7b2b1e5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh9"
  }
 },
 "S59sh44::signage": {
  "fp": "f91d3cc33f21f0c0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S59sh44": {
  "input_fingerprint": "97cba674e06393da",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 감옥 밖에서 들려오는 거대한 함성 소리에 이현우와 수빈이 일제히 천장 쪽으로 고개를 확 돌린 찰나.\n\nLOCATION (lock): Inside the underground prison cell beneath the village's public gathering area. The dim interior has no specified fixed light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the shared-reaction composition from lower-chest height on the established side of the dialogue axis, looking diagonally upward in a direct medium two-shot with generous space above both heads. Place 이현우 on the left and 수빈 on the right, catching his chin lifting as her shoulders begin to turn toward the ceiling and the off-screen cheering above; their reactions share a cause without matching exactly. Limit the upward tilt to preserve both faces and keep 앰버 below the lower crop beside 이현우, emphasizing the change of gaze rather than altering their positions or illumination.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감옥 쇠창살 (Still confining the characters) — A partial oblique section remains behind the two figures at the side edge; used as Preserves the prison setting without intercepting either upward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established cell illumination unchanged so the abrupt upward attention, not a lighting cue, carries the alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cell is still locked, with hay and pooled water remaining on its floor. 이현우: He remains imprisoned with his earlier leg bandage and recent assault injuries. He is still eating the potato. 수빈: She remains in the cell with a dirty, wounded face and radiation lesions on her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 감옥 밖에서 들려오는 거대한 함성 소리에 이현우와 수빈이 일제히 천장 쪽으로 고개를 확 돌린 찰나.\n\nLOCATION (lock): Inside the underground prison cell beneath the village's public gathering area. The dim interior has no specified fixed light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the shared-reaction composition from lower-chest height on the established side of the dialogue axis, looking diagonally upward in a direct medium two-shot with generous space above both heads. Place 이현우 on the left and 수빈 on the right, catching his chin lifting as her shoulders begin to turn toward the ceiling and the off-screen cheering above; their reactions share a cause without matching exactly. Limit the upward tilt to preserve both faces and keep 앰버 below the lower crop beside 이현우, emphasizing the change of gaze rather than altering their positions or illumination.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감옥 쇠창살 (Still confining the characters) — A partial oblique section remains behind the two figures at the side edge; used as Preserves the prison setting without intercepting either upward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established cell illumination unchanged so the abrupt upward attention, not a lighting cue, carries the alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cell is still locked, with hay and pooled water remaining on its floor. 이현우: He remains imprisoned with his earlier leg bandage and recent assault injuries. He is still eating the potato. 수빈: She remains in the cell with a dirty, wounded face and radiation lesions on her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 감옥 밖에서 들려오는 거대한 함성 소리에 이현우와 수빈이 일제히 천장 쪽으로 고개를 확 돌린 찰나.\n\nLOCATION (lock): Inside the underground prison cell beneath the village's public gathering area. The dim interior has no specified fixed light source. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the shared-reaction composition from lower-chest height on the established side of the dialogue axis, looking diagonally upward in a direct medium two-shot with generous space above both heads. Place 이현우 on the left and 수빈 on the right, catching his chin lifting as her shoulders begin to turn toward the ceiling and the off-screen cheering above; their reactions share a cause without matching exactly. Limit the upward tilt to preserve both faces and keep 앰버 below the lower crop beside 이현우, emphasizing the change of gaze rather than altering their positions or illumination.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감옥 쇠창살 (Still confining the characters) — A partial oblique section remains behind the two figures at the side edge; used as Preserves the prison setting without intercepting either upward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established cell illumination unchanged so the abrupt upward attention, not a lighting cue, carries the alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barred cell is still locked, with hay and pooled water remaining on its floor. 이현우: He remains imprisoned with his earlier leg bandage and recent assault injuries. He is still eating the potato. 수빈: She remains in the cell with a dirty, wounded face and radiation lesions on her torso.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 대각선으로 시선을 던지고 있음.",
    "built_space": "이전 샷과 동일한 돌벽과 바닥의 짚, 물기, 우측 가장자리의 감옥 쇠창살이 정확히 배치됨.",
    "entities": "이현우와 수빈 모두 참조 이미지의 외모, 복장, 상처를 정확히 따르며, 이현우는 먹다 남은 감자를 들고 있음. 앰버는 지시대로 프레임 아래로 배제됨.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 안정적으로 앉아 있으며, 이현우의 손이 감자를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 시선을 고정하고 있음.",
    "built_space": "우측 창살 등은 일치하나, 이현우 뒤쪽 배경이 이전 샷의 돌벽과 달리 볏짚단이나 거친 질감의 벽으로 이질적으로 묘사됨.",
    "entities": "두 인물의 외양과 복장, 이현우의 무전기 및 감자 모두 참조 이미지 및 텍스트와 일치함.",
    "hard_violations": [],
    "physics": "인물들의 자세가 바닥에 잘 지지되어 있고, 손가락이 감자를 자연스럽게 잡고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상승하는 시선 처리와 프레이밍을 정확히 구현했으며, 이전 샷의 돌벽 및 창살 등 공간적 배경과의 연속성이 뛰어남."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 시선, 소품은 잘 표현되었으나 이현우 뒤쪽 벽면의 질감이 이전 샷의 매끄러운 돌벽과 달라 공간 일치도가 다소 떨어짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 대각선으로 시선을 던지고 있음.",
        "built_space": "이전 샷과 동일한 돌벽과 바닥의 짚, 물기, 우측 가장자리의 감옥 쇠창살이 정확히 배치됨.",
        "entities": "이현우와 수빈 모두 참조 이미지의 외모, 복장, 상처를 정확히 따르며, 이현우는 먹다 남은 감자를 들고 있음. 앰버는 지시대로 프레임 아래로 배제됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 앉아 있으며, 이현우의 손이 감자를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 시선을 고정하고 있음.",
        "built_space": "우측 창살 등은 일치하나, 이현우 뒤쪽 배경이 이전 샷의 돌벽과 달리 볏짚단이나 거친 질감의 벽으로 이질적으로 묘사됨.",
        "entities": "두 인물의 외양과 복장, 이현우의 무전기 및 감자 모두 참조 이미지 및 텍스트와 일치함.",
        "hard_violations": [],
        "physics": "인물들의 자세가 바닥에 잘 지지되어 있고, 손가락이 감자를 자연스럽게 잡고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 상승하는 시선 처리와 프레이밍을 정확히 구현했으며, 이전 샷의 돌벽 및 창살 등 공간적 배경과의 연속성이 뛰어남."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 시선, 소품은 잘 표현되었으나 이현우 뒤쪽 벽면의 질감이 이전 샷의 매끄러운 돌벽과 달라 공간 일치도가 다소 떨어짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 대각선으로 시선을 던지고 있음.",
        "built_space": "이전 샷과 동일한 돌벽과 바닥의 짚, 물기, 우측 가장자리의 감옥 쇠창살이 정확히 배치됨.",
        "entities": "이현우와 수빈 모두 참조 이미지의 외모, 복장, 상처를 정확히 따르며, 이현우는 먹다 남은 감자를 들고 있음. 앰버는 지시대로 프레임 아래로 배제됨.",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 안정적으로 앉아 있으며, 이현우의 손이 감자를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 위쪽 천장(화면 밖)을 향해 시선을 고정하고 있음.",
        "built_space": "우측 창살 등은 일치하나, 이현우 뒤쪽 배경이 이전 샷의 돌벽과 달리 볏짚단이나 거친 질감의 벽으로 이질적으로 묘사됨.",
        "entities": "두 인물의 외양과 복장, 이현우의 무전기 및 감자 모두 참조 이미지 및 텍스트와 일치함.",
        "hard_violations": [],
        "physics": "인물들의 자세가 바닥에 잘 지지되어 있고, 손가락이 감자를 자연스럽게 잡고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 현우와 오른쪽 수빈이 함께 천장 쪽을 보는 미디엄 투숏과 가장자리 쇠창살 배치가 더 정확하지만, 두 사람 머리 위의 넉넉한 여백은 부족하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수빈의 어깨 회전과 두 사람의 다른 반응은 잘 보이지만, 현우의 머리가 상단에서 잘리고 시선이 좌우로 갈라져 A보다 공동 반응 구도가 덜 명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 턱을 들고 화면 오른쪽 위를 보며, 수빈도 고개를 더 젖혀 오른쪽 위의 화면 밖 천장을 본다. 두 시선은 같은 상부 영역을 향하되 머리 각도는 다르다. 쇠창살은 두 사람의 시선 앞을 가로막지 않는다. 손에 든 감자는 가슴 아래에 머물러 있다.",
        "built_space": "왼쪽과 중앙에는 낡고 축축한 조적벽, 오른쪽 뒤에는 닫힌 철창문 한 구획과 그 너머 올라가는 계단이 보인다. 뒤쪽 낮은 받침 한 구획에 짚이 있고, 오른쪽 바닥에는 물과 반사가 남아 있다. 가까운 벽등 한 개와 더 먼 작은 불빛 한 점은 이전 장면에도 있는 요소다. 현우는 왼쪽, 수빈은 오른쪽의 낮은 위치에 있으며 철창은 측면 배경에 남는다. 상반신 중심의 약한 올려다보기 구도지만 현우의 정수리 위 여백이 거의 없어 넉넉한 상부 공간 지시는 충족하지 못한다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 현우와 젊은 동아시아계 여성 수빈 두 명뿐이다. 현우의 헝클어진 검은 머리, 얼굴, 어두운 오염된 셔츠와 소형 인이어가 참조와 대체로 맞으며 손에는 베어 먹은 감자가 있다. 수빈의 검은 단발, 얼굴, 올리브색 낡은 재킷과 얼굴의 흙먼지·상처도 부합한다. 국적은 외모만으로 검증할 수 없다. 앰버는 보이지 않으며, 다리 붕대와 옷 안쪽 몸통 병변은 이 크롭에서 확인할 수 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 아래로 이어지는 몸통과 굽힌 팔을 가진 낮은 착석 자세로 보인다. 엉덩이와 발의 정확한 접점은 프레임 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 감자는 현우의 손가락과 손바닥이 받치고 있다. 턱을 들고 목과 어깨를 돌리는 자세는 갑자기 위쪽 소리에 반응하는 동작으로 가능하다."
       },
       {
        "label": "B",
        "direction": "현우는 턱을 들면서 화면 왼쪽 위를 보고, 수빈은 어깨와 고개를 돌려 오른쪽 위를 본다. 둘 다 화면 밖 천장 방향에 주의를 두지만 같은 지점을 바라보지는 않는다. 넓게 퍼진 함성에 대한 반응으로는 가능하나 공동 원인의 시각적 연결은 A보다 약하다. 철창은 시선 경로를 가로막지 않는다.",
        "built_space": "낡은 조적벽, 오른쪽 뒤의 닫힌 철창문 한 구획, 그 너머 위로 이어지는 계단, 중앙 아래의 짚 얹힌 낮은 받침 한 구획이 보인다. 가까운 벽등 한 개와 먼 불빛 한 점, 젖은 오른쪽 바닥도 이전 장소의 요소와 대응한다. 두 사람은 현우 왼쪽·수빈 오른쪽으로 낮게 자리한다. 다만 철창문이 측면의 부분 요소보다 넓은 배경 면적으로 드러나고, 현우의 머리 윗부분이 상단에 잘려 요구한 넉넉한 머리 위 공간이 없다.",
        "entities": "젊은 동아시아계 남녀 두 명의 얼굴과 검은 머리, 현우의 어두운 셔츠와 수빈의 올리브 재킷은 참조에 대체로 부합한다. 현우의 얼굴과 목에는 상처가 있고 손에는 먹던 감자가 있으며, 인이어는 이전 사진과 반대편 귀에 보인다. 수빈의 얼굴에도 흙먼지와 상처가 있다. 국적은 시각적으로 확정할 수 없다. 앰버는 노출되지 않는다. 다리 붕대와 몸통의 피폭 병변은 크롭과 의복 때문에 확인할 수 없으며, 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "하단에 굽힌 무릎이 드러나 두 사람이 낮게 앉아 있다는 자세가 자연스럽게 이어진다. 정확한 바닥 접점은 잘렸지만 공중 부양이나 지지 없는 신체는 보이지 않는다. 현우는 손으로 감자를 확실히 잡고 있다. 수빈의 몸통 비틀림과 올라간 어깨, 현우의 목 회전은 갑작스러운 청각 반응으로 가능한 동작이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 현우와 오른쪽 수빈이 함께 천장 쪽을 보는 미디엄 투숏과 가장자리 쇠창살 배치가 더 정확하지만, 두 사람 머리 위의 넉넉한 여백은 부족하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "수빈의 어깨 회전과 두 사람의 다른 반응은 잘 보이지만, 현우의 머리가 상단에서 잘리고 시선이 좌우로 갈라져 A보다 공동 반응 구도가 덜 명확하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 턱을 들고 화면 오른쪽 위를 보며, 수빈도 고개를 더 젖혀 오른쪽 위의 화면 밖 천장을 본다. 두 시선은 같은 상부 영역을 향하되 머리 각도는 다르다. 쇠창살은 두 사람의 시선 앞을 가로막지 않는다. 손에 든 감자는 가슴 아래에 머물러 있다.",
        "built_space": "왼쪽과 중앙에는 낡고 축축한 조적벽, 오른쪽 뒤에는 닫힌 철창문 한 구획과 그 너머 올라가는 계단이 보인다. 뒤쪽 낮은 받침 한 구획에 짚이 있고, 오른쪽 바닥에는 물과 반사가 남아 있다. 가까운 벽등 한 개와 더 먼 작은 불빛 한 점은 이전 장면에도 있는 요소다. 현우는 왼쪽, 수빈은 오른쪽의 낮은 위치에 있으며 철창은 측면 배경에 남는다. 상반신 중심의 약한 올려다보기 구도지만 현우의 정수리 위 여백이 거의 없어 넉넉한 상부 공간 지시는 충족하지 못한다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 현우와 젊은 동아시아계 여성 수빈 두 명뿐이다. 현우의 헝클어진 검은 머리, 얼굴, 어두운 오염된 셔츠와 소형 인이어가 참조와 대체로 맞으며 손에는 베어 먹은 감자가 있다. 수빈의 검은 단발, 얼굴, 올리브색 낡은 재킷과 얼굴의 흙먼지·상처도 부합한다. 국적은 외모만으로 검증할 수 없다. 앰버는 보이지 않으며, 다리 붕대와 옷 안쪽 몸통 병변은 이 크롭에서 확인할 수 없다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 아래로 이어지는 몸통과 굽힌 팔을 가진 낮은 착석 자세로 보인다. 엉덩이와 발의 정확한 접점은 프레임 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 감자는 현우의 손가락과 손바닥이 받치고 있다. 턱을 들고 목과 어깨를 돌리는 자세는 갑자기 위쪽 소리에 반응하는 동작으로 가능하다."
       },
       {
        "label": "A",
        "direction": "현우는 턱을 들면서 화면 왼쪽 위를 보고, 수빈은 어깨와 고개를 돌려 오른쪽 위를 본다. 둘 다 화면 밖 천장 방향에 주의를 두지만 같은 지점을 바라보지는 않는다. 넓게 퍼진 함성에 대한 반응으로는 가능하나 공동 원인의 시각적 연결은 A보다 약하다. 철창은 시선 경로를 가로막지 않는다.",
        "built_space": "낡은 조적벽, 오른쪽 뒤의 닫힌 철창문 한 구획, 그 너머 위로 이어지는 계단, 중앙 아래의 짚 얹힌 낮은 받침 한 구획이 보인다. 가까운 벽등 한 개와 먼 불빛 한 점, 젖은 오른쪽 바닥도 이전 장소의 요소와 대응한다. 두 사람은 현우 왼쪽·수빈 오른쪽으로 낮게 자리한다. 다만 철창문이 측면의 부분 요소보다 넓은 배경 면적으로 드러나고, 현우의 머리 윗부분이 상단에 잘려 요구한 넉넉한 머리 위 공간이 없다.",
        "entities": "젊은 동아시아계 남녀 두 명의 얼굴과 검은 머리, 현우의 어두운 셔츠와 수빈의 올리브 재킷은 참조에 대체로 부합한다. 현우의 얼굴과 목에는 상처가 있고 손에는 먹던 감자가 있으며, 인이어는 이전 사진과 반대편 귀에 보인다. 수빈의 얼굴에도 흙먼지와 상처가 있다. 국적은 시각적으로 확정할 수 없다. 앰버는 노출되지 않는다. 다리 붕대와 몸통의 피폭 병변은 크롭과 의복 때문에 확인할 수 없으며, 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "하단에 굽힌 무릎이 드러나 두 사람이 낮게 앉아 있다는 자세가 자연스럽게 이어진다. 정확한 바닥 접점은 잘렸지만 공중 부양이나 지지 없는 신체는 보이지 않는다. 현우는 손으로 감자를 확실히 잡고 있다. 수빈의 몸통 비틀림과 올라간 어깨, 현우의 목 회전은 갑작스러운 청각 반응으로 가능한 동작이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 상승하는 시선 처리와 프레이밍을 정확히 구현했으며, 이전 샷의 돌벽 및 창살 등 공간적 배경과의 연속성이 뛰어남."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "인물과 시선, 소품은 잘 표현되었으나 이현우 뒤쪽 벽면의 질감이 이전 샷의 매끄러운 돌벽과 달라 공간 일치도가 다소 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S59sh9_sel.png",
    "asset_id": "91817dd5-83bf-4e96-8d4b-b1167d1a8625",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c53-bd68-71dc-b519-4c3021b53cc5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh9"
  }
 },
 "S60sh57::signage": {
  "fp": "7b92033992151576",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S60sh57": {
  "input_fingerprint": "79fb3625c6e97f6d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 격투장 관중석 한가운데로 거대한 포탄이 떨어지며 엄청난 화염과 파편이 솟구치는 폭발의 찰나.\n\nLOCATION (lock): In the spectator section of the open-air fighting arena in front of the village church, at dusk. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the elevated arena perimeter, pan diagonally across the fighting area at a shallow downward angle and catch the instant the view reaches the mortar impact in the spectator stands. Place the impact just right of center, the arena floor across the lower foreground and substantial clear space above the rising flame and fragments, retaining enough undamaged stand structure to establish scale. This direct environmental wide shot contains no readable people or robots; the pan's arrival at the explosion, rather than a simultaneous push-in, provides the visual accent.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mortar impact in the spectator stands in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 격투장 관중석 (Struck by a mortar explosion, with fragments erupting from the impact area) — Viewed diagonally across the arena from its perimeter; used as Middle-distance impact setting with intact portions maintaining scale; 폭발 화염과 파편 (Erupting upward at the instant of impact); used as A contained center-right burst with headroom for its upward expansion; 격투장 바닥 (Visible between the perimeter viewpoint and the stands) — Recedes diagonally from the lower foreground toward the impact; used as Maintains the established viewing axis and distance from the explosion.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-day ambient light is briefly overtaken by the explosion's intense flare, with controlled highlight roll-off preserving surrounding structural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torches surround the arena in the fading daylight; Charlie lies battered on the arena floor, while B-200 has its hand-mounted guns deployed and aimed. Several underground cell doors are already open following the escape. An explosion erupts in a section of the spectator stands, leaving smoke over the impact area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 격투장 관중석 한가운데로 거대한 포탄이 떨어지며 엄청난 화염과 파편이 솟구치는 폭발의 찰나.\n\nLOCATION (lock): In the spectator section of the open-air fighting arena in front of the village church, at dusk. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the elevated arena perimeter, pan diagonally across the fighting area at a shallow downward angle and catch the instant the view reaches the mortar impact in the spectator stands. Place the impact just right of center, the arena floor across the lower foreground and substantial clear space above the rising flame and fragments, retaining enough undamaged stand structure to establish scale. This direct environmental wide shot contains no readable people or robots; the pan's arrival at the explosion, rather than a simultaneous push-in, provides the visual accent.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mortar impact in the spectator stands in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 격투장 관중석 (Struck by a mortar explosion, with fragments erupting from the impact area) — Viewed diagonally across the arena from its perimeter; used as Middle-distance impact setting with intact portions maintaining scale; 폭발 화염과 파편 (Erupting upward at the instant of impact); used as A contained center-right burst with headroom for its upward expansion; 격투장 바닥 (Visible between the perimeter viewpoint and the stands) — Recedes diagonally from the lower foreground toward the impact; used as Maintains the established viewing axis and distance from the explosion.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-day ambient light is briefly overtaken by the explosion's intense flare, with controlled highlight roll-off preserving surrounding structural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torches surround the arena in the fading daylight; Charlie lies battered on the arena floor, while B-200 has its hand-mounted guns deployed and aimed. Several underground cell doors are already open following the escape. An explosion erupts in a section of the spectator stands, leaving smoke over the impact area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 격투장 관중석 한가운데로 거대한 포탄이 떨어지며 엄청난 화염과 파편이 솟구치는 폭발의 찰나.\n\nLOCATION (lock): In the spectator section of the open-air fighting arena in front of the village church, at dusk. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the elevated arena perimeter, pan diagonally across the fighting area at a shallow downward angle and catch the instant the view reaches the mortar impact in the spectator stands. Place the impact just right of center, the arena floor across the lower foreground and substantial clear space above the rising flame and fragments, retaining enough undamaged stand structure to establish scale. This direct environmental wide shot contains no readable people or robots; the pan's arrival at the explosion, rather than a simultaneous push-in, provides the visual accent.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mortar impact in the spectator stands in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 격투장 관중석 (Struck by a mortar explosion, with fragments erupting from the impact area) — Viewed diagonally across the arena from its perimeter; used as Middle-distance impact setting with intact portions maintaining scale; 폭발 화염과 파편 (Erupting upward at the instant of impact); used as A contained center-right burst with headroom for its upward expansion; 격투장 바닥 (Visible between the perimeter viewpoint and the stands) — Recedes diagonally from the lower foreground toward the impact; used as Maintains the established viewing axis and distance from the explosion.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-day ambient light is briefly overtaken by the explosion's intense flare, with controlled highlight roll-off preserving surrounding structural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torches surround the arena in the fading daylight; Charlie lies battered on the arena floor, while B-200 has its hand-mounted guns deployed and aimed. Several underground cell doors are already open following the escape. An explosion erupts in a section of the spectator stands, leaving smoke over the impact area.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "폭발 화염과 파편이 우측 중단의 관중석에서 위를 향해 솟구침.",
    "built_space": "경기장 외곽의 높은 뷰. 전경에 난간이 있고 경기장 바닥 너머로 관중석과 교회가 배치됨. 감옥 문은 닫혀 있는 것으로 보임.",
    "entities": "지시에 맞는 화염, 잔해, 교회가 있음. 바닥에 식별 가능한 인물이나 로봇이 없어 '인물 배제' 지시를 충족함.",
    "hard_violations": [
     "[gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다."
    ],
    "physics": "강력한 폭발의 힘으로 파편들이 공중으로 튀어 오름."
   },
   {
    "label": "B",
    "direction": "화염과 파편이 우측 관중석에서 공중으로 솟구침.",
    "built_space": "전경 난간 너머로 경기장 바닥과 관중석, 교회가 보임. 벽면의 감옥 문 중 일부가 열려 있음.",
    "entities": "폭발과 잔해, 교회가 있음. 경기장 바닥에 식별 가능한 여러 명의 사람 시체들이 흩어져 있음.",
    "hard_violations": [
     "[gemini-pro] 창조된 인물 (지시문에 명시되지 않은 사람의 시체들을 경기장 바닥에 여러 구 추가함)",
     "[gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다.",
     "[gpt-high] 경기장 바닥에 여러 인간 신체와 하단의 일으켜 앉은 인물을 추가하여, 지정되지 않은 인물과 배치를 만들었습니다."
    ],
    "physics": "폭발 충격으로 구조물 파편들이 허공으로 날아감."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 넓은 샷의 카메라 구도와 폭발 스케일을 잘 구현했으며, 식별 가능한 인물을 배제하라는 제한을 충실히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "환경과 폭발 연출은 우수하나, 경기장 바닥에 지시문에 없는 식별 가능한 사람 시체들을 임의로 추가하여 핵심 규칙을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "폭발 화염과 파편이 우측 중단의 관중석에서 위를 향해 솟구침.",
        "built_space": "경기장 외곽의 높은 뷰. 전경에 난간이 있고 경기장 바닥 너머로 관중석과 교회가 배치됨. 감옥 문은 닫혀 있는 것으로 보임.",
        "entities": "지시에 맞는 화염, 잔해, 교회가 있음. 바닥에 식별 가능한 인물이나 로봇이 없어 '인물 배제' 지시를 충족함.",
        "hard_violations": [],
        "physics": "강력한 폭발의 힘으로 파편들이 공중으로 튀어 오름."
       },
       {
        "label": "B",
        "direction": "화염과 파편이 우측 관중석에서 공중으로 솟구침.",
        "built_space": "전경 난간 너머로 경기장 바닥과 관중석, 교회가 보임. 벽면의 감옥 문 중 일부가 열려 있음.",
        "entities": "폭발과 잔해, 교회가 있음. 경기장 바닥에 식별 가능한 여러 명의 사람 시체들이 흩어져 있음.",
        "hard_violations": [
         "창조된 인물 (지시문에 명시되지 않은 사람의 시체들을 경기장 바닥에 여러 구 추가함)"
        ],
        "physics": "폭발 충격으로 구조물 파편들이 허공으로 날아감."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 넓은 샷의 카메라 구도와 폭발 스케일을 잘 구현했으며, 식별 가능한 인물을 배제하라는 제한을 충실히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "환경과 폭발 연출은 우수하나, 경기장 바닥에 지시문에 없는 식별 가능한 사람 시체들을 임의로 추가하여 핵심 규칙을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "폭발 화염과 파편이 우측 중단의 관중석에서 위를 향해 솟구침.",
        "built_space": "경기장 외곽의 높은 뷰. 전경에 난간이 있고 경기장 바닥 너머로 관중석과 교회가 배치됨. 감옥 문은 닫혀 있는 것으로 보임.",
        "entities": "지시에 맞는 화염, 잔해, 교회가 있음. 바닥에 식별 가능한 인물이나 로봇이 없어 '인물 배제' 지시를 충족함.",
        "hard_violations": [],
        "physics": "강력한 폭발의 힘으로 파편들이 공중으로 튀어 오름."
       },
       {
        "label": "B",
        "direction": "화염과 파편이 우측 관중석에서 공중으로 솟구침.",
        "built_space": "전경 난간 너머로 경기장 바닥과 관중석, 교회가 보임. 벽면의 감옥 문 중 일부가 열려 있음.",
        "entities": "폭발과 잔해, 교회가 있음. 경기장 바닥에 식별 가능한 여러 명의 사람 시체들이 흩어져 있음.",
        "hard_violations": [
         "창조된 인물 (지시문에 명시되지 않은 사람의 시체들을 경기장 바닥에 여러 구 추가함)"
        ],
        "physics": "폭발 충격으로 구조물 파편들이 허공으로 날아감."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "관중석 중앙 오른쪽의 폭발과 대각선 전경은 맞지만, 식별 가능한 군중과 경기장 바닥의 추가 인물들이 인물 없는 환경 와이드숏 지시를 위반하고 파편 위 여백도 부족합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "바닥에 추가 인물 없이 높은 둘레에서 관중석 폭발을 보는 구도를 더 충실히 구현했지만, 다수의 식별 가능한 관중과 상단에 잘린 파편 때문에 역시 재생성이 필요합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "폭발의 발원점은 화면 중앙 오른쪽 관중석이며, 화염과 파편이 그곳에서 위쪽과 좌우로 퍼집니다. 경기장 바닥은 아래 전경에서 맞은편 관중석으로 이어집니다. 관중 상당수의 몸과 얼굴은 경기장 쪽을 향하고, 일부는 폭발 근처에서 몸을 돌리거나 숙이고 있습니다. 조준 중인 총구나 낙하 중인 포탄 자체는 보이지 않아 입사 방향은 확인할 수 없습니다.",
        "built_space": "하나의 경기장 바닥을 연속된 계단식 관중석과 난간이 둘러싸고, 왼쪽 뒤에는 십자가가 있는 종탑 하나와 교회 건물 하나가 있습니다. 관중석 하부에는 적어도 다섯 개의 문 또는 개구부가 보이며, 둘레와 전경에 다수의 횃불·화로가 설치되어 있습니다. 전경 난간은 높은 경기장 둘레에서 내려다보는 위치를 뒷받침하고, 폭발 양옆의 온전한 좌석은 규모를 보여 줍니다. 다만 경기장 바닥 여러 곳에 사람들이 놓여 있어 지정된 비인물 환경숏과 맞지 않습니다. 최상단 파편과 연기가 화면 가장자리에 가까워 충분한 상부 여백이 없습니다.",
        "entities": "거대한 화염, 연기, 부서진 관중석과 파편, 야외 경기장 바닥, 교회, 일몰과 둘레의 불빛은 확인됩니다. 개별 몸과 옷차림을 알아볼 수 있는 관중이 다수 있고, 바닥에도 누운 사람 여러 명과 하단의 상체를 일으킨 사람이 보입니다. 이들은 사람이나 로봇이 읽히지 않아야 한다는 프레임 지시와 충돌합니다. 찰리의 고릴라형 기계 몸체나 B-200의 전개된 총은 확인되지 않지만, 원래 인물을 배제하는 구도이므로 그 부재는 감점 사유가 아닙니다. 출처가 제시되지 않은 문양 깃발도 여러 장 추가되어 있습니다.",
        "hard_violations": [
         "사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다.",
         "경기장 바닥에 여러 인간 신체와 하단의 일으켜 앉은 인물을 추가하여, 지정되지 않은 인물과 배치를 만들었습니다."
        ],
        "physics": "공중의 목재와 구조물 파편은 관중석 폭발 지점에서 솟아나므로 비행을 일으킨 힘이 보입니다. 일부 큰 골조는 무너진 관중석에 걸쳐 있고, 나머지는 폭발로 튀어 오르는 순간으로 읽힙니다. 관중은 좌석이나 계단에 지지되어 있으며, 바닥의 누운 인물들은 지면에 닿아 있습니다. 명백히 아무 지지나 발사 원인 없이 떠 있는 물체는 확인되지 않습니다."
       },
       {
        "label": "B",
        "direction": "화염과 파편은 중앙 오른쪽 중경의 관중석에서 위로 분출하며, 일부 파편은 오른쪽 바깥으로 퍼집니다. 넓은 바닥이 왼쪽 아래 전경에서 오른쪽 중경의 충돌 지점으로 이어져 대각선 관측축이 분명합니다. 관중은 대체로 경기장을 향해 앉아 있고 일부는 폭발 쪽으로 몸을 돌립니다. 총구나 비행 중인 포탄은 보이지 않아 조준 또는 낙하 궤적은 판별할 수 없습니다.",
        "built_space": "하나의 경기장과 이를 둘러싼 연속 계단식 관중석, 왼쪽 뒤의 교회 한 채와 종탑 하나가 보입니다. 관중석 하부에는 적어도 여덟 개의 문·철창 구획이 있으며, 일부는 어둡게 열린 출입구로 보입니다. 둘레의 여러 횃불과 오른쪽 전경의 화로들이 난간에 설치되어 있습니다. 가까운 난간과 아래로 펼쳐지는 바닥은 높은 둘레에서 얕게 내려다보는 카메라 위치에 부합합니다. 폭발 양쪽의 온전한 관중석도 충분하지만, 상승 파편 일부가 상단 밖으로 잘려 요구된 넉넉한 머리 위 공간은 충족하지 못합니다.",
        "entities": "관중석 충돌 폭발, 거대한 화염과 연기, 구조물 파편, 경기장 바닥, 교회와 일몰이 모두 보입니다. 바닥에는 뚜렷한 사람이나 로봇이 없지만, 관중석에는 머리·팔다리·옷차림을 구분할 수 있는 인간 관중이 다수 있습니다. 따라서 인물 없는 환경숏 조건은 지켜지지 않았습니다. 찰리와 B-200은 식별되지 않으며, 이 프레임에서 그들을 추가하지 않은 것은 적절합니다. 문양 깃발과 멀리 보이는 수변·폐허 경관은 장소 설명이 직접 지정하지 않은 확장입니다.",
        "hard_violations": [
         "사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다."
        ],
        "physics": "떠오르는 파편은 화염의 발원점과 연결된 방사형 분출을 보여 폭발이 비행의 원인으로 읽힙니다. 무너진 좌석과 골조는 관중석 잔해에 걸쳐 있고, 관중은 계단과 좌석 위에 지지되어 있습니다. 전경 화로와 난간에도 가시적인 받침이 있습니다. 명백히 원인이나 지지 없이 공중에 정지한 인체·물체는 확인되지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "관중석 중앙 오른쪽의 폭발과 대각선 전경은 맞지만, 식별 가능한 군중과 경기장 바닥의 추가 인물들이 인물 없는 환경 와이드숏 지시를 위반하고 파편 위 여백도 부족합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "바닥에 추가 인물 없이 높은 둘레에서 관중석 폭발을 보는 구도를 더 충실히 구현했지만, 다수의 식별 가능한 관중과 상단에 잘린 파편 때문에 역시 재생성이 필요합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "폭발의 발원점은 화면 중앙 오른쪽 관중석이며, 화염과 파편이 그곳에서 위쪽과 좌우로 퍼집니다. 경기장 바닥은 아래 전경에서 맞은편 관중석으로 이어집니다. 관중 상당수의 몸과 얼굴은 경기장 쪽을 향하고, 일부는 폭발 근처에서 몸을 돌리거나 숙이고 있습니다. 조준 중인 총구나 낙하 중인 포탄 자체는 보이지 않아 입사 방향은 확인할 수 없습니다.",
        "built_space": "하나의 경기장 바닥을 연속된 계단식 관중석과 난간이 둘러싸고, 왼쪽 뒤에는 십자가가 있는 종탑 하나와 교회 건물 하나가 있습니다. 관중석 하부에는 적어도 다섯 개의 문 또는 개구부가 보이며, 둘레와 전경에 다수의 횃불·화로가 설치되어 있습니다. 전경 난간은 높은 경기장 둘레에서 내려다보는 위치를 뒷받침하고, 폭발 양옆의 온전한 좌석은 규모를 보여 줍니다. 다만 경기장 바닥 여러 곳에 사람들이 놓여 있어 지정된 비인물 환경숏과 맞지 않습니다. 최상단 파편과 연기가 화면 가장자리에 가까워 충분한 상부 여백이 없습니다.",
        "entities": "거대한 화염, 연기, 부서진 관중석과 파편, 야외 경기장 바닥, 교회, 일몰과 둘레의 불빛은 확인됩니다. 개별 몸과 옷차림을 알아볼 수 있는 관중이 다수 있고, 바닥에도 누운 사람 여러 명과 하단의 상체를 일으킨 사람이 보입니다. 이들은 사람이나 로봇이 읽히지 않아야 한다는 프레임 지시와 충돌합니다. 찰리의 고릴라형 기계 몸체나 B-200의 전개된 총은 확인되지 않지만, 원래 인물을 배제하는 구도이므로 그 부재는 감점 사유가 아닙니다. 출처가 제시되지 않은 문양 깃발도 여러 장 추가되어 있습니다.",
        "hard_violations": [
         "사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다.",
         "경기장 바닥에 여러 인간 신체와 하단의 일으켜 앉은 인물을 추가하여, 지정되지 않은 인물과 배치를 만들었습니다."
        ],
        "physics": "공중의 목재와 구조물 파편은 관중석 폭발 지점에서 솟아나므로 비행을 일으킨 힘이 보입니다. 일부 큰 골조는 무너진 관중석에 걸쳐 있고, 나머지는 폭발로 튀어 오르는 순간으로 읽힙니다. 관중은 좌석이나 계단에 지지되어 있으며, 바닥의 누운 인물들은 지면에 닿아 있습니다. 명백히 아무 지지나 발사 원인 없이 떠 있는 물체는 확인되지 않습니다."
       },
       {
        "label": "A",
        "direction": "화염과 파편은 중앙 오른쪽 중경의 관중석에서 위로 분출하며, 일부 파편은 오른쪽 바깥으로 퍼집니다. 넓은 바닥이 왼쪽 아래 전경에서 오른쪽 중경의 충돌 지점으로 이어져 대각선 관측축이 분명합니다. 관중은 대체로 경기장을 향해 앉아 있고 일부는 폭발 쪽으로 몸을 돌립니다. 총구나 비행 중인 포탄은 보이지 않아 조준 또는 낙하 궤적은 판별할 수 없습니다.",
        "built_space": "하나의 경기장과 이를 둘러싼 연속 계단식 관중석, 왼쪽 뒤의 교회 한 채와 종탑 하나가 보입니다. 관중석 하부에는 적어도 여덟 개의 문·철창 구획이 있으며, 일부는 어둡게 열린 출입구로 보입니다. 둘레의 여러 횃불과 오른쪽 전경의 화로들이 난간에 설치되어 있습니다. 가까운 난간과 아래로 펼쳐지는 바닥은 높은 둘레에서 얕게 내려다보는 카메라 위치에 부합합니다. 폭발 양쪽의 온전한 관중석도 충분하지만, 상승 파편 일부가 상단 밖으로 잘려 요구된 넉넉한 머리 위 공간은 충족하지 못합니다.",
        "entities": "관중석 충돌 폭발, 거대한 화염과 연기, 구조물 파편, 경기장 바닥, 교회와 일몰이 모두 보입니다. 바닥에는 뚜렷한 사람이나 로봇이 없지만, 관중석에는 머리·팔다리·옷차림을 구분할 수 있는 인간 관중이 다수 있습니다. 따라서 인물 없는 환경숏 조건은 지켜지지 않았습니다. 찰리와 B-200은 식별되지 않으며, 이 프레임에서 그들을 추가하지 않은 것은 적절합니다. 문양 깃발과 멀리 보이는 수변·폐허 경관은 장소 설명이 직접 지정하지 않은 확장입니다.",
        "hard_violations": [
         "사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다."
        ],
        "physics": "떠오르는 파편은 화염의 발원점과 연결된 방사형 분출을 보여 폭발이 비행의 원인으로 읽힙니다. 무너진 좌석과 골조는 관중석 잔해에 걸쳐 있고, 관중은 계단과 좌석 위에 지지되어 있습니다. 전경 화로와 난간에도 가시적인 받침이 있습니다. 명백히 원인이나 지지 없이 공중에 정지한 인체·물체는 확인되지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.095
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.845
   },
   "violations": {
    "B": [
     "[gemini-pro] 창조된 인물 (지시문에 명시되지 않은 사람의 시체들을 경기장 바닥에 여러 구 추가함)",
     "[gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다.",
     "[gpt-high] 경기장 바닥에 여러 인간 신체와 하단의 일으켜 앉은 인물을 추가하여, 지정되지 않은 인물과 배치를 만들었습니다."
    ],
    "A": [
     "[gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 845
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 넓은 샷의 카메라 구도와 폭발 스케일을 잘 구현했으며, 식별 가능한 인물을 배제하라는 제한을 충실히 준수했습니다.  ★위반: [gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다."
   },
   {
    "label": "B",
    "score": 845,
    "verdict_ko": "환경과 폭발 연출은 우수하나, 경기장 바닥에 지시문에 없는 식별 가능한 사람 시체들을 임의로 추가하여 핵심 규칙을 위반했습니다.  ★위반: [gemini-pro] 창조된 인물 (지시문에 명시되지 않은 사람의 시체들을 경기장 바닥에 여러 구 추가함) / [gpt-high] 사람이 식별되지 않는 직접적인 환경 와이드숏이어야 하는데, 개별 인물로 식별 가능한 관중을 다수 추가했습니다. / [gpt-high] 경기장 바닥에 여러 인간 신체와 하단의 일으켜 앉은 인물을 추가하여, 지정되지 않은 인물과 배치를 만들었습니다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c59-06af-75a9-b8ca-9b7287f3ece0",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh95::signage": {
  "fp": "702c5b3087532e56",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::water_tower_base": {
  "input_fingerprint": "345d8b407a465b3c",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "water_tower_base",
    "tags": [
     "S60sh95"
    ]
   },
   "context_sig": "c3698d4ad5680e55"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the soaked open ground beside the collapsed village water tower, in the evening light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 물탱크 부근까지 도망친 수빈과 현우 일행!\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the soaked open ground beside the collapsed village water tower, in the evening light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 물탱크 부근까지 도망친 수빈과 현우 일행!\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_water_tower_base_726006.png",
  "asset_id": "c4fbde8b-7617-47c7-8bd4-d06a90da9340",
  "input_asset_ids": [
   "e0360747-fa50-4ce5-88f8-53279b8ecf0c"
  ],
  "origin_tag": "S60sh95",
  "place_text": "On the soaked open ground beside the collapsed village water tower, in the evening light.",
  "origin_inputs": {
   "place_text": "On the soaked open ground beside the collapsed village water tower, in the evening light.",
   "time_of_day_en": "sunset",
   "conti_asset_id": "e0360747-fa50-4ce5-88f8-53279b8ecf0c"
  }
 },
 "S60sh95::bgfirst_bg": {
  "input_fingerprint": "a7111c5fef8f8913",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장백산의 황금 가면이 반쯤 부서져 나가 피복으로 흉측하게 녹아내린 끔찍한 맨얼굴이 노출된 근접 구도.\n\nLOCATION (lock): On the soaked open ground beside the collapsed village water tower, in the evening light.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in on the exposed side of 장백산's face, looking slightly downward at his partially raised head in an oblique close-up. His disfigured face occupies the center-right, the surviving golden mask remains to the left, and his wet shoulders and rising hands border the lower frame without concealing the damage. Observe him directly as he looks toward the soldiers outside the frame, emphasizing only the final reduction in camera distance before he covers his face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 가면 (Half broken, exposing the previously concealed face) — The remaining exterior face and broken edge are visible obliquely beside the exposed skin; used as Provides an immediate visual boundary between concealed authority and exposed vulnerability.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening light preserves the wet facial texture and broken golden mask with controlled highlights, presenting the exposure without expressionistic distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장백산의 황금 가면이 반쯤 부서져 나가 피복으로 흉측하게 녹아내린 끔찍한 맨얼굴이 노출된 근접 구도.\n\nLOCATION (lock): On the soaked open ground beside the collapsed village water tower, in the evening light.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in on the exposed side of 장백산's face, looking slightly downward at his partially raised head in an oblique close-up. His disfigured face occupies the center-right, the surviving golden mask remains to the left, and his wet shoulders and rising hands border the lower frame without concealing the damage. Observe him directly as he looks toward the soldiers outside the frame, emphasizing only the final reduction in camera distance before he covers his face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 가면 (Half broken, exposing the previously concealed face) — The remaining exterior face and broken edge are visible obliquely beside the exposed skin; used as Provides an immediate visual boundary between concealed authority and exposed vulnerability.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening light preserves the wet facial texture and broken golden mask with controlled highlights, presenting the exposure without expressionistic distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh95__bgfirst_bg.png",
  "asset_id": "fee3b2c1-8e03-433f-8c93-12a0d29778da",
  "input_asset_ids": [
   "e0360747-fa50-4ce5-88f8-53279b8ecf0c",
   "c4fbde8b-7617-47c7-8bd4-d06a90da9340"
  ]
 },
 "S60sh95": {
  "input_fingerprint": "1aa1ab844d217c70",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 장백산의 황금 가면이 반쯤 부서져 나가 피복으로 흉측하게 녹아내린 끔찍한 맨얼굴이 노출된 근접 구도.\n\nLOCATION (lock): On the soaked open ground beside the collapsed village water tower, in the evening light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in on the exposed side of 장백산's face, looking slightly downward at his partially raised head in an oblique close-up. His disfigured face occupies the center-right, the surviving golden mask remains to the left, and his wet shoulders and rising hands border the lower frame without concealing the damage. Observe him directly as he looks toward the soldiers outside the frame, emphasizing only the final reduction in camera distance before he covers his face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 가면 (Half broken, exposing the previously concealed face) — The remaining exterior face and broken edge are visible obliquely beside the exposed skin; used as Provides an immediate visual boundary between concealed authority and exposed vulnerability.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening light preserves the wet facial texture and broken golden mask with controlled highlights, presenting the exposure without expressionistic distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water-tank tower has been shot through and toppled, releasing a great volume of water across the area in the evening light. Charlie is battered, and B-200 still has its hand-mounted guns deployed. 장백산: He is soaked, and his golden mask is half broken. The exposed portion of his face is severely disfigured by radiation; his regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 장백산의 황금 가면이 반쯤 부서져 나가 피복으로 흉측하게 녹아내린 끔찍한 맨얼굴이 노출된 근접 구도.\n\nLOCATION (lock): On the soaked open ground beside the collapsed village water tower, in the evening light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in on the exposed side of 장백산's face, looking slightly downward at his partially raised head in an oblique close-up. His disfigured face occupies the center-right, the surviving golden mask remains to the left, and his wet shoulders and rising hands border the lower frame without concealing the damage. Observe him directly as he looks toward the soldiers outside the frame, emphasizing only the final reduction in camera distance before he covers his face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 가면 (Half broken, exposing the previously concealed face) — The remaining exterior face and broken edge are visible obliquely beside the exposed skin; used as Provides an immediate visual boundary between concealed authority and exposed vulnerability.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening light preserves the wet facial texture and broken golden mask with controlled highlights, presenting the exposure without expressionistic distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water-tank tower has been shot through and toppled, releasing a great volume of water across the area in the evening light. Charlie is battered, and B-200 still has its hand-mounted guns deployed. 장백산: He is soaked, and his golden mask is half broken. The exposed portion of his face is severely disfigured by radiation; his regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 장백산의 황금 가면이 반쯤 부서져 나가 피복으로 흉측하게 녹아내린 끔찍한 맨얼굴이 노출된 근접 구도.\n\nLOCATION (lock): On the soaked open ground beside the collapsed village water tower, in the evening light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in on the exposed side of 장백산's face, looking slightly downward at his partially raised head in an oblique close-up. His disfigured face occupies the center-right, the surviving golden mask remains to the left, and his wet shoulders and rising hands border the lower frame without concealing the damage. Observe him directly as he looks toward the soldiers outside the frame, emphasizing only the final reduction in camera distance before he covers his face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 황금 가면 (Half broken, exposing the previously concealed face) — The remaining exterior face and broken edge are visible obliquely beside the exposed skin; used as Provides an immediate visual boundary between concealed authority and exposed vulnerability.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening light preserves the wet facial texture and broken golden mask with controlled highlights, presenting the exposure without expressionistic distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The water-tank tower has been shot through and toppled, releasing a great volume of water across the area in the evening light. Charlie is battered, and B-200 still has its hand-mounted guns deployed. 장백산: He is soaked, and his golden mask is half broken. The exposed portion of his face is severely disfigured by radiation; his regal clothing has not yet been stripped away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 장백산 (한국인 남성, 성숙한 얼굴, 짙은색 머리카락) — wearing: 화려하고 육중한 왕과 같은 붉은색 톤의 예복, 얼굴에는 황금빛으로 칠해진 하회탈 가면. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh95__bgfirst_bg.png",
     "asset_id": "fee3b2c1-8e03-433f-8c93-12a0d29778da",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S60sh95.png",
     "asset_id": "e0360747-fa50-4ce5-88f8-53279b8ecf0c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:750246>",
     "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_water_tower_base_726006.png",
     "asset_id": "c4fbde8b-7617-47c7-8bd4-d06a90da9340",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:750246>",
     "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 우측 밖을 향하며, 두 손은 위로 올려져 있다. 배경의 군인들은 앞을 향해 총구를 겨눈다.",
    "built_space": "참조 이미지와 동일한 구조. 좌측 급수탑에서 물이 쏟아지고 우측에 폐허가 된 교회가 있다.",
    "entities": "장백산의 훼손된 얼굴과 젖은 예복은 일치하나, 하회탈이 아닌 일반 금속 가면을 썼다. 지시되지 않은 군인들이 배경에 존재한다.",
    "hard_violations": [
     "[gemini-pro] 발명된 인물 (등장하지 않아야 할 군인들이 배경에 포함됨)",
     "[gpt-high] 장백산만 보여야 하는 장면에 무장 인물 네 명을 추가했고, 명시적으로 화면 밖에 있어야 할 병사들을 화면 안에 배치했다."
    ],
    "physics": "올려진 손과 팔의 자세, 급수탑에서 쏟아지는 물의 흐름이 자연스럽게 지탱된다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 화면 우측을 향한다.",
    "built_space": "참조된 폐허 배경과 쓰러진 급수탑, 일몰이 올바르게 배치되어 있다.",
    "entities": "장백산이 예복을 입고 있으나, 온전한 형태의 하회탈이 얼굴 정면이 아닌 우측 귀 부근에 기괴하게 붙어 있다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 및 연출 (가면이 지지대 없이 얼굴 측면에 비정상적으로 융합됨)"
    ],
    "physics": "가면이 어떤 끈이나 손의 지지 없이 얼굴 측면에 불가능하게 붙어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 근접 구도와 인물의 배치는 정확히 구현했으나, 화면에 없어야 할 군인들을 배경에 추가하는 치명적인 오류를 범했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "가면이 얼굴 측면에 불가능한 형태로 붙어 있으며, 근접 구도라는 명확한 프레이밍 지시를 완전히 무시했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 우측 밖을 향하며, 두 손은 위로 올려져 있다. 배경의 군인들은 앞을 향해 총구를 겨눈다.",
        "built_space": "참조 이미지와 동일한 구조. 좌측 급수탑에서 물이 쏟아지고 우측에 폐허가 된 교회가 있다.",
        "entities": "장백산의 훼손된 얼굴과 젖은 예복은 일치하나, 하회탈이 아닌 일반 금속 가면을 썼다. 지시되지 않은 군인들이 배경에 존재한다.",
        "hard_violations": [
         "발명된 인물 (등장하지 않아야 할 군인들이 배경에 포함됨)"
        ],
        "physics": "올려진 손과 팔의 자세, 급수탑에서 쏟아지는 물의 흐름이 자연스럽게 지탱된다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측을 향한다.",
        "built_space": "참조된 폐허 배경과 쓰러진 급수탑, 일몰이 올바르게 배치되어 있다.",
        "entities": "장백산이 예복을 입고 있으나, 온전한 형태의 하회탈이 얼굴 정면이 아닌 우측 귀 부근에 기괴하게 붙어 있다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 및 연출 (가면이 지지대 없이 얼굴 측면에 비정상적으로 융합됨)"
        ],
        "physics": "가면이 어떤 끈이나 손의 지지 없이 얼굴 측면에 불가능하게 붙어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "지시된 근접 구도와 인물의 배치는 정확히 구현했으나, 화면에 없어야 할 군인들을 배경에 추가하는 치명적인 오류를 범했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "가면이 얼굴 측면에 불가능한 형태로 붙어 있으며, 근접 구도라는 명확한 프레이밍 지시를 완전히 무시했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 우측 밖을 향하며, 두 손은 위로 올려져 있다. 배경의 군인들은 앞을 향해 총구를 겨눈다.",
        "built_space": "참조 이미지와 동일한 구조. 좌측 급수탑에서 물이 쏟아지고 우측에 폐허가 된 교회가 있다.",
        "entities": "장백산의 훼손된 얼굴과 젖은 예복은 일치하나, 하회탈이 아닌 일반 금속 가면을 썼다. 지시되지 않은 군인들이 배경에 존재한다.",
        "hard_violations": [
         "발명된 인물 (등장하지 않아야 할 군인들이 배경에 포함됨)"
        ],
        "physics": "올려진 손과 팔의 자세, 급수탑에서 쏟아지는 물의 흐름이 자연스럽게 지탱된다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 우측을 향한다.",
        "built_space": "참조된 폐허 배경과 쓰러진 급수탑, 일몰이 올바르게 배치되어 있다.",
        "entities": "장백산이 예복을 입고 있으나, 온전한 형태의 하회탈이 얼굴 정면이 아닌 우측 귀 부근에 기괴하게 붙어 있다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 및 연출 (가면이 지지대 없이 얼굴 측면에 비정상적으로 융합됨)"
        ],
        "physics": "가면이 어떤 끈이나 손의 지지 없이 얼굴 측면에 불가능하게 붙어 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "추가 인물 없이 장소·예복·하회탈 형태는 살렸지만, 허리까지 보이는 넓은 구도로 얼굴 클로즈업 지시를 크게 벗어나며 가면도 절반보다 많이 남아 있다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴 크기와 올라오는 양손은 지시에 가깝지만, 화면 밖에 있어야 할 무장 인물 네 명을 추가해 실격이며 가면도 하회탈이 아닌 각진 장갑 형태다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장백산은 고개와 눈을 화면 오른쪽 바깥으로 돌리고 있어 화면 밖 병사들을 본다는 지시와 양립한다. 양손은 얼굴 아래 허리 높이에 머물며, 얼굴을 가리기 직전까지 올라온 동작은 약하다. 보이는 무기는 없다.",
        "built_space": "왼쪽 뒤에 파손된 원통형 물탱크 한 개와 쓰러진 철제 지지대가 있고 물이 쏟아진다. 오른쪽 뒤에는 종탑 한 개, 중앙에는 폐허와 산이 보인다. 젖은 공터와 저녁 하늘은 장소 참조에 부합한다. 다만 얼굴뿐 아니라 몸통과 넓은 배경까지 담아 요구된 근접 구도를 놓쳤다.",
        "entities": "짙은 젖은 머리의 성숙한 동아시아계 남성 한 명만 보인다. 붉은색 바탕의 무거운 금색 자수 예복은 참조와 가깝고, 노출된 얼굴에는 심한 조직 손상이 있다. 화면 왼쪽의 금색 가면은 웃는 하회탈 형태지만 양쪽 눈구멍과 코·입 대부분이 남아 있어 반파 상태가 약하다. 화면 아래 오른쪽에도 금색 가면 파편으로 보이는 물체가 있다. 얼굴 손상 때문에 참조 인물과의 정확한 동일성은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "머리와 어깨, 양손은 몸에 자연스럽게 연결되어 있으며 공중에 뜬 신체는 없다. 가면은 얼굴 옆에 걸쳐 있으나 고정 방식은 가려져 있다. 아래쪽 금색 조각은 무릎 또는 의복 위에 놓인 것으로 보인다. 머리 위 작은 파편들은 가면이 막 깨진 순간의 비산으로 해석할 수 있다. 물탱크는 무너진 철골에 받쳐져 있고 물은 아래로 떨어진다."
       },
       {
        "label": "B",
        "direction": "장백산의 눈은 화면 왼쪽 바깥을 향하고 양손은 얼굴을 향해 올라오되 상처를 가리지 않는다. 오른쪽 배경의 무장 인물 네 명은 대체로 화면 왼쪽 전경을 향해 총구를 내민다. 장백산은 이 보이는 인물들을 바라보지 않으며, 병사들을 화면 밖에 둔다는 지시도 지켜지지 않았다.",
        "built_space": "왼쪽 뒤에 파손된 물탱크 한 개와 붕괴한 철골, 아래에는 물이 고인 공터가 있다. 오른쪽 끝 종탑과 뒤쪽 폐허·산은 장소 참조의 주요 요소를 따른다. 얼굴이 중앙 오른쪽, 가면이 왼쪽, 어깨와 양손이 하단을 차지하는 클로즈업은 요구에 가깝다. 그러나 오른쪽 공터에 불필요한 인물 네 명을 배치했다.",
        "entities": "전경에는 짙은 머리의 성숙한 동아시아계 남성이 있고 붉은색·금색 자수 예복, 젖은 피부, 심한 얼굴 흉터가 보인다. 눈은 정상적인 홍채와 동공을 유지한다. 금색 반쪽 가면과 파손 경계는 분명하지만 웃는 하회탈이 아니라 각진 금속 전투 가면처럼 생겼다. 배경에는 소총을 든 인물 세 명과 덩치 큰 장갑 인물 한 명이 추가되어 있다.",
        "hard_violations": [
         "장백산만 보여야 하는 장면에 무장 인물 네 명을 추가했고, 명시적으로 화면 밖에 있어야 할 병사들을 화면 안에 배치했다."
        ],
        "physics": "양손은 하단의 팔에 연결된 채 얼굴 쪽으로 올라와 있으며 손가락과 손목의 자세는 가능한 동작이다. 가면은 얼굴에 밀착되어 있고 떠 있는 물체로 보이지 않는다. 배경 인물들은 발로 지면을 딛고 무기를 손으로 지탱한다. 물탱크는 붕괴한 철골에 걸려 있고 흘러나오는 물은 지면으로 떨어진다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "추가 인물 없이 장소·예복·하회탈 형태는 살렸지만, 허리까지 보이는 넓은 구도로 얼굴 클로즈업 지시를 크게 벗어나며 가면도 절반보다 많이 남아 있다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴 크기와 올라오는 양손은 지시에 가깝지만, 화면 밖에 있어야 할 무장 인물 네 명을 추가해 실격이며 가면도 하회탈이 아닌 각진 장갑 형태다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "장백산은 고개와 눈을 화면 오른쪽 바깥으로 돌리고 있어 화면 밖 병사들을 본다는 지시와 양립한다. 양손은 얼굴 아래 허리 높이에 머물며, 얼굴을 가리기 직전까지 올라온 동작은 약하다. 보이는 무기는 없다.",
        "built_space": "왼쪽 뒤에 파손된 원통형 물탱크 한 개와 쓰러진 철제 지지대가 있고 물이 쏟아진다. 오른쪽 뒤에는 종탑 한 개, 중앙에는 폐허와 산이 보인다. 젖은 공터와 저녁 하늘은 장소 참조에 부합한다. 다만 얼굴뿐 아니라 몸통과 넓은 배경까지 담아 요구된 근접 구도를 놓쳤다.",
        "entities": "짙은 젖은 머리의 성숙한 동아시아계 남성 한 명만 보인다. 붉은색 바탕의 무거운 금색 자수 예복은 참조와 가깝고, 노출된 얼굴에는 심한 조직 손상이 있다. 화면 왼쪽의 금색 가면은 웃는 하회탈 형태지만 양쪽 눈구멍과 코·입 대부분이 남아 있어 반파 상태가 약하다. 화면 아래 오른쪽에도 금색 가면 파편으로 보이는 물체가 있다. 얼굴 손상 때문에 참조 인물과의 정확한 동일성은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "머리와 어깨, 양손은 몸에 자연스럽게 연결되어 있으며 공중에 뜬 신체는 없다. 가면은 얼굴 옆에 걸쳐 있으나 고정 방식은 가려져 있다. 아래쪽 금색 조각은 무릎 또는 의복 위에 놓인 것으로 보인다. 머리 위 작은 파편들은 가면이 막 깨진 순간의 비산으로 해석할 수 있다. 물탱크는 무너진 철골에 받쳐져 있고 물은 아래로 떨어진다."
       },
       {
        "label": "A",
        "direction": "장백산의 눈은 화면 왼쪽 바깥을 향하고 양손은 얼굴을 향해 올라오되 상처를 가리지 않는다. 오른쪽 배경의 무장 인물 네 명은 대체로 화면 왼쪽 전경을 향해 총구를 내민다. 장백산은 이 보이는 인물들을 바라보지 않으며, 병사들을 화면 밖에 둔다는 지시도 지켜지지 않았다.",
        "built_space": "왼쪽 뒤에 파손된 물탱크 한 개와 붕괴한 철골, 아래에는 물이 고인 공터가 있다. 오른쪽 끝 종탑과 뒤쪽 폐허·산은 장소 참조의 주요 요소를 따른다. 얼굴이 중앙 오른쪽, 가면이 왼쪽, 어깨와 양손이 하단을 차지하는 클로즈업은 요구에 가깝다. 그러나 오른쪽 공터에 불필요한 인물 네 명을 배치했다.",
        "entities": "전경에는 짙은 머리의 성숙한 동아시아계 남성이 있고 붉은색·금색 자수 예복, 젖은 피부, 심한 얼굴 흉터가 보인다. 눈은 정상적인 홍채와 동공을 유지한다. 금색 반쪽 가면과 파손 경계는 분명하지만 웃는 하회탈이 아니라 각진 금속 전투 가면처럼 생겼다. 배경에는 소총을 든 인물 세 명과 덩치 큰 장갑 인물 한 명이 추가되어 있다.",
        "hard_violations": [
         "장백산만 보여야 하는 장면에 무장 인물 네 명을 추가했고, 명시적으로 화면 밖에 있어야 할 병사들을 화면 안에 배치했다."
        ],
        "physics": "양손은 하단의 팔에 연결된 채 얼굴 쪽으로 올라와 있으며 손가락과 손목의 자세는 가능한 동작이다. 가면은 얼굴에 밀착되어 있고 떠 있는 물체로 보이지 않는다. 배경 인물들은 발로 지면을 딛고 무기를 손으로 지탱한다. 물탱크는 붕괴한 철골에 걸려 있고 흘러나오는 물은 지면으로 떨어진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.35
   },
   "violations": {
    "A": [
     "[gemini-pro] 발명된 인물 (등장하지 않아야 할 군인들이 배경에 포함됨)",
     "[gpt-high] 장백산만 보여야 하는 장면에 무장 인물 네 명을 추가했고, 명시적으로 화면 밖에 있어야 할 병사들을 화면 안에 배치했다."
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 및 연출 (가면이 지지대 없이 얼굴 측면에 비정상적으로 융합됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1250,
   "B": 1350
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "지시된 근접 구도와 인물의 배치는 정확히 구현했으나, 화면에 없어야 할 군인들을 배경에 추가하는 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 발명된 인물 (등장하지 않아야 할 군인들이 배경에 포함됨) / [gpt-high] 장백산만 보여야 하는 장면에 무장 인물 네 명을 추가했고, 명시적으로 화면 밖에 있어야 할 병사들을 화면 안에 배치했다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "가면이 얼굴 측면에 불가능한 형태로 붙어 있으며, 근접 구도라는 명확한 프레이밍 지시를 완전히 무시했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 및 연출 (가면이 지지대 없이 얼굴 측면에 비정상적으로 융합됨)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_water_tower_base_726006.png",
    "asset_id": "c4fbde8b-7617-47c7-8bd4-d06a90da9340",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 장백산: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:750246>",
    "asset_id": "b05c1cb4-89d3-4fd5-b0d9-6b59f61e6c24",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c5f-03d2-7c06-bf1d-213e07f1ca5d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh95__bgfirst_bg.png",
   "bg_asset_id": "fee3b2c1-8e03-433f-8c93-12a0d29778da",
   "bg_record_key": "S60sh95::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "water_tower_base",
   "groupbg_asset_id": "c4fbde8b-7617-47c7-8bd4-d06a90da9340"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh115::signage": {
  "fp": "be171f6a17f2be01",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::village_ridge": {
  "input_fingerprint": "b2e1ac9155a2f4e0",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "village_ridge",
    "tags": [
     "S60sh115"
    ]
   },
   "context_sig": "b0208948dad4fb75"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On an exposed ridge overlooking the damaged village, beneath the red sunset.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 서서히 카메라 내려오면 능선에 앉아있는 두 사람의 뒷모습\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On an exposed ridge overlooking the damaged village, beneath the red sunset.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 서서히 카메라 내려오면 능선에 앉아있는 두 사람의 뒷모습\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_ridge_5f6fbe.png",
  "asset_id": "8b8172cc-c7fe-4f3d-b4bf-57184ca13cb6",
  "input_asset_ids": [
   "7a63e964-607f-496c-9a3d-466b4b1fb551"
  ],
  "origin_tag": "S60sh115",
  "place_text": "On an exposed ridge overlooking the damaged village, beneath the red sunset.",
  "origin_inputs": {
   "place_text": "On an exposed ridge overlooking the damaged village, beneath the red sunset.",
   "time_of_day_en": "sunset",
   "conti_asset_id": "7a63e964-607f-496c-9a3d-466b4b1fb551"
  }
 },
 "S60sh115::bgfirst_bg": {
  "input_fingerprint": "747a2f6764207871",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉은 노을을 배경으로 손을 맞잡은 채 나란히 앉아 먼 풍경을 응시하는 이현우와 찰리의 따뜻한 전경.\n\nLOCATION (lock): On an exposed ridge overlooking the damaged village, beneath the red sunset.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal behind and to the side of 이현우 and 찰리, holding a wide rear-oblique view with a slight downward tilt toward their joined hands. Their complete seated figures occupy the lower central third, 이현우 on the left and 찰리 on the right, with their clasp visible in the gap between their bodies and the distant landscape extending above them. Both remain absorbed in that distant landscape rather than looking back, and the retreat settles into stillness without rearranging their positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (The two friends remain seated together on it) — Seen from behind their seated position, extending beneath both figures; used as Grounds the complete figures and their unobstructed handclasp; 먼 풍경과 노을 하늘 (Visible beyond the ridge as the sun sets) — Extends ahead of the pair in their shared viewing direction; used as Leaves generous visual space above the figures for their shared contemplation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued red sunset light supplies gentle warmth while controlled contrast retains detail in the injured pair and 찰리's dented, perforated surfaces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉은 노을을 배경으로 손을 맞잡은 채 나란히 앉아 먼 풍경을 응시하는 이현우와 찰리의 따뜻한 전경.\n\nLOCATION (lock): On an exposed ridge overlooking the damaged village, beneath the red sunset.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal behind and to the side of 이현우 and 찰리, holding a wide rear-oblique view with a slight downward tilt toward their joined hands. Their complete seated figures occupy the lower central third, 이현우 on the left and 찰리 on the right, with their clasp visible in the gap between their bodies and the distant landscape extending above them. Both remain absorbed in that distant landscape rather than looking back, and the retreat settles into stillness without rearranging their positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (The two friends remain seated together on it) — Seen from behind their seated position, extending beneath both figures; used as Grounds the complete figures and their unobstructed handclasp; 먼 풍경과 노을 하늘 (Visible beyond the ridge as the sun sets) — Extends ahead of the pair in their shared viewing direction; used as Leaves generous visual space above the figures for their shared contemplation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued red sunset light supplies gentle warmth while controlled contrast retains detail in the injured pair and 찰리's dented, perforated surfaces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh115__bgfirst_bg.png",
  "asset_id": "167c6b68-db86-4ec7-aec1-c13327c38905",
  "input_asset_ids": [
   "7a63e964-607f-496c-9a3d-466b4b1fb551",
   "8b8172cc-c7fe-4f3d-b4bf-57184ca13cb6"
  ]
 },
 "S60sh115": {
  "input_fingerprint": "a48d657040b787a5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 붉은 노을을 배경으로 손을 맞잡은 채 나란히 앉아 먼 풍경을 응시하는 이현우와 찰리의 따뜻한 전경.\n\nLOCATION (lock): On an exposed ridge overlooking the damaged village, beneath the red sunset. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal behind and to the side of 이현우 and 찰리, holding a wide rear-oblique view with a slight downward tilt toward their joined hands. Their complete seated figures occupy the lower central third, 이현우 on the left and 찰리 on the right, with their clasp visible in the gap between their bodies and the distant landscape extending above them. Both remain absorbed in that distant landscape rather than looking back, and the retreat settles into stillness without rearranging their positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (The two friends remain seated together on it) — Seen from behind their seated position, extending beneath both figures; used as Grounds the complete figures and their unobstructed handclasp; 먼 풍경과 노을 하늘 (Visible beyond the ridge as the sun sets) — Extends ahead of the pair in their shared viewing direction; used as Leaves generous visual space above the figures for their shared contemplation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued red sunset light supplies gentle warmth while controlled contrast retains detail in the injured pair and 찰리's dented, perforated surfaces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light falls across the devastated village and ridge. Charlie's body is dented and punctured, with malfunctioning sensors; the earlier raincoat, oversized hat and boots have no scripted removal. 이현우: He sits on the ridge, battered from the fighting but with a brightened expression, one hand extended in a clasp. His earlier treated leg injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 붉은 노을을 배경으로 손을 맞잡은 채 나란히 앉아 먼 풍경을 응시하는 이현우와 찰리의 따뜻한 전경.\n\nLOCATION (lock): On an exposed ridge overlooking the damaged village, beneath the red sunset. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal behind and to the side of 이현우 and 찰리, holding a wide rear-oblique view with a slight downward tilt toward their joined hands. Their complete seated figures occupy the lower central third, 이현우 on the left and 찰리 on the right, with their clasp visible in the gap between their bodies and the distant landscape extending above them. Both remain absorbed in that distant landscape rather than looking back, and the retreat settles into stillness without rearranging their positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (The two friends remain seated together on it) — Seen from behind their seated position, extending beneath both figures; used as Grounds the complete figures and their unobstructed handclasp; 먼 풍경과 노을 하늘 (Visible beyond the ridge as the sun sets) — Extends ahead of the pair in their shared viewing direction; used as Leaves generous visual space above the figures for their shared contemplation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued red sunset light supplies gentle warmth while controlled contrast retains detail in the injured pair and 찰리's dented, perforated surfaces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light falls across the devastated village and ridge. Charlie's body is dented and punctured, with malfunctioning sensors; the earlier raincoat, oversized hat and boots have no scripted removal. 이현우: He sits on the ridge, battered from the fighting but with a brightened expression, one hand extended in a clasp. His earlier treated leg injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 붉은 노을을 배경으로 손을 맞잡은 채 나란히 앉아 먼 풍경을 응시하는 이현우와 찰리의 따뜻한 전경.\n\nLOCATION (lock): On an exposed ridge overlooking the damaged village, beneath the red sunset. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the withdrawal behind and to the side of 이현우 and 찰리, holding a wide rear-oblique view with a slight downward tilt toward their joined hands. Their complete seated figures occupy the lower central third, 이현우 on the left and 찰리 on the right, with their clasp visible in the gap between their bodies and the distant landscape extending above them. Both remain absorbed in that distant landscape rather than looking back, and the retreat settles into stillness without rearranging their positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (The two friends remain seated together on it) — Seen from behind their seated position, extending beneath both figures; used as Grounds the complete figures and their unobstructed handclasp; 먼 풍경과 노을 하늘 (Visible beyond the ridge as the sun sets) — Extends ahead of the pair in their shared viewing direction; used as Leaves generous visual space above the figures for their shared contemplation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued red sunset light supplies gentle warmth while controlled contrast retains detail in the injured pair and 찰리's dented, perforated surfaces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light falls across the devastated village and ridge. Charlie's body is dented and punctured, with malfunctioning sensors; the earlier raincoat, oversized hat and boots have no scripted removal. 이현우: He sits on the ridge, battered from the fighting but with a brightened expression, one hand extended in a clasp. His earlier treated leg injury remains.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh115__bgfirst_bg.png",
     "asset_id": "167c6b68-db86-4ec7-aec1-c13327c38905",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S60sh115.png",
     "asset_id": "7a63e964-607f-496c-9a3d-466b4b1fb551",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_ridge_5f6fbe.png",
     "asset_id": "8b8172cc-c7fe-4f3d-b4bf-57184ca13cb6",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 화면 중앙 너머의 노을과 마을 풍경을 바라보고 있습니다.",
    "built_space": "지정된 폐허 마을과 일치하는 배경 위 능선에 두 인물이 위치해 있습니다.",
    "entities": "이현우와 찰리의 특징(복장, 기계 장갑 색상)이 잘 반영되었으나, 찰리의 팔 굵기가 고릴라형이라는 묘사에 비해 약간 얇게 표현되었습니다.",
    "hard_violations": [],
    "physics": "인물들은 바닥에 닿아 앉아 있고, 잡고 있는 손의 위치와 무게감도 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 정면의 먼 폐허 마을과 붉은 노을을 향해 시선을 고정하고 있습니다.",
    "built_space": "레퍼런스와 동일한 수조탑과 교회가 있는 폐허 마을을 내려다보는 능선 위에 두 인물이 나란히 앉아 있습니다.",
    "entities": "이현우(어두운 셔츠를 입은 흑발 남성의 뒷모습)와 찰리(샌드 베이지색 장갑을 두른 고릴라 형태의 기계 뒷모습)가 정확히 배치되었습니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 능선의 바위 바닥에 안정적으로 체중을 싣고 앉아 있으며, 맞잡은 손 역시 중력에 맞게 자연스럽게 아래로 향해 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요청된 후면 구도와 맞잡은 손, 캐릭터의 특징(특히 찰리의 육중한 기계 체형)과 지정된 배경을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 배경과 인물의 배치, 맞잡은 손 등 전반적인 지시사항을 잘 따랐으나 찰리의 체형이 B에 비해 다소 덜 육중해 보입니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "두 인물 모두 정면의 먼 폐허 마을과 붉은 노을을 향해 시선을 고정하고 있습니다.",
        "built_space": "레퍼런스와 동일한 수조탑과 교회가 있는 폐허 마을을 내려다보는 능선 위에 두 인물이 나란히 앉아 있습니다.",
        "entities": "이현우(어두운 셔츠를 입은 흑발 남성의 뒷모습)와 찰리(샌드 베이지색 장갑을 두른 고릴라 형태의 기계 뒷모습)가 정확히 배치되었습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 능선의 바위 바닥에 안정적으로 체중을 싣고 앉아 있으며, 맞잡은 손 역시 중력에 맞게 자연스럽게 아래로 향해 있습니다."
       },
       {
        "label": "A",
        "direction": "두 인물 모두 화면 중앙 너머의 노을과 마을 풍경을 바라보고 있습니다.",
        "built_space": "지정된 폐허 마을과 일치하는 배경 위 능선에 두 인물이 위치해 있습니다.",
        "entities": "이현우와 찰리의 특징(복장, 기계 장갑 색상)이 잘 반영되었으나, 찰리의 팔 굵기가 고릴라형이라는 묘사에 비해 약간 얇게 표현되었습니다.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 닿아 앉아 있고, 잡고 있는 손의 위치와 무게감도 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요청된 후면 구도와 맞잡은 손, 캐릭터의 특징(특히 찰리의 육중한 기계 체형)과 지정된 배경을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지정된 배경과 인물의 배치, 맞잡은 손 등 전반적인 지시사항을 잘 따랐으나 찰리의 체형이 B에 비해 다소 덜 육중해 보입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 인물 모두 정면의 먼 폐허 마을과 붉은 노을을 향해 시선을 고정하고 있습니다.",
        "built_space": "레퍼런스와 동일한 수조탑과 교회가 있는 폐허 마을을 내려다보는 능선 위에 두 인물이 나란히 앉아 있습니다.",
        "entities": "이현우(어두운 셔츠를 입은 흑발 남성의 뒷모습)와 찰리(샌드 베이지색 장갑을 두른 고릴라 형태의 기계 뒷모습)가 정확히 배치되었습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 능선의 바위 바닥에 안정적으로 체중을 싣고 앉아 있으며, 맞잡은 손 역시 중력에 맞게 자연스럽게 아래로 향해 있습니다."
       },
       {
        "label": "A",
        "direction": "두 인물 모두 화면 중앙 너머의 노을과 마을 풍경을 바라보고 있습니다.",
        "built_space": "지정된 폐허 마을과 일치하는 배경 위 능선에 두 인물이 위치해 있습니다.",
        "entities": "이현우와 찰리의 특징(복장, 기계 장갑 색상)이 잘 반영되었으나, 찰리의 팔 굵기가 고릴라형이라는 묘사에 비해 약간 얇게 표현되었습니다.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 닿아 앉아 있고, 잡고 있는 손의 위치와 무게감도 자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "좌우 배치와 먼 풍경을 향한 손잡기는 맞지만, 거의 정후면인 구도와 큰 인물 비중이 지정된 후측면 와이드숏에 덜 맞고 찰리의 우비·모자 유지 조건도 빠졌다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "후측면 각도와 두 몸 사이의 분명한 손맞잡기가 A보다 지시에 가깝지만, 인물이 하단 3분의 1보다 다소 크고 찰리의 우비·모자 유지 조건은 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 이현우와 오른쪽 찰리 모두 카메라에 등을 보이고 폐허 마을과 노을 진 산줄기를 향한다. 돌아보는 인물은 없다. 이현우의 오른팔과 찰리의 왼팔이 두 몸 사이로 내려와 손을 맞잡는다.",
        "built_space": "두 인물은 흙과 암석으로 된 같은 능선에 나란히 앉아 있다. 중앙의 고가 물탱크 1기, 오른쪽의 첨탑 달린 교회 1채, 중앙 오른쪽의 원형 골조 1개, 오른쪽 아래 아치 구조물과 기와지붕 군집이 장소 참조와 대응한다. 배경 시설이 전경 크기로 과장되지 않았다. 다만 카메라는 후측면보다는 정후면에 가깝고, 앉은 인물들이 하단 중앙 3분의 1보다 위로 크게 올라온다.",
        "entities": "등장 개체는 젊은 남성 이현우와 기계 찰리뿐이다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 피와 흙이 묻은 어두운 셔츠·바지는 참조와 부합한다. 후면이라 정확한 얼굴, 한국계 외모, 밝아진 표정과 인이어 무전기는 확인하기 어렵다. 찰리는 베이지 장갑판, 육중한 팔, 짧게 접힌 다리를 가진 기계이며 표면 마모가 보인다. 흰 얼굴과 가슴 원자로는 정상적으로 가려져 있다. 계속 착용해야 할 우비와 큰 모자는 없으며, 부츠·다리 치료 흔적·센서 고장은 이 시점에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 몸의 골반과 접힌 다리가 능선 지면에 닿아 체중을 지지한다. 찰리의 바깥쪽 팔도 지면에 내려놓은 형태다. 맞잡은 손은 각자의 팔과 연결되어 있고 지면 가까이에 놓여 있다. 떠 있는 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우는 마을 너머 산줄기와 노을을 바라보고, 찰리도 머리를 약간 왼쪽으로 둔 채 같은 전방 풍경을 향한다. 둘 다 카메라를 돌아보지 않는다. 이현우의 오른손과 찰리의 왼손이 두 몸 사이의 열린 틈에서 분명하게 맞잡혀 있다.",
        "built_space": "두 인물은 연속된 바위·흙 능선 위에 왼쪽 이현우, 오른쪽 찰리 순서로 앉아 있다. 중앙 물탱크 1기, 오른쪽 교회와 첨탑 1개, 중앙 오른쪽 원형 골조 1개, 오른쪽 아래 아치 구조물 및 낮은 기와지붕들이 참조 장소의 배치를 유지한다. 후면에서 약간 옆으로 비껴 내려다보는 시점이며 손맞잡기와 지면 접촉이 잘 드러난다. 인물은 A보다 조금 낮고 작지만 여전히 하단 3분의 1을 다소 넘어선다.",
        "entities": "이현우 1명과 찰리 1대만 보인다. 이현우의 젊은 남성 체형, 헝클어진 검은 머리, 피와 먼지가 묻은 낡은 어두운 옷이 참조에 부합한다. 얼굴과 귀가 충분히 드러나지 않아 정확한 얼굴 일치, 표정과 무전기는 판정할 수 없다. 찰리는 긴 기계 팔, 짧은 다리, 베이지 장갑판과 긁힘·찍힘이 보이는 비인간 기계다. 얼굴과 가슴 원자로를 보여주려고 몸을 돌리지 않은 점은 지시에 맞는다. 우비와 큰 모자는 누락되었고, 부츠·치료된 다리 부상·센서 고장은 가시적으로 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 인물 모두 골반과 접힌 다리를 지면에 두고 안정적으로 앉아 있다. 찰리의 오른손은 몸 옆 땅을 짚고, 왼손은 이현우의 오른손을 잡는다. 손목과 팔의 연결 및 손맞잡기의 높이가 자연스럽고, 공중에 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "좌우 배치와 먼 풍경을 향한 손잡기는 맞지만, 거의 정후면인 구도와 큰 인물 비중이 지정된 후측면 와이드숏에 덜 맞고 찰리의 우비·모자 유지 조건도 빠졌다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "후측면 각도와 두 몸 사이의 분명한 손맞잡기가 A보다 지시에 가깝지만, 인물이 하단 3분의 1보다 다소 크고 찰리의 우비·모자 유지 조건은 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 이현우와 오른쪽 찰리 모두 카메라에 등을 보이고 폐허 마을과 노을 진 산줄기를 향한다. 돌아보는 인물은 없다. 이현우의 오른팔과 찰리의 왼팔이 두 몸 사이로 내려와 손을 맞잡는다.",
        "built_space": "두 인물은 흙과 암석으로 된 같은 능선에 나란히 앉아 있다. 중앙의 고가 물탱크 1기, 오른쪽의 첨탑 달린 교회 1채, 중앙 오른쪽의 원형 골조 1개, 오른쪽 아래 아치 구조물과 기와지붕 군집이 장소 참조와 대응한다. 배경 시설이 전경 크기로 과장되지 않았다. 다만 카메라는 후측면보다는 정후면에 가깝고, 앉은 인물들이 하단 중앙 3분의 1보다 위로 크게 올라온다.",
        "entities": "등장 개체는 젊은 남성 이현우와 기계 찰리뿐이다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 피와 흙이 묻은 어두운 셔츠·바지는 참조와 부합한다. 후면이라 정확한 얼굴, 한국계 외모, 밝아진 표정과 인이어 무전기는 확인하기 어렵다. 찰리는 베이지 장갑판, 육중한 팔, 짧게 접힌 다리를 가진 기계이며 표면 마모가 보인다. 흰 얼굴과 가슴 원자로는 정상적으로 가려져 있다. 계속 착용해야 할 우비와 큰 모자는 없으며, 부츠·다리 치료 흔적·센서 고장은 이 시점에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 몸의 골반과 접힌 다리가 능선 지면에 닿아 체중을 지지한다. 찰리의 바깥쪽 팔도 지면에 내려놓은 형태다. 맞잡은 손은 각자의 팔과 연결되어 있고 지면 가까이에 놓여 있다. 떠 있는 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우는 마을 너머 산줄기와 노을을 바라보고, 찰리도 머리를 약간 왼쪽으로 둔 채 같은 전방 풍경을 향한다. 둘 다 카메라를 돌아보지 않는다. 이현우의 오른손과 찰리의 왼손이 두 몸 사이의 열린 틈에서 분명하게 맞잡혀 있다.",
        "built_space": "두 인물은 연속된 바위·흙 능선 위에 왼쪽 이현우, 오른쪽 찰리 순서로 앉아 있다. 중앙 물탱크 1기, 오른쪽 교회와 첨탑 1개, 중앙 오른쪽 원형 골조 1개, 오른쪽 아래 아치 구조물 및 낮은 기와지붕들이 참조 장소의 배치를 유지한다. 후면에서 약간 옆으로 비껴 내려다보는 시점이며 손맞잡기와 지면 접촉이 잘 드러난다. 인물은 A보다 조금 낮고 작지만 여전히 하단 3분의 1을 다소 넘어선다.",
        "entities": "이현우 1명과 찰리 1대만 보인다. 이현우의 젊은 남성 체형, 헝클어진 검은 머리, 피와 먼지가 묻은 낡은 어두운 옷이 참조에 부합한다. 얼굴과 귀가 충분히 드러나지 않아 정확한 얼굴 일치, 표정과 무전기는 판정할 수 없다. 찰리는 긴 기계 팔, 짧은 다리, 베이지 장갑판과 긁힘·찍힘이 보이는 비인간 기계다. 얼굴과 가슴 원자로를 보여주려고 몸을 돌리지 않은 점은 지시에 맞는다. 우비와 큰 모자는 누락되었고, 부츠·치료된 다리 부상·센서 고장은 가시적으로 확인되지 않는다.",
        "hard_violations": [],
        "physics": "두 인물 모두 골반과 접힌 다리를 지면에 두고 안정적으로 앉아 있다. 찰리의 오른손은 몸 옆 땅을 짚고, 왼손은 이현우의 오른손을 잡는다. 손목과 팔의 연결 및 손맞잡기의 높이가 자연스럽고, 공중에 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1875,
   "A": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "요청된 후면 구도와 맞잡은 손, 캐릭터의 특징(특히 찰리의 육중한 기계 체형)과 지정된 배경을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "지정된 배경과 인물의 배치, 맞잡은 손 등 전반적인 지시사항을 잘 따랐으나 찰리의 체형이 B에 비해 다소 덜 육중해 보입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_ridge_5f6fbe.png",
    "asset_id": "8b8172cc-c7fe-4f3d-b4bf-57184ca13cb6",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c6b-b1ea-7cb8-8461-194a9f1a3348",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S60sh115__bgfirst_bg.png",
   "bg_asset_id": "167c6b68-db86-4ec7-aec1-c13327c38905",
   "bg_record_key": "S60sh115::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "village_ridge",
   "groupbg_asset_id": "8b8172cc-c7fe-4f3d-b4bf-57184ca13cb6"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S61sh1::signage": {
  "fp": "64ffc7cbc7fbb70d",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "방사능 구역 팻말"
   }
  ],
  "dropped": []
 },
 "groupbg::swamp_track_edge": {
  "input_fingerprint": "1f33adfdca0cfe0e",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "swamp_track_edge",
    "tags": [
     "S55sh8",
     "S61sh1",
     "S61sh3"
    ]
   },
   "context_sig": "67e874cd22fb0129"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 부웅~~~~ 한참 산길을 달리다 끼익! 철퍼덕!\n- 현우의 차가 늪지대 앞에서 멈추려다 풍덩 빠진다.\n- 늪지대. 박철진과 민병대원들이 만지고 둘러보고 있는 건...\n현우 일행이 탔던 자동차다. 녹슨 방사능 구역 팻말.\n\nTIME OF DAY (lock): night, moon visible.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 늪지대·방사능 구역 표지판 주변, 숲속·찰리 포획 지점, 함정 구덩이: 수분이 많은 진흙 지대와 부식된 출입 금지 철책이 있는 황폐한 숲이다. (특징: 차량 바퀴가 푹 빠진 검은 진흙 늪; 녹슬어 글씨 식별이 어려운 방사능 구역 표지판; 바닥에 찍힌 거대한 기계 발바닥(찰리의 신발) 자국; 나무 기둥에 꽂히는 깃털 달린 마취 화살촉; 현우와 앰버가 추락한 깊은 흙 구덩이와 덮쳐지는 그물망; 형형색색 하회탈 형태의 가죽 가면을 쓰고 도끼와 활을 든 병사들) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 부웅~~~~ 한참 산길을 달리다 끼익! 철퍼덕!\n- 현우의 차가 늪지대 앞에서 멈추려다 풍덩 빠진다.\n- 늪지대. 박철진과 민병대원들이 만지고 둘러보고 있는 건...\n현우 일행이 탔던 자동차다. 녹슨 방사능 구역 팻말.\n\nTIME OF DAY (lock): night, moon visible.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_swamp_track_edge_bb38ef.png",
  "asset_id": "197f971e-0b5f-4235-a233-ba02070b091a",
  "input_asset_ids": [
   "a37d2f20-6bcd-4a6f-8b15-37d140bde384"
  ],
  "origin_tag": "S61sh1",
  "place_text": "At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.",
  "origin_inputs": {
   "place_text": "At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.",
   "time_of_day_en": "night, moon visible",
   "conti_asset_id": "a37d2f20-6bcd-4a6f-8b15-37d140bde384"
  }
 },
 "S61sh1::bgfirst_bg": {
  "input_fingerprint": "58d095c1b5329f6f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 늪지대, 버려진 자동차와 녹슨 방사능 구역 팻말 주변을 둘러싸고 선 박철진과 민병대원들의 넓은 전경.\n\nLOCATION (lock): At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.\n\nTIME OF DAY (lock): night, moon visible.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high opening position outside the group, looking diagonally down across the front of the abandoned car before beginning the descent. Keep the car below two-fifths of the frame in the lower center, 박철진 near its front on the left, and the rusty radiation-zone sign legible on the right, with the swamp continuing behind them. 박철진 and the militia examine the car rather than the lens; distribute the men in unequal gaps, some leaning closer and others touching it with different arm angles and lowered head positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Abandoned and being touched and examined by the group) — Its front and one side are visible from the elevated diagonal viewpoint; used as Organizes the group's irregular arrangement while the men provide scale; 방사능 구역 팻말 (Rusty) — The warning-bearing face is turned sufficiently toward the camera for the radiation-zone designation to remain readable; used as Establishes the danger associated with the discovered vehicle; 늪지대 (Surrounds the abandoned vehicle and the searching group); used as Provides the spatial setting beyond the car without added atmospheric effects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime ambient illumination maintains readable faces and the warning sign without introducing visible lamps or exaggerated colored light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 늪지대, 버려진 자동차와 녹슨 방사능 구역 팻말 주변을 둘러싸고 선 박철진과 민병대원들의 넓은 전경.\n\nLOCATION (lock): At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night.\n\nTIME OF DAY (lock): night, moon visible.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high opening position outside the group, looking diagonally down across the front of the abandoned car before beginning the descent. Keep the car below two-fifths of the frame in the lower center, 박철진 near its front on the left, and the rusty radiation-zone sign legible on the right, with the swamp continuing behind them. 박철진 and the militia examine the car rather than the lens; distribute the men in unequal gaps, some leaning closer and others touching it with different arm angles and lowered head positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Abandoned and being touched and examined by the group) — Its front and one side are visible from the elevated diagonal viewpoint; used as Organizes the group's irregular arrangement while the men provide scale; 방사능 구역 팻말 (Rusty) — The warning-bearing face is turned sufficiently toward the camera for the radiation-zone designation to remain readable; used as Establishes the danger associated with the discovered vehicle; 늪지대 (Surrounds the abandoned vehicle and the searching group); used as Provides the spatial setting beyond the car without added atmospheric effects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime ambient illumination maintains readable faces and the warning sign without introducing visible lamps or exaggerated colored light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S61sh1__bgfirst_bg.png",
  "asset_id": "482f4b1b-1a67-4e01-9a7b-04db46ee5bae",
  "input_asset_ids": [
   "a37d2f20-6bcd-4a6f-8b15-37d140bde384",
   "197f971e-0b5f-4235-a233-ba02070b091a"
  ]
 },
 "S61sh1": {
  "input_fingerprint": "8a231f6c6b2161b4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 버려진 자동차와 녹슨 방사능 구역 팻말 주변을 둘러싸고 선 박철진과 민병대원들의 넓은 전경.\n\nLOCATION (lock): At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high opening position outside the group, looking diagonally down across the front of the abandoned car before beginning the descent. Keep the car below two-fifths of the frame in the lower center, 박철진 near its front on the left, and the rusty radiation-zone sign legible on the right, with the swamp continuing behind them. 박철진 and the militia examine the car rather than the lens; distribute the men in unequal gaps, some leaning closer and others touching it with different arm angles and lowered head positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Abandoned and being touched and examined by the group) — Its front and one side are visible from the elevated diagonal viewpoint; used as Organizes the group's irregular arrangement while the men provide scale; 방사능 구역 팻말 (Rusty) — The warning-bearing face is turned sufficiently toward the camera for the radiation-zone designation to remain readable; used as Establishes the danger associated with the discovered vehicle; 늪지대 (Surrounds the abandoned vehicle and the searching group); used as Provides the spatial setting beyond the car without added atmospheric effects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime ambient illumination maintains readable faces and the warning sign without introducing visible lamps or exaggerated colored light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp, beside the corroded radiation-zone warning sign. It is night, with a visible moon. 박철진: He is at the stranded vehicle inspecting the site.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 버려진 자동차와 녹슨 방사능 구역 팻말 주변을 둘러싸고 선 박철진과 민병대원들의 넓은 전경.\n\nLOCATION (lock): At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high opening position outside the group, looking diagonally down across the front of the abandoned car before beginning the descent. Keep the car below two-fifths of the frame in the lower center, 박철진 near its front on the left, and the rusty radiation-zone sign legible on the right, with the swamp continuing behind them. 박철진 and the militia examine the car rather than the lens; distribute the men in unequal gaps, some leaning closer and others touching it with different arm angles and lowered head positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Abandoned and being touched and examined by the group) — Its front and one side are visible from the elevated diagonal viewpoint; used as Organizes the group's irregular arrangement while the men provide scale; 방사능 구역 팻말 (Rusty) — The warning-bearing face is turned sufficiently toward the camera for the radiation-zone designation to remain readable; used as Establishes the danger associated with the discovered vehicle; 늪지대 (Surrounds the abandoned vehicle and the searching group); used as Provides the spatial setting beyond the car without added atmospheric effects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime ambient illumination maintains readable faces and the warning sign without introducing visible lamps or exaggerated colored light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp, beside the corroded radiation-zone warning sign. It is night, with a visible moon. 박철진: He is at the stranded vehicle inspecting the site.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 어두운 늪지대, 버려진 자동차와 녹슨 방사능 구역 팻말 주변을 둘러싸고 선 박철진과 민병대원들의 넓은 전경.\n\nLOCATION (lock): At the marsh-side stopping point of the abandoned camper, beside a rusted radiation-warning sign at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the crane's high opening position outside the group, looking diagonally down across the front of the abandoned car before beginning the descent. Keep the car below two-fifths of the frame in the lower center, 박철진 near its front on the left, and the rusty radiation-zone sign legible on the right, with the swamp continuing behind them. 박철진 and the militia examine the car rather than the lens; distribute the men in unequal gaps, some leaning closer and others touching it with different arm angles and lowered head positions.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Abandoned and being touched and examined by the group) — Its front and one side are visible from the elevated diagonal viewpoint; used as Organizes the group's irregular arrangement while the men provide scale; 방사능 구역 팻말 (Rusty) — The warning-bearing face is turned sufficiently toward the camera for the radiation-zone designation to remain readable; used as Establishes the danger associated with the discovered vehicle; 늪지대 (Surrounds the abandoned vehicle and the searching group); used as Provides the spatial setting beyond the car without added atmospheric effects.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low nighttime ambient illumination maintains readable faces and the warning sign without introducing visible lamps or exaggerated colored light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned camper remains stuck in the swamp, beside the corroded radiation-zone warning sign. It is night, with a visible moon. 박철진: He is at the stranded vehicle inspecting the site.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S61sh1__bgfirst_bg.png",
     "asset_id": "482f4b1b-1a67-4e01-9a7b-04db46ee5bae",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S61sh1.png",
     "asset_id": "a37d2f20-6bcd-4a6f-8b15-37d140bde384",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_swamp_track_edge_bb38ef.png",
     "asset_id": "197f971e-0b5f-4235-a233-ba02070b091a",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:999958>",
     "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 모두 하단 중앙의 자동차에 시선을 고정하고 있으며, 왼쪽 앞의 인물은 차체 보닛 위로 팔을 뻗어 만지고 있음.",
    "built_space": "위치 참조와 동일한 늪지대 환경. 버려진 자동차는 하단 중앙에, 방사능 팻말은 우측에 위치함. 인물들은 차 주변 진흙에 간격을 두고 서 있음.",
    "entities": "남색 전투복, 부니햇, 십자 마크가 있는 붉은 완장을 착용한 4명의 남성. 시선 방향과 모자로 인해 얼굴은 가려져 있으나 복장은 참조와 일치함. 자동차와 팻말이 지시대로 존재함.",
    "hard_violations": [],
    "physics": "모든 인물이 깊은 진흙 바닥에 안정적으로 체중을 싣고 서 있으며, 자동차를 짚은 손도 차체 표면에 물리적으로 맞닿아 있음."
   },
   {
    "label": "B",
    "direction": "인물들이 자동차를 둘러싸고 살피고 있으며, 좌측 앞의 쪼그려 앉은 남성이 손전등 불빛을 차 앞면 그릴 쪽에 비추고 있음.",
    "built_space": "지정된 늪지대 배경. 하단 중앙의 자동차와 우측의 팻말 배치가 올바르며, 인물들이 차 주변 진흙에 자리 잡고 있음.",
    "entities": "남색 전투복과 붉은 완장을 찬 6명의 남성. 좌측 서 있는 인물의 얼굴 형태가 박철진 참조와 유사함. 인위적인 광원인 손전등이 등장함.",
    "hard_violations": [
     "[gemini-pro] invented objects: 금지된 가시적 램프(손전등)가 추가되어 사용됨"
    ],
    "physics": "서 있거나 쪼그려 앉은 인물들 모두 진흙 바닥에 발을 단단히 지지하고 있으며 부자연스러운 부유 현상은 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "조명 제한을 정확히 준수하고, 요구된 카메라 구도 안에서 인물들의 행동과 배치를 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "명시적으로 금지된 가시적인 조명(손전등)을 화면에 무단으로 추가하여 조명 및 무드 지침을 크게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 모두 하단 중앙의 자동차에 시선을 고정하고 있으며, 왼쪽 앞의 인물은 차체 보닛 위로 팔을 뻗어 만지고 있음.",
        "built_space": "위치 참조와 동일한 늪지대 환경. 버려진 자동차는 하단 중앙에, 방사능 팻말은 우측에 위치함. 인물들은 차 주변 진흙에 간격을 두고 서 있음.",
        "entities": "남색 전투복, 부니햇, 십자 마크가 있는 붉은 완장을 착용한 4명의 남성. 시선 방향과 모자로 인해 얼굴은 가려져 있으나 복장은 참조와 일치함. 자동차와 팻말이 지시대로 존재함.",
        "hard_violations": [],
        "physics": "모든 인물이 깊은 진흙 바닥에 안정적으로 체중을 싣고 서 있으며, 자동차를 짚은 손도 차체 표면에 물리적으로 맞닿아 있음."
       },
       {
        "label": "B",
        "direction": "인물들이 자동차를 둘러싸고 살피고 있으며, 좌측 앞의 쪼그려 앉은 남성이 손전등 불빛을 차 앞면 그릴 쪽에 비추고 있음.",
        "built_space": "지정된 늪지대 배경. 하단 중앙의 자동차와 우측의 팻말 배치가 올바르며, 인물들이 차 주변 진흙에 자리 잡고 있음.",
        "entities": "남색 전투복과 붉은 완장을 찬 6명의 남성. 좌측 서 있는 인물의 얼굴 형태가 박철진 참조와 유사함. 인위적인 광원인 손전등이 등장함.",
        "hard_violations": [
         "invented objects: 금지된 가시적 램프(손전등)가 추가되어 사용됨"
        ],
        "physics": "서 있거나 쪼그려 앉은 인물들 모두 진흙 바닥에 발을 단단히 지지하고 있으며 부자연스러운 부유 현상은 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "조명 제한을 정확히 준수하고, 요구된 카메라 구도 안에서 인물들의 행동과 배치를 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "명시적으로 금지된 가시적인 조명(손전등)을 화면에 무단으로 추가하여 조명 및 무드 지침을 크게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 모두 하단 중앙의 자동차에 시선을 고정하고 있으며, 왼쪽 앞의 인물은 차체 보닛 위로 팔을 뻗어 만지고 있음.",
        "built_space": "위치 참조와 동일한 늪지대 환경. 버려진 자동차는 하단 중앙에, 방사능 팻말은 우측에 위치함. 인물들은 차 주변 진흙에 간격을 두고 서 있음.",
        "entities": "남색 전투복, 부니햇, 십자 마크가 있는 붉은 완장을 착용한 4명의 남성. 시선 방향과 모자로 인해 얼굴은 가려져 있으나 복장은 참조와 일치함. 자동차와 팻말이 지시대로 존재함.",
        "hard_violations": [],
        "physics": "모든 인물이 깊은 진흙 바닥에 안정적으로 체중을 싣고 서 있으며, 자동차를 짚은 손도 차체 표면에 물리적으로 맞닿아 있음."
       },
       {
        "label": "B",
        "direction": "인물들이 자동차를 둘러싸고 살피고 있으며, 좌측 앞의 쪼그려 앉은 남성이 손전등 불빛을 차 앞면 그릴 쪽에 비추고 있음.",
        "built_space": "지정된 늪지대 배경. 하단 중앙의 자동차와 우측의 팻말 배치가 올바르며, 인물들이 차 주변 진흙에 자리 잡고 있음.",
        "entities": "남색 전투복과 붉은 완장을 찬 6명의 남성. 좌측 서 있는 인물의 얼굴 형태가 박철진 참조와 유사함. 인위적인 광원인 손전등이 등장함.",
        "hard_violations": [
         "invented objects: 금지된 가시적 램프(손전등)가 추가되어 사용됨"
        ],
        "physics": "서 있거나 쪼그려 앉은 인물들 모두 진흙 바닥에 발을 단단히 지지하고 있으며 부자연스러운 부유 현상은 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "더 높은 사선 시점과 불균등한 인물 간격, 다양한 조사 자세가 크레인 시작 전경에 더 충실하지만, 왼쪽 전조등 부근의 국소적인 밝은 빛은 조명 제한에서 벗어난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "차량을 직접 만지는 동작과 의상은 충실하지만, 시점이 상대적으로 낮아 요구된 높은 크레인 시작 구도보다 장소 참고사진의 구도에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 앞의 박철진으로 보이는 남성은 차량 앞부분을 내려다본다. 앞쪽에 웅크린 두 명은 왼쪽 전조등과 범퍼를, 오른쪽에서 숙인 남성은 차 안을 살핀다. 나머지 남성들도 차량 쪽으로 머리와 몸을 돌리며 렌즈를 응시하지 않는다. 뒤쪽 남성의 소총은 아래쪽 차량 옆 지면을 향하며, 오른쪽 전경의 총도 아래로 내려져 있다. 오른쪽 표지판의 방사능 기호는 카메라 쪽으로 보인다.",
        "built_space": "폐차 한 대가 하단 중앙에 있고, 두 기둥으로 지지된 녹슨 표지판 한 개가 오른쪽에 있다. 남성 일곱 명이 차량 왼쪽 앞과 뒤, 앞 범퍼 주변, 오른쪽 측면과 전경에 불균등하게 배치되어 있다. 차량 전면·오른쪽 측면·지붕이 함께 보여 높은 외부 사선 시점이 읽힌다. 참고사진의 진흙, 갈대, 왼쪽 고사목, 먼 산과 넓은 수면이 유지되며 달 아래 수면 반사도 자연스럽다.",
        "entities": "녹슨 캠퍼형 차량, 방사능 기호가 있는 부식된 표지판, 늪, 밤하늘의 달이 모두 있다. 박철진으로 보이는 왼쪽 앞 남성은 중년 동아시아계 남성의 외관이며 남색 전투복, 검은 챙모자, 붉은 완장이 참고와 부합한다. 얼굴이 작고 숙여져 있어 정확한 동일 인물 여부와 머리 모양은 확인하기 어렵다. 민병대원 여섯 명도 실체 있는 성인 남성으로 보이며, 여러 명에게 참고 의상보다 많은 전술 조끼와 장비가 있다. 왼쪽 전조등 부근에는 별도의 광원처럼 보이는 밝은 빛이 있다.",
        "hard_violations": [],
        "physics": "서 있는 인물들은 진흙 지면에 발을 딛고 있으며, 앞의 두 인물은 무릎과 엉덩이를 굽혀 낮춘 자세다. 오른쪽에서 차 안을 보는 인물은 발로 체중을 지탱하며 상체를 숙인다. 차량은 바퀴와 하부가 진흙에 잠겨 지지되고, 표지판은 지면에 박힌 두 기둥에 고정되어 있다. 총기와 장비는 손이나 몸의 끈에 의해 지지되며, 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 앞의 박철진으로 보이는 남성은 보닛에 손을 대고 그 표면을 내려다본다. 오른쪽 앞 남성은 열린 측면 쪽으로 몸을 숙여 내부를 본다. 왼쪽 바깥과 차량 뒤의 두 남성도 차량 또는 그 주변 진흙을 향해 고개를 낮춘다. 렌즈를 보는 인물은 없다. 등에 멘 총기는 몸을 따라 놓여 특정 대상을 조준하지 않는다. 오른쪽 표지판의 방사능 기호는 카메라에서 명확히 식별된다.",
        "built_space": "폐차 한 대와 오른쪽의 두 기둥식 표지판 한 개가 있으며, 남성 다섯 명이 차량 왼쪽 바깥, 왼쪽 앞, 오른쪽 앞, 뒤쪽 두 자리에 서 있다. 차량은 하단 중앙에 놓이고 전면과 오른쪽 측면이 보인다. 다만 지붕을 내려다보는 각도가 A보다 얕아 높은 크레인 시작 위치의 인상이 약하다. 갈대 늪, 왼쪽 고사목, 산 능선, 달과 수면 반사 등 장소의 주요 요소는 참고사진과 잘 맞는다.",
        "entities": "버려진 녹슨 캠퍼, 부식된 방사능 경고 표지판, 늪과 달이 모두 보인다. 왼쪽 앞 남성은 남색 전투복과 붉은 완장, 검은 챙모자를 착용한 성인 동아시아계 남성으로 보인다. 얼굴이 아래로 향해 박철진의 정확한 얼굴과 나이는 판별하기 어렵다. 다른 네 명도 남색 계열 복장과 붉은 완장을 착용한 민병대원으로 읽힌다. 별도의 발광 장치나 화면 위 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "다섯 인물 모두 지면에 발을 딛고 있다. 보닛을 살피는 남성은 벌린 다리로 몸을 지탱하면서 손바닥을 보닛에 얹고, 측면을 보는 남성은 무릎과 허리를 약간 굽힌다. 차량의 바퀴와 하부는 진흙에 묻혀 있고 표지판은 두 기둥으로 지지된다. 휴대 장비는 몸에 부착되어 있으며, 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "더 높은 사선 시점과 불균등한 인물 간격, 다양한 조사 자세가 크레인 시작 전경에 더 충실하지만, 왼쪽 전조등 부근의 국소적인 밝은 빛은 조명 제한에서 벗어난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "차량을 직접 만지는 동작과 의상은 충실하지만, 시점이 상대적으로 낮아 요구된 높은 크레인 시작 구도보다 장소 참고사진의 구도에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 앞의 박철진으로 보이는 남성은 차량 앞부분을 내려다본다. 앞쪽에 웅크린 두 명은 왼쪽 전조등과 범퍼를, 오른쪽에서 숙인 남성은 차 안을 살핀다. 나머지 남성들도 차량 쪽으로 머리와 몸을 돌리며 렌즈를 응시하지 않는다. 뒤쪽 남성의 소총은 아래쪽 차량 옆 지면을 향하며, 오른쪽 전경의 총도 아래로 내려져 있다. 오른쪽 표지판의 방사능 기호는 카메라 쪽으로 보인다.",
        "built_space": "폐차 한 대가 하단 중앙에 있고, 두 기둥으로 지지된 녹슨 표지판 한 개가 오른쪽에 있다. 남성 일곱 명이 차량 왼쪽 앞과 뒤, 앞 범퍼 주변, 오른쪽 측면과 전경에 불균등하게 배치되어 있다. 차량 전면·오른쪽 측면·지붕이 함께 보여 높은 외부 사선 시점이 읽힌다. 참고사진의 진흙, 갈대, 왼쪽 고사목, 먼 산과 넓은 수면이 유지되며 달 아래 수면 반사도 자연스럽다.",
        "entities": "녹슨 캠퍼형 차량, 방사능 기호가 있는 부식된 표지판, 늪, 밤하늘의 달이 모두 있다. 박철진으로 보이는 왼쪽 앞 남성은 중년 동아시아계 남성의 외관이며 남색 전투복, 검은 챙모자, 붉은 완장이 참고와 부합한다. 얼굴이 작고 숙여져 있어 정확한 동일 인물 여부와 머리 모양은 확인하기 어렵다. 민병대원 여섯 명도 실체 있는 성인 남성으로 보이며, 여러 명에게 참고 의상보다 많은 전술 조끼와 장비가 있다. 왼쪽 전조등 부근에는 별도의 광원처럼 보이는 밝은 빛이 있다.",
        "hard_violations": [],
        "physics": "서 있는 인물들은 진흙 지면에 발을 딛고 있으며, 앞의 두 인물은 무릎과 엉덩이를 굽혀 낮춘 자세다. 오른쪽에서 차 안을 보는 인물은 발로 체중을 지탱하며 상체를 숙인다. 차량은 바퀴와 하부가 진흙에 잠겨 지지되고, 표지판은 지면에 박힌 두 기둥에 고정되어 있다. 총기와 장비는 손이나 몸의 끈에 의해 지지되며, 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 앞의 박철진으로 보이는 남성은 보닛에 손을 대고 그 표면을 내려다본다. 오른쪽 앞 남성은 열린 측면 쪽으로 몸을 숙여 내부를 본다. 왼쪽 바깥과 차량 뒤의 두 남성도 차량 또는 그 주변 진흙을 향해 고개를 낮춘다. 렌즈를 보는 인물은 없다. 등에 멘 총기는 몸을 따라 놓여 특정 대상을 조준하지 않는다. 오른쪽 표지판의 방사능 기호는 카메라에서 명확히 식별된다.",
        "built_space": "폐차 한 대와 오른쪽의 두 기둥식 표지판 한 개가 있으며, 남성 다섯 명이 차량 왼쪽 바깥, 왼쪽 앞, 오른쪽 앞, 뒤쪽 두 자리에 서 있다. 차량은 하단 중앙에 놓이고 전면과 오른쪽 측면이 보인다. 다만 지붕을 내려다보는 각도가 A보다 얕아 높은 크레인 시작 위치의 인상이 약하다. 갈대 늪, 왼쪽 고사목, 산 능선, 달과 수면 반사 등 장소의 주요 요소는 참고사진과 잘 맞는다.",
        "entities": "버려진 녹슨 캠퍼, 부식된 방사능 경고 표지판, 늪과 달이 모두 보인다. 왼쪽 앞 남성은 남색 전투복과 붉은 완장, 검은 챙모자를 착용한 성인 동아시아계 남성으로 보인다. 얼굴이 아래로 향해 박철진의 정확한 얼굴과 나이는 판별하기 어렵다. 다른 네 명도 남색 계열 복장과 붉은 완장을 착용한 민병대원으로 읽힌다. 별도의 발광 장치나 화면 위 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "다섯 인물 모두 지면에 발을 딛고 있다. 보닛을 살피는 남성은 벌린 다리로 몸을 지탱하면서 손바닥을 보닛에 얹고, 측면을 보는 남성은 무릎과 허리를 약간 굽힌다. 차량의 바퀴와 하부는 진흙에 묻혀 있고 표지판은 두 기둥으로 지지된다. 휴대 장비는 몸에 부착되어 있으며, 지지 없이 떠 있는 물체나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gemini-pro] invented objects: 금지된 가시적 램프(손전등)가 추가되어 사용됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "조명 제한을 정확히 준수하고, 요구된 카메라 구도 안에서 인물들의 행동과 배치를 충실하게 구현함."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "명시적으로 금지된 가시적인 조명(손전등)을 화면에 무단으로 추가하여 조명 및 무드 지침을 크게 위반함.  ★위반: [gemini-pro] invented objects: 금지된 가시적 램프(손전등)가 추가되어 사용됨"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_swamp_track_edge_bb38ef.png",
    "asset_id": "197f971e-0b5f-4235-a233-ba02070b091a",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c74-dda9-7d51-bf67-6ccd8db39390",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S61sh1__bgfirst_bg.png",
   "bg_asset_id": "482f4b1b-1a67-4e01-9a7b-04db46ee5bae",
   "bg_record_key": "S61sh1::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "swamp_track_edge",
   "groupbg_asset_id": "197f971e-0b5f-4235-a233-ba02070b091a"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S61sh3::signage": {
  "fp": "4bbf36f22ea39748",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S61sh3": {
  "input_fingerprint": "ddbb77220b8cbc1a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 버려진 차 앞에서 입꼬리를 한쪽으로 비틀어 올린 채 비릿하게 미소 짓는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): Immediately beside the abandoned camper at the dark marsh's edge, near the radiation-warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the approach beside the car in a facial close-up of 박철진, retaining the established diagonal axis and the endpoint's slight upward view. Place his face left of center, with a soft fragment of the car at the lower-right edge, as he lifts his chin and twists one corner of his mouth toward the subordinate outside the frame. Let the tighter distance emphasize his predatory amusement without a new lighting cue or a turn toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Remains beside 박철진 as the inspection continues) — Only an oblique fragment of its front remains visible; used as Provides a soft continuity anchor beneath the face rather than competing with it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued nighttime illumination, keeping the crooked smile readable through restrained facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the abandoned camper, corroded warning sign, swamp ground, and nighttime appearance from the reference. Exclude the capture net and pit-trap features from the separate forest locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in the dark swamp beside the rusty radiation-zone sign. 박철진: He remains at the abandoned vehicle during the inspection.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 버려진 차 앞에서 입꼬리를 한쪽으로 비틀어 올린 채 비릿하게 미소 짓는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): Immediately beside the abandoned camper at the dark marsh's edge, near the radiation-warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the approach beside the car in a facial close-up of 박철진, retaining the established diagonal axis and the endpoint's slight upward view. Place his face left of center, with a soft fragment of the car at the lower-right edge, as he lifts his chin and twists one corner of his mouth toward the subordinate outside the frame. Let the tighter distance emphasize his predatory amusement without a new lighting cue or a turn toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Remains beside 박철진 as the inspection continues) — Only an oblique fragment of its front remains visible; used as Provides a soft continuity anchor beneath the face rather than competing with it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued nighttime illumination, keeping the crooked smile readable through restrained facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the abandoned camper, corroded warning sign, swamp ground, and nighttime appearance from the reference. Exclude the capture net and pit-trap features from the separate forest locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in the dark swamp beside the rusty radiation-zone sign. 박철진: He remains at the abandoned vehicle during the inspection.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, moon visible.\n\nSHOT TEXT (authoritative, Korean): 버려진 차 앞에서 입꼬리를 한쪽으로 비틀어 올린 채 비릿하게 미소 짓는 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): Immediately beside the abandoned camper at the dark marsh's edge, near the radiation-warning sign. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the approach beside the car in a facial close-up of 박철진, retaining the established diagonal axis and the endpoint's slight upward view. Place his face left of center, with a soft fragment of the car at the lower-right edge, as he lifts his chin and twists one corner of his mouth toward the subordinate outside the frame. Let the tighter distance emphasize his predatory amusement without a new lighting cue or a turn toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 버려진 자동차 (Remains beside 박철진 as the inspection continues) — Only an oblique fragment of its front remains visible; used as Provides a soft continuity anchor beneath the face rather than competing with it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued nighttime illumination, keeping the crooked smile readable through restrained facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the abandoned camper, corroded warning sign, swamp ground, and nighttime appearance from the reference. Exclude the capture net and pit-trap features from the separate forest locations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camper remains abandoned and stuck in the dark swamp beside the rusty radiation-zone sign. 박철진: He remains at the abandoned vehicle during the inspection.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — wearing: 조악하게 맞춰 입은 어두운 남색 계열의 전투복, 팔에 붉은색 민병대 완장이 차여 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "프레임 밖 우측을 향한 시선과 한쪽 입꼬리가 올라간 미소가 뚜렷함.",
    "built_space": "야간 늪지대 배경, 우측 원경에 방사능 표지판, 우측 하단에 차량 후드(전면부) 파편이 프레임에 걸쳐 있음.",
    "entities": "박철진의 얼굴, 남색 전투복, 붉은 완장, 챙 모자 모두 레퍼런스와 일치. 표지판과 차량 파편의 질감도 잘 구현됨.",
    "hard_violations": [],
    "physics": "자연스럽게 서 있는 자세로 물리적 오류 없음."
   },
   {
    "label": "B",
    "direction": "프레임 밖 우측을 향한 시선과 옅은 미소가 확인됨.",
    "built_space": "야간 늪지대, 우측 뒤편에 표지판, 우측 하단에 차량의 일부가 보임.",
    "entities": "인물의 이목구비와 복장은 레퍼런스와 일치하지만, 우측 하단의 차량 파편이 레퍼런스의 각진 차량 형태(앞부분)가 아닌 둥근 형태로 왜곡되어 보임.",
    "hard_violations": [],
    "physics": "안정적인 자세를 유지하고 있으며 오류 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 비릿한 미소 표정을 훌륭하게 연출했으며, 우측 하단의 차량 전면부 파편도 레퍼런스와 일치하게 배치되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 분위기와 인물 묘사는 좋으나, 우측 하단의 차량 파편 형태가 레퍼런스의 각진 전면부가 아닌 둥근 측면부처럼 묘사되어 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "프레임 밖 우측을 향한 시선과 한쪽 입꼬리가 올라간 미소가 뚜렷함.",
        "built_space": "야간 늪지대 배경, 우측 원경에 방사능 표지판, 우측 하단에 차량 후드(전면부) 파편이 프레임에 걸쳐 있음.",
        "entities": "박철진의 얼굴, 남색 전투복, 붉은 완장, 챙 모자 모두 레퍼런스와 일치. 표지판과 차량 파편의 질감도 잘 구현됨.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세로 물리적 오류 없음."
       },
       {
        "label": "B",
        "direction": "프레임 밖 우측을 향한 시선과 옅은 미소가 확인됨.",
        "built_space": "야간 늪지대, 우측 뒤편에 표지판, 우측 하단에 차량의 일부가 보임.",
        "entities": "인물의 이목구비와 복장은 레퍼런스와 일치하지만, 우측 하단의 차량 파편이 레퍼런스의 각진 차량 형태(앞부분)가 아닌 둥근 형태로 왜곡되어 보임.",
        "hard_violations": [],
        "physics": "안정적인 자세를 유지하고 있으며 오류 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 비릿한 미소 표정을 훌륭하게 연출했으며, 우측 하단의 차량 전면부 파편도 레퍼런스와 일치하게 배치되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 분위기와 인물 묘사는 좋으나, 우측 하단의 차량 파편 형태가 레퍼런스의 각진 전면부가 아닌 둥근 측면부처럼 묘사되어 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "프레임 밖 우측을 향한 시선과 한쪽 입꼬리가 올라간 미소가 뚜렷함.",
        "built_space": "야간 늪지대 배경, 우측 원경에 방사능 표지판, 우측 하단에 차량 후드(전면부) 파편이 프레임에 걸쳐 있음.",
        "entities": "박철진의 얼굴, 남색 전투복, 붉은 완장, 챙 모자 모두 레퍼런스와 일치. 표지판과 차량 파편의 질감도 잘 구현됨.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세로 물리적 오류 없음."
       },
       {
        "label": "B",
        "direction": "프레임 밖 우측을 향한 시선과 옅은 미소가 확인됨.",
        "built_space": "야간 늪지대, 우측 뒤편에 표지판, 우측 하단에 차량의 일부가 보임.",
        "entities": "인물의 이목구비와 복장은 레퍼런스와 일치하지만, 우측 하단의 차량 파편이 레퍼런스의 각진 차량 형태(앞부분)가 아닌 둥근 형태로 왜곡되어 보임.",
        "hard_violations": [],
        "physics": "안정적인 자세를 유지하고 있으며 오류 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 얼굴 배치와 화면 밖을 향한 비틀린 미소, 우하단 차량 조각이 더 충실하지만, 얼굴 클로즈업치고 가슴과 배경이 많이 보인다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 야간 장소는 이어지지만, 상체를 앞으로 숙인 자세와 크게 드러난 차량 전면이 턱을 든 얼굴 중심의 밀착 구도에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "눈은 렌즈가 아니라 화면 오른쪽 바깥을 향해 있어 프레임 밖 부하를 보는 설정과 맞는다. 고개를 약간 기울이고 오른쪽으로 보이는 입꼬리를 올렸으며, 비릿한 미소가 분명하다. 턱을 들어 올린 동작은 미약하지만 아래로 숙이지는 않았다. 무기나 별도로 겨냥하는 물체는 없다.",
        "built_space": "왼쪽 전경에 인물 한 명, 우하단 가장자리에 녹슨 차량의 창틀과 전면 일부, 오른쪽 뒤에 방사능 표지판 한 개가 보인다. 늪의 물과 갈대, 먼 산, 달 한 개가 이어진다. 차량은 비스듬한 조각으로 제한되어 있으나 완전히 흐리지는 않다. 표지판은 참조처럼 차량 오른쪽 배경에 놓여 있으며 중복 시설이나 불가능한 반사는 보이지 않는다. 얼굴은 중심 왼쪽이지만 가슴까지 포함되어 요구한 밀착도보다 넓다.",
        "entities": "참조와 유사한 중년 한국인 남성 외형의 인물 한 명만 있다. 얼굴 윤곽과 눈매가 대체로 이어지며, 검은 머리는 남색 챙모자 아래 일부만 보인다. 남색 전투복, 모자 턱끈, 흰 표식이 있는 붉은 완장이 참조와 맞는다. 부식된 차량과 방사능 표지판, 달이 있는 밤의 늪도 확인된다. 차량 전체 형식과 늪에 박힌 바퀴는 프레임 밖이므로 확인할 수 없다. 추가 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 모자는 머리에 얹혀 있으며 완장은 소매를 감싸고 있다. 몸통은 화면 아래로 이어져 발의 지지는 확인되지 않지만 떠 있는 징후는 없다. 표지판에는 아래로 이어지는 지지대가 보이고, 차량 조각도 차체에 연결되어 있다. 공중에 떠 있거나 손의 지지가 필요한 물체는 없다."
       },
       {
        "label": "B",
        "direction": "눈은 화면 오른쪽 바깥을 향하며 렌즈를 직접 보지 않아 부하의 화면 밖 위치와 양립한다. 입꼬리에는 약한 비대칭 미소가 있지만 A보다 얌전하다. 상체와 고개를 앞으로 기울이고 턱을 당긴 인상이라 턱을 들어 비웃는 지정 동작과 차이가 난다. 겨냥하는 무기나 이동 물체는 없다.",
        "built_space": "왼쪽에 상체를 기울인 인물 한 명, 오른쪽에 차량 앞유리와 보닛의 상당 부분, 중앙 오른쪽 배경에 녹슨 방사능 표지판 한 개가 보인다. 늪과 갈대, 먼 산, 달 한 개도 있다. 차량이 우하단의 부드러운 작은 조각에 머무르지 않고 오른쪽 가장자리 위쪽까지 올라와 얼굴과 경쟁한다. 표지판의 화면상 위치는 참조와 달라졌지만 이 구도만으로 실제 이동이나 불가능한 배치를 단정할 수 없다. 중복 시설은 없다.",
        "entities": "참조와 대체로 유사한 중년 한국인 남성 외형의 인물 한 명이며 정상적인 눈과 얼굴을 갖는다. 남색 모자와 턱끈, 낡은 남색 전투복, 붉은 완장 일부가 참조 복장을 따른다. 머리카락은 모자 아래 조금만 드러난다. 부식된 차량 전면과 방사능 기호가 있는 표지판, 달이 뜬 늪은 요구한 종류에 맞는다. 추가 인물이나 불필요한 글자는 없다.",
        "hard_violations": [],
        "physics": "기울어진 머리와 몸통은 목과 어깨를 통해 자연스럽게 연결되어 있고, 허리에서 앞으로 굽힌 자세로 가능한 범위다. 하체와 손은 프레임 밖이므로 지면 접촉이나 차량을 짚었는지는 확인할 수 없지만, 지지 없이 떠 있다고 볼 근거는 없다. 모자는 머리에 놓이고 완장은 팔에 둘러져 있다. 표지판은 지지대가 아래로 이어지며 차량 부품도 차체에 붙어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 얼굴 배치와 화면 밖을 향한 비틀린 미소, 우하단 차량 조각이 더 충실하지만, 얼굴 클로즈업치고 가슴과 배경이 많이 보인다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물과 야간 장소는 이어지지만, 상체를 앞으로 숙인 자세와 크게 드러난 차량 전면이 턱을 든 얼굴 중심의 밀착 구도에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "눈은 렌즈가 아니라 화면 오른쪽 바깥을 향해 있어 프레임 밖 부하를 보는 설정과 맞는다. 고개를 약간 기울이고 오른쪽으로 보이는 입꼬리를 올렸으며, 비릿한 미소가 분명하다. 턱을 들어 올린 동작은 미약하지만 아래로 숙이지는 않았다. 무기나 별도로 겨냥하는 물체는 없다.",
        "built_space": "왼쪽 전경에 인물 한 명, 우하단 가장자리에 녹슨 차량의 창틀과 전면 일부, 오른쪽 뒤에 방사능 표지판 한 개가 보인다. 늪의 물과 갈대, 먼 산, 달 한 개가 이어진다. 차량은 비스듬한 조각으로 제한되어 있으나 완전히 흐리지는 않다. 표지판은 참조처럼 차량 오른쪽 배경에 놓여 있으며 중복 시설이나 불가능한 반사는 보이지 않는다. 얼굴은 중심 왼쪽이지만 가슴까지 포함되어 요구한 밀착도보다 넓다.",
        "entities": "참조와 유사한 중년 한국인 남성 외형의 인물 한 명만 있다. 얼굴 윤곽과 눈매가 대체로 이어지며, 검은 머리는 남색 챙모자 아래 일부만 보인다. 남색 전투복, 모자 턱끈, 흰 표식이 있는 붉은 완장이 참조와 맞는다. 부식된 차량과 방사능 표지판, 달이 있는 밤의 늪도 확인된다. 차량 전체 형식과 늪에 박힌 바퀴는 프레임 밖이므로 확인할 수 없다. 추가 인물이나 덧씌운 문구는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 모자는 머리에 얹혀 있으며 완장은 소매를 감싸고 있다. 몸통은 화면 아래로 이어져 발의 지지는 확인되지 않지만 떠 있는 징후는 없다. 표지판에는 아래로 이어지는 지지대가 보이고, 차량 조각도 차체에 연결되어 있다. 공중에 떠 있거나 손의 지지가 필요한 물체는 없다."
       },
       {
        "label": "A",
        "direction": "눈은 화면 오른쪽 바깥을 향하며 렌즈를 직접 보지 않아 부하의 화면 밖 위치와 양립한다. 입꼬리에는 약한 비대칭 미소가 있지만 A보다 얌전하다. 상체와 고개를 앞으로 기울이고 턱을 당긴 인상이라 턱을 들어 비웃는 지정 동작과 차이가 난다. 겨냥하는 무기나 이동 물체는 없다.",
        "built_space": "왼쪽에 상체를 기울인 인물 한 명, 오른쪽에 차량 앞유리와 보닛의 상당 부분, 중앙 오른쪽 배경에 녹슨 방사능 표지판 한 개가 보인다. 늪과 갈대, 먼 산, 달 한 개도 있다. 차량이 우하단의 부드러운 작은 조각에 머무르지 않고 오른쪽 가장자리 위쪽까지 올라와 얼굴과 경쟁한다. 표지판의 화면상 위치는 참조와 달라졌지만 이 구도만으로 실제 이동이나 불가능한 배치를 단정할 수 없다. 중복 시설은 없다.",
        "entities": "참조와 대체로 유사한 중년 한국인 남성 외형의 인물 한 명이며 정상적인 눈과 얼굴을 갖는다. 남색 모자와 턱끈, 낡은 남색 전투복, 붉은 완장 일부가 참조 복장을 따른다. 머리카락은 모자 아래 조금만 드러난다. 부식된 차량 전면과 방사능 기호가 있는 표지판, 달이 뜬 늪은 요구한 종류에 맞는다. 추가 인물이나 불필요한 글자는 없다.",
        "hard_violations": [],
        "physics": "기울어진 머리와 몸통은 목과 어깨를 통해 자연스럽게 연결되어 있고, 허리에서 앞으로 굽힌 자세로 가능한 범위다. 하체와 손은 프레임 밖이므로 지면 접촉이나 차량을 짚었는지는 확인할 수 없지만, 지지 없이 떠 있다고 볼 근거는 없다. 모자는 머리에 놓이고 완장은 팔에 둘러져 있다. 표지판은 지지대가 아래로 이어지며 차량 부품도 차체에 붙어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 앵글과 비릿한 미소 표정을 훌륭하게 연출했으며, 우측 하단의 차량 전면부 파편도 레퍼런스와 일치하게 배치되었습니다."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "전반적인 분위기와 인물 묘사는 좋으나, 우측 하단의 차량 파편 형태가 레퍼런스의 각진 전면부가 아닌 둥근 측면부처럼 묘사되어 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S61sh1_sel.png",
    "asset_id": "6c423e75-2401-4c56-90d0-b652b733ab69",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:999958>",
    "asset_id": "f98fc9a7-e0b0-468b-bb71-57ec11f0d6f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c7e-54c7-70fb-99f2-7428973f8c49",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S61sh1"
  }
 },
 "S62sh4::signage": {
  "fp": "25e9dc97c8ab0ef0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::food_storage": {
  "input_fingerprint": "cc6be73b457e674e",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "food_storage",
    "tags": [
     "S62sh4"
    ]
   },
   "context_sig": "c12c78e98614c563"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 텅 빈 식량창고. 쌀 몇 포대와 감자 한 포대 정도.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 텅 빈 식량창고. 쌀 몇 포대와 감자 한 포대 정도.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_food_storage_8124be.png",
  "asset_id": "317f75ce-9f95-4c6f-a830-9ec3f2d7c95b",
  "input_asset_ids": [
   "7f8167b8-7da2-4ac6-88dd-bb51e4640391"
  ],
  "origin_tag": "S62sh4",
  "place_text": "Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.",
  "origin_inputs": {
   "place_text": "Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.",
   "time_of_day_en": "day",
   "conti_asset_id": "7f8167b8-7da2-4ac6-88dd-bb51e4640391"
  }
 },
 "S62sh4::bgfirst_bg": {
  "input_fingerprint": "ceefc4b6485ab43a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 텅 빈 창고 안을 바라보며 충격받은 표정으로 굳어버린 이현우, 수빈, 마을총무의 구도.\n\nLOCATION (lock): Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the storage-room approach at upper-chest camera height, offset from the group's entry path, with 이현우 left, 수빈 center, and 마을총무 right in a staggered medium three-shot. Keep all three faces readable at oblique angles as their interrupted steps and uneven shoulder positions express shock, their eyes lowered toward the meager provisions below the frame rather than toward the camera. Retain the open doorway behind them and a narrow band of empty interior around their bodies, making distance the approach's only emphatic change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 식량창고 내부 (Nearly empty, with only a few rice sacks and one potato sack remaining below the reaction framing) — Seen from within the room, looking obliquely toward the entering group; used as Empty space around the figures supports the scarcity revealed by their downward eyelines; 낡은 철문 (Open) — The open door leaf is seen obliquely behind the group; used as Marks the entrance and preserves the direction of their interrupted movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light keeps all three reactions readable with subdued brightness and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 텅 빈 창고 안을 바라보며 충격받은 표정으로 굳어버린 이현우, 수빈, 마을총무의 구도.\n\nLOCATION (lock): Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the storage-room approach at upper-chest camera height, offset from the group's entry path, with 이현우 left, 수빈 center, and 마을총무 right in a staggered medium three-shot. Keep all three faces readable at oblique angles as their interrupted steps and uneven shoulder positions express shock, their eyes lowered toward the meager provisions below the frame rather than toward the camera. Retain the open doorway behind them and a narrow band of empty interior around their bodies, making distance the approach's only emphatic change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 식량창고 내부 (Nearly empty, with only a few rice sacks and one potato sack remaining below the reaction framing) — Seen from within the room, looking obliquely toward the entering group; used as Empty space around the figures supports the scarcity revealed by their downward eyelines; 낡은 철문 (Open) — The open door leaf is seen obliquely behind the group; used as Marks the entrance and preserves the direction of their interrupted movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light keeps all three reactions readable with subdued brightness and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh4__bgfirst_bg.png",
  "asset_id": "dcfb0afa-7db7-4b43-b5f5-5880aecf1588",
  "input_asset_ids": [
   "7f8167b8-7da2-4ac6-88dd-bb51e4640391",
   "317f75ce-9f95-4c6f-a830-9ec3f2d7c95b"
  ]
 },
 "S62sh4": {
  "input_fingerprint": "2df4f6dc5c1431a1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 창고 안을 바라보며 충격받은 표정으로 굳어버린 이현우, 수빈, 마을총무의 구도.\n\nLOCATION (lock): Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the storage-room approach at upper-chest camera height, offset from the group's entry path, with 이현우 left, 수빈 center, and 마을총무 right in a staggered medium three-shot. Keep all three faces readable at oblique angles as their interrupted steps and uneven shoulder positions express shock, their eyes lowered toward the meager provisions below the frame rather than toward the camera. Retain the open doorway behind them and a narrow band of empty interior around their bodies, making distance the approach's only emphatic change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 식량창고 내부 (Nearly empty, with only a few rice sacks and one potato sack remaining below the reaction framing) — Seen from within the room, looking obliquely toward the entering group; used as Empty space around the figures supports the scarcity revealed by their downward eyelines; 낡은 철문 (Open) — The open door leaf is seen obliquely behind the group; used as Marks the entrance and preserves the direction of their interrupted movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light keeps all three reactions readable with subdued brightness and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door is open, revealing only a few sacks of rice and one sack of potatoes. In the village clearing, an old truck is under attempted repair; Charlie remains dented and punctured from the battle. 이현우: He is inside the food store, still bearing the previous day's combat injuries and his earlier treated leg injury. 수빈: She is inside the food store with persistent facial wounds, grime and radiation lesions on her torso. 마을총무: He is inside the newly opened food store.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 창고 안을 바라보며 충격받은 표정으로 굳어버린 이현우, 수빈, 마을총무의 구도.\n\nLOCATION (lock): Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the storage-room approach at upper-chest camera height, offset from the group's entry path, with 이현우 left, 수빈 center, and 마을총무 right in a staggered medium three-shot. Keep all three faces readable at oblique angles as their interrupted steps and uneven shoulder positions express shock, their eyes lowered toward the meager provisions below the frame rather than toward the camera. Retain the open doorway behind them and a narrow band of empty interior around their bodies, making distance the approach's only emphatic change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 식량창고 내부 (Nearly empty, with only a few rice sacks and one potato sack remaining below the reaction framing) — Seen from within the room, looking obliquely toward the entering group; used as Empty space around the figures supports the scarcity revealed by their downward eyelines; 낡은 철문 (Open) — The open door leaf is seen obliquely behind the group; used as Marks the entrance and preserves the direction of their interrupted movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light keeps all three reactions readable with subdued brightness and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door is open, revealing only a few sacks of rice and one sack of potatoes. In the village clearing, an old truck is under attempted repair; Charlie remains dented and punctured from the battle. 이현우: He is inside the food store, still bearing the previous day's combat injuries and his earlier treated leg injury. 수빈: She is inside the food store with persistent facial wounds, grime and radiation lesions on her torso. 마을총무: He is inside the newly opened food store.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 창고 안을 바라보며 충격받은 표정으로 굳어버린 이현우, 수빈, 마을총무의 구도.\n\nLOCATION (lock): Inside the village food storehouse just beyond its opened iron door. Daylight reveals an almost empty space with only a few sacks of rice and potatoes. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the storage-room approach at upper-chest camera height, offset from the group's entry path, with 이현우 left, 수빈 center, and 마을총무 right in a staggered medium three-shot. Keep all three faces readable at oblique angles as their interrupted steps and uneven shoulder positions express shock, their eyes lowered toward the meager provisions below the frame rather than toward the camera. Retain the open doorway behind them and a narrow band of empty interior around their bodies, making distance the approach's only emphatic change.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 식량창고 내부 (Nearly empty, with only a few rice sacks and one potato sack remaining below the reaction framing) — Seen from within the room, looking obliquely toward the entering group; used as Empty space around the figures supports the scarcity revealed by their downward eyelines; 낡은 철문 (Open) — The open door leaf is seen obliquely behind the group; used as Marks the entrance and preserves the direction of their interrupted movement.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light keeps all three reactions readable with subdued brightness and no stylized distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food store's old iron door is open, revealing only a few sacks of rice and one sack of potatoes. In the village clearing, an old truck is under attempted repair; Charlie remains dented and punctured from the battle. 이현우: He is inside the food store, still bearing the previous day's combat injuries and his earlier treated leg injury. 수빈: She is inside the food store with persistent facial wounds, grime and radiation lesions on her torso. 마을총무: He is inside the newly opened food store.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh4__bgfirst_bg.png",
     "asset_id": "dcfb0afa-7db7-4b43-b5f5-5880aecf1588",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S62sh4.png",
     "asset_id": "7f8167b8-7da2-4ac6-88dd-bb51e4640391",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:741399>",
     "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_food_storage_8124be.png",
     "asset_id": "317f75ce-9f95-4c6f-a830-9ec3f2d7c95b",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:741399>",
     "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "세 인물 모두 프레임 내부 하단의 감자와 쌀포대를 향해 시선을 고정함.",
    "built_space": "창고 내부. 인물들 뒤로 열린 철문과 외부의 급수탑, 트럭이 보임.",
    "entities": "왼쪽부터 이현우, 수빈, 마을총무 순으로 위치하며 지정된 인상착의와 복장 요소를 갖춤.",
    "hard_violations": [],
    "physics": "세 사람 모두 바닥에 서서 체중을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "세 인물 모두 화면에 보이지 않는 프레임 아래쪽을 향해 시선을 떨구고 있음.",
    "built_space": "창고 내부. 뒤편의 열린 철문 밖으로 급수탑과 트럭 등 배경 요소가 올바르게 배치됨.",
    "entities": "세 인물의 엇갈린 배치, 인상착의, 더러워진 복장 등이 설정에 잘 부합함.",
    "hard_violations": [],
    "physics": "세 인물 모두 바닥을 딛고 자연스럽게 멈춰 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 미디엄 샷 구도를 유지하고, 식량을 프레임 아래로 배제한 점이 프롬프트와 가장 잘 일치함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레임 바깥에 있어야 할 식량을 화면에 담기 위해 샷을 과도하게 넓혀 미디엄 샷 구도 지시를 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 프레임 내부 하단의 감자와 쌀포대를 향해 시선을 고정함.",
        "built_space": "창고 내부. 인물들 뒤로 열린 철문과 외부의 급수탑, 트럭이 보임.",
        "entities": "왼쪽부터 이현우, 수빈, 마을총무 순으로 위치하며 지정된 인상착의와 복장 요소를 갖춤.",
        "hard_violations": [],
        "physics": "세 사람 모두 바닥에 서서 체중을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 화면에 보이지 않는 프레임 아래쪽을 향해 시선을 떨구고 있음.",
        "built_space": "창고 내부. 뒤편의 열린 철문 밖으로 급수탑과 트럭 등 배경 요소가 올바르게 배치됨.",
        "entities": "세 인물의 엇갈린 배치, 인상착의, 더러워진 복장 등이 설정에 잘 부합함.",
        "hard_violations": [],
        "physics": "세 인물 모두 바닥을 딛고 자연스럽게 멈춰 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 미디엄 샷 구도를 유지하고, 식량을 프레임 아래로 배제한 점이 프롬프트와 가장 잘 일치함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레임 바깥에 있어야 할 식량을 화면에 담기 위해 샷을 과도하게 넓혀 미디엄 샷 구도 지시를 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 프레임 내부 하단의 감자와 쌀포대를 향해 시선을 고정함.",
        "built_space": "창고 내부. 인물들 뒤로 열린 철문과 외부의 급수탑, 트럭이 보임.",
        "entities": "왼쪽부터 이현우, 수빈, 마을총무 순으로 위치하며 지정된 인상착의와 복장 요소를 갖춤.",
        "hard_violations": [],
        "physics": "세 사람 모두 바닥에 서서 체중을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 화면에 보이지 않는 프레임 아래쪽을 향해 시선을 떨구고 있음.",
        "built_space": "창고 내부. 뒤편의 열린 철문 밖으로 급수탑과 트럭 등 배경 요소가 올바르게 배치됨.",
        "entities": "세 인물의 엇갈린 배치, 인상착의, 더러워진 복장 등이 설정에 잘 부합함.",
        "hard_violations": [],
        "physics": "세 인물 모두 바닥을 딛고 자연스럽게 멈춰 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "좌·중·우 배치와 엇갈린 깊이의 미디엄 3인숏, 화면 아래로 향한 시선, 열린 철문을 잘 지켰으며, 걸음을 멈춘 비대칭 자세의 표현만 다소 약하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물과 장소 및 하향 시선은 맞지만, 무릎 부근까지 넓혀 전경의 식량 포대를 보여주므로 식량을 화면 아래에 두는 미디엄 반응숏 지시에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 오른쪽 아래, 수빈은 정면 아래, 총무는 왼쪽 아래의 화면 밖 공간을 바라본다. 세 시선 모두 카메라가 아니라 앞쪽 낮은 식량 위치를 향하는 것으로 읽힌다. 몸은 뒤의 출입구에서 실내로 들어온 방향이며, 무기나 겨누는 물체는 없다.",
        "built_space": "실내에서 출입구를 비스듬히 바라본다. 뒤쪽 왼편에 출입구 하나와 열린 철문 문짝 하나, 위쪽에 매달린 조명 하나, 오른쪽 끝에 금속 선반 하나가 보인다. 낡은 벽체와 목재 기둥·보, 바깥 급수탑과 트럭이 장소 참조와 일치한다. 이현우는 왼쪽 가까이, 수빈은 중앙, 총무는 오른쪽에 있으며 어깨 높이와 깊이가 엇갈린다. 허리에서 골반 부근을 자르는 구도로 인물 주변 실내는 좁게 남는다. 오른쪽 선반의 흰 포대 일부는 보이지만 주요 식량은 화면 아래에 있다.",
        "entities": "인물은 지정된 세 명뿐이다. 이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 짧은 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠 및 귀의 인이어를 갖춘다. 수빈은 젊은 동아시아계 여성으로 검은 단발, 올리브색 실용 재킷, 장갑과 허리 장비가 참조에 부합하며 얼굴 상처와 오염이 보인다. 총무는 중년 동아시아계 남성으로 참조의 모자, 붉은 체크 셔츠, 작업 바지와 얼굴 특징을 따른다. 국적은 외양만으로 확인할 수 없다. 다리 부상과 옷 아래 몸통 병변, 감자 포대는 이 구도에서 확인할 수 없다.",
        "hard_violations": [],
        "physics": "세 사람의 하체는 화면 아래로 이어지는 자연스러운 직립 자세이며 부유하거나 지지 없이 기운 모습은 없다. 발은 잘려 있어 실제 접지는 확인되지 않는다. 팔은 몸 옆에 있고 수빈의 장비는 허리띠에, 인이어는 귀에 고정되어 있다. 철문은 문틀의 경첩으로, 선반의 포대는 선반으로 지지된다. 놀라 멈춘 자세는 가능하지만 팔과 몸의 비대칭은 비교적 약하다."
       },
       {
        "label": "B",
        "direction": "이현우와 수빈은 오른쪽 아래의 전경 식량을 보고, 총무도 고개와 눈을 낮춰 바로 앞 쌀 포대 쪽을 바라본다. 카메라를 보는 인물은 없다. 시선의 실제 목표는 명확하지만, 그 목표를 화면 밖에 두라는 지시와 달리 포대들이 화면 안에 크게 들어온다.",
        "built_space": "뒤쪽 왼편의 출입구 하나, 열린 철문 문짝 하나, 천장 조명 하나, 오른쪽 금속 선반 하나가 보이며 벽과 목재 구조도 장소 참조에 부합한다. 세 사람의 좌·중·우 순서는 맞는다. 그러나 무릎 부근까지 보이는 넓은 구도로 바닥과 천장 면적이 늘었고, 전경에 쌀 포대 두 개와 열린 감자 포대 하나가 나타난다. 선반에도 흰 포대가 겹쳐 보인다. 요청한 좁은 주변 공간의 미디엄 반응숏보다 장소 참조의 넓은 구도에 가깝다.",
        "entities": "지정된 세 사람만 등장한다. 이현우의 젊은 얼굴, 검은 헝클어진 머리, 인이어, 오염된 어두운 셔츠가 맞는다. 수빈의 젊은 얼굴, 검은 단발, 올리브색 재킷과 카고 바지, 장갑 및 얼굴 상처도 부합한다. 총무는 참조와 같은 중년 남성의 얼굴, 모자, 붉은 체크 셔츠와 낡은 작업 바지를 착용한다. 세 사람은 참조와 유사한 동아시아계 외양이며 국적 자체는 확인할 수 없다. 전경의 열린 포대에는 실제 감자들이 보이고 흰 포대는 쌀 포대로 읽히지만 내용물은 드러나지 않는다.",
        "hard_violations": [],
        "physics": "이현우와 총무가 상체를 앞으로 기울인 자세는 화면 아래로 이어지는 다리로 지탱할 수 있는 범위다. 발은 보이지 않지만 부유의 징후는 없다. 수빈은 한쪽 팔을 몸통 앞에 접고 다른 손을 입에 대어 손과 팔의 지지가 자연스럽다. 다만 이 동작은 걸음이 끊긴 순간보다는 이미 충격에 반응한 자세로 읽힌다. 전경 포대는 바닥과 하단 받침 위에 놓여 있고 감자는 포대 안에 담겨 있다. 문과 조명, 선반에도 정상적인 구조적 지지가 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "좌·중·우 배치와 엇갈린 깊이의 미디엄 3인숏, 화면 아래로 향한 시선, 열린 철문을 잘 지켰으며, 걸음을 멈춘 비대칭 자세의 표현만 다소 약하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물과 장소 및 하향 시선은 맞지만, 무릎 부근까지 넓혀 전경의 식량 포대를 보여주므로 식량을 화면 아래에 두는 미디엄 반응숏 지시에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 오른쪽 아래, 수빈은 정면 아래, 총무는 왼쪽 아래의 화면 밖 공간을 바라본다. 세 시선 모두 카메라가 아니라 앞쪽 낮은 식량 위치를 향하는 것으로 읽힌다. 몸은 뒤의 출입구에서 실내로 들어온 방향이며, 무기나 겨누는 물체는 없다.",
        "built_space": "실내에서 출입구를 비스듬히 바라본다. 뒤쪽 왼편에 출입구 하나와 열린 철문 문짝 하나, 위쪽에 매달린 조명 하나, 오른쪽 끝에 금속 선반 하나가 보인다. 낡은 벽체와 목재 기둥·보, 바깥 급수탑과 트럭이 장소 참조와 일치한다. 이현우는 왼쪽 가까이, 수빈은 중앙, 총무는 오른쪽에 있으며 어깨 높이와 깊이가 엇갈린다. 허리에서 골반 부근을 자르는 구도로 인물 주변 실내는 좁게 남는다. 오른쪽 선반의 흰 포대 일부는 보이지만 주요 식량은 화면 아래에 있다.",
        "entities": "인물은 지정된 세 명뿐이다. 이현우는 젊은 동아시아계 남성의 얼굴, 헝클어진 짧은 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠 및 귀의 인이어를 갖춘다. 수빈은 젊은 동아시아계 여성으로 검은 단발, 올리브색 실용 재킷, 장갑과 허리 장비가 참조에 부합하며 얼굴 상처와 오염이 보인다. 총무는 중년 동아시아계 남성으로 참조의 모자, 붉은 체크 셔츠, 작업 바지와 얼굴 특징을 따른다. 국적은 외양만으로 확인할 수 없다. 다리 부상과 옷 아래 몸통 병변, 감자 포대는 이 구도에서 확인할 수 없다.",
        "hard_violations": [],
        "physics": "세 사람의 하체는 화면 아래로 이어지는 자연스러운 직립 자세이며 부유하거나 지지 없이 기운 모습은 없다. 발은 잘려 있어 실제 접지는 확인되지 않는다. 팔은 몸 옆에 있고 수빈의 장비는 허리띠에, 인이어는 귀에 고정되어 있다. 철문은 문틀의 경첩으로, 선반의 포대는 선반으로 지지된다. 놀라 멈춘 자세는 가능하지만 팔과 몸의 비대칭은 비교적 약하다."
       },
       {
        "label": "A",
        "direction": "이현우와 수빈은 오른쪽 아래의 전경 식량을 보고, 총무도 고개와 눈을 낮춰 바로 앞 쌀 포대 쪽을 바라본다. 카메라를 보는 인물은 없다. 시선의 실제 목표는 명확하지만, 그 목표를 화면 밖에 두라는 지시와 달리 포대들이 화면 안에 크게 들어온다.",
        "built_space": "뒤쪽 왼편의 출입구 하나, 열린 철문 문짝 하나, 천장 조명 하나, 오른쪽 금속 선반 하나가 보이며 벽과 목재 구조도 장소 참조에 부합한다. 세 사람의 좌·중·우 순서는 맞는다. 그러나 무릎 부근까지 보이는 넓은 구도로 바닥과 천장 면적이 늘었고, 전경에 쌀 포대 두 개와 열린 감자 포대 하나가 나타난다. 선반에도 흰 포대가 겹쳐 보인다. 요청한 좁은 주변 공간의 미디엄 반응숏보다 장소 참조의 넓은 구도에 가깝다.",
        "entities": "지정된 세 사람만 등장한다. 이현우의 젊은 얼굴, 검은 헝클어진 머리, 인이어, 오염된 어두운 셔츠가 맞는다. 수빈의 젊은 얼굴, 검은 단발, 올리브색 재킷과 카고 바지, 장갑 및 얼굴 상처도 부합한다. 총무는 참조와 같은 중년 남성의 얼굴, 모자, 붉은 체크 셔츠와 낡은 작업 바지를 착용한다. 세 사람은 참조와 유사한 동아시아계 외양이며 국적 자체는 확인할 수 없다. 전경의 열린 포대에는 실제 감자들이 보이고 흰 포대는 쌀 포대로 읽히지만 내용물은 드러나지 않는다.",
        "hard_violations": [],
        "physics": "이현우와 총무가 상체를 앞으로 기울인 자세는 화면 아래로 이어지는 다리로 지탱할 수 있는 범위다. 발은 보이지 않지만 부유의 징후는 없다. 수빈은 한쪽 팔을 몸통 앞에 접고 다른 손을 입에 대어 손과 팔의 지지가 자연스럽다. 다만 이 동작은 걸음이 끊긴 순간보다는 이미 충격에 반응한 자세로 읽힌다. 전경 포대는 바닥과 하단 받침 위에 놓여 있고 감자는 포대 안에 담겨 있다. 문과 조명, 선반에도 정상적인 구조적 지지가 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.238,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.238,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1238
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 미디엄 샷 구도를 유지하고, 식량을 프레임 아래로 배제한 점이 프롬프트와 가장 잘 일치함."
   },
   {
    "label": "A",
    "score": 1238,
    "verdict_ko": "프레임 바깥에 있어야 할 식량을 화면에 담기 위해 샷을 과도하게 넓혀 미디엄 샷 구도 지시를 위반함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_food_storage_8124be.png",
    "asset_id": "317f75ce-9f95-4c6f-a830-9ec3f2d7c95b",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:741399>",
    "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c83-abef-7cef-bc3b-aeef55cbae7b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh4__bgfirst_bg.png",
   "bg_asset_id": "dcfb0afa-7db7-4b43-b5f5-5880aecf1588",
   "bg_record_key": "S62sh4::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "food_storage",
   "groupbg_asset_id": "317f75ce-9f95-4c6f-a830-9ec3f2d7c95b"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S62sh17::signage": {
  "fp": "78047226bee7da96",
  "inscriptions": [
   {
    "text_native": "마트",
    "source": "scene_text_quoted",
    "reason_ko": "찰리의 1인칭 시점 HUD 지형도 그래픽에 붉게 점등된 목표 지점 라벨이다.",
    "source_quote": "마트"
   },
   {
    "text_native": "25.6km",
    "source": "scene_text_quoted",
    "reason_ko": "찰리의 내비게이션 그래픽 화면에 표시된 목표 지점까지의 거리 수치이다.",
    "source_quote": "25.6km"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S62sh17": {
  "input_fingerprint": "e85241fdb49dc52a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 1인칭 시점 화면(POV) 속 주변 지형도 그래픽 위에 서남쪽 25.6km 지점의 마트 위치가 붉게 켜진 구도.\n\nLOCATION (lock): In the village's outdoor communal clearing beside the old truck; the robot's navigation display overlays this location rather than placing it at the distant store. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static first-person navigation image aligned with 찰리's visual field at his viewing height, with no external screen angle, device housing, or visible body. Present the terrain graphic as an internal image plane, placing the current-position reference near center and the red supermarket marker toward the lower left so the southwest relationship and the distance label '25.6 km' read immediately. Keep the marker small relative to the surrounding terrain, allowing the map rather than an oversized icon to carry the spatial information.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: current-position reference in the middle-center of the frame, background; red supermarket marker, southwest, 25.6 km in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주변 지형도 그래픽 (Active in 찰리's first-person navigation view) — The graphic is presented directly in the image plane, with the current position near center and the southwest supermarket location marked in red beside the distance label; used as Supplies geographic context without a physical monitor or external observer viewpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained graphic luminance with a localized red destination highlight, clearly separating the internal navigation view from photographed daylight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's navigation display identifies a department store and market 25.6 km to the southwest, while his body retains its battle dents and holes. The old truck remains in the clearing under repair, with a worn football in use nearby.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"마트\"\n- \"25.6km\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 1인칭 시점 화면(POV) 속 주변 지형도 그래픽 위에 서남쪽 25.6km 지점의 마트 위치가 붉게 켜진 구도.\n\nLOCATION (lock): In the village's outdoor communal clearing beside the old truck; the robot's navigation display overlays this location rather than placing it at the distant store. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static first-person navigation image aligned with 찰리's visual field at his viewing height, with no external screen angle, device housing, or visible body. Present the terrain graphic as an internal image plane, placing the current-position reference near center and the red supermarket marker toward the lower left so the southwest relationship and the distance label '25.6 km' read immediately. Keep the marker small relative to the surrounding terrain, allowing the map rather than an oversized icon to carry the spatial information.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: current-position reference in the middle-center of the frame, background; red supermarket marker, southwest, 25.6 km in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주변 지형도 그래픽 (Active in 찰리's first-person navigation view) — The graphic is presented directly in the image plane, with the current position near center and the southwest supermarket location marked in red beside the distance label; used as Supplies geographic context without a physical monitor or external observer viewpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained graphic luminance with a localized red destination highlight, clearly separating the internal navigation view from photographed daylight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's navigation display identifies a department store and market 25.6 km to the southwest, while his body retains its battle dents and holes. The old truck remains in the clearing under repair, with a worn football in use nearby.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"마트\"\n- \"25.6km\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 1인칭 시점 화면(POV) 속 주변 지형도 그래픽 위에 서남쪽 25.6km 지점의 마트 위치가 붉게 켜진 구도.\n\nLOCATION (lock): In the village's outdoor communal clearing beside the old truck; the robot's navigation display overlays this location rather than placing it at the distant store. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static first-person navigation image aligned with 찰리's visual field at his viewing height, with no external screen angle, device housing, or visible body. Present the terrain graphic as an internal image plane, placing the current-position reference near center and the red supermarket marker toward the lower left so the southwest relationship and the distance label '25.6 km' read immediately. Keep the marker small relative to the surrounding terrain, allowing the map rather than an oversized icon to carry the spatial information.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: current-position reference in the middle-center of the frame, background; red supermarket marker, southwest, 25.6 km in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주변 지형도 그래픽 (Active in 찰리's first-person navigation view) — The graphic is presented directly in the image plane, with the current position near center and the southwest supermarket location marked in red beside the distance label; used as Supplies geographic context without a physical monitor or external observer viewpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained graphic luminance with a localized red destination highlight, clearly separating the internal navigation view from photographed daylight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's navigation display identifies a department store and market 25.6 km to the southwest, while his body retains its battle dents and holes. The old truck remains in the clearing under repair, with a worn football in use nearby.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"마트\"\n- \"25.6km\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 중앙의 현재 위치 마커에서 남서쪽 하단으로 점선이 직선으로 이어져 붉은색 타겟 마커를 정확히 가리킴.",
    "built_space": "실내 구조물이 없으며, 상공에서 내려다본 산, 강, 마을 등의 지형도 배경 위에 그래픽 인터페이스가 오버레이됨.",
    "entities": "프롬프트가 요구한 '마트'와 '25.6km' 텍스트가 남서쪽 마커 옆에 정확히 렌더링되었으며, 불필요한 추가 텍스트나 인물이 등장하지 않음.",
    "hard_violations": [],
    "physics": "물리적 객체가 아닌 1인칭 시점의 시각적 HUD 그래픽이므로 물리 법칙이나 지지대 분석이 적용되지 않음."
   },
   {
    "label": "B",
    "direction": "중앙부 위치 마커에서 구불구불한 붉은 점선 경로가 남서쪽 하단의 붉은 마커를 향해 이어짐.",
    "built_space": "상공에서 내려다본 지형도(산, 강, 농경지 등) 배경 위에 내비게이션 그래픽, 나침반, 미니맵 등이 배치됨.",
    "entities": "'마트'와 '25.6km' 텍스트가 지정된 위치에 렌더링되었으나, 좌측 상단 나침반 그래픽에 지시되지 않은 영어 알파벳(N, W, S, E)이 포함됨.",
    "hard_violations": [
     "[gemini-pro] 지시되지 않은 추가 텍스트(N, W, S, E)가 나침반에 삽입됨 (어떤 읽을 수 있는 텍스트도 추가하지 말라는 지시 위반)",
     "[gpt-high] 허용된 두 문구 외에 나침반의 'N', 'W', 'E', 'S'와 눈금의 '0'을 추가하여, 다른 읽을 수 있는 문자를 금지한 지시를 위반했다."
    ],
    "physics": "1인칭 시점 화면 내의 그래픽 오버레이 형태이므로 중력이나 지지대에 관한 물리적 분석은 해당 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 남서쪽 위치에 붉은색 마커와 정확한 텍스트('마트', '25.6km')만을 깔끔하게 배치하여 프롬프트의 지시사항을 완벽히 수행함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 텍스트를 잘 표현했으나, 좌측 상단 나침반 그래픽에 프롬프트에 없는 알파벳(N, W, S, E)을 추가하여 텍스트 제한 규정을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 중앙의 현재 위치 마커에서 남서쪽 하단으로 점선이 직선으로 이어져 붉은색 타겟 마커를 정확히 가리킴.",
        "built_space": "실내 구조물이 없으며, 상공에서 내려다본 산, 강, 마을 등의 지형도 배경 위에 그래픽 인터페이스가 오버레이됨.",
        "entities": "프롬프트가 요구한 '마트'와 '25.6km' 텍스트가 남서쪽 마커 옆에 정확히 렌더링되었으며, 불필요한 추가 텍스트나 인물이 등장하지 않음.",
        "hard_violations": [],
        "physics": "물리적 객체가 아닌 1인칭 시점의 시각적 HUD 그래픽이므로 물리 법칙이나 지지대 분석이 적용되지 않음."
       },
       {
        "label": "B",
        "direction": "중앙부 위치 마커에서 구불구불한 붉은 점선 경로가 남서쪽 하단의 붉은 마커를 향해 이어짐.",
        "built_space": "상공에서 내려다본 지형도(산, 강, 농경지 등) 배경 위에 내비게이션 그래픽, 나침반, 미니맵 등이 배치됨.",
        "entities": "'마트'와 '25.6km' 텍스트가 지정된 위치에 렌더링되었으나, 좌측 상단 나침반 그래픽에 지시되지 않은 영어 알파벳(N, W, S, E)이 포함됨.",
        "hard_violations": [
         "지시되지 않은 추가 텍스트(N, W, S, E)가 나침반에 삽입됨 (어떤 읽을 수 있는 텍스트도 추가하지 말라는 지시 위반)"
        ],
        "physics": "1인칭 시점 화면 내의 그래픽 오버레이 형태이므로 중력이나 지지대에 관한 물리적 분석은 해당 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 남서쪽 위치에 붉은색 마커와 정확한 텍스트('마트', '25.6km')만을 깔끔하게 배치하여 프롬프트의 지시사항을 완벽히 수행함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 텍스트를 잘 표현했으나, 좌측 상단 나침반 그래픽에 프롬프트에 없는 알파벳(N, W, S, E)을 추가하여 텍스트 제한 규정을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 중앙의 현재 위치 마커에서 남서쪽 하단으로 점선이 직선으로 이어져 붉은색 타겟 마커를 정확히 가리킴.",
        "built_space": "실내 구조물이 없으며, 상공에서 내려다본 산, 강, 마을 등의 지형도 배경 위에 그래픽 인터페이스가 오버레이됨.",
        "entities": "프롬프트가 요구한 '마트'와 '25.6km' 텍스트가 남서쪽 마커 옆에 정확히 렌더링되었으며, 불필요한 추가 텍스트나 인물이 등장하지 않음.",
        "hard_violations": [],
        "physics": "물리적 객체가 아닌 1인칭 시점의 시각적 HUD 그래픽이므로 물리 법칙이나 지지대 분석이 적용되지 않음."
       },
       {
        "label": "B",
        "direction": "중앙부 위치 마커에서 구불구불한 붉은 점선 경로가 남서쪽 하단의 붉은 마커를 향해 이어짐.",
        "built_space": "상공에서 내려다본 지형도(산, 강, 농경지 등) 배경 위에 내비게이션 그래픽, 나침반, 미니맵 등이 배치됨.",
        "entities": "'마트'와 '25.6km' 텍스트가 지정된 위치에 렌더링되었으나, 좌측 상단 나침반 그래픽에 지시되지 않은 영어 알파벳(N, W, S, E)이 포함됨.",
        "hard_violations": [
         "지시되지 않은 추가 텍스트(N, W, S, E)가 나침반에 삽입됨 (어떤 읽을 수 있는 텍스트도 추가하지 말라는 지시 위반)"
        ],
        "physics": "1인칭 시점 화면 내의 그래픽 오버레이 형태이므로 중력이나 지지대에 관한 물리적 분석은 해당 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "중앙 현재 위치와 좌하단의 작은 붉은 마트 표시는 맞지만, 허용되지 않은 나침반 영문과 눈금 숫자를 추가해 문자 제한을 위반한다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "외부 기기나 신체 없이 내부 지형도를 채우고, 중앙 현재 위치에서 좌하단의 작은 붉은 마트 표식까지의 관계와 지정 문구를 명확히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현재 위치를 나타내는 청백색 원이 화면 중앙 부근에 있고, 붉은 마트 점은 그 왼쪽 아래에 있다. 위쪽을 북쪽으로 표시한 나침반에 따라 목적지는 서남쪽으로 읽힌다. 굽은 붉은 점선은 중앙 위치 부근에서 마트 표식까지 이어진다. 사람의 시선이나 무기는 없다.",
        "built_space": "산림, 도로, 경작지, 하천과 건물군을 담은 지형도가 화면 전체를 채운다. 중앙 위치 표식 하나, 목적지 표식 하나, 좌상단 나침반 하나, 우상단 보조 지도 하나가 보인다. 모니터 외함이나 외부에서 화면을 보는 각도는 없다. 실제 공터와 트럭은 직접 보이지 않으며, 내부 지도 인서트라는 구도에서 이들의 부재는 결함이 아니다.",
        "entities": "목적지 옆에 '마트'와 '25.6km'가 붉게 표시되어 있고 목적지 점은 주변 지형에 비해 작다. 다만 나침반의 'N', 'W', 'E', 'S'와 좌하단 눈금의 '0'이 추가로 읽힌다. 찰리의 신체나 다른 인물은 보이지 않아 요구된 무신체 일인칭 구도와 맞는다. 별도 참조 이미지는 제공되지 않았다.",
        "hard_violations": [
         "허용된 두 문구 외에 나침반의 'N', 'W', 'E', 'S'와 눈금의 '0'을 추가하여, 다른 읽을 수 있는 문자를 금지한 지시를 위반했다."
        ],
        "physics": "떠 있는 신체나 손에 들린 물체는 없다. 발광점과 선은 명시적으로 요구된 내부 내비게이션 그래픽이므로 물리적 지지물이 필요한 공중 물체가 아니다. 지도 속 건물과 도로는 지형 위에 놓여 있으며, 불가능한 반사나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "청백색 현재 위치 원은 중앙 부근에 있고, 작은 붉은 목적지 원은 왼쪽 아래에 있다. 두 표식을 연결하는 대각선 점선이 중앙에서 좌하단 마트를 향해 이어져 서남쪽 관계를 즉시 읽을 수 있다. 별도의 시선, 무기 또는 이동하는 몸은 없다.",
        "built_space": "산지, 하천, 도로, 농경지와 건물군의 지형 영상이 화면을 가득 채운다. 현재 위치 표식 하나와 목적지 표식 하나가 있으며, 격자와 가장자리 곡선 눈금이 내부 내비게이션 영상으로 읽힌다. 외부 화면 외함, 관찰자 또는 실제 매장으로 이동한 장면은 없다. 공터의 실제 설비 수나 배치는 지도 인서트만으로 판별할 수 없다.",
        "entities": "좌하단 목적지에 '마트'와 '25.6km'가 정확히 표시되어 있다. 목적지는 작은 붉은 점과 원으로 강조되고, 문구와 연결선은 절제된 청백색이다. 그 밖의 읽을 수 있는 문자는 보이지 않는다. 찰리와 장면 수준에서 언급된 다른 사람들은 노출되지 않아 지정된 프레이밍을 유지한다.",
        "hard_violations": [],
        "physics": "공중에 놓인 실물이나 움직이는 신체가 없다. 위치 표식과 점선은 요구된 내부 영상의 구성 요소이며 실물처럼 떠 있는 물체가 아니다. 지형 속 도로와 건물은 지면에 놓여 있고, 물리적으로 불가능한 자세나 반사는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "중앙 현재 위치와 좌하단의 작은 붉은 마트 표시는 맞지만, 허용되지 않은 나침반 영문과 눈금 숫자를 추가해 문자 제한을 위반한다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "외부 기기나 신체 없이 내부 지형도를 채우고, 중앙 현재 위치에서 좌하단의 작은 붉은 마트 표식까지의 관계와 지정 문구를 명확히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현재 위치를 나타내는 청백색 원이 화면 중앙 부근에 있고, 붉은 마트 점은 그 왼쪽 아래에 있다. 위쪽을 북쪽으로 표시한 나침반에 따라 목적지는 서남쪽으로 읽힌다. 굽은 붉은 점선은 중앙 위치 부근에서 마트 표식까지 이어진다. 사람의 시선이나 무기는 없다.",
        "built_space": "산림, 도로, 경작지, 하천과 건물군을 담은 지형도가 화면 전체를 채운다. 중앙 위치 표식 하나, 목적지 표식 하나, 좌상단 나침반 하나, 우상단 보조 지도 하나가 보인다. 모니터 외함이나 외부에서 화면을 보는 각도는 없다. 실제 공터와 트럭은 직접 보이지 않으며, 내부 지도 인서트라는 구도에서 이들의 부재는 결함이 아니다.",
        "entities": "목적지 옆에 '마트'와 '25.6km'가 붉게 표시되어 있고 목적지 점은 주변 지형에 비해 작다. 다만 나침반의 'N', 'W', 'E', 'S'와 좌하단 눈금의 '0'이 추가로 읽힌다. 찰리의 신체나 다른 인물은 보이지 않아 요구된 무신체 일인칭 구도와 맞는다. 별도 참조 이미지는 제공되지 않았다.",
        "hard_violations": [
         "허용된 두 문구 외에 나침반의 'N', 'W', 'E', 'S'와 눈금의 '0'을 추가하여, 다른 읽을 수 있는 문자를 금지한 지시를 위반했다."
        ],
        "physics": "떠 있는 신체나 손에 들린 물체는 없다. 발광점과 선은 명시적으로 요구된 내부 내비게이션 그래픽이므로 물리적 지지물이 필요한 공중 물체가 아니다. 지도 속 건물과 도로는 지형 위에 놓여 있으며, 불가능한 반사나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "청백색 현재 위치 원은 중앙 부근에 있고, 작은 붉은 목적지 원은 왼쪽 아래에 있다. 두 표식을 연결하는 대각선 점선이 중앙에서 좌하단 마트를 향해 이어져 서남쪽 관계를 즉시 읽을 수 있다. 별도의 시선, 무기 또는 이동하는 몸은 없다.",
        "built_space": "산지, 하천, 도로, 농경지와 건물군의 지형 영상이 화면을 가득 채운다. 현재 위치 표식 하나와 목적지 표식 하나가 있으며, 격자와 가장자리 곡선 눈금이 내부 내비게이션 영상으로 읽힌다. 외부 화면 외함, 관찰자 또는 실제 매장으로 이동한 장면은 없다. 공터의 실제 설비 수나 배치는 지도 인서트만으로 판별할 수 없다.",
        "entities": "좌하단 목적지에 '마트'와 '25.6km'가 정확히 표시되어 있다. 목적지는 작은 붉은 점과 원으로 강조되고, 문구와 연결선은 절제된 청백색이다. 그 밖의 읽을 수 있는 문자는 보이지 않는다. 찰리와 장면 수준에서 언급된 다른 사람들은 노출되지 않아 지정된 프레이밍을 유지한다.",
        "hard_violations": [],
        "physics": "공중에 놓인 실물이나 움직이는 신체가 없다. 위치 표식과 점선은 요구된 내부 영상의 구성 요소이며 실물처럼 떠 있는 물체가 아니다. 지형 속 도로와 건물은 지면에 놓여 있고, 물리적으로 불가능한 자세나 반사는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.111
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.861
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시되지 않은 추가 텍스트(N, W, S, E)가 나침반에 삽입됨 (어떤 읽을 수 있는 텍스트도 추가하지 말라는 지시 위반)",
     "[gpt-high] 허용된 두 문구 외에 나침반의 'N', 'W', 'E', 'S'와 눈금의 '0'을 추가하여, 다른 읽을 수 있는 문자를 금지한 지시를 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 861
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 남서쪽 위치에 붉은색 마커와 정확한 텍스트('마트', '25.6km')만을 깔끔하게 배치하여 프롬프트의 지시사항을 완벽히 수행함."
   },
   {
    "label": "B",
    "score": 861,
    "verdict_ko": "지정된 텍스트를 잘 표현했으나, 좌측 상단 나침반 그래픽에 프롬프트에 없는 알파벳(N, W, S, E)을 추가하여 텍스트 제한 규정을 위반함.  ★위반: [gemini-pro] 지시되지 않은 추가 텍스트(N, W, S, E)가 나침반에 삽입됨 (어떤 읽을 수 있는 텍스트도 추가하지 말라는 지시 위반) / [gpt-high] 허용된 두 문구 외에 나침반의 'N', 'W', 'E', 'S'와 눈금의 '0'을 추가하여, 다른 읽을 수 있는 문자를 금지한 지시를 위반했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c8c-bc43-7e74-994b-093248b27e33",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S62sh19::signage": {
  "fp": "ce817723938c1295",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::village_repair_clearing": {
  "input_fingerprint": "8154004f41892e6d",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "village_repair_clearing",
    "tags": [
     "S62sh19"
    ]
   },
   "context_sig": "e9a1139ea6677daf"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the village communal clearing beside the old truck being repaired.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 마을 공터\n- 한편, 공터에서 낡은 트럭의 시동을 부르릉 부르릉 걸고 있는 앰버..\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the village communal clearing beside the old truck being repaired.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 마을 공터\n- 한편, 공터에서 낡은 트럭의 시동을 부르릉 부르릉 걸고 있는 앰버..\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_repair_clearing_b31b85.png",
  "asset_id": "87ef2d30-fa84-40f3-8625-fbc7845f6736",
  "input_asset_ids": [
   "ed980390-e6f6-4352-bc2a-816f460b25a8"
  ],
  "origin_tag": "S62sh19",
  "place_text": "In the village communal clearing beside the old truck being repaired.",
  "origin_inputs": {
   "place_text": "In the village communal clearing beside the old truck being repaired.",
   "time_of_day_en": "day",
   "conti_asset_id": "ed980390-e6f6-4352-bc2a-816f460b25a8"
  }
 },
 "S62sh19::bgfirst_bg": {
  "input_fingerprint": "8f079e528efdc2de",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 극구 반대하며 두 손을 거세게 엑스(X) 자로 교차한 마을총무의 다급한 상반신.\n\nLOCATION (lock): In the village communal clearing beside the old truck being repaired.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the short lateral track beside 이현우 at shoulder height, settling on 마을총무 in an oblique upper-body composition without crossing the conversation axis. Place the administrator slightly right of center and retain both elbows and the complete X of his forearms below his face as he leans into his urgent refusal. His eyes and crossed-arm appeal address 이현우 and 수빈 just outside the left edge, while the clearing remains a quiet background rather than a new focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마을 공터 (The group remains gathered here during the discussion); used as Provides unobtrusive spatial continuity behind the fully visible crossed-arm gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral daytime illumination and controlled contrast, letting the forceful gesture rather than a lighting change convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 극구 반대하며 두 손을 거세게 엑스(X) 자로 교차한 마을총무의 다급한 상반신.\n\nLOCATION (lock): In the village communal clearing beside the old truck being repaired.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the short lateral track beside 이현우 at shoulder height, settling on 마을총무 in an oblique upper-body composition without crossing the conversation axis. Place the administrator slightly right of center and retain both elbows and the complete X of his forearms below his face as he leans into his urgent refusal. His eyes and crossed-arm appeal address 이현우 and 수빈 just outside the left edge, while the clearing remains a quiet background rather than a new focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마을 공터 (The group remains gathered here during the discussion); used as Provides unobtrusive spatial continuity behind the fully visible crossed-arm gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral daytime illumination and controlled contrast, letting the forceful gesture rather than a lighting change convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh19__bgfirst_bg.png",
  "asset_id": "cea8326e-546a-4d42-8338-7e8292087943",
  "input_asset_ids": [
   "ed980390-e6f6-4352-bc2a-816f460b25a8",
   "87ef2d30-fa84-40f3-8625-fbc7845f6736"
  ]
 },
 "S62sh19": {
  "input_fingerprint": "3e047858835dcce2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 극구 반대하며 두 손을 거세게 엑스(X) 자로 교차한 마을총무의 다급한 상반신.\n\nLOCATION (lock): In the village communal clearing beside the old truck being repaired. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the short lateral track beside 이현우 at shoulder height, settling on 마을총무 in an oblique upper-body composition without crossing the conversation axis. Place the administrator slightly right of center and retain both elbows and the complete X of his forearms below his face as he leans into his urgent refusal. His eyes and crossed-arm appeal address 이현우 and 수빈 just outside the left edge, while the clearing remains a quiet background rather than a new focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마을 공터 (The group remains gathered here during the discussion); used as Provides unobtrusive spatial continuity behind the fully visible crossed-arm gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral daytime illumination and controlled contrast, letting the forceful gesture rather than a lighting change convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the village clearing under repair, and the worn football remains in use nearby. Charlie retains his dented, punctured body and malfunctioning sensors. 마을총무: He remains in the clearing during the discussion of the supply trip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 극구 반대하며 두 손을 거세게 엑스(X) 자로 교차한 마을총무의 다급한 상반신.\n\nLOCATION (lock): In the village communal clearing beside the old truck being repaired. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the short lateral track beside 이현우 at shoulder height, settling on 마을총무 in an oblique upper-body composition without crossing the conversation axis. Place the administrator slightly right of center and retain both elbows and the complete X of his forearms below his face as he leans into his urgent refusal. His eyes and crossed-arm appeal address 이현우 and 수빈 just outside the left edge, while the clearing remains a quiet background rather than a new focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마을 공터 (The group remains gathered here during the discussion); used as Provides unobtrusive spatial continuity behind the fully visible crossed-arm gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral daytime illumination and controlled contrast, letting the forceful gesture rather than a lighting change convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the village clearing under repair, and the worn football remains in use nearby. Charlie retains his dented, punctured body and malfunctioning sensors. 마을총무: He remains in the clearing during the discussion of the supply trip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 극구 반대하며 두 손을 거세게 엑스(X) 자로 교차한 마을총무의 다급한 상반신.\n\nLOCATION (lock): In the village communal clearing beside the old truck being repaired. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the short lateral track beside 이현우 at shoulder height, settling on 마을총무 in an oblique upper-body composition without crossing the conversation axis. Place the administrator slightly right of center and retain both elbows and the complete X of his forearms below his face as he leans into his urgent refusal. His eyes and crossed-arm appeal address 이현우 and 수빈 just outside the left edge, while the clearing remains a quiet background rather than a new focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마을 공터 (The group remains gathered here during the discussion); used as Provides unobtrusive spatial continuity behind the fully visible crossed-arm gesture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral daytime illumination and controlled contrast, letting the forceful gesture rather than a lighting change convey alarm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck remains in the village clearing under repair, and the worn football remains in use nearby. Charlie retains his dented, punctured body and malfunctioning sensors. 마을총무: He remains in the clearing during the discussion of the supply trip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 마을총무 (한국인 성인, 50대의 얼굴, 짙은색 머리카락) — wearing: 색이 바래고 여기저기 해진 체크무늬 셔츠와 낡고 투박한 작업용 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh19__bgfirst_bg.png",
     "asset_id": "cea8326e-546a-4d42-8338-7e8292087943",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S62sh19.png",
     "asset_id": "ed980390-e6f6-4352-bc2a-816f460b25a8",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:741399>",
     "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_repair_clearing_b31b85.png",
     "asset_id": "87ef2d30-fa84-40f3-8625-fbc7845f6736",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:741399>",
     "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "마을총무의 시선과 자세가 화면 왼쪽 전경에 위치한 두 인물을 향하고 있음.",
    "built_space": "위치 레퍼런스의 건축물과 배경이 일치하며, 우측에 낡은 트럭이 있음.",
    "entities": "마을총무의 외모와 의상은 레퍼런스와 일치하지만, 샷 텍스트에 언급되지 않은 추가 인물 3명(전경 2명, 배경의 정비공 1명)이 등장함.",
    "hard_violations": [
     "[gemini-pro] invented people (샷 텍스트에 없는 전경 및 배경 인물 추가)",
     "[gpt-high] 화면 밖 왼쪽에 있어야 하는 대화 상대 두 명을 전경에 노출해, 총무만 보이도록 한 인물 제한과 배치를 위반했다.",
     "[gpt-high] 트럭 엔진룸 앞에 샷에서 허용하지 않은 정비공 한 명을 추가했다."
    ],
    "physics": "모든 인물이 지면에 안정적으로 서 있으며, 팔을 교차한 동작에 물리적 오류가 없음."
   },
   {
    "label": "B",
    "direction": "마을총무의 시선이 카메라 프레임 왼쪽 바깥의 보이지 않는 대상을 향하고 있음.",
    "built_space": "지정된 마을 공터와 배경 건축물, 우측의 트럭이 레퍼런스대로 정확히 배치됨.",
    "entities": "지시문대로 마을총무가 단독으로 등장하며, 얼굴과 복장이 캐릭터 레퍼런스와 완벽히 일치함.",
    "hard_violations": [],
    "physics": "지면에 몸을 지탱하고 서서 상반신을 약간 기울인 채 양팔을 안정적으로 교차하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 마을총무 단독 등장 원칙을 지켰으며, 요구된 X자 팔 교차와 우측 치우친 구도를 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 구도와 동작은 준수했으나, 샷 텍스트에 없는 인물들(전경 2명, 배경 1명)을 임의로 추가하여 심각한 규정 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "마을총무의 시선과 자세가 화면 왼쪽 전경에 위치한 두 인물을 향하고 있음.",
        "built_space": "위치 레퍼런스의 건축물과 배경이 일치하며, 우측에 낡은 트럭이 있음.",
        "entities": "마을총무의 외모와 의상은 레퍼런스와 일치하지만, 샷 텍스트에 언급되지 않은 추가 인물 3명(전경 2명, 배경의 정비공 1명)이 등장함.",
        "hard_violations": [
         "invented people (샷 텍스트에 없는 전경 및 배경 인물 추가)"
        ],
        "physics": "모든 인물이 지면에 안정적으로 서 있으며, 팔을 교차한 동작에 물리적 오류가 없음."
       },
       {
        "label": "B",
        "direction": "마을총무의 시선이 카메라 프레임 왼쪽 바깥의 보이지 않는 대상을 향하고 있음.",
        "built_space": "지정된 마을 공터와 배경 건축물, 우측의 트럭이 레퍼런스대로 정확히 배치됨.",
        "entities": "지시문대로 마을총무가 단독으로 등장하며, 얼굴과 복장이 캐릭터 레퍼런스와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "지면에 몸을 지탱하고 서서 상반신을 약간 기울인 채 양팔을 안정적으로 교차하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 마을총무 단독 등장 원칙을 지켰으며, 요구된 X자 팔 교차와 우측 치우친 구도를 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 구도와 동작은 준수했으나, 샷 텍스트에 없는 인물들(전경 2명, 배경 1명)을 임의로 추가하여 심각한 규정 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "마을총무의 시선과 자세가 화면 왼쪽 전경에 위치한 두 인물을 향하고 있음.",
        "built_space": "위치 레퍼런스의 건축물과 배경이 일치하며, 우측에 낡은 트럭이 있음.",
        "entities": "마을총무의 외모와 의상은 레퍼런스와 일치하지만, 샷 텍스트에 언급되지 않은 추가 인물 3명(전경 2명, 배경의 정비공 1명)이 등장함.",
        "hard_violations": [
         "invented people (샷 텍스트에 없는 전경 및 배경 인물 추가)"
        ],
        "physics": "모든 인물이 지면에 안정적으로 서 있으며, 팔을 교차한 동작에 물리적 오류가 없음."
       },
       {
        "label": "B",
        "direction": "마을총무의 시선이 카메라 프레임 왼쪽 바깥의 보이지 않는 대상을 향하고 있음.",
        "built_space": "지정된 마을 공터와 배경 건축물, 우측의 트럭이 레퍼런스대로 정확히 배치됨.",
        "entities": "지시문대로 마을총무가 단독으로 등장하며, 얼굴과 복장이 캐릭터 레퍼런스와 완벽히 일치함.",
        "hard_violations": [],
        "physics": "지면에 몸을 지탱하고 서서 상반신을 약간 기울인 채 양팔을 안정적으로 교차하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "총무만을 중앙 오른쪽 상반신으로 잡고 양 팔꿈치와 완전한 X, 화면 밖 왼쪽을 향한 다급한 호소를 구현해 지정 구도에 충실하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "총무의 X 동작은 맞지만 왼쪽 대화 상대 두 명과 뒤쪽 정비공을 추가해 단독 상반신 구도와 명시적인 인물 제한을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "총무의 얼굴과 눈은 화면 밖 왼쪽 상대를 향한다. 두 팔뚝은 얼굴 아래에서 X로 교차하고 손끝은 각각 좌우 위쪽을 향해 거절 동작을 이룬다. 이현우와 수빈을 왼쪽 프레임 밖에 두라는 지시와 부합한다.",
        "built_space": "왼쪽 교회 첨탑 한 곳, 중앙 왼쪽 철제 급수탑 한 기, 오른쪽 기와 건물군과 전신주 한 기가 보인다. 총무 뒤 오른쪽에는 보닛을 연 낡은 트럭 한 대와 공구함, 낮은 작업 받침들이 있다. 흙 공터와 시설의 배치·재질은 장소 참조를 따른다. 총무는 중앙보다 조금 오른쪽에 있고 양 팔꿈치가 모두 프레임 안에 있으며, 어깨 높이의 비스듬한 상반신 구도로 읽힌다.",
        "entities": "보이는 사람은 총무 한 명뿐이다. 중년 한국인 남성으로 보이는 얼굴, 짙은 머리, 올리브색 모자, 해진 적갈색 체크 셔츠와 회색 속옷, 손목시계가 인물 참조와 부합한다. 작업 바지는 대부분 화면 밖이다. 오른쪽 아래에는 참조의 소형 바퀴 로봇과 유사한 물체가 보이지만 센서 고장이나 구멍은 판독하기 어렵다. 축구공은 보이지 않으며, 요구된 상반신 구도에서 반드시 포함할 대상은 아니다.",
        "hard_violations": [],
        "physics": "두 손은 각각 팔목과 팔뚝에 자연스럽게 연결되고, 굽힌 팔꿈치와 어깨가 교차한 팔을 지탱한다. 몸을 약간 앞으로 기울인 거절 자세는 실제로 가능한 동작이다. 하체의 지면 접촉은 화면 밖이지만 공중에 뜬 정황은 없다. 트럭과 작업 물품, 바퀴 로봇은 지면에 놓여 있으며 열린 보닛은 차체에 연결되어 있다."
       },
       {
        "label": "B",
        "direction": "총무는 왼쪽 전경 남성을 바라보며 팔뚝을 얼굴 아래 X로 교차한다. 왼쪽 전경의 두 사람도 총무 쪽으로 몸과 머리를 돌리고 있다. 거절의 대상 방향은 맞지만, 화면 밖에 있어야 할 상대들이 화면 안으로 들어왔다. 뒤쪽 정비공은 트럭 엔진룸을 향해 숙이고 있다.",
        "built_space": "급수탑 한 기가 중앙 왼쪽에, 기와 건물군과 전신주 한 기가 뒤쪽에 있으며 보닛을 연 트럭 한 대가 오른쪽에 있다. 왼쪽 교회는 전경 인물에 상당 부분 가려져 있다. 공터와 트럭 주변의 공간은 장소 참조와 대체로 이어지지만, 전경 두 인물이 화면 왼쪽을 크게 차지해 지정된 총무 단독 구도가 대화 상대 너머로 보는 구도로 바뀌었다.",
        "entities": "총무의 중년 남성 얼굴, 모자, 체크 셔츠, 회색 속옷과 드러난 갈색 작업 바지는 참조와 대체로 부합한다. 그러나 왼쪽에 남성 한 명과 머리를 묶은 여성 한 명, 트럭 앞에 정비공 한 명이 추가되어 총 네 사람이 보인다. 이 세 명은 이 샷에서 허용된 인물이 아니다. 낡은 트럭, 지면의 축구공 한 개, 소형 바퀴 로봇은 보이지만 로봇의 구멍과 센서 오작동은 확인하기 어렵다.",
        "hard_violations": [
         "화면 밖 왼쪽에 있어야 하는 대화 상대 두 명을 전경에 노출해, 총무만 보이도록 한 인물 제한과 배치를 위반했다.",
         "트럭 엔진룸 앞에 샷에서 허용하지 않은 정비공 한 명을 추가했다."
        ],
        "physics": "총무의 양팔은 어깨와 팔꿈치로 지탱되며 X 동작은 해부학적으로 가능하다. 전경 인물들은 하체가 잘렸을 뿐 부유하는 모습은 아니다. 정비공은 트럭 쪽으로 몸을 숙여 팔을 엔진룸 가장자리에 대고 있다. 축구공은 땅에 닿아 있고 트럭과 로봇은 바퀴로 지면에 지지된다. 명백히 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "총무만을 중앙 오른쪽 상반신으로 잡고 양 팔꿈치와 완전한 X, 화면 밖 왼쪽을 향한 다급한 호소를 구현해 지정 구도에 충실하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "총무의 X 동작은 맞지만 왼쪽 대화 상대 두 명과 뒤쪽 정비공을 추가해 단독 상반신 구도와 명시적인 인물 제한을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "총무의 얼굴과 눈은 화면 밖 왼쪽 상대를 향한다. 두 팔뚝은 얼굴 아래에서 X로 교차하고 손끝은 각각 좌우 위쪽을 향해 거절 동작을 이룬다. 이현우와 수빈을 왼쪽 프레임 밖에 두라는 지시와 부합한다.",
        "built_space": "왼쪽 교회 첨탑 한 곳, 중앙 왼쪽 철제 급수탑 한 기, 오른쪽 기와 건물군과 전신주 한 기가 보인다. 총무 뒤 오른쪽에는 보닛을 연 낡은 트럭 한 대와 공구함, 낮은 작업 받침들이 있다. 흙 공터와 시설의 배치·재질은 장소 참조를 따른다. 총무는 중앙보다 조금 오른쪽에 있고 양 팔꿈치가 모두 프레임 안에 있으며, 어깨 높이의 비스듬한 상반신 구도로 읽힌다.",
        "entities": "보이는 사람은 총무 한 명뿐이다. 중년 한국인 남성으로 보이는 얼굴, 짙은 머리, 올리브색 모자, 해진 적갈색 체크 셔츠와 회색 속옷, 손목시계가 인물 참조와 부합한다. 작업 바지는 대부분 화면 밖이다. 오른쪽 아래에는 참조의 소형 바퀴 로봇과 유사한 물체가 보이지만 센서 고장이나 구멍은 판독하기 어렵다. 축구공은 보이지 않으며, 요구된 상반신 구도에서 반드시 포함할 대상은 아니다.",
        "hard_violations": [],
        "physics": "두 손은 각각 팔목과 팔뚝에 자연스럽게 연결되고, 굽힌 팔꿈치와 어깨가 교차한 팔을 지탱한다. 몸을 약간 앞으로 기울인 거절 자세는 실제로 가능한 동작이다. 하체의 지면 접촉은 화면 밖이지만 공중에 뜬 정황은 없다. 트럭과 작업 물품, 바퀴 로봇은 지면에 놓여 있으며 열린 보닛은 차체에 연결되어 있다."
       },
       {
        "label": "A",
        "direction": "총무는 왼쪽 전경 남성을 바라보며 팔뚝을 얼굴 아래 X로 교차한다. 왼쪽 전경의 두 사람도 총무 쪽으로 몸과 머리를 돌리고 있다. 거절의 대상 방향은 맞지만, 화면 밖에 있어야 할 상대들이 화면 안으로 들어왔다. 뒤쪽 정비공은 트럭 엔진룸을 향해 숙이고 있다.",
        "built_space": "급수탑 한 기가 중앙 왼쪽에, 기와 건물군과 전신주 한 기가 뒤쪽에 있으며 보닛을 연 트럭 한 대가 오른쪽에 있다. 왼쪽 교회는 전경 인물에 상당 부분 가려져 있다. 공터와 트럭 주변의 공간은 장소 참조와 대체로 이어지지만, 전경 두 인물이 화면 왼쪽을 크게 차지해 지정된 총무 단독 구도가 대화 상대 너머로 보는 구도로 바뀌었다.",
        "entities": "총무의 중년 남성 얼굴, 모자, 체크 셔츠, 회색 속옷과 드러난 갈색 작업 바지는 참조와 대체로 부합한다. 그러나 왼쪽에 남성 한 명과 머리를 묶은 여성 한 명, 트럭 앞에 정비공 한 명이 추가되어 총 네 사람이 보인다. 이 세 명은 이 샷에서 허용된 인물이 아니다. 낡은 트럭, 지면의 축구공 한 개, 소형 바퀴 로봇은 보이지만 로봇의 구멍과 센서 오작동은 확인하기 어렵다.",
        "hard_violations": [
         "화면 밖 왼쪽에 있어야 하는 대화 상대 두 명을 전경에 노출해, 총무만 보이도록 한 인물 제한과 배치를 위반했다.",
         "트럭 엔진룸 앞에 샷에서 허용하지 않은 정비공 한 명을 추가했다."
        ],
        "physics": "총무의 양팔은 어깨와 팔꿈치로 지탱되며 X 동작은 해부학적으로 가능하다. 전경 인물들은 하체가 잘렸을 뿐 부유하는 모습은 아니다. 정비공은 트럭 쪽으로 몸을 숙여 팔을 엔진룸 가장자리에 대고 있다. 축구공은 땅에 닿아 있고 트럭과 로봇은 바퀴로 지면에 지지된다. 명백히 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.619,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.369,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (샷 텍스트에 없는 전경 및 배경 인물 추가)",
     "[gpt-high] 화면 밖 왼쪽에 있어야 하는 대화 상대 두 명을 전경에 노출해, 총무만 보이도록 한 인물 제한과 배치를 위반했다.",
     "[gpt-high] 트럭 엔진룸 앞에 샷에서 허용하지 않은 정비공 한 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 369
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "샷 텍스트에 명시된 마을총무 단독 등장 원칙을 지켰으며, 요구된 X자 팔 교차와 우측 치우친 구도를 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 369,
    "verdict_ko": "지정된 구도와 동작은 준수했으나, 샷 텍스트에 없는 인물들(전경 2명, 배경 1명)을 임의로 추가하여 심각한 규정 위반이 발생했습니다.  ★위반: [gemini-pro] invented people (샷 텍스트에 없는 전경 및 배경 인물 추가) / [gpt-high] 화면 밖 왼쪽에 있어야 하는 대화 상대 두 명을 전경에 노출해, 총무만 보이도록 한 인물 제한과 배치를 위반했다. / [gpt-high] 트럭 엔진룸 앞에 샷에서 허용하지 않은 정비공 한 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_repair_clearing_b31b85.png",
    "asset_id": "87ef2d30-fa84-40f3-8625-fbc7845f6736",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 마을총무: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:741399>",
    "asset_id": "c867938c-50ab-4b35-a9c4-c14582d504b0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c91-8158-7a76-971c-4e197c354e91",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S62sh19__bgfirst_bg.png",
   "bg_asset_id": "cea8326e-546a-4d42-8338-7e8292087943",
   "bg_record_key": "S62sh19::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "village_repair_clearing",
   "groupbg_asset_id": "87ef2d30-fa84-40f3-8625-fbc7845f6736"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S63sh7::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 숏은 트럭 운전석 내부라는 좁고 제한된 공간에서 진행되며, 조수석에 앉은 수빈이 운전석의 인물을 향해 소리치는 상황입니다. 운전석과 조수석의 정확한 위치 관계와 인물의 시선 방향이 어긋나면 치명적인 오류가 발생하므로 평면도 보조가 필요합니다.",
  "input_fingerprint": "7a508dabad99e116"
 },
 "S63sh7::signage": {
  "fp": "7a7471943e1d96cc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::d3c11ae68730": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_d3c11ae68730.png",
  "place_text": "At the passenger seat inside the old truck's driving cab, beside the driver. Daylight enters through the windshield and side windows.",
  "input_fingerprint": "0f388b6077df89bc"
 },
 "S63sh7::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located at the front left driver station.",
   "mirrors": "No mirrors are indicated in the diagram.",
   "camera": "Positioned behind the driver-side edge of the central gap between the seats, facing straight forward toward the windshield.",
   "occupants": "수빈 (Subin) occupies the right passenger seat."
  },
  "mismatches": [],
  "scene_description_en": "The camera captures a medium shot facing forward from the central gap behind the truck seats. Subin occupies the passenger seat on the right two-thirds of the frame, twisting her torso leftward with her mouth wide open in an angry shout. The inner edge of the driver's seat occupies the near left foreground. In the far left background, the steering wheel is attached to the driver's station. Natural daylight enters through the front windshield in the center-to-far-right background and the right side window on the far right edge. No mirrors are present in the vehicle cabin.",
  "fixed": true,
  "input_fingerprint": "495ca3bfcaf289de"
 },
 "era_assess::7a976082bd710ea5": {
  "subjects": [],
  "subject_text": "수빈의 트럭 운전실\n낡고 간소한 운전 공간. 앞쪽에 운전대와 계기판이 있고, 전면 유리와 옆문 창을 통해 바깥이 보인다.",
  "identity": "canonical",
  "scope_id": "L110",
  "scope_role": "location_exterior",
  "scope_sha": "8352dab7c7221dfe"
 },
 "S63sh7": {
  "input_fingerprint": "24ce23d26308ec60",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 정한수를 배신자라 칭하는 이현우를 향해 목에 핏대를 세운 채 입을 크게 벌리고 버럭 소리치는 수빈의 역동적인 상반신.\n\nLOCATION (lock): At the passenger seat inside the old truck's driving cab, beside the driver. Daylight enters through the windshield and side windows. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the driver-side edge of the gap between the seats at 수빈's seated eye level, remaining behind the occupants' connecting line and holding enough distance for her upper body. Frame 수빈 across the right two-thirds with a narrow seat edge on the left, catching her torso twisting toward 이현우 outside the left edge, her mouth open and neck taut in protest. Preserve this oblique axis through the inward move so her anger travels across the cabin toward him, never toward the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 트럭 좌석 (The driver and passenger positions remain occupied, although the driver is outside this framing) — A narrow driver-seat edge borders the left foreground while the passenger seat remains behind 수빈; used as Locates the camera within the central gap and contains the outburst within the cabin.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime cabin light preserves the tension in 수빈's neck and face without introducing a dramatic color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck is traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and torso radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera captures a medium shot facing forward from the central gap behind the truck seats. Subin occupies the passenger seat on the right two-thirds of the frame, twisting her torso leftward with her mouth wide open in an angry shout. The inner edge of the driver's seat occupies the near left foreground. In the far left background, the steering wheel is attached to the driver's station. Natural daylight enters through the front windshield in the center-to-far-right background and the right side window on the far right edge. No mirrors are present in the vehicle cabin.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 정한수를 배신자라 칭하는 이현우를 향해 목에 핏대를 세운 채 입을 크게 벌리고 버럭 소리치는 수빈의 역동적인 상반신.\n\nLOCATION (lock): At the passenger seat inside the old truck's driving cab, beside the driver. Daylight enters through the windshield and side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime cabin light preserves the tension in 수빈's neck and face without introducing a dramatic color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck is traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and torso radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera captures a medium shot facing forward from the central gap behind the truck seats. Subin occupies the passenger seat on the right two-thirds of the frame, twisting her torso leftward with her mouth wide open in an angry shout. The inner edge of the driver's seat occupies the near left foreground. In the far left background, the steering wheel is attached to the driver's station. Natural daylight enters through the front windshield in the center-to-far-right background and the right side window on the far right edge. No mirrors are present in the vehicle cabin.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 정한수를 배신자라 칭하는 이현우를 향해 목에 핏대를 세운 채 입을 크게 벌리고 버럭 소리치는 수빈의 역동적인 상반신.\n\nLOCATION (lock): At the passenger seat inside the old truck's driving cab, beside the driver. Daylight enters through the windshield and side windows. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime cabin light preserves the tension in 수빈's neck and face without introducing a dramatic color shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck is traveling along the road in daylight. 수빈: She occupies the passenger seat. Her facial wounds and torso radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "수빈",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh7_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "수빈",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선과 외침의 방향이 화면 왼쪽 밖(이현우가 있는 방향)을 명확히 향하고 있음.",
    "built_space": "트럭 내부. 화면 왼쪽 전경에 운전석 등받이 가장자리가 있고, 수빈은 우측 조수석에 위치하며 전면 유리창과 우측 창문이 보임.",
    "entities": "수빈(단발, 올리브색 테크웨어, 얼굴 상처 등 레퍼런스와 일치). 다른 인물은 없음.",
    "hard_violations": [],
    "physics": "조수석 시트에 앉아 안전벨트로 몸이 지지된 상태에서 상체를 왼쪽으로 튼 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선과 외침이 화면 왼쪽 앞을 향하고 있음.",
    "built_space": "트럭 내부 조수석. 좌석 사이 구도를 의도했으나 왼쪽 전경을 운전자가 크게 가리고 있음.",
    "entities": "수빈(복장과 외양 일치). 화면 왼쪽 전경에 프롬프트에 없는 후드를 쓴 인물이 등장함.",
    "hard_violations": [
     "[gemini-pro] Invented people: 프롬프트에 명시되지 않은 인물(운전자)의 뒷모습이 화면에 나타남."
    ],
    "physics": "조수석에 앉아 앞으로 기울인 상체가 시트와 중력에 의해 정상적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 구도와 좌석 배치를 정확히 따랐으며, 지시되지 않은 인물을 배제하는 엄격한 규정을 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "표정과 위치는 좋으나, 허용되지 않은 인물의 신체 일부가 화면을 차지하여 치명적 규정 위반이 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 외침의 방향이 화면 왼쪽 밖(이현우가 있는 방향)을 명확히 향하고 있음.",
        "built_space": "트럭 내부. 화면 왼쪽 전경에 운전석 등받이 가장자리가 있고, 수빈은 우측 조수석에 위치하며 전면 유리창과 우측 창문이 보임.",
        "entities": "수빈(단발, 올리브색 테크웨어, 얼굴 상처 등 레퍼런스와 일치). 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "조수석 시트에 앉아 안전벨트로 몸이 지지된 상태에서 상체를 왼쪽으로 튼 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 외침이 화면 왼쪽 앞을 향하고 있음.",
        "built_space": "트럭 내부 조수석. 좌석 사이 구도를 의도했으나 왼쪽 전경을 운전자가 크게 가리고 있음.",
        "entities": "수빈(복장과 외양 일치). 화면 왼쪽 전경에 프롬프트에 없는 후드를 쓴 인물이 등장함.",
        "hard_violations": [
         "Invented people: 프롬프트에 명시되지 않은 인물(운전자)의 뒷모습이 화면에 나타남."
        ],
        "physics": "조수석에 앉아 앞으로 기울인 상체가 시트와 중력에 의해 정상적으로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "카메라 구도와 좌석 배치를 정확히 따랐으며, 지시되지 않은 인물을 배제하는 엄격한 규정을 잘 준수함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "표정과 위치는 좋으나, 허용되지 않은 인물의 신체 일부가 화면을 차지하여 치명적 규정 위반이 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 외침의 방향이 화면 왼쪽 밖(이현우가 있는 방향)을 명확히 향하고 있음.",
        "built_space": "트럭 내부. 화면 왼쪽 전경에 운전석 등받이 가장자리가 있고, 수빈은 우측 조수석에 위치하며 전면 유리창과 우측 창문이 보임.",
        "entities": "수빈(단발, 올리브색 테크웨어, 얼굴 상처 등 레퍼런스와 일치). 다른 인물은 없음.",
        "hard_violations": [],
        "physics": "조수석 시트에 앉아 안전벨트로 몸이 지지된 상태에서 상체를 왼쪽으로 튼 자세가 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선과 외침이 화면 왼쪽 앞을 향하고 있음.",
        "built_space": "트럭 내부 조수석. 좌석 사이 구도를 의도했으나 왼쪽 전경을 운전자가 크게 가리고 있음.",
        "entities": "수빈(복장과 외양 일치). 화면 왼쪽 전경에 프롬프트에 없는 후드를 쓴 인물이 등장함.",
        "hard_violations": [
         "Invented people: 프롬프트에 명시되지 않은 인물(운전자)의 뒷모습이 화면에 나타남."
        ],
        "physics": "조수석에 앉아 앞으로 기울인 상체가 시트와 중력에 의해 정상적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "왼쪽 이현우를 향한 고함과 목의 긴장은 맞지만, 운전석 등받이가 전경을 너무 크게 차지하고 조수석 등받이와 수빈의 착석 관계가 B보다 불명확하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "좌석 사이 뒤쪽의 사선 시점, 오른쪽 상반신 배치, 왼쪽 상대를 향한 고함과 목의 긴장, 조수석 지지 및 노출된 피폭 병변을 더 충실하게 구현했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈의 눈과 벌어진 입은 렌즈가 아니라 화면 왼쪽 운전석 쪽을 향한다. 몸통도 왼쪽으로 비틀고 있어 화면 밖 이현우에게 항의하는 방향과 부합한다. 무기나 겨누는 소품은 없다.",
        "built_space": "왼쪽 전경에 운전석 등받이 하나, 그 앞에 운전대 하나, 중앙에 대시보드와 변속 레버 하나가 보인다. 앞유리 하나와 오른쪽 측면 창 하나, 외부 거울 하나가 보이며 낡은 트럭 실내로 읽힌다. 운전석 등받이는 좁은 가장자리라기보다 왼쪽 아래를 크게 가린다. 조수석의 일부는 수빈 아래와 뒤에 보이지만 등받이와 몸의 관계가 선명하지 않다. 운전자는 등받이와 화면 경계에 가려져 점유 여부를 직접 확인할 수 없다. 불가능한 반사나 명백한 중복 설비는 보이지 않는다.",
        "entities": "보이는 인물은 젊은 동아시아계 여성 수빈 한 명이다. 검은 단발, 마른 체형, 올리브색의 낡고 먼지 묻은 실용적 재킷과 회색 안쪽 상의가 참조와 대체로 맞는다. 볼의 상처가 보이며 목의 힘줄도 드러난다. 몸통은 옷에 가려 피폭 병변의 유지 여부를 판단하기 어렵다. 하의와 부츠는 프레임 밖이므로 평가 대상이 아니다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "하체가 잘렸지만 몸통 아래의 조수석과 이어지는 착석 자세로 읽히며, 앉은 채 허리를 비틀고 앞으로 기울이는 동작은 가능하다. 팔은 아래쪽 프레임 밖으로 이어져 손의 접촉점은 확인되지 않는다. 공중에 뜬 신체나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "수빈의 시선과 얼굴은 화면 왼쪽 운전석의 보이지 않는 상대를 향하고, 입을 크게 벌린 채 몸통을 그쪽으로 돌린다. 렌즈를 향해 외치는 모습이 아니어서 이현우에게 전달되는 분노의 방향이 명확하다. 겨누는 물체는 없다.",
        "built_space": "왼쪽 전경의 운전석 등받이 하나와 수빈 뒤 조수석 등받이 하나가 구분된다. 운전대 하나는 왼쪽 좌석 앞에 있고 중앙 대시보드, 앞유리 하나, 오른쪽 측면 창 하나와 외부 거울 하나가 보인다. 좌석 사이 뒤쪽에서 조수석을 비스듬히 보는 시점이며 수빈이 오른쪽 영역을 차지한다. 운전석 가장자리는 여전히 다소 넓지만 A보다 덜 침범한다. 운전자는 가려져 있으며 빈 운전석이라고 단정할 근거는 없다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 여성 한 명으로, 참조의 검은 단발과 얼굴 인상, 마른 체형, 낡은 올리브색 재킷 및 회색 안쪽 상의에 대체로 부합한다. 볼의 상처와 목 아래 노출된 윗가슴의 붉은 병변이 보인다. 정상적인 눈과 얼굴 근육으로 분노를 연기하며 목의 긴장도 뚜렷하다. 안전벨트가 몸통을 가로지른다. 하의와 부츠는 잘려 있어 평가하지 않는다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈의 몸은 뒤의 조수석 등받이와 아래 좌석으로 이어지는 착석 위치에 있고 안전벨트가 몸통을 가로지른다. 앉은 상태에서 어깨와 목을 왼쪽으로 돌려 고함치는 자세로 충분히 가능한 동작이다. 손은 프레임 밖이지만 팔의 연결은 자연스럽고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 이현우를 향한 고함과 목의 긴장은 맞지만, 운전석 등받이가 전경을 너무 크게 차지하고 조수석 등받이와 수빈의 착석 관계가 B보다 불명확하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "좌석 사이 뒤쪽의 사선 시점, 오른쪽 상반신 배치, 왼쪽 상대를 향한 고함과 목의 긴장, 조수석 지지 및 노출된 피폭 병변을 더 충실하게 구현했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "수빈의 눈과 벌어진 입은 렌즈가 아니라 화면 왼쪽 운전석 쪽을 향한다. 몸통도 왼쪽으로 비틀고 있어 화면 밖 이현우에게 항의하는 방향과 부합한다. 무기나 겨누는 소품은 없다.",
        "built_space": "왼쪽 전경에 운전석 등받이 하나, 그 앞에 운전대 하나, 중앙에 대시보드와 변속 레버 하나가 보인다. 앞유리 하나와 오른쪽 측면 창 하나, 외부 거울 하나가 보이며 낡은 트럭 실내로 읽힌다. 운전석 등받이는 좁은 가장자리라기보다 왼쪽 아래를 크게 가린다. 조수석의 일부는 수빈 아래와 뒤에 보이지만 등받이와 몸의 관계가 선명하지 않다. 운전자는 등받이와 화면 경계에 가려져 점유 여부를 직접 확인할 수 없다. 불가능한 반사나 명백한 중복 설비는 보이지 않는다.",
        "entities": "보이는 인물은 젊은 동아시아계 여성 수빈 한 명이다. 검은 단발, 마른 체형, 올리브색의 낡고 먼지 묻은 실용적 재킷과 회색 안쪽 상의가 참조와 대체로 맞는다. 볼의 상처가 보이며 목의 힘줄도 드러난다. 몸통은 옷에 가려 피폭 병변의 유지 여부를 판단하기 어렵다. 하의와 부츠는 프레임 밖이므로 평가 대상이 아니다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "하체가 잘렸지만 몸통 아래의 조수석과 이어지는 착석 자세로 읽히며, 앉은 채 허리를 비틀고 앞으로 기울이는 동작은 가능하다. 팔은 아래쪽 프레임 밖으로 이어져 손의 접촉점은 확인되지 않는다. 공중에 뜬 신체나 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "수빈의 시선과 얼굴은 화면 왼쪽 운전석의 보이지 않는 상대를 향하고, 입을 크게 벌린 채 몸통을 그쪽으로 돌린다. 렌즈를 향해 외치는 모습이 아니어서 이현우에게 전달되는 분노의 방향이 명확하다. 겨누는 물체는 없다.",
        "built_space": "왼쪽 전경의 운전석 등받이 하나와 수빈 뒤 조수석 등받이 하나가 구분된다. 운전대 하나는 왼쪽 좌석 앞에 있고 중앙 대시보드, 앞유리 하나, 오른쪽 측면 창 하나와 외부 거울 하나가 보인다. 좌석 사이 뒤쪽에서 조수석을 비스듬히 보는 시점이며 수빈이 오른쪽 영역을 차지한다. 운전석 가장자리는 여전히 다소 넓지만 A보다 덜 침범한다. 운전자는 가려져 있으며 빈 운전석이라고 단정할 근거는 없다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 여성 한 명으로, 참조의 검은 단발과 얼굴 인상, 마른 체형, 낡은 올리브색 재킷 및 회색 안쪽 상의에 대체로 부합한다. 볼의 상처와 목 아래 노출된 윗가슴의 붉은 병변이 보인다. 정상적인 눈과 얼굴 근육으로 분노를 연기하며 목의 긴장도 뚜렷하다. 안전벨트가 몸통을 가로지른다. 하의와 부츠는 잘려 있어 평가하지 않는다. 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "수빈의 몸은 뒤의 조수석 등받이와 아래 좌석으로 이어지는 착석 위치에 있고 안전벨트가 몸통을 가로지른다. 앉은 상태에서 어깨와 목을 왼쪽으로 돌려 고함치는 자세로 충분히 가능한 동작이다. 손은 프레임 밖이지만 팔의 연결은 자연스럽고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.206
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.956
   },
   "violations": {
    "B": [
     "[gemini-pro] Invented people: 프롬프트에 명시되지 않은 인물(운전자)의 뒷모습이 화면에 나타남."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 956
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "카메라 구도와 좌석 배치를 정확히 따랐으며, 지시되지 않은 인물을 배제하는 엄격한 규정을 잘 준수함."
   },
   {
    "label": "B",
    "score": 956,
    "verdict_ko": "표정과 위치는 좋으나, 허용되지 않은 인물의 신체 일부가 화면을 차지하여 치명적 규정 위반이 발생함.  ★위반: [gemini-pro] Invented people: 프롬프트에 명시되지 않은 인물(운전자)의 뒷모습이 화면에 나타남."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh7_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "수빈",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0c9a-b212-78e0-bc36-928f120d6f45",
  "confined_fp": {
   "base_key": "confinedfp::d3c11ae68730",
   "apt_reason": "이 숏은 트럭 운전석 내부라는 좁고 제한된 공간에서 진행되며, 조수석에 앉은 수빈이 운전석의 인물을 향해 소리치는 상황입니다. 운전석과 조수석의 정확한 위치 관계와 인물의 시선 방향이 어긋나면 치명적인 오류가 발생하므로 평면도 보조가 필요합니다.",
   "fixed": true,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S63sh13::signage": {
  "fp": "8aab197e8abd3edf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S63sh13": {
  "input_fingerprint": "4c40c839d8cdba79",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 바닷물에 젖은 판자 위에서 얼굴에 피가 흐른 채 힘겹게 눈을 반쯤 뜬 미연의 창백한 얼굴 클로즈업.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, shown as a monochrome memory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the explicitly separated flashback with a static high-oblique close-up from above and to one side of 미연's face, never aligning directly with her facial centerline. Her pale, blood-streaked face occupies the central area, with her neck, shoulder, and a narrow portion of the seawater-wet plank retaining physical context. Catch her half-open eyes directed weakly beyond the frame toward the person she is addressing as she struggles to speak, treating this as a remembered moment rather than a present-tense view from the truck.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 판자 (Wet with seawater beneath 미연) — Its upper supporting surface is visible obliquely beside her head and shoulder; used as Anchors the recalled physical situation without widening away from her final effort to speak.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the memory in restrained black and white with soft tonal separation that preserves the blood streaks and half-open eyes without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is recumbent on a floating container panel, critically wounded and bleeding from her abdomen, with her eyes opening only with difficulty. The panel supports her body, but the exact orientation of her head and torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the flood memory, a container panel floats on the sea in the dawn light. 미연: She lies on the floating panel with a bleeding puncture wound in her abdomen, close to death. Her face retains the swelling from the earlier beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 바닷물에 젖은 판자 위에서 얼굴에 피가 흐른 채 힘겹게 눈을 반쯤 뜬 미연의 창백한 얼굴 클로즈업.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, shown as a monochrome memory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the explicitly separated flashback with a static high-oblique close-up from above and to one side of 미연's face, never aligning directly with her facial centerline. Her pale, blood-streaked face occupies the central area, with her neck, shoulder, and a narrow portion of the seawater-wet plank retaining physical context. Catch her half-open eyes directed weakly beyond the frame toward the person she is addressing as she struggles to speak, treating this as a remembered moment rather than a present-tense view from the truck.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 판자 (Wet with seawater beneath 미연) — Its upper supporting surface is visible obliquely beside her head and shoulder; used as Anchors the recalled physical situation without widening away from her final effort to speak.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the memory in restrained black and white with soft tonal separation that preserves the blood streaks and half-open eyes without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is recumbent on a floating container panel, critically wounded and bleeding from her abdomen, with her eyes opening only with difficulty. The panel supports her body, but the exact orientation of her head and torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the flood memory, a container panel floats on the sea in the dawn light. 미연: She lies on the floating panel with a bleeding puncture wound in her abdomen, close to death. Her face retains the swelling from the earlier beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흑백 화면 속, 바닷물에 젖은 판자 위에서 얼굴에 피가 흐른 채 힘겹게 눈을 반쯤 뜬 미연의 창백한 얼굴 클로즈업.\n\nLOCATION (lock): On a floating container panel amid the flooded refugee settlement at dawn, shown as a monochrome memory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Enter the explicitly separated flashback with a static high-oblique close-up from above and to one side of 미연's face, never aligning directly with her facial centerline. Her pale, blood-streaked face occupies the central area, with her neck, shoulder, and a narrow portion of the seawater-wet plank retaining physical context. Catch her half-open eyes directed weakly beyond the frame toward the person she is addressing as she struggles to speak, treating this as a remembered moment rather than a present-tense view from the truck.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 판자 (Wet with seawater beneath 미연) — Its upper supporting surface is visible obliquely beside her head and shoulder; used as Anchors the recalled physical situation without widening away from her final effort to speak.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Render the memory in restrained black and white with soft tonal separation that preserves the blood streaks and half-open eyes without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Miyeon is recumbent on a floating container panel, critically wounded and bleeding from her abdomen, with her eyes opening only with difficulty. The panel supports her body, but the exact orientation of her head and torso and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the flood memory, a container panel floats on the sea in the dawn light. 미연: She lies on the floating panel with a bleeding puncture wound in her abdomen, close to death. Her face retains the swelling from the earlier beating.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 미연 (한국인 여성, 40대 후반의 얼굴, 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "반쯤 뜬 눈으로 화면 밖 우측을 향해 시선을 둠.",
    "built_space": "젖은 컨테이너 패널 위, 지정된 하이 앵글 클로즈업 구도.",
    "entities": "미연의 인상착의, 의상, 상처가 레퍼런스와 일치하며 흑백으로 묘사됨.",
    "hard_violations": [],
    "physics": "머리와 어깨가 바닥에 자연스럽게 밀착되어 중력을 따름."
   },
   {
    "label": "B",
    "direction": "반쯤 뜬 눈으로 화면 밖 우측을 바라봄.",
    "built_space": "물기 있는 패널 위, 비스듬한 하이 앵글 클로즈업.",
    "entities": "미연의 외모와 의상 조건이 흑백으로 무난하게 표현됨.",
    "hard_violations": [],
    "physics": "신체가 바닥 표면에 안정적으로 닿아 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 흑백 클로즈업 구도와 레퍼런스의 인물 특징, 젖은 패널의 질감을 매우 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지침을 전반적으로 잘 따랐으나, 인물의 디테일과 질감 표현이 A에 비해 미세하게 부족함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "반쯤 뜬 눈으로 화면 밖 우측을 향해 시선을 둠.",
        "built_space": "젖은 컨테이너 패널 위, 지정된 하이 앵글 클로즈업 구도.",
        "entities": "미연의 인상착의, 의상, 상처가 레퍼런스와 일치하며 흑백으로 묘사됨.",
        "hard_violations": [],
        "physics": "머리와 어깨가 바닥에 자연스럽게 밀착되어 중력을 따름."
       },
       {
        "label": "B",
        "direction": "반쯤 뜬 눈으로 화면 밖 우측을 바라봄.",
        "built_space": "물기 있는 패널 위, 비스듬한 하이 앵글 클로즈업.",
        "entities": "미연의 외모와 의상 조건이 흑백으로 무난하게 표현됨.",
        "hard_violations": [],
        "physics": "신체가 바닥 표면에 안정적으로 닿아 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 흑백 클로즈업 구도와 레퍼런스의 인물 특징, 젖은 패널의 질감을 매우 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지침을 전반적으로 잘 따랐으나, 인물의 디테일과 질감 표현이 A에 비해 미세하게 부족함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "반쯤 뜬 눈으로 화면 밖 우측을 향해 시선을 둠.",
        "built_space": "젖은 컨테이너 패널 위, 지정된 하이 앵글 클로즈업 구도.",
        "entities": "미연의 인상착의, 의상, 상처가 레퍼런스와 일치하며 흑백으로 묘사됨.",
        "hard_violations": [],
        "physics": "머리와 어깨가 바닥에 자연스럽게 밀착되어 중력을 따름."
       },
       {
        "label": "B",
        "direction": "반쯤 뜬 눈으로 화면 밖 우측을 바라봄.",
        "built_space": "물기 있는 패널 위, 비스듬한 하이 앵글 클로즈업.",
        "entities": "미연의 외모와 의상 조건이 흑백으로 무난하게 표현됨.",
        "hard_violations": [],
        "physics": "신체가 바닥 표면에 안정적으로 닿아 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴 중심축을 비껴 내려다보는 밀착 클로즈업, 반쯤 열린 눈과 화면 밖 시선, 흑백 표현이 요구된 마지막 발화 순간에 더 충실하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "젖은 패널에 누운 미연과 부상은 잘 구현했지만, 얼굴을 더 정면으로 바라보는 각도와 상대적으로 또렷하게 열린 눈이 지정된 비스듬한 시점과 쇠약한 순간에서 조금 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "미연의 반쯤 열린 두 눈은 화면 오른쪽 바깥을 약하게 향한다. 응시 대상인 상대방은 보이지 않으며, 렌즈를 똑바로 보는 표정은 아니다. 입술이 조금 벌어져 힘겹게 말하려는 순간으로 읽힌다.",
        "built_space": "얼굴이 중앙을 크게 차지하고 목과 양쪽 어깨 일부만 포함된다. 카메라는 얼굴 위쪽의 한편에서 비스듬히 내려다본다. 머리 아래에는 젖고 도장이 벗겨진 금속 컨테이너 패널 하나가 있으며, 오른쪽의 길게 솟은 보강부와 원형 체결부가 보인다. 오른쪽 위에는 좁은 수면이 보인다. 목재보다는 금속으로 읽히지만 이전 장면의 실제 지지면 재질과 맞으며, 중복 설비나 불가능한 반사는 없다.",
        "entities": "검은 단발머리의 중년 동아시아 여성 한 명만 보이며, 미연 참고 이미지의 얼굴 윤곽과 연령대에 부합한다. 젖고 오염된 밝은 작업 셔츠와 어두운 안쪽 옷도 이어진다. 창백한 얼굴에 이마에서 내려온 피와 입가의 피, 타박 흔적이 있다. 눈은 정상적인 홍채와 동공을 유지한다. 복부 상처는 올바른 클로즈업 범위 밖이므로 확인할 수 없다. 화면은 흑백이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 패널 위에 놓이고 어깨와 상체도 같은 지지면에 기대어 있다. 목이 몸통과 자연스럽게 이어지며 머리를 허공에 들고 있는 모습은 아니다. 물은 패널 표면에 맺히고 주변 수면은 패널의 부유 상황을 뒷받침한다. 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 눈은 화면 오른쪽 위 바깥을 향하며, 보이지 않는 상대에게 시선을 보내는 관계는 성립한다. 입은 조금 열려 있다. 다만 눈꺼풀이 A보다 더 올라가 있어 힘겹게 반쯤 눈을 뜬 상태보다는 상대를 비교적 또렷이 보는 인상이 강하다.",
        "built_space": "중앙의 얼굴과 목, 어깨 및 가슴 윗부분이 포함된 클로즈업이다. 위에서 내려다보지만 얼굴의 정면 중심축에 A보다 가깝다. 지지면은 벗겨진 도장과 젖은 표면을 가진 컨테이너 패널 하나다. 오른쪽 가장자리에 직사각형 체결 구조 하나와 원형 체결부가 보이고, 왼쪽 위에는 두 개의 어두운 홈이 보인다. 오른쪽 위 수면과 패널의 위치 관계는 자연스럽다. 참고 장소와 재질은 부합하며 설비 중복이나 불가능한 반사는 없다.",
        "entities": "미연에 해당하는 검은 단발머리의 중년 동아시아 여성 한 명만 있다. 얼굴과 젖은 작업 셔츠, 어두운 안쪽 옷이 참고 이미지에 대체로 부합한다. 이마의 흐른 피, 양 볼의 타박 흔적과 입가의 피가 선명하다. 눈의 해부학적 형태는 정상이다. 복부는 프레임 밖이므로 상처 노출 여부는 감점 대상이 아니다. 전체는 거의 흑백이나 일부 혈흔에는 미세한 갈색 기운이 남아 보인다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락, 어깨와 등이 패널에 받쳐진 누운 자세다. 목 아래로 이어지는 몸통에도 자연스러운 하중 관계가 있으며, 공중에 떠 있는 신체 부위는 보이지 않는다. 젖은 옷은 몸에 붙고 주름이 잡혀 있으며 패널 위 물방울과 주변 바닷물도 물리적으로 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴 중심축을 비껴 내려다보는 밀착 클로즈업, 반쯤 열린 눈과 화면 밖 시선, 흑백 표현이 요구된 마지막 발화 순간에 더 충실하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "젖은 패널에 누운 미연과 부상은 잘 구현했지만, 얼굴을 더 정면으로 바라보는 각도와 상대적으로 또렷하게 열린 눈이 지정된 비스듬한 시점과 쇠약한 순간에서 조금 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "미연의 반쯤 열린 두 눈은 화면 오른쪽 바깥을 약하게 향한다. 응시 대상인 상대방은 보이지 않으며, 렌즈를 똑바로 보는 표정은 아니다. 입술이 조금 벌어져 힘겹게 말하려는 순간으로 읽힌다.",
        "built_space": "얼굴이 중앙을 크게 차지하고 목과 양쪽 어깨 일부만 포함된다. 카메라는 얼굴 위쪽의 한편에서 비스듬히 내려다본다. 머리 아래에는 젖고 도장이 벗겨진 금속 컨테이너 패널 하나가 있으며, 오른쪽의 길게 솟은 보강부와 원형 체결부가 보인다. 오른쪽 위에는 좁은 수면이 보인다. 목재보다는 금속으로 읽히지만 이전 장면의 실제 지지면 재질과 맞으며, 중복 설비나 불가능한 반사는 없다.",
        "entities": "검은 단발머리의 중년 동아시아 여성 한 명만 보이며, 미연 참고 이미지의 얼굴 윤곽과 연령대에 부합한다. 젖고 오염된 밝은 작업 셔츠와 어두운 안쪽 옷도 이어진다. 창백한 얼굴에 이마에서 내려온 피와 입가의 피, 타박 흔적이 있다. 눈은 정상적인 홍채와 동공을 유지한다. 복부 상처는 올바른 클로즈업 범위 밖이므로 확인할 수 없다. 화면은 흑백이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 패널 위에 놓이고 어깨와 상체도 같은 지지면에 기대어 있다. 목이 몸통과 자연스럽게 이어지며 머리를 허공에 들고 있는 모습은 아니다. 물은 패널 표면에 맺히고 주변 수면은 패널의 부유 상황을 뒷받침한다. 지지 없는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 눈은 화면 오른쪽 위 바깥을 향하며, 보이지 않는 상대에게 시선을 보내는 관계는 성립한다. 입은 조금 열려 있다. 다만 눈꺼풀이 A보다 더 올라가 있어 힘겹게 반쯤 눈을 뜬 상태보다는 상대를 비교적 또렷이 보는 인상이 강하다.",
        "built_space": "중앙의 얼굴과 목, 어깨 및 가슴 윗부분이 포함된 클로즈업이다. 위에서 내려다보지만 얼굴의 정면 중심축에 A보다 가깝다. 지지면은 벗겨진 도장과 젖은 표면을 가진 컨테이너 패널 하나다. 오른쪽 가장자리에 직사각형 체결 구조 하나와 원형 체결부가 보이고, 왼쪽 위에는 두 개의 어두운 홈이 보인다. 오른쪽 위 수면과 패널의 위치 관계는 자연스럽다. 참고 장소와 재질은 부합하며 설비 중복이나 불가능한 반사는 없다.",
        "entities": "미연에 해당하는 검은 단발머리의 중년 동아시아 여성 한 명만 있다. 얼굴과 젖은 작업 셔츠, 어두운 안쪽 옷이 참고 이미지에 대체로 부합한다. 이마의 흐른 피, 양 볼의 타박 흔적과 입가의 피가 선명하다. 눈의 해부학적 형태는 정상이다. 복부는 프레임 밖이므로 상처 노출 여부는 감점 대상이 아니다. 전체는 거의 흑백이나 일부 혈흔에는 미세한 갈색 기운이 남아 보인다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락, 어깨와 등이 패널에 받쳐진 누운 자세다. 목 아래로 이어지는 몸통에도 자연스러운 하중 관계가 있으며, 공중에 떠 있는 신체 부위는 보이지 않는다. 젖은 옷은 몸에 붙고 주름이 잡혀 있으며 패널 위 물방울과 주변 바닷물도 물리적으로 자연스럽다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "지정된 흑백 클로즈업 구도와 레퍼런스의 인물 특징, 젖은 패널의 질감을 매우 충실하게 구현함."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "지침을 전반적으로 잘 따랐으나, 인물의 디테일과 질감 표현이 A에 비해 미세하게 부족함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 미연 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S38sh8_sel.png",
    "asset_id": "c5910d68-60ae-4983-9d76-bf47827a68ec",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 미연: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1244810>",
    "asset_id": "b568747f-d513-4086-8aa5-eb33d2e7eb6f",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ca7-bb1e-7b0f-9256-13b07f2d6846",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh8"
  }
 },
 "S63sh16::confined_fp_apt": {
  "applies": true,
  "reason_ko": "트럭 운전석 내부를 배경으로 하는 샷으로, 인물이 운전대 및 차창과 맺는 공간적 위치 관계가 시각적으로 정확하게 표현되어야 하기 때문입니다.",
  "input_fingerprint": "474a93dd3b090098"
 },
 "S63sh16::signage": {
  "fp": "ba38f3736cf1034f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::388a94bc37e2": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_388a94bc37e2.png",
  "place_text": "At the driver's position inside the old truck's compact cab. Strong sunlight enters through the side window and falls across the driver's face.",
  "input_fingerprint": "01c1d64f9c631ca3"
 },
 "S63sh16::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located at the front left station (driver's seat).",
   "mirrors": "No mirrors or reflective surfaces are indicated in the diagram.",
   "camera": "The camera stands in the central space between the two seats, angled forward and to the left, pointing directly at the driver's seat.",
   "occupants": "이현우 is seated in the left station (driver's seat). The right station (passenger seat) is empty."
  },
  "mismatches": [],
  "scene_description_en": "The camera is positioned in the center of the truck cabin, pointing forward and leftward toward the driver's seat. 이현우 occupies this seat and appears on the right side of the close-up frame, facing toward the left. A steering wheel is located immediately in front of him, partially occupying the lower edge. On the left side of the frame, a narrow edge of the side window is visible, allowing bright sunlight to shine directly onto his face. The empty passenger seat remains out of view behind and to the right of the camera. There are no mirrors or reflective surfaces depicted in this scene.",
  "fixed": false,
  "input_fingerprint": "0684318b673a06ad"
 },
 "S63sh16": {
  "input_fingerprint": "f277717eef03d55f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차창 밖에서 비쳐 든 눈부신 햇살 아래, 입술을 굳게 다문 채 결연한 의지를 띠고 있는 이현우의 얼굴 구도.\n\nLOCATION (lock): At the driver's position inside the old truck's compact cab. Strong sunlight enters through the side window and falls across the driver's face. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit return from memory, finish the approach from the passenger-side edge of the central gap, just below 이현우's cheek level with a slight upward angle and the established lateral view. Keep his face on the right with looking room to the left toward the road outside the frame, his lips pressed together and his driving shoulders held taut. A narrow window edge retains the cabin context while the reduced distance, not a new turn of his head, makes his resolve legible.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 트럭 차창 (Visible as a narrow portion of the cabin boundary) — Seen obliquely beside the driver's face, with the exterior beyond it left indistinct; used as Retains the present-tense driving environment without pulling attention from his expression.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright sunlight entering through the truck window washes across 이현우's face with controlled highlights, restoring natural color after the monochrome memory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck continues along the road in daylight. 이현우: Back in the present, he remains in the driver's seat with his recent injuries. His expression has become resolute.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the center of the truck cabin, pointing forward and leftward toward the driver's seat. 이현우 occupies this seat and appears on the right side of the close-up frame, facing toward the left. A steering wheel is located immediately in front of him, partially occupying the lower edge. On the left side of the frame, a narrow edge of the side window is visible, allowing bright sunlight to shine directly onto his face. The empty passenger seat remains out of view behind and to the right of the camera. There are no mirrors or reflective surfaces depicted in this scene.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차창 밖에서 비쳐 든 눈부신 햇살 아래, 입술을 굳게 다문 채 결연한 의지를 띠고 있는 이현우의 얼굴 구도.\n\nLOCATION (lock): At the driver's position inside the old truck's compact cab. Strong sunlight enters through the side window and falls across the driver's face. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright sunlight entering through the truck window washes across 이현우's face with controlled highlights, restoring natural color after the monochrome memory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck continues along the road in daylight. 이현우: Back in the present, he remains in the driver's seat with his recent injuries. His expression has become resolute.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the center of the truck cabin, pointing forward and leftward toward the driver's seat. 이현우 occupies this seat and appears on the right side of the close-up frame, facing toward the left. A steering wheel is located immediately in front of him, partially occupying the lower edge. On the left side of the frame, a narrow edge of the side window is visible, allowing bright sunlight to shine directly onto his face. The empty passenger seat remains out of view behind and to the right of the camera. There are no mirrors or reflective surfaces depicted in this scene.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차창 밖에서 비쳐 든 눈부신 햇살 아래, 입술을 굳게 다문 채 결연한 의지를 띠고 있는 이현우의 얼굴 구도.\n\nLOCATION (lock): At the driver's position inside the old truck's compact cab. Strong sunlight enters through the side window and falls across the driver's face. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright sunlight entering through the truck window washes across 이현우's face with controlled highlights, restoring natural color after the monochrome memory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck continues along the road in daylight. 이현우: Back in the present, he remains in the driver's seat with his recent injuries. His expression has become resolute.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh16_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh16_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 왼쪽의 창문 너머 도로를 향하고 있음.",
    "built_space": "트럭 운전석 내부. 화면 왼쪽에 빛이 들어오는 창문이 있고, 하단 전경에 스티어링 휠의 일부가 보이며 카메라가 조수석 위치에서 운전자를 바라보는 구도.",
    "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 흙과 핏자국이 묻은 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
    "hard_violations": [],
    "physics": "운전석에 앉아 양어깨에 긴장감이 들어간 자세를 취하고 있으며, 보이지 않는 손으로 운전대를 잡고 지탱하는 모습이 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선은 화면 왼쪽의 창문 너머를 향하고 있음.",
    "built_space": "트럭 운전석 내부. 왼쪽에 창문이 있고 카메라가 조수석 측에서 운전자를 바라보며, 하단 전경에 스티어링 휠이 위치함.",
    "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 오염된 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
    "hard_violations": [],
    "physics": "운전석에 앉아 왼손으로 스티어링 휠을 쥐고 있으나 손의 형태와 밀착된 모습이 다소 불분명함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍, 카메라 앵글, 그리고 창문을 통해 들어오는 빛의 방향을 매우 정확하게 구현했으며, 인물의 결연한 표정과 옷의 질감이 사실적입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 프레이밍과 조명 등 지시사항을 잘 따랐으나, 운전대를 잡고 있는 손의 형태와 운전대의 질감이 다소 어색하게 표현되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽의 창문 너머 도로를 향하고 있음.",
        "built_space": "트럭 운전석 내부. 화면 왼쪽에 빛이 들어오는 창문이 있고, 하단 전경에 스티어링 휠의 일부가 보이며 카메라가 조수석 위치에서 운전자를 바라보는 구도.",
        "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 흙과 핏자국이 묻은 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
        "hard_violations": [],
        "physics": "운전석에 앉아 양어깨에 긴장감이 들어간 자세를 취하고 있으며, 보이지 않는 손으로 운전대를 잡고 지탱하는 모습이 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 화면 왼쪽의 창문 너머를 향하고 있음.",
        "built_space": "트럭 운전석 내부. 왼쪽에 창문이 있고 카메라가 조수석 측에서 운전자를 바라보며, 하단 전경에 스티어링 휠이 위치함.",
        "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 오염된 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
        "hard_violations": [],
        "physics": "운전석에 앉아 왼손으로 스티어링 휠을 쥐고 있으나 손의 형태와 밀착된 모습이 다소 불분명함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍, 카메라 앵글, 그리고 창문을 통해 들어오는 빛의 방향을 매우 정확하게 구현했으며, 인물의 결연한 표정과 옷의 질감이 사실적입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 프레이밍과 조명 등 지시사항을 잘 따랐으나, 운전대를 잡고 있는 손의 형태와 운전대의 질감이 다소 어색하게 표현되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 왼쪽의 창문 너머 도로를 향하고 있음.",
        "built_space": "트럭 운전석 내부. 화면 왼쪽에 빛이 들어오는 창문이 있고, 하단 전경에 스티어링 휠의 일부가 보이며 카메라가 조수석 위치에서 운전자를 바라보는 구도.",
        "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 흙과 핏자국이 묻은 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
        "hard_violations": [],
        "physics": "운전석에 앉아 양어깨에 긴장감이 들어간 자세를 취하고 있으며, 보이지 않는 손으로 운전대를 잡고 지탱하는 모습이 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선은 화면 왼쪽의 창문 너머를 향하고 있음.",
        "built_space": "트럭 운전석 내부. 왼쪽에 창문이 있고 카메라가 조수석 측에서 운전자를 바라보며, 하단 전경에 스티어링 휠이 위치함.",
        "entities": "이현우(레퍼런스와 일치하는 젊은 아시아인 남성), 오염된 어두운 셔츠, 오른쪽 귀에 소형 인이어 무전기 착용.",
        "hard_violations": [],
        "physics": "운전석에 앉아 왼손으로 스티어링 휠을 쥐고 있으나 손의 형태와 밀착된 모습이 다소 불분명함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "얼굴이 더 크게 오른쪽에 배치되고 약한 올려다보기 각도가 살아 있어 우세하지만, 지정된 측면 접근 구도보다 정면에 가깝고 차창과 운전대가 지나치게 많이 보인다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "다문 입술과 결연한 도로 쪽 시선은 적절하지만, 얼굴 크기가 더 작고 시점도 정면에 가까워 얼굴 중심의 밀착된 측면 클로즈업에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈은 화면 왼쪽의 프레임 밖 전방을 향하며 카메라를 응시하지 않는다. 운전 중 도로를 주시하는 것으로 읽힌다. 다만 얼굴의 양쪽 눈과 앞면이 넓게 보여, 지정된 측면 시점보다 운전자 앞쪽에서 본 구도에 가깝다.",
        "built_space": "왼쪽에 운전석 측창 하나와 잠금장치 하나, 인물 뒤에 좌석 등받이 하나, 아래 전경에 운전대 하나가 보인다. 인물은 측창 옆 등받이 앞에 앉아 있어 운전석 배치는 성립하며 중복 설비나 불가능한 반사는 없다. 그러나 차창이 화면 왼쪽의 상당 부분을 차지하고 운전대도 크게 들어와, 좁은 창 가장자리만 남기라는 지시와 다르다. 첨부된 공간 자료는 평면도이므로 마모 상태의 정확한 연속성은 확인할 수 없다.",
        "entities": "인물은 한 명이며, 참고 인물과 대체로 부합하는 동아시아계의 10대 후반으로 보이는 젊은 남성이다. 짧고 헝클어진 검은 머리, 마른 체형, 어둡고 낡은 셔츠의 먼지와 혈흔, 얼굴과 목의 상처, 귀의 검은 소형 인이어 장치가 보인다. 국적은 외관만으로 확인할 수 없다. 입술은 닫혀 있으나 굳게 압착한 긴장감은 비교적 약하다. 바지는 구도 밖이다.",
        "hard_violations": [],
        "physics": "등 뒤에 좌석 등받이가 있고 상체는 자연스럽게 앉은 자세다. 아래 왼쪽의 손이 운전대 테두리를 실제로 감싸며, 운전대는 차량 조향 장치로 이어져 화면 아래로 잘린다. 인이어 장치는 귀에 끼워져 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않으며, 어깨와 팔의 자세도 운전 동작으로 가능하다."
       },
       {
        "label": "B",
        "direction": "눈과 코는 화면 왼쪽의 프레임 밖 전방을 향한다. 도로를 집중해서 보는 시선과 다문 입술이 결연한 상태를 전달한다. 다만 카메라는 운전자의 측면보다 앞쪽에 가까워, 중앙 틈의 조수석 쪽 가장자리에서 유지해야 할 측면 관찰 각도가 충분히 구현되지 않았다.",
        "built_space": "왼쪽 측창 하나, 그 뒤 기둥의 벨트 고정부, 인물 뒤 좌석 등받이 하나, 아래 전경 운전대 하나, 오른쪽 뒤 작은 창 일부가 보인다. 운전자는 측창 옆 좌석 앞에 정상적으로 위치하며 설비 중복이나 불가능한 반사는 없다. 머리 전체와 넓은 어깨, 큰 차창 면적까지 보여 지정된 밀착 클로즈업보다 여유가 많다. 평면도만으로 실제 이전 장면의 내장재와 마모 일치 여부는 확인할 수 없다.",
        "entities": "참고 인물과 대체로 부합하는 동아시아계의 젊은 남성 한 명이다. 10대 후반에 가까운 얼굴, 헝클어진 짧은 검은 머리, 어두운 낡은 셔츠, 흙먼지와 혈흔, 뺨과 목의 상처, 귀의 소형 검은 인이어 장치가 확인된다. 국적은 외관으로 판별할 수 없다. 입술과 눈썹의 긴장은 결연한 표정에 맞는다. 하의는 화면 밖이다.",
        "hard_violations": [],
        "physics": "좌석 등받이가 상체 뒤에 있고 목과 어깨는 정상적인 착석 자세로 연결된다. 팔은 운전대 방향으로 내려가지만 손과 접촉점은 화면 아래에 가려져 확인되지 않으며, 이 가림 자체는 물리적 오류가 아니다. 운전대는 차량에 설치된 부품으로 보이고 인이어 장치는 귀에 지지된다. 무지지 부유나 불가능한 관절 자세는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "얼굴이 더 크게 오른쪽에 배치되고 약한 올려다보기 각도가 살아 있어 우세하지만, 지정된 측면 접근 구도보다 정면에 가깝고 차창과 운전대가 지나치게 많이 보인다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "다문 입술과 결연한 도로 쪽 시선은 적절하지만, 얼굴 크기가 더 작고 시점도 정면에 가까워 얼굴 중심의 밀착된 측면 클로즈업에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈은 화면 왼쪽의 프레임 밖 전방을 향하며 카메라를 응시하지 않는다. 운전 중 도로를 주시하는 것으로 읽힌다. 다만 얼굴의 양쪽 눈과 앞면이 넓게 보여, 지정된 측면 시점보다 운전자 앞쪽에서 본 구도에 가깝다.",
        "built_space": "왼쪽에 운전석 측창 하나와 잠금장치 하나, 인물 뒤에 좌석 등받이 하나, 아래 전경에 운전대 하나가 보인다. 인물은 측창 옆 등받이 앞에 앉아 있어 운전석 배치는 성립하며 중복 설비나 불가능한 반사는 없다. 그러나 차창이 화면 왼쪽의 상당 부분을 차지하고 운전대도 크게 들어와, 좁은 창 가장자리만 남기라는 지시와 다르다. 첨부된 공간 자료는 평면도이므로 마모 상태의 정확한 연속성은 확인할 수 없다.",
        "entities": "인물은 한 명이며, 참고 인물과 대체로 부합하는 동아시아계의 10대 후반으로 보이는 젊은 남성이다. 짧고 헝클어진 검은 머리, 마른 체형, 어둡고 낡은 셔츠의 먼지와 혈흔, 얼굴과 목의 상처, 귀의 검은 소형 인이어 장치가 보인다. 국적은 외관만으로 확인할 수 없다. 입술은 닫혀 있으나 굳게 압착한 긴장감은 비교적 약하다. 바지는 구도 밖이다.",
        "hard_violations": [],
        "physics": "등 뒤에 좌석 등받이가 있고 상체는 자연스럽게 앉은 자세다. 아래 왼쪽의 손이 운전대 테두리를 실제로 감싸며, 운전대는 차량 조향 장치로 이어져 화면 아래로 잘린다. 인이어 장치는 귀에 끼워져 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않으며, 어깨와 팔의 자세도 운전 동작으로 가능하다."
       },
       {
        "label": "A",
        "direction": "눈과 코는 화면 왼쪽의 프레임 밖 전방을 향한다. 도로를 집중해서 보는 시선과 다문 입술이 결연한 상태를 전달한다. 다만 카메라는 운전자의 측면보다 앞쪽에 가까워, 중앙 틈의 조수석 쪽 가장자리에서 유지해야 할 측면 관찰 각도가 충분히 구현되지 않았다.",
        "built_space": "왼쪽 측창 하나, 그 뒤 기둥의 벨트 고정부, 인물 뒤 좌석 등받이 하나, 아래 전경 운전대 하나, 오른쪽 뒤 작은 창 일부가 보인다. 운전자는 측창 옆 좌석 앞에 정상적으로 위치하며 설비 중복이나 불가능한 반사는 없다. 머리 전체와 넓은 어깨, 큰 차창 면적까지 보여 지정된 밀착 클로즈업보다 여유가 많다. 평면도만으로 실제 이전 장면의 내장재와 마모 일치 여부는 확인할 수 없다.",
        "entities": "참고 인물과 대체로 부합하는 동아시아계의 젊은 남성 한 명이다. 10대 후반에 가까운 얼굴, 헝클어진 짧은 검은 머리, 어두운 낡은 셔츠, 흙먼지와 혈흔, 뺨과 목의 상처, 귀의 소형 검은 인이어 장치가 확인된다. 국적은 외관으로 판별할 수 없다. 입술과 눈썹의 긴장은 결연한 표정에 맞는다. 하의는 화면 밖이다.",
        "hard_violations": [],
        "physics": "좌석 등받이가 상체 뒤에 있고 목과 어깨는 정상적인 착석 자세로 연결된다. 팔은 운전대 방향으로 내려가지만 손과 접촉점은 화면 아래에 가려져 확인되지 않으며, 이 가림 자체는 물리적 오류가 아니다. 운전대는 차량에 설치된 부품으로 보이고 인이어 장치는 귀에 지지된다. 무지지 부유나 불가능한 관절 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "지시된 클로즈업 프레이밍, 카메라 앵글, 그리고 창문을 통해 들어오는 빛의 방향을 매우 정확하게 구현했으며, 인물의 결연한 표정과 옷의 질감이 사실적입니다."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "전반적인 프레이밍과 조명 등 지시사항을 잘 따랐으나, 운전대를 잡고 있는 손의 형태와 운전대의 질감이 다소 어색하게 표현되었습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S63sh16_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cac-cdee-7297-82bf-fc16bc31da33",
  "confined_fp": {
   "base_key": "confinedfp::388a94bc37e2",
   "apt_reason": "트럭 운전석 내부를 배경으로 하는 샷으로, 인물이 운전대 및 차창과 맺는 공간적 위치 관계가 시각적으로 정확하게 표현되어야 하기 때문입니다.",
   "fixed": false,
   "mismatches": []
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S63sh7"
  }
 },
 "S64sh7::signage": {
  "fp": "88aaaf42caa0be8b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ebf454b611ff66b2": {
  "subjects": [],
  "subject_text": "익산 한옥마을 폐창고 내부\n기계 부품을 보관하는 넓고 어두운 폐창고. 한 귀퉁이에 폐로봇 부품이 산더미처럼 쌓여 있고, 주변에는 보관 상자가 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L100",
  "scope_role": "location_interior",
  "scope_sha": "c3f9ef84225ccc1c"
 },
 "S64sh7::bgfirst_bg": {
  "input_fingerprint": "e7f92d496d6dcb5c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고사포를 겨눈 B-200의 곁에 털썩 주저앉은 찰리의 구도.\n\nLOCATION (lock): In a shadowed corner of the village's large abandoned machinery warehouse, near piles of broken robots. Daylight enters through the half-open door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the pair's open side, begin the descending crane move above seated shoulder height, looking diagonally downward with both seated bodies and the raised cannon visible. Place 찰리 on the left, caught as his weight lands beside B-200 on the right; 찰리 watches his companion while B-200 turns toward him and begins drawing his torso away. Let their changing bodily proximity carry the beat, preserving the established oblique axis and leaving the piled robots behind them as spatial context.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pile of broken robots (Broken robots heaped in the warehouse corner) — Individual bodies lie at differing angles behind the seated pair; used as Provides a layered background without overlapping either character's face; B-200's cannon (Raised and aimed toward 찰리) — Seen obliquely, with its aiming direction legible rather than pointed into the lens; used as Maintains the threatening counterpoint to 찰리's friendly proximity while occupying a small portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the warehouse corner dim, with restrained contrast that preserves readable eyes and precise mechanical contours without softening B-200's guarded presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고사포를 겨눈 B-200의 곁에 털썩 주저앉은 찰리의 구도.\n\nLOCATION (lock): In a shadowed corner of the village's large abandoned machinery warehouse, near piles of broken robots. Daylight enters through the half-open door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the pair's open side, begin the descending crane move above seated shoulder height, looking diagonally downward with both seated bodies and the raised cannon visible. Place 찰리 on the left, caught as his weight lands beside B-200 on the right; 찰리 watches his companion while B-200 turns toward him and begins drawing his torso away. Let their changing bodily proximity carry the beat, preserving the established oblique axis and leaving the piled robots behind them as spatial context.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pile of broken robots (Broken robots heaped in the warehouse corner) — Individual bodies lie at differing angles behind the seated pair; used as Provides a layered background without overlapping either character's face; B-200's cannon (Raised and aimed toward 찰리) — Seen obliquely, with its aiming direction legible rather than pointed into the lens; used as Maintains the threatening counterpoint to 찰리's friendly proximity while occupying a small portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the warehouse corner dim, with restrained contrast that preserves readable eyes and precise mechanical contours without softening B-200's guarded presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S64sh7__bgfirst_bg.png",
  "asset_id": "c2dd327e-3849-4eaf-b919-4fd0ae131e40",
  "input_asset_ids": [
   "b9b0de2b-7735-4510-b88f-1796fd4716b5",
   "53ccd349-7e78-4b37-a0ac-0aa4477c1416"
  ]
 },
 "S64sh7": {
  "input_fingerprint": "f38d355b49077251",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고사포를 겨눈 B-200의 곁에 털썩 주저앉은 찰리의 구도.\n\nLOCATION (lock): In a shadowed corner of the village's large abandoned machinery warehouse, near piles of broken robots. Daylight enters through the half-open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the pair's open side, begin the descending crane move above seated shoulder height, looking diagonally downward with both seated bodies and the raised cannon visible. Place 찰리 on the left, caught as his weight lands beside B-200 on the right; 찰리 watches his companion while B-200 turns toward him and begins drawing his torso away. Let their changing bodily proximity carry the beat, preserving the established oblique axis and leaving the piled robots behind them as spatial context.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pile of broken robots (Broken robots heaped in the warehouse corner) — Individual bodies lie at differing angles behind the seated pair; used as Provides a layered background without overlapping either character's face; B-200's cannon (Raised and aimed toward 찰리) — Seen obliquely, with its aiming direction legible rather than pointed into the lens; used as Maintains the threatening counterpoint to 찰리's friendly proximity while occupying a small portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the warehouse corner dim, with restrained contrast that preserves readable eyes and precise mechanical contours without softening B-200's guarded presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door is half open, with machine parts and a heap of broken robots inside; baby birds are already concealed behind a box. Charlie sits with his battle-damaged body and faulty chest ring and language system, while B-200 remains equipped with weaponized gun-hands.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고사포를 겨눈 B-200의 곁에 털썩 주저앉은 찰리의 구도.\n\nLOCATION (lock): In a shadowed corner of the village's large abandoned machinery warehouse, near piles of broken robots. Daylight enters through the half-open door. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the pair's open side, begin the descending crane move above seated shoulder height, looking diagonally downward with both seated bodies and the raised cannon visible. Place 찰리 on the left, caught as his weight lands beside B-200 on the right; 찰리 watches his companion while B-200 turns toward him and begins drawing his torso away. Let their changing bodily proximity carry the beat, preserving the established oblique axis and leaving the piled robots behind them as spatial context.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pile of broken robots (Broken robots heaped in the warehouse corner) — Individual bodies lie at differing angles behind the seated pair; used as Provides a layered background without overlapping either character's face; B-200's cannon (Raised and aimed toward 찰리) — Seen obliquely, with its aiming direction legible rather than pointed into the lens; used as Maintains the threatening counterpoint to 찰리's friendly proximity while occupying a small portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the warehouse corner dim, with restrained contrast that preserves readable eyes and precise mechanical contours without softening B-200's guarded presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door is half open, with machine parts and a heap of broken robots inside; baby birds are already concealed behind a box. Charlie sits with his battle-damaged body and faulty chest ring and language system, while B-200 remains equipped with weaponized gun-hands.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고사포를 겨눈 B-200의 곁에 털썩 주저앉은 찰리의 구도.\n\nLOCATION (lock): In a shadowed corner of the village's large abandoned machinery warehouse, near piles of broken robots. Daylight enters through the half-open door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the pair's open side, begin the descending crane move above seated shoulder height, looking diagonally downward with both seated bodies and the raised cannon visible. Place 찰리 on the left, caught as his weight lands beside B-200 on the right; 찰리 watches his companion while B-200 turns toward him and begins drawing his torso away. Let their changing bodily proximity carry the beat, preserving the established oblique axis and leaving the piled robots behind them as spatial context.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pile of broken robots (Broken robots heaped in the warehouse corner) — Individual bodies lie at differing angles behind the seated pair; used as Provides a layered background without overlapping either character's face; B-200's cannon (Raised and aimed toward 찰리) — Seen obliquely, with its aiming direction legible rather than pointed into the lens; used as Maintains the threatening counterpoint to 찰리's friendly proximity while occupying a small portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the warehouse corner dim, with restrained contrast that preserves readable eyes and precise mechanical contours without softening B-200's guarded presence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The warehouse door is half open, with machine parts and a heap of broken robots inside; baby birds are already concealed behind a box. Charlie sits with his battle-damaged body and faulty chest ring and language system, while B-200 remains equipped with weaponized gun-hands.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by B-200 right now, so B-200's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to B-200: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S64sh7__bgfirst_bg.png",
     "asset_id": "c2dd327e-3849-4eaf-b919-4fd0ae131e40",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S64sh7.png",
     "asset_id": "b9b0de2b-7735-4510-b88f-1796fd4716b5",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176962>",
     "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L100B01.png",
     "asset_id": "53ccd349-7e78-4b37-a0ac-0aa4477c1416",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176962>",
     "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "B-200의 우측 고사포 포신과 시선이 찰리를 향함. 찰리의 시선은 B-200을 바라봄.",
    "built_space": "창고 중앙. 우측에 열린 문 1개, 좌우에 부품 선반과 더미가 있음. 카메라는 지시와 달리 눈높이에서 정면 수평을 향하며 인물들은 텅 빈 배경을 등지고 앉아 있음.",
    "entities": "찰리(고릴라 체형, 샌드 베이지 장갑, 흰 마스크, 가슴 원자로) 일치. B-200(중장비형 실루엣, 철제 외장, 양손 고사포) 일치.",
    "hard_violations": [],
    "physics": "두 캐릭터 모두 엉덩이를 바닥에 대고 앉아 체중을 지지함. B-200의 들려진 우측 팔은 몸통에 연결되어 유지됨."
   },
   {
    "label": "B",
    "direction": "B-200의 우측 고사포 포신이 찰리의 상체를 정확히 겨눔. 찰리는 고개를 돌려 B-200을 응시함.",
    "built_space": "창고 구석. 좌측 반쯤 열린 문으로 빛이 들어오며, 인물들 바로 뒤에 부서진 기계 더미가 위치함. 카메라는 샷 텍스트가 요구한 하향 대각선 뷰를 정확히 형성함.",
    "entities": "찰리(고릴라 비율, 샌드 베이지 파츠, 흰 마스크) 외형 완벽 일치. B-200(거대한 체형, 짙은 회색 철제 장갑, 가슴의 UNIT 07 마킹 및 양손 고사포) 외형 완벽 일치.",
    "hard_violations": [],
    "physics": "찰리는 바닥에 앉아 양손을 짚어 자세를 지지함. B-200은 금속 상자 위에 앉아 양 다리로 하중을 버티며, 겨눈 우측 팔은 어깨 관절로 튼튼하게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시된 하향 대각선 크레인 뷰 카메라 앵글을 정확히 따랐으며, 인물 바로 뒤에 고철 더미를 배치해 프레이밍 지시와 캐릭터 디테일을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라가 지시된 하향 뷰가 아닌 눈높이에 수평으로 위치하며, 고철 더미가 인물 뒤가 아닌 측면에 배치되어 샷 텍스트의 뼈대를 놓쳤습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 우측 고사포 포신과 시선이 찰리를 향함. 찰리의 시선은 B-200을 바라봄.",
        "built_space": "창고 중앙. 우측에 열린 문 1개, 좌우에 부품 선반과 더미가 있음. 카메라는 지시와 달리 눈높이에서 정면 수평을 향하며 인물들은 텅 빈 배경을 등지고 앉아 있음.",
        "entities": "찰리(고릴라 체형, 샌드 베이지 장갑, 흰 마스크, 가슴 원자로) 일치. B-200(중장비형 실루엣, 철제 외장, 양손 고사포) 일치.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 엉덩이를 바닥에 대고 앉아 체중을 지지함. B-200의 들려진 우측 팔은 몸통에 연결되어 유지됨."
       },
       {
        "label": "B",
        "direction": "B-200의 우측 고사포 포신이 찰리의 상체를 정확히 겨눔. 찰리는 고개를 돌려 B-200을 응시함.",
        "built_space": "창고 구석. 좌측 반쯤 열린 문으로 빛이 들어오며, 인물들 바로 뒤에 부서진 기계 더미가 위치함. 카메라는 샷 텍스트가 요구한 하향 대각선 뷰를 정확히 형성함.",
        "entities": "찰리(고릴라 비율, 샌드 베이지 파츠, 흰 마스크) 외형 완벽 일치. B-200(거대한 체형, 짙은 회색 철제 장갑, 가슴의 UNIT 07 마킹 및 양손 고사포) 외형 완벽 일치.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉아 양손을 짚어 자세를 지지함. B-200은 금속 상자 위에 앉아 양 다리로 하중을 버티며, 겨눈 우측 팔은 어깨 관절로 튼튼하게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "명시된 하향 대각선 크레인 뷰 카메라 앵글을 정확히 따랐으며, 인물 바로 뒤에 고철 더미를 배치해 프레이밍 지시와 캐릭터 디테일을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라가 지시된 하향 뷰가 아닌 눈높이에 수평으로 위치하며, 고철 더미가 인물 뒤가 아닌 측면에 배치되어 샷 텍스트의 뼈대를 놓쳤습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 우측 고사포 포신과 시선이 찰리를 향함. 찰리의 시선은 B-200을 바라봄.",
        "built_space": "창고 중앙. 우측에 열린 문 1개, 좌우에 부품 선반과 더미가 있음. 카메라는 지시와 달리 눈높이에서 정면 수평을 향하며 인물들은 텅 빈 배경을 등지고 앉아 있음.",
        "entities": "찰리(고릴라 체형, 샌드 베이지 장갑, 흰 마스크, 가슴 원자로) 일치. B-200(중장비형 실루엣, 철제 외장, 양손 고사포) 일치.",
        "hard_violations": [],
        "physics": "두 캐릭터 모두 엉덩이를 바닥에 대고 앉아 체중을 지지함. B-200의 들려진 우측 팔은 몸통에 연결되어 유지됨."
       },
       {
        "label": "B",
        "direction": "B-200의 우측 고사포 포신이 찰리의 상체를 정확히 겨눔. 찰리는 고개를 돌려 B-200을 응시함.",
        "built_space": "창고 구석. 좌측 반쯤 열린 문으로 빛이 들어오며, 인물들 바로 뒤에 부서진 기계 더미가 위치함. 카메라는 샷 텍스트가 요구한 하향 대각선 뷰를 정확히 형성함.",
        "entities": "찰리(고릴라 비율, 샌드 베이지 파츠, 흰 마스크) 외형 완벽 일치. B-200(거대한 체형, 짙은 회색 철제 장갑, 가슴의 UNIT 07 마킹 및 양손 고사포) 외형 완벽 일치.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉아 양손을 짚어 자세를 지지함. B-200은 금속 상자 위에 앉아 양 다리로 하중을 버티며, 겨눈 우측 팔은 어깨 관절로 튼튼하게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "높은 사선 시점과 찰리를 향한 조준은 맞지만, 포신의 화면 비중이 크고 B-200이 몸을 빼는 동작과 뒤편의 부서진 로봇 더미가 부족하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "작은 비중의 조준 포신, 왼쪽 찰리와 오른쪽 B-200의 착석 관계, 뒤로 물러나는 상체와 장소 재현이 더 충실하지만 로봇 잔해 더미와 털썩 앉는 순간성은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 얼굴을 오른쪽 위로 돌려 B-200을 바라보고, B-200은 고개를 왼쪽 아래의 찰리 쪽으로 돌린다. 들어 올린 포신은 왼쪽 아래로 뻗어 찰리의 어깨와 상체 쪽을 겨누며 렌즈 정면을 겨누지는 않는다. 반대쪽 포신은 화면 오른쪽 아래 바닥 쪽을 향한다.",
        "built_space": "왼쪽 위에 일부 열린 출입구 하나, 뒤편과 왼쪽에 팔레트 위 기계 부품 더미, 오른쪽 뒤에 금속 선반이 보인다. 콘크리트 바닥과 낡은 금속 부품은 장소 참고와 어울리지만, 참고의 넓은 중앙 통로와 뒤쪽 벽 구조는 잘 드러나지 않는다. 찰리는 왼쪽 바닥에, B-200은 오른쪽의 낮은 부품 받침 위에 앉아 있다. 카메라는 두 인물의 어깨보다 높은 곳에서 비스듬히 내려다보지만 인물과 포신의 화면 점유율이 비교적 크다. 뒤편 더미는 각기 다른 각도로 누운 로봇 몸체가 아니라 주로 산업용 부품으로 보인다.",
        "entities": "주요 인물은 기계 몸체 두 대뿐이다. 찰리의 베이지 장갑, 긴 팔과 짧은 다리, 흰 얼굴판, 주황색 점 눈 두 개와 입선, 파란 가슴 원형 장치가 맞는다. B-200의 짙은 회색 장갑, 주황색 눈, 탄흔, 양팔 일체형 다연장 포신과 더 큰 체격도 참고에 가깝다. 찰리의 표면 손상은 보이지만 가슴 장치의 고장은 명확하지 않으며 언어 장치 고장은 정지 화면으로 확인할 수 없다. 기계 부품은 있으나 요구된 로봇 시신 더미는 식별되지 않는다. 숨겨진 새끼 새는 노출되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 골반과 접힌 다리는 바닥에 놓이고 화면 왼쪽 손바닥이 체중을 받친다. B-200의 골반 아래에는 낮은 금속 받침이 있으며 두 발도 바닥에 닿는다. 양쪽 포신은 팔 관절과 전완에 연결되어 있어 떠 있는 물체는 없다. 찰리의 비대칭 자세는 내려앉은 직후로 읽을 수 있지만, B-200의 상체는 비교적 곧게 서 있어 찰리에게서 몸을 빼기 시작하는 움직임이 약하다."
       },
       {
        "label": "B",
        "direction": "찰리는 오른쪽의 B-200을 보고, B-200도 머리를 왼쪽 찰리 쪽으로 돌린다. 화면 왼쪽으로 뻗은 포신의 축은 찰리의 얼굴 아래와 목 부근을 향해 조준 관계가 읽힌다. 다른 포신은 무릎 위에서 왼쪽 앞쪽으로 향한다. 어느 포신도 렌즈 정면을 주된 표적으로 삼지 않는다.",
        "built_space": "왼쪽 위 창 구획 하나, 오른쪽 뒤의 일부 열린 셔터 출입구 하나, 중앙 뒤쪽 녹색 수납함 하나가 보인다. 왼쪽 기계 부품 적재대와 오른쪽 금속 선반, 낡은 벽의 수평 보강재 및 넓은 콘크리트 통로가 장소 참고와 잘 대응한다. 두 인물은 통로 왼쪽 부품 더미 옆 바닥에 나란히 앉아 있으며 찰리가 왼쪽이다. 카메라는 앉은 어깨 높이보다 위에서 사선으로 내려다보고, 두 몸 전체와 작은 비중의 포신을 포함하는 와이드 구도다. 다만 두 인물 뒤에는 부품 적재물이 있을 뿐, 서로 다른 각도로 누운 부서진 로봇 몸체들은 식별되지 않는다.",
        "entities": "찰리와 B-200에 해당하는 비인간 기계 두 대만 등장한다. 찰리의 베이지 장갑, 흰 마스크형 얼굴, 보이는 주황색 눈, 파란 원형 가슴 장치와 긴 팔이 참고에 부합한다. 얼굴이 옆으로 돌아 반대쪽 눈과 입선은 제한적으로 보인다. B-200은 회색 전신 장갑과 주황색 눈, 양팔의 포신을 갖추지만 참고보다 포신이 짧고 두 인물 사이의 체격 차이도 작다. 마모와 손상은 있으나 찰리 가슴 장치의 고장은 분명하지 않다. 기계 부품과 상자는 있지만 요구된 로봇 잔해 더미는 확인되지 않으며 숨겨진 새끼 새도 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 골반과 다리, 화면 왼쪽의 바닥에 짚은 손으로 지지된다. B-200도 골반과 양발이 바닥에 놓이고, 들어 올린 무기는 어깨와 팔꿈치 관절에 연결되어 지지된다. 몸이나 물체가 지지 없이 떠 있지는 않다. B-200의 상체가 찰리 반대편으로 약간 물러나 있어 방어적인 거리 조절이 A보다 잘 읽히지만, 찰리는 이미 안정적으로 앉아 있어 체중이 막 떨어지는 순간의 동세는 약하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "높은 사선 시점과 찰리를 향한 조준은 맞지만, 포신의 화면 비중이 크고 B-200이 몸을 빼는 동작과 뒤편의 부서진 로봇 더미가 부족하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "작은 비중의 조준 포신, 왼쪽 찰리와 오른쪽 B-200의 착석 관계, 뒤로 물러나는 상체와 장소 재현이 더 충실하지만 로봇 잔해 더미와 털썩 앉는 순간성은 부족하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 얼굴을 오른쪽 위로 돌려 B-200을 바라보고, B-200은 고개를 왼쪽 아래의 찰리 쪽으로 돌린다. 들어 올린 포신은 왼쪽 아래로 뻗어 찰리의 어깨와 상체 쪽을 겨누며 렌즈 정면을 겨누지는 않는다. 반대쪽 포신은 화면 오른쪽 아래 바닥 쪽을 향한다.",
        "built_space": "왼쪽 위에 일부 열린 출입구 하나, 뒤편과 왼쪽에 팔레트 위 기계 부품 더미, 오른쪽 뒤에 금속 선반이 보인다. 콘크리트 바닥과 낡은 금속 부품은 장소 참고와 어울리지만, 참고의 넓은 중앙 통로와 뒤쪽 벽 구조는 잘 드러나지 않는다. 찰리는 왼쪽 바닥에, B-200은 오른쪽의 낮은 부품 받침 위에 앉아 있다. 카메라는 두 인물의 어깨보다 높은 곳에서 비스듬히 내려다보지만 인물과 포신의 화면 점유율이 비교적 크다. 뒤편 더미는 각기 다른 각도로 누운 로봇 몸체가 아니라 주로 산업용 부품으로 보인다.",
        "entities": "주요 인물은 기계 몸체 두 대뿐이다. 찰리의 베이지 장갑, 긴 팔과 짧은 다리, 흰 얼굴판, 주황색 점 눈 두 개와 입선, 파란 가슴 원형 장치가 맞는다. B-200의 짙은 회색 장갑, 주황색 눈, 탄흔, 양팔 일체형 다연장 포신과 더 큰 체격도 참고에 가깝다. 찰리의 표면 손상은 보이지만 가슴 장치의 고장은 명확하지 않으며 언어 장치 고장은 정지 화면으로 확인할 수 없다. 기계 부품은 있으나 요구된 로봇 시신 더미는 식별되지 않는다. 숨겨진 새끼 새는 노출되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 골반과 접힌 다리는 바닥에 놓이고 화면 왼쪽 손바닥이 체중을 받친다. B-200의 골반 아래에는 낮은 금속 받침이 있으며 두 발도 바닥에 닿는다. 양쪽 포신은 팔 관절과 전완에 연결되어 있어 떠 있는 물체는 없다. 찰리의 비대칭 자세는 내려앉은 직후로 읽을 수 있지만, B-200의 상체는 비교적 곧게 서 있어 찰리에게서 몸을 빼기 시작하는 움직임이 약하다."
       },
       {
        "label": "A",
        "direction": "찰리는 오른쪽의 B-200을 보고, B-200도 머리를 왼쪽 찰리 쪽으로 돌린다. 화면 왼쪽으로 뻗은 포신의 축은 찰리의 얼굴 아래와 목 부근을 향해 조준 관계가 읽힌다. 다른 포신은 무릎 위에서 왼쪽 앞쪽으로 향한다. 어느 포신도 렌즈 정면을 주된 표적으로 삼지 않는다.",
        "built_space": "왼쪽 위 창 구획 하나, 오른쪽 뒤의 일부 열린 셔터 출입구 하나, 중앙 뒤쪽 녹색 수납함 하나가 보인다. 왼쪽 기계 부품 적재대와 오른쪽 금속 선반, 낡은 벽의 수평 보강재 및 넓은 콘크리트 통로가 장소 참고와 잘 대응한다. 두 인물은 통로 왼쪽 부품 더미 옆 바닥에 나란히 앉아 있으며 찰리가 왼쪽이다. 카메라는 앉은 어깨 높이보다 위에서 사선으로 내려다보고, 두 몸 전체와 작은 비중의 포신을 포함하는 와이드 구도다. 다만 두 인물 뒤에는 부품 적재물이 있을 뿐, 서로 다른 각도로 누운 부서진 로봇 몸체들은 식별되지 않는다.",
        "entities": "찰리와 B-200에 해당하는 비인간 기계 두 대만 등장한다. 찰리의 베이지 장갑, 흰 마스크형 얼굴, 보이는 주황색 눈, 파란 원형 가슴 장치와 긴 팔이 참고에 부합한다. 얼굴이 옆으로 돌아 반대쪽 눈과 입선은 제한적으로 보인다. B-200은 회색 전신 장갑과 주황색 눈, 양팔의 포신을 갖추지만 참고보다 포신이 짧고 두 인물 사이의 체격 차이도 작다. 마모와 손상은 있으나 찰리 가슴 장치의 고장은 분명하지 않다. 기계 부품과 상자는 있지만 요구된 로봇 잔해 더미는 확인되지 않으며 숨겨진 새끼 새도 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 골반과 다리, 화면 왼쪽의 바닥에 짚은 손으로 지지된다. B-200도 골반과 양발이 바닥에 놓이고, 들어 올린 무기는 어깨와 팔꿈치 관절에 연결되어 지지된다. 몸이나 물체가 지지 없이 떠 있지는 않다. B-200의 상체가 찰리 반대편으로 약간 물러나 있어 방어적인 거리 조절이 A보다 잘 읽히지만, 찰리는 이미 안정적으로 앉아 있어 체중이 막 떨어지는 순간의 동세는 약하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1857,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "명시된 하향 대각선 크레인 뷰 카메라 앵글을 정확히 따랐으며, 인물 바로 뒤에 고철 더미를 배치해 프레이밍 지시와 캐릭터 디테일을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "카메라가 지시된 하향 뷰가 아닌 눈높이에 수평으로 위치하며, 고철 더미가 인물 뒤가 아닌 측면에 배치되어 샷 텍스트의 뼈대를 놓쳤습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L100B01.png",
    "asset_id": "53ccd349-7e78-4b37-a0ac-0aa4477c1416",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1176962>",
    "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cb6-9efc-758a-a6a6-4f29aad3e6de",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S64sh7__bgfirst_bg.png",
   "bg_asset_id": "c2dd327e-3849-4eaf-b919-4fd0ae131e40",
   "bg_record_key": "S64sh7::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S64sh11::signage": {
  "fp": "40120c418eae7bef",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh11": {
  "input_fingerprint": "48d685cebb74b999",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200의 거친 금속 손가락이 찰리의 낡은 가슴 중앙에 달린 링 모양 장치를 가리키는 근접 구도.\n\nLOCATION (lock): In the same shadowed corner of the large machinery warehouse, away from the daylight at the half-open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach along the established oblique line, stopping close to 찰리's chest with the lens slightly above the ring and tilted downward. His worn chest occupies the left and central portions of the frame, while B-200's rough metal pointing finger enters from the right without covering the ring; retain 찰리's lowered chin at the upper edge to connect his downward attention to the device. Emphasize camera distance alone, keeping the ring smaller than the surrounding chest and rendering the gesture as direct observation rather than a subjective view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ring-shaped chest device (Attached to the center of 찰리's worn chest; described by 찰리 as apparently broken) — Its front and a shallow side edge are visible from above at an oblique angle; used as Small central focal anchor linking the pointing finger to 찰리's downward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim warehouse exposure and controlled contrast, separating the finger, worn chest, and ring without adding a glow to the device.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated with dents, holes and a faulty chest ring, while B-200 retains its converted gun-hands. The warehouse door remains half open, broken robots and parts remain piled inside, and the baby birds are still concealed behind the box.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200의 거친 금속 손가락이 찰리의 낡은 가슴 중앙에 달린 링 모양 장치를 가리키는 근접 구도.\n\nLOCATION (lock): In the same shadowed corner of the large machinery warehouse, away from the daylight at the half-open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach along the established oblique line, stopping close to 찰리's chest with the lens slightly above the ring and tilted downward. His worn chest occupies the left and central portions of the frame, while B-200's rough metal pointing finger enters from the right without covering the ring; retain 찰리's lowered chin at the upper edge to connect his downward attention to the device. Emphasize camera distance alone, keeping the ring smaller than the surrounding chest and rendering the gesture as direct observation rather than a subjective view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ring-shaped chest device (Attached to the center of 찰리's worn chest; described by 찰리 as apparently broken) — Its front and a shallow side edge are visible from above at an oblique angle; used as Small central focal anchor linking the pointing finger to 찰리's downward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim warehouse exposure and controlled contrast, separating the finger, worn chest, and ring without adding a glow to the device.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated with dents, holes and a faulty chest ring, while B-200 retains its converted gun-hands. The warehouse door remains half open, broken robots and parts remain piled inside, and the baby birds are still concealed behind the box.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): B-200의 거친 금속 손가락이 찰리의 낡은 가슴 중앙에 달린 링 모양 장치를 가리키는 근접 구도.\n\nLOCATION (lock): In the same shadowed corner of the large machinery warehouse, away from the daylight at the half-open doorway. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach along the established oblique line, stopping close to 찰리's chest with the lens slightly above the ring and tilted downward. His worn chest occupies the left and central portions of the frame, while B-200's rough metal pointing finger enters from the right without covering the ring; retain 찰리's lowered chin at the upper edge to connect his downward attention to the device. Emphasize camera distance alone, keeping the ring smaller than the surrounding chest and rendering the gesture as direct observation rather than a subjective view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ring-shaped chest device (Attached to the center of 찰리's worn chest; described by 찰리 as apparently broken) — Its front and a shallow side edge are visible from above at an oblique angle; used as Small central focal anchor linking the pointing finger to 찰리's downward attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim warehouse exposure and controlled contrast, separating the finger, worn chest, and ring without adding a glow to the device.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains seated with dents, holes and a faulty chest ring, while B-200 retains its converted gun-hands. The warehouse door remains half open, broken robots and parts remain piled inside, and the baby birds are still concealed behind the box.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "B-200의 손가락이 찰리의 가슴 중앙 링 장치를 향해 정확히 가리키고 있으며, 찰리의 시선도 아래쪽 장치를 향하고 있음.",
    "built_space": "어두운 창고 내부 배경으로 낡은 기계 부품과 상자들이 흐릿하게 보임. 조명은 차분하고 어두운 편임.",
    "entities": "찰리(베이지색 장갑, 하얀 마스크, 노란 눈, 가슴의 파손된 링)와 B-200(오른쪽에서 들어오는 거친 다크 그레이 금속 팔과 손가락)이 프롬프트 및 레퍼런스대로 잘 묘사됨.",
    "hard_violations": [],
    "physics": "찰리는 바닥에 앉은 자세로 다리 관절이 화면 하단에 보이며, B-200의 팔은 화면 밖 본체로부터 안정적으로 뻗어 나와 있음."
   },
   {
    "label": "B",
    "direction": "B-200의 손가락이 찰리의 가슴 링 장치를 향하고 있으며, 찰리의 시선은 아래쪽을 향함.",
    "built_space": "어두운 기계 창고 모서리로 배경에 흐릿한 구조물들이 배치되어 있음.",
    "entities": "찰리의 외형(마스크, 장갑판)과 B-200의 금속 팔은 일치하나, 링 장치 중앙에 프롬프트 지시와 달리 파란색 발광 효과가 강하게 나타남.",
    "hard_violations": [],
    "physics": "찰리는 바닥에 앉은 자세를 유지하며, B-200의 팔은 화면 우측에서 물리적 지지점을 가지고 자연스럽게 뻗어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트의 요구사항(근접 구도, 턱의 위치, 장치에 추가적인 발광 효과 배제)을 매우 충실히 이행했으며, 파손된 링의 질감과 금속 손가락의 묘사가 사실적입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물 외형은 잘 구현되었으나, 링 장치 중앙에 강한 발광 효과가 들어가 '광채를 더하지 말 것'이라는 조명 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 손가락이 찰리의 가슴 중앙 링 장치를 향해 정확히 가리키고 있으며, 찰리의 시선도 아래쪽 장치를 향하고 있음.",
        "built_space": "어두운 창고 내부 배경으로 낡은 기계 부품과 상자들이 흐릿하게 보임. 조명은 차분하고 어두운 편임.",
        "entities": "찰리(베이지색 장갑, 하얀 마스크, 노란 눈, 가슴의 파손된 링)와 B-200(오른쪽에서 들어오는 거친 다크 그레이 금속 팔과 손가락)이 프롬프트 및 레퍼런스대로 잘 묘사됨.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉은 자세로 다리 관절이 화면 하단에 보이며, B-200의 팔은 화면 밖 본체로부터 안정적으로 뻗어 나와 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 손가락이 찰리의 가슴 링 장치를 향하고 있으며, 찰리의 시선은 아래쪽을 향함.",
        "built_space": "어두운 기계 창고 모서리로 배경에 흐릿한 구조물들이 배치되어 있음.",
        "entities": "찰리의 외형(마스크, 장갑판)과 B-200의 금속 팔은 일치하나, 링 장치 중앙에 프롬프트 지시와 달리 파란색 발광 효과가 강하게 나타남.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉은 자세를 유지하며, B-200의 팔은 화면 우측에서 물리적 지지점을 가지고 자연스럽게 뻗어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트의 요구사항(근접 구도, 턱의 위치, 장치에 추가적인 발광 효과 배제)을 매우 충실히 이행했으며, 파손된 링의 질감과 금속 손가락의 묘사가 사실적입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 인물 외형은 잘 구현되었으나, 링 장치 중앙에 강한 발광 효과가 들어가 '광채를 더하지 말 것'이라는 조명 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "B-200의 손가락이 찰리의 가슴 중앙 링 장치를 향해 정확히 가리키고 있으며, 찰리의 시선도 아래쪽 장치를 향하고 있음.",
        "built_space": "어두운 창고 내부 배경으로 낡은 기계 부품과 상자들이 흐릿하게 보임. 조명은 차분하고 어두운 편임.",
        "entities": "찰리(베이지색 장갑, 하얀 마스크, 노란 눈, 가슴의 파손된 링)와 B-200(오른쪽에서 들어오는 거친 다크 그레이 금속 팔과 손가락)이 프롬프트 및 레퍼런스대로 잘 묘사됨.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉은 자세로 다리 관절이 화면 하단에 보이며, B-200의 팔은 화면 밖 본체로부터 안정적으로 뻗어 나와 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 손가락이 찰리의 가슴 링 장치를 향하고 있으며, 찰리의 시선은 아래쪽을 향함.",
        "built_space": "어두운 기계 창고 모서리로 배경에 흐릿한 구조물들이 배치되어 있음.",
        "entities": "찰리의 외형(마스크, 장갑판)과 B-200의 금속 팔은 일치하나, 링 장치 중앙에 프롬프트 지시와 달리 파란색 발광 효과가 강하게 나타남.",
        "hard_violations": [],
        "physics": "찰리는 바닥에 앉은 자세를 유지하며, B-200의 팔은 화면 우측에서 물리적 지지점을 가지고 자연스럽게 뻗어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽 금속 손가락이 링을 가리지 않고 정확히 가리키지만, 복부와 무릎까지 상대적으로 많이 포함하고 링 중앙의 푸른 발광이 무발광 지시에서 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "가슴 중심의 사선 근접 구도, 내려간 턱, 링을 향한 손끝과 억제된 장치 밝기가 더 충실하지만, 상단에는 턱뿐 아니라 얼굴 상당 부분도 들어온다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽에서 들어온 금속 검지가 왼쪽 아래로 뻗어 링의 오른쪽 가장자리와 내부를 향한다. 손끝은 링 바깥에 멈춰 장치 전면을 가리지 않는다. 찰리의 얼굴은 아래쪽으로 기울어 있지만, 발광하는 점 형태의 눈만으로 정확한 시선 도착점을 확인하기는 어렵다.",
        "built_space": "찰리의 가슴이 왼쪽과 중앙에 있고 오른쪽 전경에는 B-200의 손 일부가 있다. 배경에 어두운 기계 부품과 수직 구조물이 흐리게 보인다. 문과 고정 설비는 근접 크롭 밖이어서 개수와 배치를 검증할 수 없으며, 보이는 범위에서 중복 설비나 불가능한 반사는 없다. 링의 전면과 가장자리 두께가 위쪽 사선 시점으로 보이지만 복부와 무릎도 상당히 포함된다.",
        "entities": "찰리의 샌드 베이지 장갑, 흰 마스크, 주황색 점 눈과 선 모양 입이 참조와 부합한다. 가슴에는 구멍과 마모가 있고 중앙 링 한 개에는 깨진 유리와 푸른 중심광이 보인다. B-200의 노출 부분은 탄흔과 마모가 있는 짙은 금속 관절 손이다. 참조의 다연장 포신은 이 크롭에서 확인되지 않는다. 추가 인물, 새, 상자나 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "링은 가슴 장갑에 매립·체결되어 지지된다. 검지는 마디를 통해 오른쪽 손과 이어지고 나머지 손가락은 굽혀져 있어 지시 동작이 기계적으로 가능하다. 찰리의 머리와 팔도 목과 어깨 관절에 연결되어 있다. 앉은 하체 일부는 보이지만 바닥 접촉점은 프레임 밖이며, 떠 있거나 지지 없이 분리된 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "오른쪽 위에서 들어온 금속 검지가 왼쪽 아래의 가슴 링 내부를 향한다. 손끝과 링 오른쪽 테두리 사이에 간격이 있어 장치를 가리지 않는다. 찰리의 턱과 얼굴은 가슴 쪽으로 내려가 있어 장치를 내려다보는 관계가 읽히지만 점 눈의 세부 시선은 확정하기 어렵다.",
        "built_space": "낡은 가슴 장갑이 왼쪽과 중앙을 채우고 B-200의 손이 오른쪽에 배치된다. 링 한 개의 전면과 얕은 측면 테두리가 사선으로 보인다. 뒤쪽에는 어두운 창고의 선반형 구조와 부품이 흐려져 있으며 문이나 개별 고정 설비의 수는 이 크롭에서 검증할 수 없다. 중복 구조물이나 불가능한 반사는 보이지 않는다. 가슴 중심 구도가 유지되지만 상단에는 턱 외에 눈과 얼굴 일부도 포함된다.",
        "entities": "찰리의 베이지색 각진 장갑, 흰 마스크, 주황색 점 눈과 선 입이 참조의 기계 정체성을 유지한다. 중앙 링은 금속 테두리와 균열 난 어두운 유리로 표현되며 푸른 흔적은 있으나 뚜렷한 주변 발광은 없다. B-200의 손은 거칠고 탄흔이 있는 짙은 회색 금속으로 맞지만, 참조의 포신형 손 유지 여부는 보이는 부분만으로 확인할 수 없다. 추가 인물이나 새 문구는 없다.",
        "hard_violations": [],
        "physics": "링은 가슴 중앙의 장착부에 고정되어 있고 손끝은 장치와 접촉하지 않은 채 가리킨다. 검지와 굽힌 손가락은 오른쪽 손 구조에 연속적으로 연결되어 있어 지지 없는 부품은 없다. 머리는 목 관절에, 팔은 어깨에 연결된다. 하단의 굽힌 다리 일부는 앉은 자세와 양립하며 실제 바닥 지지는 크롭 밖이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽 금속 손가락이 링을 가리지 않고 정확히 가리키지만, 복부와 무릎까지 상대적으로 많이 포함하고 링 중앙의 푸른 발광이 무발광 지시에서 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "가슴 중심의 사선 근접 구도, 내려간 턱, 링을 향한 손끝과 억제된 장치 밝기가 더 충실하지만, 상단에는 턱뿐 아니라 얼굴 상당 부분도 들어온다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽에서 들어온 금속 검지가 왼쪽 아래로 뻗어 링의 오른쪽 가장자리와 내부를 향한다. 손끝은 링 바깥에 멈춰 장치 전면을 가리지 않는다. 찰리의 얼굴은 아래쪽으로 기울어 있지만, 발광하는 점 형태의 눈만으로 정확한 시선 도착점을 확인하기는 어렵다.",
        "built_space": "찰리의 가슴이 왼쪽과 중앙에 있고 오른쪽 전경에는 B-200의 손 일부가 있다. 배경에 어두운 기계 부품과 수직 구조물이 흐리게 보인다. 문과 고정 설비는 근접 크롭 밖이어서 개수와 배치를 검증할 수 없으며, 보이는 범위에서 중복 설비나 불가능한 반사는 없다. 링의 전면과 가장자리 두께가 위쪽 사선 시점으로 보이지만 복부와 무릎도 상당히 포함된다.",
        "entities": "찰리의 샌드 베이지 장갑, 흰 마스크, 주황색 점 눈과 선 모양 입이 참조와 부합한다. 가슴에는 구멍과 마모가 있고 중앙 링 한 개에는 깨진 유리와 푸른 중심광이 보인다. B-200의 노출 부분은 탄흔과 마모가 있는 짙은 금속 관절 손이다. 참조의 다연장 포신은 이 크롭에서 확인되지 않는다. 추가 인물, 새, 상자나 문구는 보이지 않는다.",
        "hard_violations": [],
        "physics": "링은 가슴 장갑에 매립·체결되어 지지된다. 검지는 마디를 통해 오른쪽 손과 이어지고 나머지 손가락은 굽혀져 있어 지시 동작이 기계적으로 가능하다. 찰리의 머리와 팔도 목과 어깨 관절에 연결되어 있다. 앉은 하체 일부는 보이지만 바닥 접촉점은 프레임 밖이며, 떠 있거나 지지 없이 분리된 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "오른쪽 위에서 들어온 금속 검지가 왼쪽 아래의 가슴 링 내부를 향한다. 손끝과 링 오른쪽 테두리 사이에 간격이 있어 장치를 가리지 않는다. 찰리의 턱과 얼굴은 가슴 쪽으로 내려가 있어 장치를 내려다보는 관계가 읽히지만 점 눈의 세부 시선은 확정하기 어렵다.",
        "built_space": "낡은 가슴 장갑이 왼쪽과 중앙을 채우고 B-200의 손이 오른쪽에 배치된다. 링 한 개의 전면과 얕은 측면 테두리가 사선으로 보인다. 뒤쪽에는 어두운 창고의 선반형 구조와 부품이 흐려져 있으며 문이나 개별 고정 설비의 수는 이 크롭에서 검증할 수 없다. 중복 구조물이나 불가능한 반사는 보이지 않는다. 가슴 중심 구도가 유지되지만 상단에는 턱 외에 눈과 얼굴 일부도 포함된다.",
        "entities": "찰리의 베이지색 각진 장갑, 흰 마스크, 주황색 점 눈과 선 입이 참조의 기계 정체성을 유지한다. 중앙 링은 금속 테두리와 균열 난 어두운 유리로 표현되며 푸른 흔적은 있으나 뚜렷한 주변 발광은 없다. B-200의 손은 거칠고 탄흔이 있는 짙은 회색 금속으로 맞지만, 참조의 포신형 손 유지 여부는 보이는 부분만으로 확인할 수 없다. 추가 인물이나 새 문구는 없다.",
        "hard_violations": [],
        "physics": "링은 가슴 중앙의 장착부에 고정되어 있고 손끝은 장치와 접촉하지 않은 채 가리킨다. 검지와 굽힌 손가락은 오른쪽 손 구조에 연속적으로 연결되어 있어 지지 없는 부품은 없다. 머리는 목 관절에, 팔은 어깨에 연결된다. 하단의 굽힌 다리 일부는 앉은 자세와 양립하며 실제 바닥 지지는 크롭 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.625
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트의 요구사항(근접 구도, 턱의 위치, 장치에 추가적인 발광 효과 배제)을 매우 충실히 이행했으며, 파손된 링의 질감과 금속 손가락의 묘사가 사실적입니다."
   },
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "구도와 인물 외형은 잘 구현되었으나, 링 장치 중앙에 강한 발광 효과가 들어가 '광채를 더하지 말 것'이라는 조명 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of B-200, 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S64sh7_sel.png",
    "asset_id": "a68b5ff0-ed4f-4526-a55b-032c1be0fe87",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1176962>",
    "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cc0-4c50-719f-aacf-2b1421a98bd8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh7"
  }
 },
 "S64sh17::signage": {
  "fp": "943ada81e196fb71",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh17": {
  "input_fingerprint": "9d2fc3fe74bb24f6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아기 새들을 내려다보며 부드러운 눈빛을 띠는 B-200의 굽은 상체.\n\nLOCATION (lock): At the concealed box of hatchlings in the abandoned machinery warehouse's dark corner. Faint daylight reaches in from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the box and settle the upward tilt below B-200's face, framing his bent upper body from a low three-quarter side angle. Place his curved torso on the right and softened eyes above center, with his gaze descending toward the baby birds just below the left frame edge, clearly away from the lens. Preserve the camera's planted position and let the redirected viewing angle reveal tenderness, with only a small edge of the shifted box retained below.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shifted box (Moved aside to reveal the hidden baby birds) — Only an oblique edge remains visible at the lower frame boundary; used as Anchors the camera beside the birds without obstructing B-200's bent torso; Broken robot pile (Still heaped in the warehouse corner) — Fragmentary silhouettes sit behind B-200 at varied angles; used as Subordinate background context contrasting with his careful, living attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the corner's dim ambient light and restrained contrast, allowing tenderness to register through the eyes and posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The box has been moved aside, exposing the baby birds hidden behind it; B-200 still has weaponized gun-hands. The warehouse door remains half open amid the machinery and broken-robot piles, and Charlie has left the interior with his battle damage unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아기 새들을 내려다보며 부드러운 눈빛을 띠는 B-200의 굽은 상체.\n\nLOCATION (lock): At the concealed box of hatchlings in the abandoned machinery warehouse's dark corner. Faint daylight reaches in from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the box and settle the upward tilt below B-200's face, framing his bent upper body from a low three-quarter side angle. Place his curved torso on the right and softened eyes above center, with his gaze descending toward the baby birds just below the left frame edge, clearly away from the lens. Preserve the camera's planted position and let the redirected viewing angle reveal tenderness, with only a small edge of the shifted box retained below.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shifted box (Moved aside to reveal the hidden baby birds) — Only an oblique edge remains visible at the lower frame boundary; used as Anchors the camera beside the birds without obstructing B-200's bent torso; Broken robot pile (Still heaped in the warehouse corner) — Fragmentary silhouettes sit behind B-200 at varied angles; used as Subordinate background context contrasting with his careful, living attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the corner's dim ambient light and restrained contrast, allowing tenderness to register through the eyes and posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The box has been moved aside, exposing the baby birds hidden behind it; B-200 still has weaponized gun-hands. The warehouse door remains half open amid the machinery and broken-robot piles, and Charlie has left the interior with his battle damage unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아기 새들을 내려다보며 부드러운 눈빛을 띠는 B-200의 굽은 상체.\n\nLOCATION (lock): At the concealed box of hatchlings in the abandoned machinery warehouse's dark corner. Faint daylight reaches in from the half-open entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Remain beside the box and settle the upward tilt below B-200's face, framing his bent upper body from a low three-quarter side angle. Place his curved torso on the right and softened eyes above center, with his gaze descending toward the baby birds just below the left frame edge, clearly away from the lens. Preserve the camera's planted position and let the redirected viewing angle reveal tenderness, with only a small edge of the shifted box retained below.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shifted box (Moved aside to reveal the hidden baby birds) — Only an oblique edge remains visible at the lower frame boundary; used as Anchors the camera beside the birds without obstructing B-200's bent torso; Broken robot pile (Still heaped in the warehouse corner) — Fragmentary silhouettes sit behind B-200 at varied angles; used as Subordinate background context contrasting with his careful, living attention.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the corner's dim ambient light and restrained contrast, allowing tenderness to register through the eyes and posture rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The box has been moved aside, exposing the baby birds hidden behind it; B-200 still has weaponized gun-hands. The warehouse door remains half open amid the machinery and broken-robot piles, and Charlie has left the interior with his battle damage unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 시선이 화면 좌측 하단의 상자 모서리를 향하고 있음.",
    "built_space": "폐창고 배경. 좌측에 열린 문으로 빛이 들어오며, 로봇 뒤편으로 기계 부품 더미가 배치되어 있음.",
    "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)이 존재하며 레퍼런스와 일치함. 하단에 상자 모서리가 있으나 아기 새들은 매우 흐릿함.",
    "hard_violations": [],
    "physics": "무릎을 굽히고 상체를 숙인 상태로 지면에 발을 딛고 안정적으로 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "로봇의 시선이 화면 좌측 하단 상자 안의 아기 새들을 명확히 향하고 있음.",
    "built_space": "폐창고 내부에 위치. 좌측의 문과 빛, 배경의 고철 부품 더미가 적절한 위치와 스케일로 표현됨.",
    "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)의 외형이 정확하며, 좌측 하단 상자 안에 여러 마리의 아기 새들이 뚜렷하게 묘사됨.",
    "hard_violations": [],
    "physics": "앞으로 몸을 깊숙이 숙인 채 두 발로 바닥을 딛고 자세를 유지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라 구도와 캐릭터 디자인은 준수하나, 화면 좌측 하단에 명시된 아기 새들의 형태가 흐릿하고 알아보기 어렵습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 프레이밍과 캐릭터의 시선 처리를 잘 구현했으며, 화면 좌측 하단의 상자와 아기 새들을 명확하게 표현하여 지시사항에 매우 충실합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선이 화면 좌측 하단의 상자 모서리를 향하고 있음.",
        "built_space": "폐창고 배경. 좌측에 열린 문으로 빛이 들어오며, 로봇 뒤편으로 기계 부품 더미가 배치되어 있음.",
        "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)이 존재하며 레퍼런스와 일치함. 하단에 상자 모서리가 있으나 아기 새들은 매우 흐릿함.",
        "hard_violations": [],
        "physics": "무릎을 굽히고 상체를 숙인 상태로 지면에 발을 딛고 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 시선이 화면 좌측 하단 상자 안의 아기 새들을 명확히 향하고 있음.",
        "built_space": "폐창고 내부에 위치. 좌측의 문과 빛, 배경의 고철 부품 더미가 적절한 위치와 스케일로 표현됨.",
        "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)의 외형이 정확하며, 좌측 하단 상자 안에 여러 마리의 아기 새들이 뚜렷하게 묘사됨.",
        "hard_violations": [],
        "physics": "앞으로 몸을 깊숙이 숙인 채 두 발로 바닥을 딛고 자세를 유지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라 구도와 캐릭터 디자인은 준수하나, 화면 좌측 하단에 명시된 아기 새들의 형태가 흐릿하고 알아보기 어렵습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 로우 앵글 프레이밍과 캐릭터의 시선 처리를 잘 구현했으며, 화면 좌측 하단의 상자와 아기 새들을 명확하게 표현하여 지시사항에 매우 충실합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 시선이 화면 좌측 하단의 상자 모서리를 향하고 있음.",
        "built_space": "폐창고 배경. 좌측에 열린 문으로 빛이 들어오며, 로봇 뒤편으로 기계 부품 더미가 배치되어 있음.",
        "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)이 존재하며 레퍼런스와 일치함. 하단에 상자 모서리가 있으나 아기 새들은 매우 흐릿함.",
        "hard_violations": [],
        "physics": "무릎을 굽히고 상체를 숙인 상태로 지면에 발을 딛고 안정적으로 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "로봇의 시선이 화면 좌측 하단 상자 안의 아기 새들을 명확히 향하고 있음.",
        "built_space": "폐창고 내부에 위치. 좌측의 문과 빛, 배경의 고철 부품 더미가 적절한 위치와 스케일로 표현됨.",
        "entities": "B-200(기관총 팔을 가진 짙은 회색 로봇)의 외형이 정확하며, 좌측 하단 상자 안에 여러 마리의 아기 새들이 뚜렷하게 묘사됨.",
        "hard_violations": [],
        "physics": "앞으로 몸을 깊숙이 숙인 채 두 발로 바닥을 딛고 자세를 유지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "굽은 상체와 아래로 향한 시선은 맞지만, 새끼 새들과 둥지를 전경에 크게 드러내어 새들을 왼쪽 화면 밖에 두라는 핵심 구도에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽의 굽은 상체와 낮은 사선 시점, 하단에 대부분 잘린 새들이 지정 구도에 더 가깝지만, 상자 가장자리의 비중과 입구의 밝기는 다소 큽니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200은 고개와 주황색 광학 눈을 왼쪽 아래의 보이는 새끼 새들에게 향하고 있으며 렌즈를 보지 않습니다. 새들은 부리를 위로 들어 로봇 쪽을 향합니다. 양팔 포신은 전경 하단으로 내려가며, 새들을 직접 겨누는 모습은 아닙니다.",
        "built_space": "왼쪽에 일부 열린 셔터 출입구 하나, 중앙 뒤에 여러 층으로 쌓인 낡은 기계 부품, 오른쪽 뒤에 어두운 적재 공간이 보입니다. 낮은 카메라 앞 하단에는 상자 테두리와 둥지, 새끼 새 세 마리가 상당 부분 노출됩니다. 로봇의 상체는 오른쪽에 있지만, 상자의 작은 사선 모서리만 남기라는 지시보다 전경 정보가 많습니다. 창고의 금속 구조와 마모는 참조와 유사하나 입구의 채광은 더 두드러집니다.",
        "entities": "활동하는 인물은 B-200 하나뿐이며 찰리는 없습니다. 인간 피부나 눈 대신 금속 얼굴과 주황색 광학 눈을 갖추고, 짙은 회색 중장갑과 탄흔, 양쪽 다연장 포신, 가슴의 기존 계열 표식이 참조의 정체성을 유지합니다. 새끼 새와 둥지는 식별 가능하지만, 상자를 치워 그 뒤의 새들을 드러냈다기보다 상자 안에 둥지가 있는 것처럼 보입니다.",
        "hard_violations": [],
        "physics": "상체는 허리와 골반, 하단에 일부 보이는 다리로 이어져 있으며 앞으로 굽힌 관절 배치가 가능합니다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만 공중에 떠 있는 형상은 아닙니다. 포신은 팔 관절에 붙어 있고, 새들은 상자 테두리 뒤 둥지 재료에 몸을 기대고 있습니다. 배경 부품은 적재물과 받침에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "B-200의 얼굴과 광학 눈은 왼쪽 아래, 하단에 일부만 보이는 새끼 새들 쪽으로 기울어져 있으며 관객을 응시하지 않습니다. 두 포신은 새들 양옆의 화면 하단을 향하고, 새를 명시적으로 조준하지 않습니다. 새들은 대부분 잘려 개별 시선은 판별하기 어렵습니다.",
        "built_space": "왼쪽에 일부 열린 셔터 출입구 하나가 있고, 그 안쪽 중앙에는 기계 부품 더미와 하단 받침이 보이며 오른쪽 뒤에는 어두운 선반이 있습니다. 낮은 삼사분 측면 시점에서 굽은 몸통이 오른쪽, 눈이 중앙 위에 놓입니다. 새들은 하단 경계에 대부분 잘리고 상자 테두리가 비스듬히 지나가므로 A보다 지정 구도에 가깝습니다. 다만 테두리는 하단을 넓게 가로지르고 입구의 낮빛도 요청한 희미한 채광보다 강합니다.",
        "entities": "B-200 하나만 등장하며 다른 인물이나 찰리는 없습니다. 육중한 기계 체형, 금속 안면, 주황색 광학 눈, 탄흔 있는 암회색 장갑, 양팔의 굵은 포신이 참조와 부합합니다. 하단에는 흐릿한 새끼 새들의 머리와 둥지 재료 일부가 보입니다. 상자 뒤에 노출된 둥지인지 상자 내부인지, 이 잘린 구도만으로는 명확히 확인하기 어렵습니다.",
        "hard_violations": [],
        "physics": "굽힌 상체는 기계식 허리와 골반, 하단의 다리 부분에 연결되어 자연스러운 전방 굴곡으로 읽힙니다. 발의 지지점은 프레임 밖이지만 몸이 떠 있다는 징후는 없습니다. 양쪽 포신은 팔에 구조적으로 결합되어 있고, 새들의 보이는 부분은 하단 둥지 재료에 받쳐져 있습니다. 배경의 무거운 부품들은 더미와 받침 위에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "굽은 상체와 아래로 향한 시선은 맞지만, 새끼 새들과 둥지를 전경에 크게 드러내어 새들을 왼쪽 화면 밖에 두라는 핵심 구도에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽의 굽은 상체와 낮은 사선 시점, 하단에 대부분 잘린 새들이 지정 구도에 더 가깝지만, 상자 가장자리의 비중과 입구의 밝기는 다소 큽니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "B-200은 고개와 주황색 광학 눈을 왼쪽 아래의 보이는 새끼 새들에게 향하고 있으며 렌즈를 보지 않습니다. 새들은 부리를 위로 들어 로봇 쪽을 향합니다. 양팔 포신은 전경 하단으로 내려가며, 새들을 직접 겨누는 모습은 아닙니다.",
        "built_space": "왼쪽에 일부 열린 셔터 출입구 하나, 중앙 뒤에 여러 층으로 쌓인 낡은 기계 부품, 오른쪽 뒤에 어두운 적재 공간이 보입니다. 낮은 카메라 앞 하단에는 상자 테두리와 둥지, 새끼 새 세 마리가 상당 부분 노출됩니다. 로봇의 상체는 오른쪽에 있지만, 상자의 작은 사선 모서리만 남기라는 지시보다 전경 정보가 많습니다. 창고의 금속 구조와 마모는 참조와 유사하나 입구의 채광은 더 두드러집니다.",
        "entities": "활동하는 인물은 B-200 하나뿐이며 찰리는 없습니다. 인간 피부나 눈 대신 금속 얼굴과 주황색 광학 눈을 갖추고, 짙은 회색 중장갑과 탄흔, 양쪽 다연장 포신, 가슴의 기존 계열 표식이 참조의 정체성을 유지합니다. 새끼 새와 둥지는 식별 가능하지만, 상자를 치워 그 뒤의 새들을 드러냈다기보다 상자 안에 둥지가 있는 것처럼 보입니다.",
        "hard_violations": [],
        "physics": "상체는 허리와 골반, 하단에 일부 보이는 다리로 이어져 있으며 앞으로 굽힌 관절 배치가 가능합니다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만 공중에 떠 있는 형상은 아닙니다. 포신은 팔 관절에 붙어 있고, 새들은 상자 테두리 뒤 둥지 재료에 몸을 기대고 있습니다. 배경 부품은 적재물과 받침에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "B-200의 얼굴과 광학 눈은 왼쪽 아래, 하단에 일부만 보이는 새끼 새들 쪽으로 기울어져 있으며 관객을 응시하지 않습니다. 두 포신은 새들 양옆의 화면 하단을 향하고, 새를 명시적으로 조준하지 않습니다. 새들은 대부분 잘려 개별 시선은 판별하기 어렵습니다.",
        "built_space": "왼쪽에 일부 열린 셔터 출입구 하나가 있고, 그 안쪽 중앙에는 기계 부품 더미와 하단 받침이 보이며 오른쪽 뒤에는 어두운 선반이 있습니다. 낮은 삼사분 측면 시점에서 굽은 몸통이 오른쪽, 눈이 중앙 위에 놓입니다. 새들은 하단 경계에 대부분 잘리고 상자 테두리가 비스듬히 지나가므로 A보다 지정 구도에 가깝습니다. 다만 테두리는 하단을 넓게 가로지르고 입구의 낮빛도 요청한 희미한 채광보다 강합니다.",
        "entities": "B-200 하나만 등장하며 다른 인물이나 찰리는 없습니다. 육중한 기계 체형, 금속 안면, 주황색 광학 눈, 탄흔 있는 암회색 장갑, 양팔의 굵은 포신이 참조와 부합합니다. 하단에는 흐릿한 새끼 새들의 머리와 둥지 재료 일부가 보입니다. 상자 뒤에 노출된 둥지인지 상자 내부인지, 이 잘린 구도만으로는 명확히 확인하기 어렵습니다.",
        "hard_violations": [],
        "physics": "굽힌 상체는 기계식 허리와 골반, 하단의 다리 부분에 연결되어 자연스러운 전방 굴곡으로 읽힙니다. 발의 지지점은 프레임 밖이지만 몸이 떠 있다는 징후는 없습니다. 양쪽 포신은 팔에 구조적으로 결합되어 있고, 새들의 보이는 부분은 하단 둥지 재료에 받쳐져 있습니다. 배경의 무거운 부품들은 더미와 받침 위에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "카메라 구도와 캐릭터 디자인은 준수하나, 화면 좌측 하단에 명시된 아기 새들의 형태가 흐릿하고 알아보기 어렵습니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "지정된 로우 앵글 프레이밍과 캐릭터의 시선 처리를 잘 구현했으며, 화면 좌측 하단의 상자와 아기 새들을 명확하게 표현하여 지시사항에 매우 충실합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S64sh7_sel.png",
    "asset_id": "a68b5ff0-ed4f-4526-a55b-032c1be0fe87",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1176962>",
    "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cc5-d663-71f8-8f5c-0f47d1ea81ce",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh7"
  }
 },
 "S65sh3::signage": {
  "fp": "039b198f7f7f65a5",
  "inscriptions": [
   {
    "text_native": "0",
    "source": "scene_text_quoted",
    "reason_ko": "근접 구도에서 측정기 바늘이 정확히 가리켜야 하는 눈금 수치입니다.",
    "source_quote": "0"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "era_assess::1d8d3080ee32eb4e": {
  "subjects": [],
  "subject_text": "무너진 백화점 외부·잔해 지대\n거대한 상업 건물이 반쯤 무너진 잔해 지대. 넓게 덮인 이끼와 고인 웅덩이, 무성하게 자란 나무와 작은 꽃들이 콘크리트 틈을 채운다.",
  "identity": "canonical",
  "scope_id": "L107",
  "scope_role": "location_exterior",
  "scope_sha": "a79d42d96a47b25a"
 },
 "S65sh3::bgfirst_bg": {
  "input_fingerprint": "057960d2fbdad0bb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈의 손에 든 방사능 측정기의 바늘이 0 수치를 정확히 가리키고 있는 근접 구도.\n\nLOCATION (lock): On the overgrown approach to a partially collapsed department store, among moss, puddles, and unmanaged vegetation.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the close approach beside 수빈's forearm, looking obliquely downward from just above the radiation meter so its scale remains unobstructed. The meter occupies roughly the central third, supported by her hands with one finger poised outside the scale before tapping; her forearms and a narrow portion of her torso supply scale while her attention remains lowered toward the reading. Keep the needle precisely aligned with zero in sharp focus and allow the moss-covered surroundings to recede softly, making camera distance the principal change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter scale (Needle holding at zero before 수빈 taps the instrument) — The marked face is directed upward and obliquely toward the camera, with the zero marking and needle unobstructed; used as Primary readable detail, bounded by 수빈's fingers rather than isolated from her body; Moss-covered ground (Moss spread widely around the ruined department store); used as Soft peripheral context beyond the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued daytime ambient light with controlled contrast that keeps the zero mark and needle readable without introducing an illuminated display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수빈의 손에 든 방사능 측정기의 바늘이 0 수치를 정확히 가리키고 있는 근접 구도.\n\nLOCATION (lock): On the overgrown approach to a partially collapsed department store, among moss, puddles, and unmanaged vegetation.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the close approach beside 수빈's forearm, looking obliquely downward from just above the radiation meter so its scale remains unobstructed. The meter occupies roughly the central third, supported by her hands with one finger poised outside the scale before tapping; her forearms and a narrow portion of her torso supply scale while her attention remains lowered toward the reading. Keep the needle precisely aligned with zero in sharp focus and allow the moss-covered surroundings to recede softly, making camera distance the principal change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter scale (Needle holding at zero before 수빈 taps the instrument) — The marked face is directed upward and obliquely toward the camera, with the zero marking and needle unobstructed; used as Primary readable detail, bounded by 수빈's fingers rather than isolated from her body; Moss-covered ground (Moss spread widely around the ruined department store); used as Soft peripheral context beyond the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued daytime ambient light with controlled contrast that keeps the zero mark and needle readable without introducing an illuminated display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S65sh3__bgfirst_bg.png",
  "asset_id": "07d8d562-29a5-4588-a585-0c692de737de",
  "input_asset_ids": [
   "730f8bb3-3af6-4471-9e2c-0390b59ba3fa",
   "c7199403-49df-4e25-a41c-f9817990734f"
  ]
 },
 "S65sh3": {
  "input_fingerprint": "91b79cbf1b6cca28",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 든 방사능 측정기의 바늘이 0 수치를 정확히 가리키고 있는 근접 구도.\n\nLOCATION (lock): On the overgrown approach to a partially collapsed department store, among moss, puddles, and unmanaged vegetation. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the close approach beside 수빈's forearm, looking obliquely downward from just above the radiation meter so its scale remains unobstructed. The meter occupies roughly the central third, supported by her hands with one finger poised outside the scale before tapping; her forearms and a narrow portion of her torso supply scale while her attention remains lowered toward the reading. Keep the needle precisely aligned with zero in sharp focus and allow the moss-covered surroundings to recede softly, making camera distance the principal change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter scale (Needle holding at zero before 수빈 taps the instrument) — The marked face is directed upward and obliquely toward the camera, with the zero marking and needle unobstructed; used as Primary readable detail, bounded by 수빈's fingers rather than isolated from her body; Moss-covered ground (Moss spread widely around the ruined department store); used as Soft peripheral context beyond the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued daytime ambient light with controlled contrast that keeps the zero mark and needle readable without introducing an illuminated display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store is half collapsed and overgrown with moss and trees, with standing pools of water. The radiation meter reads close to zero. 수빈: She wears a gas mask and holds the radiation meter. Her pre-existing facial wounds and torso lesions remain beneath her clothing and equipment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 든 방사능 측정기의 바늘이 0 수치를 정확히 가리키고 있는 근접 구도.\n\nLOCATION (lock): On the overgrown approach to a partially collapsed department store, among moss, puddles, and unmanaged vegetation. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the close approach beside 수빈's forearm, looking obliquely downward from just above the radiation meter so its scale remains unobstructed. The meter occupies roughly the central third, supported by her hands with one finger poised outside the scale before tapping; her forearms and a narrow portion of her torso supply scale while her attention remains lowered toward the reading. Keep the needle precisely aligned with zero in sharp focus and allow the moss-covered surroundings to recede softly, making camera distance the principal change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter scale (Needle holding at zero before 수빈 taps the instrument) — The marked face is directed upward and obliquely toward the camera, with the zero marking and needle unobstructed; used as Primary readable detail, bounded by 수빈's fingers rather than isolated from her body; Moss-covered ground (Moss spread widely around the ruined department store); used as Soft peripheral context beyond the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued daytime ambient light with controlled contrast that keeps the zero mark and needle readable without introducing an illuminated display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store is half collapsed and overgrown with moss and trees, with standing pools of water. The radiation meter reads close to zero. 수빈: She wears a gas mask and holds the radiation meter. Her pre-existing facial wounds and torso lesions remain beneath her clothing and equipment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수빈의 손에 든 방사능 측정기의 바늘이 0 수치를 정확히 가리키고 있는 근접 구도.\n\nLOCATION (lock): On the overgrown approach to a partially collapsed department store, among moss, puddles, and unmanaged vegetation. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the close approach beside 수빈's forearm, looking obliquely downward from just above the radiation meter so its scale remains unobstructed. The meter occupies roughly the central third, supported by her hands with one finger poised outside the scale before tapping; her forearms and a narrow portion of her torso supply scale while her attention remains lowered toward the reading. Keep the needle precisely aligned with zero in sharp focus and allow the moss-covered surroundings to recede softly, making camera distance the principal change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Radiation meter scale (Needle holding at zero before 수빈 taps the instrument) — The marked face is directed upward and obliquely toward the camera, with the zero marking and needle unobstructed; used as Primary readable detail, bounded by 수빈's fingers rather than isolated from her body; Moss-covered ground (Moss spread widely around the ruined department store); used as Soft peripheral context beyond the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued daytime ambient light with controlled contrast that keeps the zero mark and needle readable without introducing an illuminated display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store is half collapsed and overgrown with moss and trees, with standing pools of water. The radiation meter reads close to zero. 수빈: She wears a gas mask and holds the radiation meter. Her pre-existing facial wounds and torso lesions remain beneath her clothing and equipment.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수빈 right now, so 수빈's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수빈: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"0\"\n\nThe WORDS TO RENDER above are fixed: render them exactly as given, in the language and script they are written in.\n\nWriting that the shot text and the attached references already fix may also be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S65sh3__bgfirst_bg.png",
     "asset_id": "07d8d562-29a5-4588-a585-0c692de737de",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S65sh3.png",
     "asset_id": "730f8bb3-3af6-4471-9e2c-0390b59ba3fa",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 방사능 측정기: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1094991>",
     "asset_id": "848fb2dc-d716-420a-a6ca-9aa78b8da75c",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L107B01.png",
     "asset_id": "c7199403-49df-4e25-a41c-f9817990734f",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 방사능 측정기: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1094991>",
     "asset_id": "848fb2dc-d716-420a-a6ca-9aa78b8da75c",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물은 손에 든 측정기를 바라보며, 오른손 검지가 기기 위에 위치함.",
    "built_space": "배경에 무너진 건물과 웅덩이가 넓게 보이나, 이는 지시된 비스듬한 하향 구도가 아님.",
    "entities": "수빈(방독면, 의상)이 등장하지만, 방사능 측정기가 레퍼런스와 전혀 다른 가로형태로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 방사능 측정기 레퍼런스 형태 완전 불일치 및 임의 변형(Invented object)",
     "[gemini-pro] 지정된 카메라 시점(팔뚝 옆 하향 근접 구도) 위반"
    ],
    "physics": "양손으로 기기를 정상적으로 쥐고 있으며 지탱에 문제 없음."
   },
   {
    "label": "B",
    "direction": "카메라와 인물의 시선 모두 아래쪽의 방사능 측정기를 향하고 있음.",
    "built_space": "하향 구도에 맞춰 얕은 심도로 처리된 이끼 덮인 바닥이 자연스러운 배경을 이룸.",
    "entities": "수빈의 복장과 더불어 레퍼런스와 일치하는 세로형 측정기가 묘사되었고, 바늘이 정확히 0을 가리킴.",
    "hard_violations": [],
    "physics": "양손으로 기기를 안정적이고 자연스럽게 받쳐 들고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 하향 근접 카메라 구도와 레퍼런스 소품의 형태를 정확히 구현하여 프롬프트 충실도가 매우 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 하향 카메라 시점을 어겼으며, 측정기를 레퍼런스와 전혀 다른 형태로 임의 렌더링하여 주요 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 손에 든 측정기를 바라보며, 오른손 검지가 기기 위에 위치함.",
        "built_space": "배경에 무너진 건물과 웅덩이가 넓게 보이나, 이는 지시된 비스듬한 하향 구도가 아님.",
        "entities": "수빈(방독면, 의상)이 등장하지만, 방사능 측정기가 레퍼런스와 전혀 다른 가로형태로 묘사됨.",
        "hard_violations": [
         "방사능 측정기 레퍼런스 형태 완전 불일치 및 임의 변형(Invented object)",
         "지정된 카메라 시점(팔뚝 옆 하향 근접 구도) 위반"
        ],
        "physics": "양손으로 기기를 정상적으로 쥐고 있으며 지탱에 문제 없음."
       },
       {
        "label": "B",
        "direction": "카메라와 인물의 시선 모두 아래쪽의 방사능 측정기를 향하고 있음.",
        "built_space": "하향 구도에 맞춰 얕은 심도로 처리된 이끼 덮인 바닥이 자연스러운 배경을 이룸.",
        "entities": "수빈의 복장과 더불어 레퍼런스와 일치하는 세로형 측정기가 묘사되었고, 바늘이 정확히 0을 가리킴.",
        "hard_violations": [],
        "physics": "양손으로 기기를 안정적이고 자연스럽게 받쳐 들고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 하향 근접 카메라 구도와 레퍼런스 소품의 형태를 정확히 구현하여 프롬프트 충실도가 매우 높습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 하향 카메라 시점을 어겼으며, 측정기를 레퍼런스와 전혀 다른 형태로 임의 렌더링하여 주요 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 손에 든 측정기를 바라보며, 오른손 검지가 기기 위에 위치함.",
        "built_space": "배경에 무너진 건물과 웅덩이가 넓게 보이나, 이는 지시된 비스듬한 하향 구도가 아님.",
        "entities": "수빈(방독면, 의상)이 등장하지만, 방사능 측정기가 레퍼런스와 전혀 다른 가로형태로 묘사됨.",
        "hard_violations": [
         "방사능 측정기 레퍼런스 형태 완전 불일치 및 임의 변형(Invented object)",
         "지정된 카메라 시점(팔뚝 옆 하향 근접 구도) 위반"
        ],
        "physics": "양손으로 기기를 정상적으로 쥐고 있으며 지탱에 문제 없음."
       },
       {
        "label": "B",
        "direction": "카메라와 인물의 시선 모두 아래쪽의 방사능 측정기를 향하고 있음.",
        "built_space": "하향 구도에 맞춰 얕은 심도로 처리된 이끼 덮인 바닥이 자연스러운 배경을 이룸.",
        "entities": "수빈의 복장과 더불어 레퍼런스와 일치하는 세로형 측정기가 묘사되었고, 바늘이 정확히 0을 가리킴.",
        "hard_violations": [],
        "physics": "양손으로 기기를 안정적이고 자연스럽게 받쳐 들고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "0에 정렬된 바늘과 아래로 내려다보는 근접 시점이 우세하며 소품도 참조에 가깝지만, 어깨 비중이 크고 두드리기 직전의 손가락 동작은 약하다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "0 지시와 양손 지지는 맞지만, 건물까지 펼쳐 보이는 시점이 지정된 하향 근접 구도에서 벗어나고 측정기의 형태도 참조와 크게 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "붉은 바늘이 눈금 왼쪽 끝의 0 지점에 정렬되어 있다. 측정면은 위쪽과 카메라 쪽으로 비스듬히 향해 숫자와 바늘이 가려지지 않는다. 수빈의 눈은 보이지 않지만 고개와 방독면은 측정기 쪽으로 숙여져 있다. 오른손 엄지는 눈금 밖 테두리에 닿아 있어 두드리기보다는 잡고 있는 동작에 가깝다.",
        "built_space": "배경에는 이끼 낀 포장 바닥, 물웅덩이와 식생이 보이며 건물의 고정 설비는 프레임 밖이다. 근접 촬영에 필요한 지면 맥락은 맞고, 건물이 제외된 것은 결함이 아니다. 카메라는 측정기 위에서 아래를 보지만 팔 옆보다는 어깨 너머에 가깝고, 좁은 몸통 일부만 요구한 것에 비해 왼쪽 어깨와 머리가 크게 들어온다.",
        "entities": "수빈 한 명의 검은 단발, 방독면, 낡은 올리브색 소매와 손가락이 노출된 장갑이 보인다. 노출된 손과 목은 젊은 여성 설정과 모순되지 않지만 얼굴과 정확한 나이·출신은 확인할 수 없다. 측정기는 세로형 어두운 케이스, 큰 투명 눈금창, 붉은 바늘과 하단 회전 다이얼로 참조와 가깝다. 다만 단자와 세부 표기 배열은 다르다. 0과 다른 눈금 숫자는 물체 표면의 인쇄로 보이며, 참조에도 숫자 눈금이 있다. 하의와 신발, 가려진 상처는 평가할 수 없다.",
        "hard_violations": [],
        "physics": "왼손이 측정기 왼쪽 가장자리를 잡고 오른손이 오른쪽과 아래쪽을 받쳐 무게를 지탱한다. 손목과 팔의 연결은 자연스럽고 기기가 떠 있지 않다. 눈금 바늘은 내부 축에 연결되어 있으며 두드리기 전 정지 상태로 해석할 수 있다."
       },
       {
        "label": "B",
        "direction": "검은 바늘은 왼쪽 끝 0 아래의 시작 눈금을 향한다. 측정면은 수빈과 카메라 양쪽에서 읽을 수 있게 기울어져 있고 눈금은 가려지지 않는다. 눈은 머리와 방독면에 가려져 있으나 고개는 측정기 방향이다. 오른손 검지는 눈금 아래의 케이스를 향하지만 이미 표면에 닿아 보인다.",
        "built_space": "뒤에는 중앙 출입구 하나로 이어지는 접근로, 양옆의 물웅덩이와 식생, 무너진 전면 벽체와 상부 창문 열이 보여 장소 참조와 대체로 맞는다. 다만 카메라가 접근로와 건물 정면을 넓게 보여 주어, 측정기 바로 위에서 내려다보고 이끼 지면을 부드러운 주변부로 남기라는 구도보다 전방 지향적이다. 머리와 어깨가 차지하는 면적도 크다.",
        "entities": "한 명의 수빈에게 검은 단발, 방독면과 낡은 올리브색 상의가 보인다. 맨손은 젊은 성인의 손으로 읽히지만 참조의 손가락 노출 장갑은 없다. 얼굴이 가려져 정확한 얼굴 일치는 판정할 수 없다. 측정기는 물리적인 아날로그 장치이고 0 표시는 명료하지만, 가로로 넓은 창과 상자형 케이스, 검은 바늘, 작은 왼쪽 다이얼은 참조의 세로형 장치·붉은 바늘·큰 중앙 다이얼과 다르다. 하의, 신발과 옷 아래 병변은 보이지 않는다.",
        "hard_violations": [],
        "physics": "왼손이 기기 왼쪽과 밑면을 받치고 오른손도 하단을 지지하면서 검지를 앞면에 댄다. 끈은 측면 연결부에서 아래로 늘어져 중력에 맞으며, 측정기와 손에 지지 없는 부유는 없다. 바늘 역시 눈금창 아래쪽 내부에서 이어지는 기계 부품으로 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "0에 정렬된 바늘과 아래로 내려다보는 근접 시점이 우세하며 소품도 참조에 가깝지만, 어깨 비중이 크고 두드리기 직전의 손가락 동작은 약하다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "0 지시와 양손 지지는 맞지만, 건물까지 펼쳐 보이는 시점이 지정된 하향 근접 구도에서 벗어나고 측정기의 형태도 참조와 크게 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "붉은 바늘이 눈금 왼쪽 끝의 0 지점에 정렬되어 있다. 측정면은 위쪽과 카메라 쪽으로 비스듬히 향해 숫자와 바늘이 가려지지 않는다. 수빈의 눈은 보이지 않지만 고개와 방독면은 측정기 쪽으로 숙여져 있다. 오른손 엄지는 눈금 밖 테두리에 닿아 있어 두드리기보다는 잡고 있는 동작에 가깝다.",
        "built_space": "배경에는 이끼 낀 포장 바닥, 물웅덩이와 식생이 보이며 건물의 고정 설비는 프레임 밖이다. 근접 촬영에 필요한 지면 맥락은 맞고, 건물이 제외된 것은 결함이 아니다. 카메라는 측정기 위에서 아래를 보지만 팔 옆보다는 어깨 너머에 가깝고, 좁은 몸통 일부만 요구한 것에 비해 왼쪽 어깨와 머리가 크게 들어온다.",
        "entities": "수빈 한 명의 검은 단발, 방독면, 낡은 올리브색 소매와 손가락이 노출된 장갑이 보인다. 노출된 손과 목은 젊은 여성 설정과 모순되지 않지만 얼굴과 정확한 나이·출신은 확인할 수 없다. 측정기는 세로형 어두운 케이스, 큰 투명 눈금창, 붉은 바늘과 하단 회전 다이얼로 참조와 가깝다. 다만 단자와 세부 표기 배열은 다르다. 0과 다른 눈금 숫자는 물체 표면의 인쇄로 보이며, 참조에도 숫자 눈금이 있다. 하의와 신발, 가려진 상처는 평가할 수 없다.",
        "hard_violations": [],
        "physics": "왼손이 측정기 왼쪽 가장자리를 잡고 오른손이 오른쪽과 아래쪽을 받쳐 무게를 지탱한다. 손목과 팔의 연결은 자연스럽고 기기가 떠 있지 않다. 눈금 바늘은 내부 축에 연결되어 있으며 두드리기 전 정지 상태로 해석할 수 있다."
       },
       {
        "label": "A",
        "direction": "검은 바늘은 왼쪽 끝 0 아래의 시작 눈금을 향한다. 측정면은 수빈과 카메라 양쪽에서 읽을 수 있게 기울어져 있고 눈금은 가려지지 않는다. 눈은 머리와 방독면에 가려져 있으나 고개는 측정기 방향이다. 오른손 검지는 눈금 아래의 케이스를 향하지만 이미 표면에 닿아 보인다.",
        "built_space": "뒤에는 중앙 출입구 하나로 이어지는 접근로, 양옆의 물웅덩이와 식생, 무너진 전면 벽체와 상부 창문 열이 보여 장소 참조와 대체로 맞는다. 다만 카메라가 접근로와 건물 정면을 넓게 보여 주어, 측정기 바로 위에서 내려다보고 이끼 지면을 부드러운 주변부로 남기라는 구도보다 전방 지향적이다. 머리와 어깨가 차지하는 면적도 크다.",
        "entities": "한 명의 수빈에게 검은 단발, 방독면과 낡은 올리브색 상의가 보인다. 맨손은 젊은 성인의 손으로 읽히지만 참조의 손가락 노출 장갑은 없다. 얼굴이 가려져 정확한 얼굴 일치는 판정할 수 없다. 측정기는 물리적인 아날로그 장치이고 0 표시는 명료하지만, 가로로 넓은 창과 상자형 케이스, 검은 바늘, 작은 왼쪽 다이얼은 참조의 세로형 장치·붉은 바늘·큰 중앙 다이얼과 다르다. 하의, 신발과 옷 아래 병변은 보이지 않는다.",
        "hard_violations": [],
        "physics": "왼손이 기기 왼쪽과 밑면을 받치고 오른손도 하단을 지지하면서 검지를 앞면에 댄다. 끈은 측면 연결부에서 아래로 늘어져 중력에 맞으며, 측정기와 손에 지지 없는 부유는 없다. 바늘 역시 눈금창 아래쪽 내부에서 이어지는 기계 부품으로 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.054,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.804,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 방사능 측정기 레퍼런스 형태 완전 불일치 및 임의 변형(Invented object)",
     "[gemini-pro] 지정된 카메라 시점(팔뚝 옆 하향 근접 구도) 위반"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 804
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 하향 근접 카메라 구도와 레퍼런스 소품의 형태를 정확히 구현하여 프롬프트 충실도가 매우 높습니다."
   },
   {
    "label": "A",
    "score": 804,
    "verdict_ko": "지정된 하향 카메라 시점을 어겼으며, 측정기를 레퍼런스와 전혀 다른 형태로 임의 렌더링하여 주요 지침을 위반했습니다.  ★위반: [gemini-pro] 방사능 측정기 레퍼런스 형태 완전 불일치 및 임의 변형(Invented object) / [gemini-pro] 지정된 카메라 시점(팔뚝 옆 하향 근접 구도) 위반"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L107B01.png",
    "asset_id": "c7199403-49df-4e25-a41c-f9817990734f",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 방사능 측정기: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1094991>",
    "asset_id": "848fb2dc-d716-420a-a6ca-9aa78b8da75c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cca-f2ca-73b5-87ec-53b42145065c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S65sh3__bgfirst_bg.png",
   "bg_asset_id": "07d8d562-29a5-4588-a585-0c692de737de",
   "bg_record_key": "S65sh3::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S65sh12::signage": {
  "fp": "0dac8a74800e7f21",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh12": {
  "input_fingerprint": "926864252c394319",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방독면을 벗은 수빈과 이현우가 서로를 마주 본 채 환하게 미소 짓는 상반신 구도.\n\nLOCATION (lock): Outside the ruined department store, in the moss-covered overgrowth before the opening through the rubble. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at upper-chest height, maintaining the flow's oblique view across the space between 수빈 and 이현우 with a slight upward inclination toward their faces. Frame 수빈 on the left as her shoulders release after removing the gas mask, and 이현우 on the right leaning subtly toward her; both uncovered faces smile at each other, never at the lens. Retain a partial view of the department-store gap behind them, emphasizing their exchanged gaze before their departure begins.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Gap in the collapsed department store (Open passage through the partially collapsed building) — Seen obliquely behind the pair, with the opening readable between broken building sections; used as Quiet background reference for their forthcoming movement inside; Irregularly grown trees (Growing amid the ruined department-store surroundings); used as Soft background layering that preserves the recovered-life context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime ambient light and gentle facial contrast, letting their smiles supply warmth without changing the scene's neutral tonal treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed store remains covered in moss and vegetation, with standing water and wildlife around it. The radiation meter has registered close to zero. 이현우: His gas mask is off, leaving his face exposed. His recent combat injuries remain. 수빈: Her gas mask is off and has been cast aside; she retains the radiation meter. Her existing facial wounds and torso radiation lesions do not disappear with the environmental reading.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방독면을 벗은 수빈과 이현우가 서로를 마주 본 채 환하게 미소 짓는 상반신 구도.\n\nLOCATION (lock): Outside the ruined department store, in the moss-covered overgrowth before the opening through the rubble. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at upper-chest height, maintaining the flow's oblique view across the space between 수빈 and 이현우 with a slight upward inclination toward their faces. Frame 수빈 on the left as her shoulders release after removing the gas mask, and 이현우 on the right leaning subtly toward her; both uncovered faces smile at each other, never at the lens. Retain a partial view of the department-store gap behind them, emphasizing their exchanged gaze before their departure begins.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Gap in the collapsed department store (Open passage through the partially collapsed building) — Seen obliquely behind the pair, with the opening readable between broken building sections; used as Quiet background reference for their forthcoming movement inside; Irregularly grown trees (Growing amid the ruined department-store surroundings); used as Soft background layering that preserves the recovered-life context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime ambient light and gentle facial contrast, letting their smiles supply warmth without changing the scene's neutral tonal treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed store remains covered in moss and vegetation, with standing water and wildlife around it. The radiation meter has registered close to zero. 이현우: His gas mask is off, leaving his face exposed. His recent combat injuries remain. 수빈: Her gas mask is off and has been cast aside; she retains the radiation meter. Her existing facial wounds and torso radiation lesions do not disappear with the environmental reading.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방독면을 벗은 수빈과 이현우가 서로를 마주 본 채 환하게 미소 짓는 상반신 구도.\n\nLOCATION (lock): Outside the ruined department store, in the moss-covered overgrowth before the opening through the rubble. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at upper-chest height, maintaining the flow's oblique view across the space between 수빈 and 이현우 with a slight upward inclination toward their faces. Frame 수빈 on the left as her shoulders release after removing the gas mask, and 이현우 on the right leaning subtly toward her; both uncovered faces smile at each other, never at the lens. Retain a partial view of the department-store gap behind them, emphasizing their exchanged gaze before their departure begins.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Gap in the collapsed department store (Open passage through the partially collapsed building) — Seen obliquely behind the pair, with the opening readable between broken building sections; used as Quiet background reference for their forthcoming movement inside; Irregularly grown trees (Growing amid the ruined department-store surroundings); used as Soft background layering that preserves the recovered-life context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued daytime ambient light and gentle facial contrast, letting their smiles supply warmth without changing the scene's neutral tonal treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The half-collapsed store remains covered in moss and vegetation, with standing water and wildlife around it. The radiation meter has registered close to zero. 이현우: His gas mask is off, leaving his face exposed. His recent combat injuries remain. 수빈: Her gas mask is off and has been cast aside; she retains the radiation meter. Her existing facial wounds and torso radiation lesions do not disappear with the environmental reading.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "수빈과 이현우가 서로를 마주 보며 밝게 미소 짓고 있으며 시선 교환이 정확함.",
    "built_space": "붕괴된 백화점 잔해 사이에 하늘이 보이는 열린 틈새가 있으며, 식물과 물웅덩이가 적절히 배치됨.",
    "entities": "수빈은 단발머리에 올리브색 상의를 입고 측정기를 들고 있음. 이현우는 셔츠와 인이어 무전기를 착용했으나, 레퍼런스에 없는 배낭 끈을 메고 있음.",
    "hard_violations": [
     "[gemini-pro] 이현우에게 프롬프트와 레퍼런스에 명시되지 않은 배낭(어깨 끈)이 추가됨 (invented objects).",
     "[gemini-pro] 수빈이 측정기를 쥔 오른손의 손바닥이 바깥을 향한 채 관절이 꺾여 있어 물리적으로 불가능한 해부학적 구조를 보임 (physically impossible anatomy).",
     "[gpt-high] 이현우에게 프롬프트와 인물 참조에 없는 배낭 장비를 추가했으며, 양쪽 어깨의 두꺼운 멜빵과 버클이 뚜렷하게 보인다."
    ],
    "physics": "수빈이 두 손으로 측정기를 들고 있으나 오른팔과 손목의 연결 각도가 인체 구조상 불가능함."
   },
   {
    "label": "B",
    "direction": "수빈과 이현우가 얼굴을 마주 보며 미소 짓고 있어 프롬프트가 요구한 시선 방향을 충족함.",
    "built_space": "배경 중앙에 무너진 건물의 통로가 명확히 보이며, 이끼 낀 식물과 고인 물이 설정된 환경과 일치함.",
    "entities": "수빈은 레퍼런스와 일치하는 복장을 하고 방사능 측정기를 들고 있음. 이현우는 낡은 셔츠와 인이어 무전기를 착용하였으며 불필요한 배낭 없이 레퍼런스를 잘 따름.",
    "hard_violations": [],
    "physics": "수빈이 한 손으로 측정기를 자연스럽게 받쳐 들고 있으며 인물들의 서 있는 자세와 지지 상태가 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 인물의 복장, 시선, 그리고 폐허가 된 배경을 정확히 구현하였으며, 불필요한 사물 추가나 해부학적 오류 없이 훌륭한 구도를 완성함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 인물의 구도와 표정은 좋으나, 이현우에게 프롬프트에 없는 배낭이 추가되었고 수빈이 측정기를 쥔 손의 해부학적 형태가 왜곡되어 감점됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "수빈과 이현우가 서로를 마주 보며 밝게 미소 짓고 있으며 시선 교환이 정확함.",
        "built_space": "붕괴된 백화점 잔해 사이에 하늘이 보이는 열린 틈새가 있으며, 식물과 물웅덩이가 적절히 배치됨.",
        "entities": "수빈은 단발머리에 올리브색 상의를 입고 측정기를 들고 있음. 이현우는 셔츠와 인이어 무전기를 착용했으나, 레퍼런스에 없는 배낭 끈을 메고 있음.",
        "hard_violations": [
         "이현우에게 프롬프트와 레퍼런스에 명시되지 않은 배낭(어깨 끈)이 추가됨 (invented objects).",
         "수빈이 측정기를 쥔 오른손의 손바닥이 바깥을 향한 채 관절이 꺾여 있어 물리적으로 불가능한 해부학적 구조를 보임 (physically impossible anatomy)."
        ],
        "physics": "수빈이 두 손으로 측정기를 들고 있으나 오른팔과 손목의 연결 각도가 인체 구조상 불가능함."
       },
       {
        "label": "B",
        "direction": "수빈과 이현우가 얼굴을 마주 보며 미소 짓고 있어 프롬프트가 요구한 시선 방향을 충족함.",
        "built_space": "배경 중앙에 무너진 건물의 통로가 명확히 보이며, 이끼 낀 식물과 고인 물이 설정된 환경과 일치함.",
        "entities": "수빈은 레퍼런스와 일치하는 복장을 하고 방사능 측정기를 들고 있음. 이현우는 낡은 셔츠와 인이어 무전기를 착용하였으며 불필요한 배낭 없이 레퍼런스를 잘 따름.",
        "hard_violations": [],
        "physics": "수빈이 한 손으로 측정기를 자연스럽게 받쳐 들고 있으며 인물들의 서 있는 자세와 지지 상태가 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 인물의 복장, 시선, 그리고 폐허가 된 배경을 정확히 구현하였으며, 불필요한 사물 추가나 해부학적 오류 없이 훌륭한 구도를 완성함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 인물의 구도와 표정은 좋으나, 이현우에게 프롬프트에 없는 배낭이 추가되었고 수빈이 측정기를 쥔 손의 해부학적 형태가 왜곡되어 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "수빈과 이현우가 서로를 마주 보며 밝게 미소 짓고 있으며 시선 교환이 정확함.",
        "built_space": "붕괴된 백화점 잔해 사이에 하늘이 보이는 열린 틈새가 있으며, 식물과 물웅덩이가 적절히 배치됨.",
        "entities": "수빈은 단발머리에 올리브색 상의를 입고 측정기를 들고 있음. 이현우는 셔츠와 인이어 무전기를 착용했으나, 레퍼런스에 없는 배낭 끈을 메고 있음.",
        "hard_violations": [
         "이현우에게 프롬프트와 레퍼런스에 명시되지 않은 배낭(어깨 끈)이 추가됨 (invented objects).",
         "수빈이 측정기를 쥔 오른손의 손바닥이 바깥을 향한 채 관절이 꺾여 있어 물리적으로 불가능한 해부학적 구조를 보임 (physically impossible anatomy)."
        ],
        "physics": "수빈이 두 손으로 측정기를 들고 있으나 오른팔과 손목의 연결 각도가 인체 구조상 불가능함."
       },
       {
        "label": "B",
        "direction": "수빈과 이현우가 얼굴을 마주 보며 미소 짓고 있어 프롬프트가 요구한 시선 방향을 충족함.",
        "built_space": "배경 중앙에 무너진 건물의 통로가 명확히 보이며, 이끼 낀 식물과 고인 물이 설정된 환경과 일치함.",
        "entities": "수빈은 레퍼런스와 일치하는 복장을 하고 방사능 측정기를 들고 있음. 이현우는 낡은 셔츠와 인이어 무전기를 착용하였으며 불필요한 배낭 없이 레퍼런스를 잘 따름.",
        "hard_violations": [],
        "physics": "수빈이 한 손으로 측정기를 자연스럽게 받쳐 들고 있으며 인물들의 서 있는 자세와 지지 상태가 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "서로를 보며 환하게 웃는 좌우 상반신 구도, 이현우의 미세한 기울임, 뒤쪽 통로의 절제된 노출을 잘 구현하며, 계측기 외형은 참조와 조금 다릅니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "시선 교환과 미소는 맞지만 이현우에게 참조에 없는 배낭 장비를 추가했고, 중앙의 큰 개구부가 조용한 배경이어야 할 통로를 상대적으로 강조합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 수빈은 오른쪽 위의 이현우 눈을 보고, 오른쪽 이현우는 왼쪽 아래의 수빈 얼굴을 보며 웃는다. 둘 다 렌즈를 보지 않는다. 계측기 눈금판은 카메라 쪽으로 기울어 있지만 수빈은 이를 읽거나 조작하지 않고 내려서 들고 있다.",
        "built_space": "두 사람 사이 뒤편에 지상 통로 하나가 보이며, 그 위에는 부서진 콘크리트 바닥판과 어두운 상층 개방부가 있다. 양옆 잔존 벽체와 불규칙한 나무, 덩굴, 잔해, 고인 물이 배경을 이룬다. 두 사람은 통로 바깥 전경에 서 있다. 이전 사진의 이끼 낀 콘크리트와 식생 재질은 이어지며, 이전 사진에는 건물 전체 구조가 없어 정확한 부재 수 비교는 불가능하다. 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 여성 한 명과 남성 한 명만 보인다. 수빈의 검은 단발, 올리브색 낡은 상의, 손가락이 드러나는 장갑과 얼굴 상처가 맞는다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 어두운 피 묻은 셔츠, 인이어 무전기와 얼굴·목의 부상이 보인다. 얼굴은 각 인물 참조와 대체로 부합한다. 두 얼굴 모두 방독면으로 가려져 있지 않다. 아날로그 방사선 계측기는 수빈 손에 있고 바늘은 낮은 눈금 부근이지만, 참조보다 눈금판과 케이스 비례가 달라졌다. 몸통 병변은 옷에 가려져 확인할 수 없고, 버린 방독면과 하체 복장은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "수빈은 장갑 낀 손으로 계측기 옆면을 잡고 다른 손으로 아래쪽을 받친다. 계측기는 공중에 떠 있지 않다. 두 사람의 몸통은 프레임 아래로 자연스럽게 이어지고, 이현우가 수빈 쪽으로 조금 기운 자세도 서 있는 사람이 취할 수 있다. 발 접촉은 화면 밖이라 직접 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 수빈의 시선은 오른쪽 이현우의 눈에, 이현우의 시선은 수빈 얼굴에 향한다. 둘 다 미소 짓고 카메라를 보지 않는다. 수빈은 눈금판이 바깥으로 향한 계측기를 가슴에 안고 있으며, 화면을 읽는 동작은 아니다.",
        "built_space": "두 사람 뒤 중앙에 큰 통로 하나가 있고, 위쪽에는 부서진 보와 작은 틈이 보인다. 양옆 콘크리트 골조와 덩굴, 나무, 잔해 및 중앙 바닥의 고인 물이 이어진다. 인물들은 개구부 앞 야외에 있다. 식생과 콘크리트의 재질은 이전 사진과 어울리지만, 개구부가 화면 중앙에서 크게 드러나 조용한 부분 배경이라는 지시보다 강조된다. 물에는 밝은 하늘과 주변 구조물이 반사되며 뚜렷한 광학적 모순은 없다.",
        "entities": "수빈과 이현우로 읽히는 젊은 동아시아계 여성과 남성 두 명만 있다. 검은 단발과 올리브 상의, 검은 장갑, 이현우의 검은 머리와 어두운 혈흔 셔츠, 인이어 무전기는 대체로 맞는다. 두 사람 모두 얼굴과 목의 상처가 있고 방독면은 벗었다. 수빈은 낮은 수치를 가리키는 아날로그 계측기를 지니지만 케이스와 조작부 세부는 참조와 다르다. 이현우 양쪽 어깨에는 인물 참조에 없는 두꺼운 배낭 멜빵이 추가됐다. 몸통 병변과 하체 착장은 화면에서 확인할 수 없다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조에 없는 배낭 장비를 추가했으며, 양쪽 어깨의 두꺼운 멜빵과 버클이 뚜렷하게 보인다."
        ],
        "physics": "수빈의 두 손과 굽힌 팔이 계측기를 몸 앞에서 감싸 받치므로 무게 지지가 자연스럽다. 이현우의 기울어진 머리와 상체도 선 자세에서 가능하다. 배낭 멜빵은 어깨에 걸쳐져 있으며, 떠 있는 물체나 지지 없이 누운 몸은 없다. 두 사람의 발은 프레임 밖이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "서로를 보며 환하게 웃는 좌우 상반신 구도, 이현우의 미세한 기울임, 뒤쪽 통로의 절제된 노출을 잘 구현하며, 계측기 외형은 참조와 조금 다릅니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "시선 교환과 미소는 맞지만 이현우에게 참조에 없는 배낭 장비를 추가했고, 중앙의 큰 개구부가 조용한 배경이어야 할 통로를 상대적으로 강조합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 수빈은 오른쪽 위의 이현우 눈을 보고, 오른쪽 이현우는 왼쪽 아래의 수빈 얼굴을 보며 웃는다. 둘 다 렌즈를 보지 않는다. 계측기 눈금판은 카메라 쪽으로 기울어 있지만 수빈은 이를 읽거나 조작하지 않고 내려서 들고 있다.",
        "built_space": "두 사람 사이 뒤편에 지상 통로 하나가 보이며, 그 위에는 부서진 콘크리트 바닥판과 어두운 상층 개방부가 있다. 양옆 잔존 벽체와 불규칙한 나무, 덩굴, 잔해, 고인 물이 배경을 이룬다. 두 사람은 통로 바깥 전경에 서 있다. 이전 사진의 이끼 낀 콘크리트와 식생 재질은 이어지며, 이전 사진에는 건물 전체 구조가 없어 정확한 부재 수 비교는 불가능하다. 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 여성 한 명과 남성 한 명만 보인다. 수빈의 검은 단발, 올리브색 낡은 상의, 손가락이 드러나는 장갑과 얼굴 상처가 맞는다. 이현우의 짧고 헝클어진 검은 머리, 마른 체격, 어두운 피 묻은 셔츠, 인이어 무전기와 얼굴·목의 부상이 보인다. 얼굴은 각 인물 참조와 대체로 부합한다. 두 얼굴 모두 방독면으로 가려져 있지 않다. 아날로그 방사선 계측기는 수빈 손에 있고 바늘은 낮은 눈금 부근이지만, 참조보다 눈금판과 케이스 비례가 달라졌다. 몸통 병변은 옷에 가려져 확인할 수 없고, 버린 방독면과 하체 복장은 프레임 밖이다.",
        "hard_violations": [],
        "physics": "수빈은 장갑 낀 손으로 계측기 옆면을 잡고 다른 손으로 아래쪽을 받친다. 계측기는 공중에 떠 있지 않다. 두 사람의 몸통은 프레임 아래로 자연스럽게 이어지고, 이현우가 수빈 쪽으로 조금 기운 자세도 서 있는 사람이 취할 수 있다. 발 접촉은 화면 밖이라 직접 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 수빈의 시선은 오른쪽 이현우의 눈에, 이현우의 시선은 수빈 얼굴에 향한다. 둘 다 미소 짓고 카메라를 보지 않는다. 수빈은 눈금판이 바깥으로 향한 계측기를 가슴에 안고 있으며, 화면을 읽는 동작은 아니다.",
        "built_space": "두 사람 뒤 중앙에 큰 통로 하나가 있고, 위쪽에는 부서진 보와 작은 틈이 보인다. 양옆 콘크리트 골조와 덩굴, 나무, 잔해 및 중앙 바닥의 고인 물이 이어진다. 인물들은 개구부 앞 야외에 있다. 식생과 콘크리트의 재질은 이전 사진과 어울리지만, 개구부가 화면 중앙에서 크게 드러나 조용한 부분 배경이라는 지시보다 강조된다. 물에는 밝은 하늘과 주변 구조물이 반사되며 뚜렷한 광학적 모순은 없다.",
        "entities": "수빈과 이현우로 읽히는 젊은 동아시아계 여성과 남성 두 명만 있다. 검은 단발과 올리브 상의, 검은 장갑, 이현우의 검은 머리와 어두운 혈흔 셔츠, 인이어 무전기는 대체로 맞는다. 두 사람 모두 얼굴과 목의 상처가 있고 방독면은 벗었다. 수빈은 낮은 수치를 가리키는 아날로그 계측기를 지니지만 케이스와 조작부 세부는 참조와 다르다. 이현우 양쪽 어깨에는 인물 참조에 없는 두꺼운 배낭 멜빵이 추가됐다. 몸통 병변과 하체 착장은 화면에서 확인할 수 없다.",
        "hard_violations": [
         "이현우에게 프롬프트와 인물 참조에 없는 배낭 장비를 추가했으며, 양쪽 어깨의 두꺼운 멜빵과 버클이 뚜렷하게 보인다."
        ],
        "physics": "수빈의 두 손과 굽힌 팔이 계측기를 몸 앞에서 감싸 받치므로 무게 지지가 자연스럽다. 이현우의 기울어진 머리와 상체도 선 자세에서 가능하다. 배낭 멜빵은 어깨에 걸쳐져 있으며, 떠 있는 물체나 지지 없이 누운 몸은 없다. 두 사람의 발은 프레임 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.056,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.806,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 이현우에게 프롬프트와 레퍼런스에 명시되지 않은 배낭(어깨 끈)이 추가됨 (invented objects).",
     "[gemini-pro] 수빈이 측정기를 쥔 오른손의 손바닥이 바깥을 향한 채 관절이 꺾여 있어 물리적으로 불가능한 해부학적 구조를 보임 (physically impossible anatomy).",
     "[gpt-high] 이현우에게 프롬프트와 인물 참조에 없는 배낭 장비를 추가했으며, 양쪽 어깨의 두꺼운 멜빵과 버클이 뚜렷하게 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 806
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 인물의 복장, 시선, 그리고 폐허가 된 배경을 정확히 구현하였으며, 불필요한 사물 추가나 해부학적 오류 없이 훌륭한 구도를 완성함."
   },
   {
    "label": "A",
    "score": 806,
    "verdict_ko": "두 인물의 구도와 표정은 좋으나, 이현우에게 프롬프트에 없는 배낭이 추가되었고 수빈이 측정기를 쥔 손의 해부학적 형태가 왜곡되어 감점됨.  ★위반: [gemini-pro] 이현우에게 프롬프트와 레퍼런스에 명시되지 않은 배낭(어깨 끈)이 추가됨 (invented objects). / [gemini-pro] 수빈이 측정기를 쥔 오른손의 손바닥이 바깥을 향한 채 관절이 꺾여 있어 물리적으로 불가능한 해부학적 구조를 보임 (physically impossible anatomy). / [gpt-high] 이현우에게 프롬프트와 인물 참조에 없는 배낭 장비를 추가했으며, 양쪽 어깨의 두꺼운 멜빵과 버클이 뚜렷하게 보인다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 수빈 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S65sh3_sel.png",
    "asset_id": "e224deeb-b51c-490d-9fce-b8ec869c9a45",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cd2-f1d1-79ac-a997-37dc726ae958",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S65sh3"
  }
 },
 "S65sh14::signage": {
  "fp": "4fe49175ff1bd65b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh14": {
  "input_fingerprint": "45a1d48681e29939",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백화점의 높은 잔해 위에서 아래쪽 틈새를 향해 발을 내디딘 이현우와 수빈을 굽어보는 정체불명 사람의 어깨 너머 시점 구도.\n\nLOCATION (lock): On high exposed rubble beside the collapsed department store, overlooking the gap used to enter the building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit spatial cut to the high remains behind 정체불명 사람, with the lens just above and outside one shoulder, looking steeply downward in a static over-the-shoulder view. The unidentified observer's back and shoulder occupy the left foreground edge, head inclined toward 이현우 and 수빈 below; their full figures occupy the lower-right portion as they watch their footing and step toward the gap. Keep their steps out of phase—이현우 transferring weight forward while 수빈 lifts her trailing heel—and preserve the preceding travel direction rather than mirroring their route.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store entrance gap in the lower-right of the frame, background; 이현우 in the lower-right of the frame, background, moves toward Department-store entrance gap.\n- KEY BACKGROUND ELEMENTS: High department-store remains (Elevated remnants of the partially collapsed building) — A broken upper edge lies beneath the observer, with the lower approach visible beyond it; used as Establishes the observer's elevation without obscuring the pair below; Department-store entrance gap (Open and being entered by 이현우 and 수빈) — Seen from above and outside, below the observer's position; used as Shared spatial anchor linking foreground surveillance to the pair's route.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued daytime ambience, keeping the foreground observer's identity unreadable through back-facing composition rather than an invented lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store remains half collapsed, with an opening through the ruins, extensive moss, pools and overgrown trees. The earlier radiation reading was close to zero. 이현우: He enters the gap in the ruined store without a gas mask over his face. His recent injuries persist. 수빈: She enters the ruined store with her gas mask removed and the radiation meter in her possession. Her facial wounds and torso lesions persist. 정한수: He watches from above the entry route, with his short-haired back turned toward the viewpoint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백화점의 높은 잔해 위에서 아래쪽 틈새를 향해 발을 내디딘 이현우와 수빈을 굽어보는 정체불명 사람의 어깨 너머 시점 구도.\n\nLOCATION (lock): On high exposed rubble beside the collapsed department store, overlooking the gap used to enter the building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit spatial cut to the high remains behind 정체불명 사람, with the lens just above and outside one shoulder, looking steeply downward in a static over-the-shoulder view. The unidentified observer's back and shoulder occupy the left foreground edge, head inclined toward 이현우 and 수빈 below; their full figures occupy the lower-right portion as they watch their footing and step toward the gap. Keep their steps out of phase—이현우 transferring weight forward while 수빈 lifts her trailing heel—and preserve the preceding travel direction rather than mirroring their route.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store entrance gap in the lower-right of the frame, background; 이현우 in the lower-right of the frame, background, moves toward Department-store entrance gap.\n- KEY BACKGROUND ELEMENTS: High department-store remains (Elevated remnants of the partially collapsed building) — A broken upper edge lies beneath the observer, with the lower approach visible beyond it; used as Establishes the observer's elevation without obscuring the pair below; Department-store entrance gap (Open and being entered by 이현우 and 수빈) — Seen from above and outside, below the observer's position; used as Shared spatial anchor linking foreground surveillance to the pair's route.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued daytime ambience, keeping the foreground observer's identity unreadable through back-facing composition rather than an invented lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store remains half collapsed, with an opening through the ruins, extensive moss, pools and overgrown trees. The earlier radiation reading was close to zero. 이현우: He enters the gap in the ruined store without a gas mask over his face. His recent injuries persist. 수빈: She enters the ruined store with her gas mask removed and the radiation meter in her possession. Her facial wounds and torso lesions persist. 정한수: He watches from above the entry route, with his short-haired back turned toward the viewpoint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백화점의 높은 잔해 위에서 아래쪽 틈새를 향해 발을 내디딘 이현우와 수빈을 굽어보는 정체불명 사람의 어깨 너머 시점 구도.\n\nLOCATION (lock): On high exposed rubble beside the collapsed department store, overlooking the gap used to enter the building. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit spatial cut to the high remains behind 정체불명 사람, with the lens just above and outside one shoulder, looking steeply downward in a static over-the-shoulder view. The unidentified observer's back and shoulder occupy the left foreground edge, head inclined toward 이현우 and 수빈 below; their full figures occupy the lower-right portion as they watch their footing and step toward the gap. Keep their steps out of phase—이현우 transferring weight forward while 수빈 lifts her trailing heel—and preserve the preceding travel direction rather than mirroring their route.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Department-store entrance gap in the lower-right of the frame, background; 이현우 in the lower-right of the frame, background, moves toward Department-store entrance gap.\n- KEY BACKGROUND ELEMENTS: High department-store remains (Elevated remnants of the partially collapsed building) — A broken upper edge lies beneath the observer, with the lower approach visible beyond it; used as Establishes the observer's elevation without obscuring the pair below; Department-store entrance gap (Open and being entered by 이현우 and 수빈) — Seen from above and outside, below the observer's position; used as Shared spatial anchor linking foreground surveillance to the pair's route.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued daytime ambience, keeping the foreground observer's identity unreadable through back-facing composition rather than an invented lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The department store remains half collapsed, with an opening through the ruins, extensive moss, pools and overgrown trees. The earlier radiation reading was close to zero. 이현우: He enters the gap in the ruined store without a gas mask over his face. His recent injuries persist. 수빈: She enters the ruined store with her gas mask removed and the radiation meter in her possession. Her facial wounds and torso lesions persist. 정한수: He watches from above the entry route, with his short-haired back turned toward the viewpoint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음.; 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "정한수(어깨 너머 시점)는 우측 아래의 이현우와 수빈을 내려다보고 있으며, 두 인물은 화면 우측 하단의 백화점 입구 틈새를 향해 시선과 방향을 두고 똑바로 나아가고 있음.",
    "built_space": "프롬프트에 지시된 대로 좌측 전경에 높은 잔해가 위치하며, 그 너머로 물이 고인 폐허 바닥과 덩굴이 얽힌 백화점 건물의 틈새가 정확한 원근감으로 자연스럽게 배치됨. 이전 샷 레퍼런스의 콘크리트 구조적 특징이 잘 유지됨.",
    "entities": "정한수는 뒷모습과 함께 레퍼런스와 일치하는 갈색 점퍼를 입고 있음. 이현우와 수빈의 복장(어두운 셔츠, 올리브색 테크웨어)과 헤어스타일이 레퍼런스와 일치하며, 수빈은 왼손에 사각형 모양의 방사능 측정기를 정확히 들고 있음.",
    "hard_violations": [],
    "physics": "이현우는 앞으로 체중을 이동시키며 발을 딛고 있고, 수빈은 뒤쪽 발뒤꿈치를 들어 올리는 등 엇갈린 걸음걸이(out of phase)가 물리적으로 안정감 있게 지면에 닿아 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "전경의 관찰자는 우측 아래의 두 사람을 내려다보고 있으며, 이현우와 수빈은 우측 하단의 입구를 향해 걷고 있음.",
    "built_space": "좌측 전경의 높은 잔해와 배경의 건물 틈새가 존재하나, 건물 외벽의 디테일이 이전 샷 레퍼런스의 특징과 다소 차이가 있으며 덜 뚜렷함.",
    "entities": "인물들의 기본 복장은 비슷하나, 정한수의 점퍼 뒷면 어깨 부분에 레퍼런스에 존재하지 않는 검은색 기계 장치가 부착되어 있음. 수빈의 손에 든 물체가 방사능 측정기(사각형)가 아닌 정체불명의 둥근 원반(또는 통) 형태로 왜곡됨.",
    "hard_violations": [
     "[gemini-pro] invented objects: 정한수의 등에 레퍼런스에 없는 기계 장치(무전기 형태)가 부착됨",
     "[gemini-pro] invented objects: 수빈이 손에 든 물건이 방사능 측정기가 아닌 정체불명의 원반형 물체로 묘사됨"
    ],
    "physics": "두 인물이 지면에 발을 딛고 걷는 자세는 물리적으로 자연스럽게 지지되고 있으나, A 후보에 비해 지시된 엇갈린 걸음걸이와 체중 이동의 역동성이 다소 약함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 오버더숄더 앵글과 엇갈린 걸음걸이 묘사를 정확히 구현했으며, 인물들의 의상과 방사능 측정기의 디테일, 이전 샷의 공간적 특징을 가장 충실하게 재현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 배경은 무난하나, 정한수의 등에 레퍼런스에 없는 장치가 추가되고 수빈이 든 방사능 측정기가 원반 형태로 완전히 왜곡되어 프롬프트 충실도가 떨어짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "정한수(어깨 너머 시점)는 우측 아래의 이현우와 수빈을 내려다보고 있으며, 두 인물은 화면 우측 하단의 백화점 입구 틈새를 향해 시선과 방향을 두고 똑바로 나아가고 있음.",
        "built_space": "프롬프트에 지시된 대로 좌측 전경에 높은 잔해가 위치하며, 그 너머로 물이 고인 폐허 바닥과 덩굴이 얽힌 백화점 건물의 틈새가 정확한 원근감으로 자연스럽게 배치됨. 이전 샷 레퍼런스의 콘크리트 구조적 특징이 잘 유지됨.",
        "entities": "정한수는 뒷모습과 함께 레퍼런스와 일치하는 갈색 점퍼를 입고 있음. 이현우와 수빈의 복장(어두운 셔츠, 올리브색 테크웨어)과 헤어스타일이 레퍼런스와 일치하며, 수빈은 왼손에 사각형 모양의 방사능 측정기를 정확히 들고 있음.",
        "hard_violations": [],
        "physics": "이현우는 앞으로 체중을 이동시키며 발을 딛고 있고, 수빈은 뒤쪽 발뒤꿈치를 들어 올리는 등 엇갈린 걸음걸이(out of phase)가 물리적으로 안정감 있게 지면에 닿아 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "전경의 관찰자는 우측 아래의 두 사람을 내려다보고 있으며, 이현우와 수빈은 우측 하단의 입구를 향해 걷고 있음.",
        "built_space": "좌측 전경의 높은 잔해와 배경의 건물 틈새가 존재하나, 건물 외벽의 디테일이 이전 샷 레퍼런스의 특징과 다소 차이가 있으며 덜 뚜렷함.",
        "entities": "인물들의 기본 복장은 비슷하나, 정한수의 점퍼 뒷면 어깨 부분에 레퍼런스에 존재하지 않는 검은색 기계 장치가 부착되어 있음. 수빈의 손에 든 물체가 방사능 측정기(사각형)가 아닌 정체불명의 둥근 원반(또는 통) 형태로 왜곡됨.",
        "hard_violations": [
         "invented objects: 정한수의 등에 레퍼런스에 없는 기계 장치(무전기 형태)가 부착됨",
         "invented objects: 수빈이 손에 든 물건이 방사능 측정기가 아닌 정체불명의 원반형 물체로 묘사됨"
        ],
        "physics": "두 인물이 지면에 발을 딛고 걷는 자세는 물리적으로 자연스럽게 지지되고 있으나, A 후보에 비해 지시된 엇갈린 걸음걸이와 체중 이동의 역동성이 다소 약함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 오버더숄더 앵글과 엇갈린 걸음걸이 묘사를 정확히 구현했으며, 인물들의 의상과 방사능 측정기의 디테일, 이전 샷의 공간적 특징을 가장 충실하게 재현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "구도와 배경은 무난하나, 정한수의 등에 레퍼런스에 없는 장치가 추가되고 수빈이 든 방사능 측정기가 원반 형태로 완전히 왜곡되어 프롬프트 충실도가 떨어짐."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "정한수(어깨 너머 시점)는 우측 아래의 이현우와 수빈을 내려다보고 있으며, 두 인물은 화면 우측 하단의 백화점 입구 틈새를 향해 시선과 방향을 두고 똑바로 나아가고 있음.",
        "built_space": "프롬프트에 지시된 대로 좌측 전경에 높은 잔해가 위치하며, 그 너머로 물이 고인 폐허 바닥과 덩굴이 얽힌 백화점 건물의 틈새가 정확한 원근감으로 자연스럽게 배치됨. 이전 샷 레퍼런스의 콘크리트 구조적 특징이 잘 유지됨.",
        "entities": "정한수는 뒷모습과 함께 레퍼런스와 일치하는 갈색 점퍼를 입고 있음. 이현우와 수빈의 복장(어두운 셔츠, 올리브색 테크웨어)과 헤어스타일이 레퍼런스와 일치하며, 수빈은 왼손에 사각형 모양의 방사능 측정기를 정확히 들고 있음.",
        "hard_violations": [],
        "physics": "이현우는 앞으로 체중을 이동시키며 발을 딛고 있고, 수빈은 뒤쪽 발뒤꿈치를 들어 올리는 등 엇갈린 걸음걸이(out of phase)가 물리적으로 안정감 있게 지면에 닿아 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "전경의 관찰자는 우측 아래의 두 사람을 내려다보고 있으며, 이현우와 수빈은 우측 하단의 입구를 향해 걷고 있음.",
        "built_space": "좌측 전경의 높은 잔해와 배경의 건물 틈새가 존재하나, 건물 외벽의 디테일이 이전 샷 레퍼런스의 특징과 다소 차이가 있으며 덜 뚜렷함.",
        "entities": "인물들의 기본 복장은 비슷하나, 정한수의 점퍼 뒷면 어깨 부분에 레퍼런스에 존재하지 않는 검은색 기계 장치가 부착되어 있음. 수빈의 손에 든 물체가 방사능 측정기(사각형)가 아닌 정체불명의 둥근 원반(또는 통) 형태로 왜곡됨.",
        "hard_violations": [
         "invented objects: 정한수의 등에 레퍼런스에 없는 기계 장치(무전기 형태)가 부착됨",
         "invented objects: 수빈이 손에 든 물건이 방사능 측정기가 아닌 정체불명의 원반형 물체로 묘사됨"
        ],
        "physics": "두 인물이 지면에 발을 딛고 걷는 자세는 물리적으로 자연스럽게 지지되고 있으나, A 후보에 비해 지시된 엇갈린 걸음걸이와 체중 이동의 역동성이 다소 약함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "높은 잔해 위 어깨 너머 시점과 오른쪽 아래 두 사람의 배치는 맞지만, 수빈의 뒤꿈치 들림과 두 사람의 엇갈린 보행 순간이 B보다 불명확하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 전경 감시자 너머로 진입하는 두 사람의 전신을 내려다보며, 이현우의 전방 체중 이동과 수빈의 들린 뒷발을 더 명확하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 남자는 고개를 오른쪽 아래 두 사람 쪽으로 숙인다. 이현우와 수빈은 등을 보인 채 오른쪽의 어두운 입구를 향하며, 고개는 발밑 잔해로 내려가 있다. 두 사람의 이동 방향과 입구 위치는 연결되며 반대 방향으로 걷지 않는다.",
        "built_space": "왼쪽 전경에 관찰자의 등과 어깨, 그 아래에 깨진 상층 콘크리트 가장자리가 있다. 그 너머 낮은 접근로의 오른쪽 아래에 두 사람의 전신이 보인다. 주된 진입 개구부는 하나이며 양옆 콘크리트 벽체와 위쪽 파손 보, 노출 철근으로 둘러싸인다. 입구 안에는 기울어진 콘크리트 판들이 있고 접근로 왼쪽에는 물웅덩이가 있다. 덩굴과 나무, 낡은 콘크리트는 이전 장면과 부합하지만 구체적인 파손 형태까지 같은지는 확인하기 어렵다.",
        "entities": "사람은 세 명이다. 관찰자는 짧은 검은 머리와 갈색 후드 점퍼를 지닌 성인 남성의 뒷모습이며 얼굴은 드러나지 않는다. 등에 추가 배낭과 끈이 보인다. 이현우는 검은 머리, 마른 체격, 더럽혀진 어두운 셔츠와 바지를 유지한다. 수빈은 검은 단발, 올리브색 실용복, 카고 바지와 배낭을 유지하며 오른손 근처에 작은 계기형 물체가 있다. 두 사람 모두 얼굴에 방독면을 쓰지 않았다. 얼굴의 정확한 동일성, 상처와 인이어는 이 거리와 후면 각도에서 판별하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽 발은 잔해 위에 놓이고 다른 발은 뒤에 있어 전진 보행으로 읽힌다. 수빈 역시 잔해에 발을 디디고 있으나 뒷발의 뒤꿈치가 들리는 순간은 뚜렷하지 않다. 손 근처 물체는 손으로 지지되고 배낭은 어깨끈으로 지지된다. 관찰자의 하체는 화면 밖이지만 몸 앞에 상층 잔해가 있어 떠 있는 인물로 보이지 않는다. 붕괴 판재도 주변 잔해에 걸쳐 있다."
       },
       {
        "label": "B",
        "direction": "관찰자는 머리를 숙여 오른쪽 아래 이현우와 수빈을 향한다. 두 사람은 발밑을 살피며 화면 오른쪽 안쪽의 입구로 이동한다. 이현우가 앞서고 수빈이 뒤따르며, 몸과 발의 진행 방향이 실제로 입구에 이어진다.",
        "built_space": "카메라는 왼쪽 전경 관찰자의 어깨 바로 위쪽에서 낮은 접근로를 가파르게 내려다본다. 관찰자 아래의 깨진 상층 가장자리가 높이 차를 설명하고 두 사람의 전신을 가리지 않는다. 오른쪽에는 큰 진입 개구부 하나가 있고 양옆 벽체, 상부의 부서진 보와 매달린 금속 부재가 보인다. 두 사람은 개구부 아래쪽 문턱으로 이어지는 잔해 위에 있다. 왼쪽 접근로와 입구 내부에 물이 고여 있으며 덩굴, 나무, 콘크리트의 마모와 낮의 조명은 참조 장소에 부합한다.",
        "entities": "세 사람만 보인다. 관찰자는 짧은 검은 머리와 갈색 후드 점퍼를 가진 성인 남성으로 얼굴을 숨긴다. 참조에 없던 배낭 끈도 보인다. 이현우의 검은 머리, 어두운 얼룩진 셔츠와 바지, 수빈의 단발과 올리브색 실용복 및 배낭은 앞 장면의 외형을 이어간다. 수빈의 왼손에는 직사각형 계기 장치가 들려 있어 방사선 측정기 소지 상태가 읽힌다. 두 사람 모두 방독면을 착용하지 않았다. 얼굴 세부와 상처, 인이어는 후면 원경이라 검증하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우는 앞쪽 발을 콘크리트에 디딘 채 몸을 입구 방향으로 옮긴다. 수빈은 한쪽 다리로 체중을 지탱하고 다른 무릎을 굽혀 뒷발과 뒤꿈치를 들어 올려, 두 사람의 보행 단계가 분명히 다르다. 측정기는 손에 잡혀 있고 배낭은 끈에 매달려 있다. 관찰자의 발은 구도 밖이나 상층 잔해 뒤에 선 상체로 자연스럽게 읽힌다. 잔해와 금속 부재는 바닥 또는 남은 구조체에 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "높은 잔해 위 어깨 너머 시점과 오른쪽 아래 두 사람의 배치는 맞지만, 수빈의 뒤꿈치 들림과 두 사람의 엇갈린 보행 순간이 B보다 불명확하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 전경 감시자 너머로 진입하는 두 사람의 전신을 내려다보며, 이현우의 전방 체중 이동과 수빈의 들린 뒷발을 더 명확하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "전경 남자는 고개를 오른쪽 아래 두 사람 쪽으로 숙인다. 이현우와 수빈은 등을 보인 채 오른쪽의 어두운 입구를 향하며, 고개는 발밑 잔해로 내려가 있다. 두 사람의 이동 방향과 입구 위치는 연결되며 반대 방향으로 걷지 않는다.",
        "built_space": "왼쪽 전경에 관찰자의 등과 어깨, 그 아래에 깨진 상층 콘크리트 가장자리가 있다. 그 너머 낮은 접근로의 오른쪽 아래에 두 사람의 전신이 보인다. 주된 진입 개구부는 하나이며 양옆 콘크리트 벽체와 위쪽 파손 보, 노출 철근으로 둘러싸인다. 입구 안에는 기울어진 콘크리트 판들이 있고 접근로 왼쪽에는 물웅덩이가 있다. 덩굴과 나무, 낡은 콘크리트는 이전 장면과 부합하지만 구체적인 파손 형태까지 같은지는 확인하기 어렵다.",
        "entities": "사람은 세 명이다. 관찰자는 짧은 검은 머리와 갈색 후드 점퍼를 지닌 성인 남성의 뒷모습이며 얼굴은 드러나지 않는다. 등에 추가 배낭과 끈이 보인다. 이현우는 검은 머리, 마른 체격, 더럽혀진 어두운 셔츠와 바지를 유지한다. 수빈은 검은 단발, 올리브색 실용복, 카고 바지와 배낭을 유지하며 오른손 근처에 작은 계기형 물체가 있다. 두 사람 모두 얼굴에 방독면을 쓰지 않았다. 얼굴의 정확한 동일성, 상처와 인이어는 이 거리와 후면 각도에서 판별하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽 발은 잔해 위에 놓이고 다른 발은 뒤에 있어 전진 보행으로 읽힌다. 수빈 역시 잔해에 발을 디디고 있으나 뒷발의 뒤꿈치가 들리는 순간은 뚜렷하지 않다. 손 근처 물체는 손으로 지지되고 배낭은 어깨끈으로 지지된다. 관찰자의 하체는 화면 밖이지만 몸 앞에 상층 잔해가 있어 떠 있는 인물로 보이지 않는다. 붕괴 판재도 주변 잔해에 걸쳐 있다."
       },
       {
        "label": "A",
        "direction": "관찰자는 머리를 숙여 오른쪽 아래 이현우와 수빈을 향한다. 두 사람은 발밑을 살피며 화면 오른쪽 안쪽의 입구로 이동한다. 이현우가 앞서고 수빈이 뒤따르며, 몸과 발의 진행 방향이 실제로 입구에 이어진다.",
        "built_space": "카메라는 왼쪽 전경 관찰자의 어깨 바로 위쪽에서 낮은 접근로를 가파르게 내려다본다. 관찰자 아래의 깨진 상층 가장자리가 높이 차를 설명하고 두 사람의 전신을 가리지 않는다. 오른쪽에는 큰 진입 개구부 하나가 있고 양옆 벽체, 상부의 부서진 보와 매달린 금속 부재가 보인다. 두 사람은 개구부 아래쪽 문턱으로 이어지는 잔해 위에 있다. 왼쪽 접근로와 입구 내부에 물이 고여 있으며 덩굴, 나무, 콘크리트의 마모와 낮의 조명은 참조 장소에 부합한다.",
        "entities": "세 사람만 보인다. 관찰자는 짧은 검은 머리와 갈색 후드 점퍼를 가진 성인 남성으로 얼굴을 숨긴다. 참조에 없던 배낭 끈도 보인다. 이현우의 검은 머리, 어두운 얼룩진 셔츠와 바지, 수빈의 단발과 올리브색 실용복 및 배낭은 앞 장면의 외형을 이어간다. 수빈의 왼손에는 직사각형 계기 장치가 들려 있어 방사선 측정기 소지 상태가 읽힌다. 두 사람 모두 방독면을 착용하지 않았다. 얼굴 세부와 상처, 인이어는 후면 원경이라 검증하기 어렵다.",
        "hard_violations": [],
        "physics": "이현우는 앞쪽 발을 콘크리트에 디딘 채 몸을 입구 방향으로 옮긴다. 수빈은 한쪽 다리로 체중을 지탱하고 다른 무릎을 굽혀 뒷발과 뒤꿈치를 들어 올려, 두 사람의 보행 단계가 분명히 다르다. 측정기는 손에 잡혀 있고 배낭은 끈에 매달려 있다. 관찰자의 발은 구도 밖이나 상층 잔해 뒤에 선 상체로 자연스럽게 읽힌다. 잔해와 금속 부재는 바닥 또는 남은 구조체에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.556
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.306
   },
   "violations": {
    "B": [
     "[gemini-pro] invented objects: 정한수의 등에 레퍼런스에 없는 기계 장치(무전기 형태)가 부착됨",
     "[gemini-pro] invented objects: 수빈이 손에 든 물건이 방사능 측정기가 아닌 정체불명의 원반형 물체로 묘사됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1306
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 오버더숄더 앵글과 엇갈린 걸음걸이 묘사를 정확히 구현했으며, 인물들의 의상과 방사능 측정기의 디테일, 이전 샷의 공간적 특징을 가장 충실하게 재현함."
   },
   {
    "label": "B",
    "score": 1306,
    "verdict_ko": "구도와 배경은 무난하나, 정한수의 등에 레퍼런스에 없는 장치가 추가되고 수빈이 든 방사능 측정기가 원반 형태로 완전히 왜곡되어 프롬프트 충실도가 떨어짐.  ★위반: [gemini-pro] invented objects: 정한수의 등에 레퍼런스에 없는 기계 장치(무전기 형태)가 부착됨 / [gemini-pro] invented objects: 수빈이 손에 든 물건이 방사능 측정기가 아닌 정체불명의 원반형 물체로 묘사됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 수빈, 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S65sh12_sel.png",
    "asset_id": "a0d66d06-b946-444d-8770-18c13c8deda1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 정한수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:836640>",
    "asset_id": "be570610-37f0-4bda-b6e5-2e4fdaac528b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cda-2f40-7ae1-b5c6-1e1fc4d6466f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S65sh12"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S66sh13::signage": {
  "fp": "9dc4926d05b26e54",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::c6e47b88b5a425a0": {
  "subjects": [],
  "subject_text": "무너진 백화점 상층 내부·에스컬레이터·바닥 구멍\n어둡고 위험한 실내로 붕괴된 콘크리트 슬라브가 겹쳐져 있다.",
  "identity": "canonical",
  "scope_id": "L108",
  "scope_role": "location_interior",
  "scope_sha": "6eb425edf6ad000a"
 },
 "S66sh13::bgfirst_bg": {
  "input_fingerprint": "fbc88ca7ae585ea0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 부서진 바닥 잔해들과 함께 거대한 어둠 속을 향해 허공으로 떨어지고 있는 이현우와 수빈의 역동적인 찰나.\n\nLOCATION (lock): Within the collapsing floor opening beside the ruined department store's damaged escalator route, above the lower food section. Handheld flashlights cut through the dark interior.\n\nTIME OF DAY (lock): day, rain later.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Descend parallel to the falling pair from the established lateral position, keeping the lens slightly above their upper bodies and angled downward in an oblique side view. Place 이현우 lower-left with his legs folding beneath him and 수빈 upper-right with her torso pitching after him, both looking down into the space they are falling through; retain their full bodies, scattered floor fragments, and the broken opening across the upper background. Emphasize their downward displacement rather than rotating the camera, using restrained motion blur on the fragments while keeping the human silhouettes readable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed floor opening (Freshly broken by the floor collapse) — The underside and irregular opening edge remain visible above the falling pair; used as Upper spatial reference that makes the descent legible; Falling floor fragments (Falling alongside 이현우 and 수빈) — Separate fragments turn at different angles and remain smaller than the figures; used as Staggered foreground and background motion accents, kept clear of faces.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the department store's dimness and readable tonal separation between bodies, falling fragments, and the deeper darkness below, without adding a new source or atmospheric haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 부서진 바닥 잔해들과 함께 거대한 어둠 속을 향해 허공으로 떨어지고 있는 이현우와 수빈의 역동적인 찰나.\n\nLOCATION (lock): Within the collapsing floor opening beside the ruined department store's damaged escalator route, above the lower food section. Handheld flashlights cut through the dark interior.\n\nTIME OF DAY (lock): day, rain later.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Descend parallel to the falling pair from the established lateral position, keeping the lens slightly above their upper bodies and angled downward in an oblique side view. Place 이현우 lower-left with his legs folding beneath him and 수빈 upper-right with her torso pitching after him, both looking down into the space they are falling through; retain their full bodies, scattered floor fragments, and the broken opening across the upper background. Emphasize their downward displacement rather than rotating the camera, using restrained motion blur on the fragments while keeping the human silhouettes readable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed floor opening (Freshly broken by the floor collapse) — The underside and irregular opening edge remain visible above the falling pair; used as Upper spatial reference that makes the descent legible; Falling floor fragments (Falling alongside 이현우 and 수빈) — Separate fragments turn at different angles and remain smaller than the figures; used as Staggered foreground and background motion accents, kept clear of faces.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the department store's dimness and readable tonal separation between bodies, falling fragments, and the deeper darkness below, without adding a new source or atmospheric haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S66sh13__bgfirst_bg.png",
  "asset_id": "71ff7f4b-20d5-4169-b01b-878ec9dafcb1",
  "input_asset_ids": [
   "795e42d5-f005-4aee-ace8-39c3c8f0693b",
   "6666d6da-9c59-433c-b0e6-a6cd393ec81e"
  ]
 },
 "S66sh13": {
  "input_fingerprint": "27f301a067f181d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 부서진 바닥 잔해들과 함께 거대한 어둠 속을 향해 허공으로 떨어지고 있는 이현우와 수빈의 역동적인 찰나.\n\nLOCATION (lock): Within the collapsing floor opening beside the ruined department store's damaged escalator route, above the lower food section. Handheld flashlights cut through the dark interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Descend parallel to the falling pair from the established lateral position, keeping the lens slightly above their upper bodies and angled downward in an oblique side view. Place 이현우 lower-left with his legs folding beneath him and 수빈 upper-right with her torso pitching after him, both looking down into the space they are falling through; retain their full bodies, scattered floor fragments, and the broken opening across the upper background. Emphasize their downward displacement rather than rotating the camera, using restrained motion blur on the fragments while keeping the human silhouettes readable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed floor opening (Freshly broken by the floor collapse) — The underside and irregular opening edge remain visible above the falling pair; used as Upper spatial reference that makes the descent legible; Falling floor fragments (Falling alongside 이현우 and 수빈) — Separate fragments turn at different angles and remain smaller than the figures; used as Staggered foreground and background motion accents, kept clear of faces.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the department store's dimness and readable tonal separation between bodies, falling fragments, and the deeper darkness below, without adding a new source or atmospheric haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined store interior is dark, and the floor has collapsed into a sinkhole leading down to the food section. Broken walls surround the partly visible descending escalator. 이현우: He is falling through the collapsed floor, with his flashlight still in his possession and his gas mask off. His earlier injuries remain. 수빈: She is falling through the collapsed floor with her flashlight and radiation meter in her possession, her gas mask off. Her earlier wounds and radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 부서진 바닥 잔해들과 함께 거대한 어둠 속을 향해 허공으로 떨어지고 있는 이현우와 수빈의 역동적인 찰나.\n\nLOCATION (lock): Within the collapsing floor opening beside the ruined department store's damaged escalator route, above the lower food section. Handheld flashlights cut through the dark interior. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Descend parallel to the falling pair from the established lateral position, keeping the lens slightly above their upper bodies and angled downward in an oblique side view. Place 이현우 lower-left with his legs folding beneath him and 수빈 upper-right with her torso pitching after him, both looking down into the space they are falling through; retain their full bodies, scattered floor fragments, and the broken opening across the upper background. Emphasize their downward displacement rather than rotating the camera, using restrained motion blur on the fragments while keeping the human silhouettes readable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed floor opening (Freshly broken by the floor collapse) — The underside and irregular opening edge remain visible above the falling pair; used as Upper spatial reference that makes the descent legible; Falling floor fragments (Falling alongside 이현우 and 수빈) — Separate fragments turn at different angles and remain smaller than the figures; used as Staggered foreground and background motion accents, kept clear of faces.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the department store's dimness and readable tonal separation between bodies, falling fragments, and the deeper darkness below, without adding a new source or atmospheric haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined store interior is dark, and the floor has collapsed into a sinkhole leading down to the food section. Broken walls surround the partly visible descending escalator. 이현우: He is falling through the collapsed floor, with his flashlight still in his possession and his gas mask off. His earlier injuries remain. 수빈: She is falling through the collapsed floor with her flashlight and radiation meter in her possession, her gas mask off. Her earlier wounds and radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 부서진 바닥 잔해들과 함께 거대한 어둠 속을 향해 허공으로 떨어지고 있는 이현우와 수빈의 역동적인 찰나.\n\nLOCATION (lock): Within the collapsing floor opening beside the ruined department store's damaged escalator route, above the lower food section. Handheld flashlights cut through the dark interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Descend parallel to the falling pair from the established lateral position, keeping the lens slightly above their upper bodies and angled downward in an oblique side view. Place 이현우 lower-left with his legs folding beneath him and 수빈 upper-right with her torso pitching after him, both looking down into the space they are falling through; retain their full bodies, scattered floor fragments, and the broken opening across the upper background. Emphasize their downward displacement rather than rotating the camera, using restrained motion blur on the fragments while keeping the human silhouettes readable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Collapsed floor opening (Freshly broken by the floor collapse) — The underside and irregular opening edge remain visible above the falling pair; used as Upper spatial reference that makes the descent legible; Falling floor fragments (Falling alongside 이현우 and 수빈) — Separate fragments turn at different angles and remain smaller than the figures; used as Staggered foreground and background motion accents, kept clear of faces.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the department store's dimness and readable tonal separation between bodies, falling fragments, and the deeper darkness below, without adding a new source or atmospheric haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined store interior is dark, and the floor has collapsed into a sinkhole leading down to the food section. Broken walls surround the partly visible descending escalator. 이현우: He is falling through the collapsed floor, with his flashlight still in his possession and his gas mask off. His earlier injuries remain. 수빈: She is falling through the collapsed floor with her flashlight and radiation meter in her possession, her gas mask off. Her earlier wounds and radiation lesions remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S66sh13__bgfirst_bg.png",
     "asset_id": "71ff7f4b-20d5-4169-b01b-878ec9dafcb1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S66sh13.png",
     "asset_id": "795e42d5-f005-4aee-ace8-39c3c8f0693b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L108B02.png",
     "asset_id": "6666d6da-9c59-433c-b0e6-a6cd393ec81e",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 아래쪽 공간으로 시선과 손전등을 향하고 있음.",
    "built_space": "상단에 부서진 구멍과 에스컬레이터가 존재하나, 하단 공간에 에스컬레이터가 불필요하게 중복 생성됨.",
    "entities": "이현우는 지시된 소지품을 갖췄으나, 수빈은 필수 소지품인 방사능 측정기가 누락된 채 손전등(무전기 형태)만 들고 있음.",
    "hard_violations": [
     "[gemini-pro] 지탱하는 것 없음 (두 인물 모두 허공에 떠 있음)",
     "[gemini-pro] 단순 회전된 자세 (이현우의 포즈는 기어가는 자세를 물리적 변화 없이 회전시킨 형태)",
     "[gemini-pro] 구조물 중복 (프레임 하단에 에스컬레이터가 추가로 배치됨)",
     "[gpt-high] 참조 장소의 에스컬레이터 외에 구멍 하부 좌우로 별도 에스컬레이터 두 경로를 추가해, 고정된 장소의 주요 설비와 동선 구조를 변경했다."
    ],
    "physics": "지탱하는 것 없음. 특히 이현우는 추락하는 힘을 받지 않고 바닥에 엎드린 포즈가 허공에 떠 있는 상태로 부자연스러움."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 아래쪽 어둠을 향해 시선과 손전등 빛을 정확히 향하고 있음.",
    "built_space": "상단에 부서진 바닥 구멍과 에스컬레이터가 레퍼런스와 일치하게 위치하며, 하단은 깊은 어둠으로 표현됨.",
    "entities": "이현우(무전기, 손전등, 의상)와 수빈(오른손 손전등, 왼손 방사능 측정기, 의상) 모두 지시된 외형 및 소지품과 정확히 일치함.",
    "hard_violations": [
     "[gemini-pro] 지탱하는 것 없음 (두 인물 모두 물리적 지지대나 도약점 없이 허공에 떠 있음)"
    ],
    "physics": "지탱하는 것 없음. 허공을 떨어지는 중이며, 머리카락과 옷자락이 위로 쏠려 추락의 물리적 힘은 잘 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수빈의 소지품(손전등, 방사능 측정기)과 추락하는 역동성(머리카락, 옷자락)을 훌륭히 구현했으나, 지탱하는 요소 없이 허공에 떠 있어 엄격한 물리 법칙 지침을 위반했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수빈의 방사능 측정기가 누락되었고, 이현우의 포즈가 단순히 엎드린 자세를 회전시킨 형태이며, 공간 하단에 에스컬레이터가 중복 생성되어 치명적 오류가 다수 존재합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "두 인물 모두 아래쪽 어둠을 향해 시선과 손전등 빛을 정확히 향하고 있음.",
        "built_space": "상단에 부서진 바닥 구멍과 에스컬레이터가 레퍼런스와 일치하게 위치하며, 하단은 깊은 어둠으로 표현됨.",
        "entities": "이현우(무전기, 손전등, 의상)와 수빈(오른손 손전등, 왼손 방사능 측정기, 의상) 모두 지시된 외형 및 소지품과 정확히 일치함.",
        "hard_violations": [
         "지탱하는 것 없음 (두 인물 모두 물리적 지지대나 도약점 없이 허공에 떠 있음)"
        ],
        "physics": "지탱하는 것 없음. 허공을 떨어지는 중이며, 머리카락과 옷자락이 위로 쏠려 추락의 물리적 힘은 잘 표현됨."
       },
       {
        "label": "A",
        "direction": "두 인물 모두 아래쪽 공간으로 시선과 손전등을 향하고 있음.",
        "built_space": "상단에 부서진 구멍과 에스컬레이터가 존재하나, 하단 공간에 에스컬레이터가 불필요하게 중복 생성됨.",
        "entities": "이현우는 지시된 소지품을 갖췄으나, 수빈은 필수 소지품인 방사능 측정기가 누락된 채 손전등(무전기 형태)만 들고 있음.",
        "hard_violations": [
         "지탱하는 것 없음 (두 인물 모두 허공에 떠 있음)",
         "단순 회전된 자세 (이현우의 포즈는 기어가는 자세를 물리적 변화 없이 회전시킨 형태)",
         "구조물 중복 (프레임 하단에 에스컬레이터가 추가로 배치됨)"
        ],
        "physics": "지탱하는 것 없음. 특히 이현우는 추락하는 힘을 받지 않고 바닥에 엎드린 포즈가 허공에 떠 있는 상태로 부자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "수빈의 소지품(손전등, 방사능 측정기)과 추락하는 역동성(머리카락, 옷자락)을 훌륭히 구현했으나, 지탱하는 요소 없이 허공에 떠 있어 엄격한 물리 법칙 지침을 위반했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수빈의 방사능 측정기가 누락되었고, 이현우의 포즈가 단순히 엎드린 자세를 회전시킨 형태이며, 공간 하단에 에스컬레이터가 중복 생성되어 치명적 오류가 다수 존재합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 인물 모두 아래쪽 어둠을 향해 시선과 손전등 빛을 정확히 향하고 있음.",
        "built_space": "상단에 부서진 바닥 구멍과 에스컬레이터가 레퍼런스와 일치하게 위치하며, 하단은 깊은 어둠으로 표현됨.",
        "entities": "이현우(무전기, 손전등, 의상)와 수빈(오른손 손전등, 왼손 방사능 측정기, 의상) 모두 지시된 외형 및 소지품과 정확히 일치함.",
        "hard_violations": [
         "지탱하는 것 없음 (두 인물 모두 물리적 지지대나 도약점 없이 허공에 떠 있음)"
        ],
        "physics": "지탱하는 것 없음. 허공을 떨어지는 중이며, 머리카락과 옷자락이 위로 쏠려 추락의 물리적 힘은 잘 표현됨."
       },
       {
        "label": "A",
        "direction": "두 인물 모두 아래쪽 공간으로 시선과 손전등을 향하고 있음.",
        "built_space": "상단에 부서진 구멍과 에스컬레이터가 존재하나, 하단 공간에 에스컬레이터가 불필요하게 중복 생성됨.",
        "entities": "이현우는 지시된 소지품을 갖췄으나, 수빈은 필수 소지품인 방사능 측정기가 누락된 채 손전등(무전기 형태)만 들고 있음.",
        "hard_violations": [
         "지탱하는 것 없음 (두 인물 모두 허공에 떠 있음)",
         "단순 회전된 자세 (이현우의 포즈는 기어가는 자세를 물리적 변화 없이 회전시킨 형태)",
         "구조물 중복 (프레임 하단에 에스컬레이터가 추가로 배치됨)"
        ],
        "physics": "지탱하는 것 없음. 특히 이현우는 추락하는 힘을 받지 않고 바닥에 엎드린 포즈가 허공에 떠 있는 상태로 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 인물의 전신·좌하단/우상단 배치와 어둠 속 낙하, 손전등 및 측정기 소지가 충실하나, 카메라 수평이 기울고 하부의 추가 조명이 남는다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물 배치와 추락 동작은 맞지만 하부에 에스컬레이터 두 경로를 추가하고 밝은 상층 바닥을 크게 보여, 장소 고정과 구멍 내부의 측면 하강 시점에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 좌하단에서 고개를 아래로 숙이고, 수빈은 우상단에서 그보다 아래쪽 낙하 공간을 본다. 두 손전등의 발광면과 빛줄기는 하부 공간을 향한다. 이현우의 접힌 다리와 수빈의 앞으로 기운 상체가 아래로 떨어지는 관계를 만든다.",
        "built_space": "상단 중앙에 에스컬레이터 한 경로가 일부 보이고, 주변에 유리 난간과 기둥이 있다. 불규칙하게 깨진 슬래브 가장자리와 밑면이 인물들 위를 가로지르며, 양옆에는 노출 철근과 파손된 벽체가 이어진다. 카메라는 개구부 안에서 두 사람보다 조금 높은 위치로 읽힌다. 다만 구조물의 수평선이 기울어 카메라 회전을 억제하라는 지시에는 덜 충실하며, 하부에는 별도의 점광원들이 보인다.",
        "entities": "인물은 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 어두운 셔츠와 바지, 얼굴과 옷의 상처·혈흔, 귀의 소형 장치가 보인다. 수빈은 검은 단발의 젊은 동아시아계 여성으로 올리브색 상의, 카고 바지, 부츠와 장갑을 착용했다. 두 사람 모두 방독면을 쓰지 않았으며 각자 손전등을 쥐고 있다. 수빈의 반대 손에는 화면이 있는 소형 측정기가 보인다. 국적과 정확한 나이, 피폭 병변의 성격은 외관만으로 확정할 수 없다. 여러 바닥 파편이 얼굴을 가리지 않고 배치되어 있다.",
        "hard_violations": [],
        "physics": "두 사람의 발은 바닥에 닿지 않지만, 바로 위의 붕괴한 개구부와 함께 떨어지는 파편이 지지면을 잃은 추락의 출발점을 설명한다. 아래에는 낙하가 계속될 공간이 열려 있어 근거 없는 공중 부양으로 보이지 않는다. 접힌 무릎, 벌어진 팔과 들린 머리카락도 순간적인 추락에 가능하다. 손전등과 측정기는 손에 잡혀 있고, 파편에는 붕괴한 슬래브라는 분명한 발생원이 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 좌하단에서 얼굴을 오른쪽 아래 낙하 공간으로 돌리고, 수빈은 우상단에서 아래쪽으로 시선을 두며 한 손을 이현우 쪽으로 뻗는다. 손전등 빛은 모두 아래로 향한다. 수빈의 다리는 위에 남고 상체가 먼저 내려가므로 뒤따라 고꾸라지는 방향은 읽힌다.",
        "built_space": "상단 중앙 에스컬레이터 한 경로와 그 양쪽 유리 난간, 기둥, 오른쪽 창벽이 보인다. 추가로 구멍 아래 좌우에 각각 에스컬레이터 경로가 하나씩 보여 총 세 경로가 구성된다. 이는 참조의 고정 장소에 없는 별도 하부 동선을 크게 도입한 것이다. 깨진 가장자리는 인물 위에 있지만 슬래브 밑면보다 상층 바닥 윗면이 넓게 보여, 구멍 안에서 나란히 하강하는 측면 시점보다는 위에서 내려다보는 시점이 강하다. 하부 판매대와 조명도 비교적 밝게 드러난다.",
        "entities": "인물은 이현우와 수빈으로 읽히는 두 명이며, 젊은 얼굴과 검은 머리, 마른 체격, 어두운 남성복과 올리브색 여성 전술복은 대체로 일치한다. 이현우의 인이어 장치와 두 사람의 피부 상처가 보이고 방독면은 없다. 각자의 손전등은 보이지만 수빈이 함께 든 기기는 긴 안테나가 달린 무전기처럼 보여 방사선 측정기로 명확하게 식별되지 않는다. 파편은 여러 거리와 각도로 흩어져 있고 얼굴을 가리지 않는다.",
        "hard_violations": [
         "참조 장소의 에스컬레이터 외에 구멍 하부 좌우로 별도 에스컬레이터 두 경로를 추가해, 고정된 장소의 주요 설비와 동선 구조를 변경했다."
        ],
        "physics": "두 사람 모두 바닥 지지를 잃고 깨진 개구부 아래로 떨어지는 상황이며, 붕괴한 상층 바닥이 추락의 출발점을 제공한다. 수빈의 거의 엎드린 자세도 상체가 먼저 쏠린 추락으로 가능하고, 단순히 보행 자세를 회전시킨 모습은 아니다. 손전등과 소형 기기는 손으로 지지되며 주변 파편은 슬래브 붕괴로 설명된다. 인물이나 소품이 원인 없이 떠 있다고 판단할 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 인물의 전신·좌하단/우상단 배치와 어둠 속 낙하, 손전등 및 측정기 소지가 충실하나, 카메라 수평이 기울고 하부의 추가 조명이 남는다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물 배치와 추락 동작은 맞지만 하부에 에스컬레이터 두 경로를 추가하고 밝은 상층 바닥을 크게 보여, 장소 고정과 구멍 내부의 측면 하강 시점에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 좌하단에서 고개를 아래로 숙이고, 수빈은 우상단에서 그보다 아래쪽 낙하 공간을 본다. 두 손전등의 발광면과 빛줄기는 하부 공간을 향한다. 이현우의 접힌 다리와 수빈의 앞으로 기운 상체가 아래로 떨어지는 관계를 만든다.",
        "built_space": "상단 중앙에 에스컬레이터 한 경로가 일부 보이고, 주변에 유리 난간과 기둥이 있다. 불규칙하게 깨진 슬래브 가장자리와 밑면이 인물들 위를 가로지르며, 양옆에는 노출 철근과 파손된 벽체가 이어진다. 카메라는 개구부 안에서 두 사람보다 조금 높은 위치로 읽힌다. 다만 구조물의 수평선이 기울어 카메라 회전을 억제하라는 지시에는 덜 충실하며, 하부에는 별도의 점광원들이 보인다.",
        "entities": "인물은 두 명뿐이다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 어두운 셔츠와 바지, 얼굴과 옷의 상처·혈흔, 귀의 소형 장치가 보인다. 수빈은 검은 단발의 젊은 동아시아계 여성으로 올리브색 상의, 카고 바지, 부츠와 장갑을 착용했다. 두 사람 모두 방독면을 쓰지 않았으며 각자 손전등을 쥐고 있다. 수빈의 반대 손에는 화면이 있는 소형 측정기가 보인다. 국적과 정확한 나이, 피폭 병변의 성격은 외관만으로 확정할 수 없다. 여러 바닥 파편이 얼굴을 가리지 않고 배치되어 있다.",
        "hard_violations": [],
        "physics": "두 사람의 발은 바닥에 닿지 않지만, 바로 위의 붕괴한 개구부와 함께 떨어지는 파편이 지지면을 잃은 추락의 출발점을 설명한다. 아래에는 낙하가 계속될 공간이 열려 있어 근거 없는 공중 부양으로 보이지 않는다. 접힌 무릎, 벌어진 팔과 들린 머리카락도 순간적인 추락에 가능하다. 손전등과 측정기는 손에 잡혀 있고, 파편에는 붕괴한 슬래브라는 분명한 발생원이 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 좌하단에서 얼굴을 오른쪽 아래 낙하 공간으로 돌리고, 수빈은 우상단에서 아래쪽으로 시선을 두며 한 손을 이현우 쪽으로 뻗는다. 손전등 빛은 모두 아래로 향한다. 수빈의 다리는 위에 남고 상체가 먼저 내려가므로 뒤따라 고꾸라지는 방향은 읽힌다.",
        "built_space": "상단 중앙 에스컬레이터 한 경로와 그 양쪽 유리 난간, 기둥, 오른쪽 창벽이 보인다. 추가로 구멍 아래 좌우에 각각 에스컬레이터 경로가 하나씩 보여 총 세 경로가 구성된다. 이는 참조의 고정 장소에 없는 별도 하부 동선을 크게 도입한 것이다. 깨진 가장자리는 인물 위에 있지만 슬래브 밑면보다 상층 바닥 윗면이 넓게 보여, 구멍 안에서 나란히 하강하는 측면 시점보다는 위에서 내려다보는 시점이 강하다. 하부 판매대와 조명도 비교적 밝게 드러난다.",
        "entities": "인물은 이현우와 수빈으로 읽히는 두 명이며, 젊은 얼굴과 검은 머리, 마른 체격, 어두운 남성복과 올리브색 여성 전술복은 대체로 일치한다. 이현우의 인이어 장치와 두 사람의 피부 상처가 보이고 방독면은 없다. 각자의 손전등은 보이지만 수빈이 함께 든 기기는 긴 안테나가 달린 무전기처럼 보여 방사선 측정기로 명확하게 식별되지 않는다. 파편은 여러 거리와 각도로 흩어져 있고 얼굴을 가리지 않는다.",
        "hard_violations": [
         "참조 장소의 에스컬레이터 외에 구멍 하부 좌우로 별도 에스컬레이터 두 경로를 추가해, 고정된 장소의 주요 설비와 동선 구조를 변경했다."
        ],
        "physics": "두 사람 모두 바닥 지지를 잃고 깨진 개구부 아래로 떨어지는 상황이며, 붕괴한 상층 바닥이 추락의 출발점을 제공한다. 수빈의 거의 엎드린 자세도 상체가 먼저 쏠린 추락으로 가능하고, 단순히 보행 자세를 회전시킨 모습은 아니다. 손전등과 소형 기기는 손으로 지지되며 주변 파편은 슬래브 붕괴로 설명된다. 인물이나 소품이 원인 없이 떠 있다고 판단할 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.054,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.804,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 지탱하는 것 없음 (두 인물 모두 물리적 지지대나 도약점 없이 허공에 떠 있음)"
    ],
    "A": [
     "[gemini-pro] 지탱하는 것 없음 (두 인물 모두 허공에 떠 있음)",
     "[gemini-pro] 단순 회전된 자세 (이현우의 포즈는 기어가는 자세를 물리적 변화 없이 회전시킨 형태)",
     "[gemini-pro] 구조물 중복 (프레임 하단에 에스컬레이터가 추가로 배치됨)",
     "[gpt-high] 참조 장소의 에스컬레이터 외에 구멍 하부 좌우로 별도 에스컬레이터 두 경로를 추가해, 고정된 장소의 주요 설비와 동선 구조를 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 804
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "수빈의 소지품(손전등, 방사능 측정기)과 추락하는 역동성(머리카락, 옷자락)을 훌륭히 구현했으나, 지탱하는 요소 없이 허공에 떠 있어 엄격한 물리 법칙 지침을 위반했습니다.  ★위반: [gemini-pro] 지탱하는 것 없음 (두 인물 모두 물리적 지지대나 도약점 없이 허공에 떠 있음)"
   },
   {
    "label": "A",
    "score": 804,
    "verdict_ko": "수빈의 방사능 측정기가 누락되었고, 이현우의 포즈가 단순히 엎드린 자세를 회전시킨 형태이며, 공간 하단에 에스컬레이터가 중복 생성되어 치명적 오류가 다수 존재합니다.  ★위반: [gemini-pro] 지탱하는 것 없음 (두 인물 모두 허공에 떠 있음) / [gemini-pro] 단순 회전된 자세 (이현우의 포즈는 기어가는 자세를 물리적 변화 없이 회전시킨 형태) / [gemini-pro] 구조물 중복 (프레임 하단에 에스컬레이터가 추가로 배치됨) / [gpt-high] 참조 장소의 에스컬레이터 외에 구멍 하부 좌우로 별도 에스컬레이터 두 경로를 추가해, 고정된 장소의 주요 설비와 동선 구조를 변경했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L108B02.png",
    "asset_id": "6666d6da-9c59-433c-b0e6-a6cd393ec81e",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ce0-6ca1-71ba-aef5-4dfe750627d1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S66sh13__bgfirst_bg.png",
   "bg_asset_id": "71ff7f4b-20d5-4169-b01b-878ec9dafcb1",
   "bg_record_key": "S66sh13::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh38::signage": {
  "fp": "32ea7cd97235e3e4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh38": {
  "input_fingerprint": "4660163ed37ff46a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 어두운 천장 구멍에서 굵은 밧줄 하나가 바닥을 향해 허공으로 늘어진 찰나.\n\nLOCATION (lock): In the department store's buried lower food section, directly beneath the high hole in the collapsed floor above. The opening admits the limited outside light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at the crane's low starting position above the lower floor, laterally offset from the rope and tilted steeply upward toward the ceiling opening. Place the opening near upper center and let the thick rope descend through the central space, its free lower end clearly visible below center with ample dark room around it. Keep all people outside the frame and delay the upward camera travel, making the rope's arrival the only moving emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Ceiling opening above the suspended rope in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair fell through the collapsed floor) — Seen from below at a steep oblique angle, with its underside and broken perimeter visible; used as Upper anchor establishing where the rescue rope originates; Descending rope (Hanging down from the opening, with its free end still suspended) — Its vertical length is seen from the side rather than directly beneath it; used as Narrow central line connecting the distant opening to the lower space without occupying a large portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dark interior with restrained ambient separation sufficient to distinguish the rope and broken opening, without turning the opening into an unsupported shaft of light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food section below the collapsed floor contains dusty wine, canned food and snacks, while the other exits are blocked. An open parasol is present near the sleeping place, and rain drips through the opening overhead. A rope descends through the ceiling opening into the food section.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 어두운 천장 구멍에서 굵은 밧줄 하나가 바닥을 향해 허공으로 늘어진 찰나.\n\nLOCATION (lock): In the department store's buried lower food section, directly beneath the high hole in the collapsed floor above. The opening admits the limited outside light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at the crane's low starting position above the lower floor, laterally offset from the rope and tilted steeply upward toward the ceiling opening. Place the opening near upper center and let the thick rope descend through the central space, its free lower end clearly visible below center with ample dark room around it. Keep all people outside the frame and delay the upward camera travel, making the rope's arrival the only moving emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Ceiling opening above the suspended rope in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair fell through the collapsed floor) — Seen from below at a steep oblique angle, with its underside and broken perimeter visible; used as Upper anchor establishing where the rescue rope originates; Descending rope (Hanging down from the opening, with its free end still suspended) — Its vertical length is seen from the side rather than directly beneath it; used as Narrow central line connecting the distant opening to the lower space without occupying a large portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dark interior with restrained ambient separation sufficient to distinguish the rope and broken opening, without turning the opening into an unsupported shaft of light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food section below the collapsed floor contains dusty wine, canned food and snacks, while the other exits are blocked. An open parasol is present near the sleeping place, and rain drips through the opening overhead. A rope descends through the ceiling opening into the food section.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 어두운 천장 구멍에서 굵은 밧줄 하나가 바닥을 향해 허공으로 늘어진 찰나.\n\nLOCATION (lock): In the department store's buried lower food section, directly beneath the high hole in the collapsed floor above. The opening admits the limited outside light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at the crane's low starting position above the lower floor, laterally offset from the rope and tilted steeply upward toward the ceiling opening. Place the opening near upper center and let the thick rope descend through the central space, its free lower end clearly visible below center with ample dark room around it. Keep all people outside the frame and delay the upward camera travel, making the rope's arrival the only moving emphasis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Ceiling opening above the suspended rope in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Ceiling opening (Open where the pair fell through the collapsed floor) — Seen from below at a steep oblique angle, with its underside and broken perimeter visible; used as Upper anchor establishing where the rescue rope originates; Descending rope (Hanging down from the opening, with its free end still suspended) — Its vertical length is seen from the side rather than directly beneath it; used as Narrow central line connecting the distant opening to the lower space without occupying a large portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dark interior with restrained ambient separation sufficient to distinguish the rope and broken opening, without turning the opening into an unsupported shaft of light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The food section below the collapsed floor contains dusty wine, canned food and snacks, while the other exits are blocked. An open parasol is present near the sleeping place, and rain drips through the opening overhead. A rope descends through the ceiling opening into the food section.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 아래에서 천장 구멍을 가파르게 올려다보며, 밧줄이 중앙을 관통해 하단으로 향함.",
    "built_space": "붕괴된 천장 위로 레퍼런스와 일치하는 에스컬레이터가 보임. 좌측 진열대, 우측에 붉은 우산이 적절히 배치됨.",
    "entities": "끝이 묶인 굵은 밧줄, 와인병, 통조림, 펼쳐진 붉은 우산, 떨어지는 빗물 모두 존재함.",
    "hard_violations": [
     "[gpt-high] 프롬프트와 참조에 없는 켜진 랜턴을 오른쪽에 추가하여 새로운 소품과 실내 광원을 만들었다."
    ],
    "physics": "밧줄과 빗물이 중력에 맞게 구멍에서 수직으로 떨어지며 허공에 자연스럽게 매달려 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 하층부에서 천장 구멍을 향해 위로 꺾여 있으며, 중앙으로 밧줄이 떨어짐.",
    "built_space": "천장 구멍 위로 상층부 난간이 보이나 에스컬레이터는 명확하지 않음. 좌측에 우산과 와인, 우측에 진열대 위치.",
    "entities": "끝단이 묶인 밧줄, 검은 우산, 와인병, 식료품, 빗물 등 요구된 사물들이 확인됨.",
    "hard_violations": [],
    "physics": "밧줄이 허공을 가로질러 수직으로 안정감 있게 늘어져 있으며 빗방울이 자연스럽게 떨어짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 건축적 특징(상단 에스컬레이터)을 훌륭하게 반영하였으며, 요구된 카메라 구도와 소품(와인, 우산, 밧줄)을 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 카메라 각도와 공간의 분위기는 잘 연출되었으나, 상층부 구멍 너머의 레퍼런스 디테일이 A에 비해 다소 생략됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 아래에서 천장 구멍을 가파르게 올려다보며, 밧줄이 중앙을 관통해 하단으로 향함.",
        "built_space": "붕괴된 천장 위로 레퍼런스와 일치하는 에스컬레이터가 보임. 좌측 진열대, 우측에 붉은 우산이 적절히 배치됨.",
        "entities": "끝이 묶인 굵은 밧줄, 와인병, 통조림, 펼쳐진 붉은 우산, 떨어지는 빗물 모두 존재함.",
        "hard_violations": [],
        "physics": "밧줄과 빗물이 중력에 맞게 구멍에서 수직으로 떨어지며 허공에 자연스럽게 매달려 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 하층부에서 천장 구멍을 향해 위로 꺾여 있으며, 중앙으로 밧줄이 떨어짐.",
        "built_space": "천장 구멍 위로 상층부 난간이 보이나 에스컬레이터는 명확하지 않음. 좌측에 우산과 와인, 우측에 진열대 위치.",
        "entities": "끝단이 묶인 밧줄, 검은 우산, 와인병, 식료품, 빗물 등 요구된 사물들이 확인됨.",
        "hard_violations": [],
        "physics": "밧줄이 허공을 가로질러 수직으로 안정감 있게 늘어져 있으며 빗방울이 자연스럽게 떨어짐."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 건축적 특징(상단 에스컬레이터)을 훌륭하게 반영하였으며, 요구된 카메라 구도와 소품(와인, 우산, 밧줄)을 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지시된 카메라 각도와 공간의 분위기는 잘 연출되었으나, 상층부 구멍 너머의 레퍼런스 디테일이 A에 비해 다소 생략됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 아래에서 천장 구멍을 가파르게 올려다보며, 밧줄이 중앙을 관통해 하단으로 향함.",
        "built_space": "붕괴된 천장 위로 레퍼런스와 일치하는 에스컬레이터가 보임. 좌측 진열대, 우측에 붉은 우산이 적절히 배치됨.",
        "entities": "끝이 묶인 굵은 밧줄, 와인병, 통조림, 펼쳐진 붉은 우산, 떨어지는 빗물 모두 존재함.",
        "hard_violations": [],
        "physics": "밧줄과 빗물이 중력에 맞게 구멍에서 수직으로 떨어지며 허공에 자연스럽게 매달려 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 하층부에서 천장 구멍을 향해 위로 꺾여 있으며, 중앙으로 밧줄이 떨어짐.",
        "built_space": "천장 구멍 위로 상층부 난간이 보이나 에스컬레이터는 명확하지 않음. 좌측에 우산과 와인, 우측에 진열대 위치.",
        "entities": "끝단이 묶인 밧줄, 검은 우산, 와인병, 식료품, 빗물 등 요구된 사물들이 확인됨.",
        "hard_violations": [],
        "physics": "밧줄이 허공을 가로질러 수직으로 안정감 있게 늘어져 있으며 빗방울이 자연스럽게 떨어짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "더 가파른 상향 시점과 넓은 암부 속 중앙 밧줄이 지정된 순간과 구도를 충실히 구현하지만, 참조의 에스컬레이터는 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "상부 에스컬레이터로 장소는 잘 연결되지만, 바닥과 진열대의 비중이 커지고 지시되지 않은 켜진 랜턴이 추가되어 밧줄만 강조하는 장면에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "밧줄 하나가 상단 중앙의 개구부에서 바닥 방향으로 거의 수직으로 내려오며, 풀어진 끝은 화면 중심 아래에서 멈춘다. 카메라는 밧줄 옆에서 천장 아래쪽을 가파르게 올려다본다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "큰 붕괴 개구부 하나와 그 위 유리 난간이 보인다. 둘레에는 콘크리트 단면, 철근과 배관이 드러나고, 하부에는 여러 기둥과 양옆 식품 진열대가 있다. 펼친 파라솔은 왼쪽에 하나 있다. 참조의 콘크리트 구조와 유리 난간은 연결되지만 에스컬레이터는 이 구도에서 확인되지 않는다. 바닥은 하단의 좁은 띠로 남아 상향 구도가 강하다.",
        "entities": "굵게 꼬인 밧줄 하나, 자유롭게 늘어진 하단, 붕괴 구멍 하나가 분명하다. 진열대에 와인병, 통조림과 포장 식품이 보이며 펼친 파라솔도 있다. 사람과 얼굴은 없다. 구멍 아래 물방울과 젖은 바닥이 보이고, 침구와 다른 출구의 폐쇄 상태는 확인되지 않는다. 읽을 수 있는 추가 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "밧줄은 구멍 위 화면 밖으로 이어지며 팽팽한 상단을 통해 위에서 매달린 상태로 읽힌다. 실제 고정점은 보이지 않지만 밧줄 전체가 공중에 분리되어 뜬 모습은 아니다. 하단은 바닥에 닿지 않는다. 매달린 콘크리트 조각에는 철근 연결이 보이고, 상품은 선반에, 잔해는 바닥에 놓여 있다. 파라솔에는 아래로 이어지는 지주가 보인다."
       },
       {
        "label": "B",
        "direction": "밧줄은 상단 중앙 개구부 가장자리에서 수직으로 내려오고 자유단은 중심 아래에 떠 있다. 카메라는 아래에서 비스듬히 위를 향하지만 A보다 바닥과 매장 내부가 더 많이 들어온다. 상부 에스컬레이터는 개구부 뒤에서 위쪽으로 이어진다. 사람이나 시선은 없다.",
        "built_space": "붕괴 개구부 하나, 그 뒤 에스컬레이터 하나와 양옆 유리 난간이 보여 참조 장소의 고정 시설을 잘 연결한다. 구멍 둘레의 노출 철근과 깨진 슬래브, 하부 기둥들이 보인다. 왼쪽 전경과 중앙 및 오른쪽에 진열대가 있고, 오른쪽에 펼친 파라솔 하나가 있다. 하단의 넓은 바닥과 중앙 진열대가 공간을 크게 차지한다. 오른쪽에는 켜진 랜턴이 추가되어 있다.",
        "entities": "굵은 밧줄 하나와 매듭 아래 풀어진 끝, 와인병, 통조림, 포장 식품, 펼친 파라솔이 보인다. 인물과 얼굴은 없다. 물방울과 젖은 잔해 바닥도 있다. 침구와 모든 출구의 차단 여부는 확인되지 않는다. 오른쪽의 불 켜진 랜턴은 프롬프트나 참조에 없는 소품이다.",
        "hard_violations": [
         "프롬프트와 참조에 없는 켜진 랜턴을 오른쪽에 추가하여 새로운 소품과 실내 광원을 만들었다."
        ],
        "physics": "밧줄 상단은 개구부 가장자리 너머로 이어지며 위에서 지지되는 매달림으로 읽힌다. 고정 장치 자체는 가려져 있고, 자유단은 바닥 위에 남아 있다. 큰 콘크리트 파편들은 노출 철근에 연결되어 있다. 진열대와 잔해는 바닥에, 병과 캔은 선반에 놓여 있으며 파라솔에는 지주가 있다. 랜턴도 물체 위에 놓여 있어 부유하는 소품은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "더 가파른 상향 시점과 넓은 암부 속 중앙 밧줄이 지정된 순간과 구도를 충실히 구현하지만, 참조의 에스컬레이터는 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "상부 에스컬레이터로 장소는 잘 연결되지만, 바닥과 진열대의 비중이 커지고 지시되지 않은 켜진 랜턴이 추가되어 밧줄만 강조하는 장면에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "밧줄 하나가 상단 중앙의 개구부에서 바닥 방향으로 거의 수직으로 내려오며, 풀어진 끝은 화면 중심 아래에서 멈춘다. 카메라는 밧줄 옆에서 천장 아래쪽을 가파르게 올려다본다. 사람이나 시선, 조준 대상은 없다.",
        "built_space": "큰 붕괴 개구부 하나와 그 위 유리 난간이 보인다. 둘레에는 콘크리트 단면, 철근과 배관이 드러나고, 하부에는 여러 기둥과 양옆 식품 진열대가 있다. 펼친 파라솔은 왼쪽에 하나 있다. 참조의 콘크리트 구조와 유리 난간은 연결되지만 에스컬레이터는 이 구도에서 확인되지 않는다. 바닥은 하단의 좁은 띠로 남아 상향 구도가 강하다.",
        "entities": "굵게 꼬인 밧줄 하나, 자유롭게 늘어진 하단, 붕괴 구멍 하나가 분명하다. 진열대에 와인병, 통조림과 포장 식품이 보이며 펼친 파라솔도 있다. 사람과 얼굴은 없다. 구멍 아래 물방울과 젖은 바닥이 보이고, 침구와 다른 출구의 폐쇄 상태는 확인되지 않는다. 읽을 수 있는 추가 문구나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "밧줄은 구멍 위 화면 밖으로 이어지며 팽팽한 상단을 통해 위에서 매달린 상태로 읽힌다. 실제 고정점은 보이지 않지만 밧줄 전체가 공중에 분리되어 뜬 모습은 아니다. 하단은 바닥에 닿지 않는다. 매달린 콘크리트 조각에는 철근 연결이 보이고, 상품은 선반에, 잔해는 바닥에 놓여 있다. 파라솔에는 아래로 이어지는 지주가 보인다."
       },
       {
        "label": "A",
        "direction": "밧줄은 상단 중앙 개구부 가장자리에서 수직으로 내려오고 자유단은 중심 아래에 떠 있다. 카메라는 아래에서 비스듬히 위를 향하지만 A보다 바닥과 매장 내부가 더 많이 들어온다. 상부 에스컬레이터는 개구부 뒤에서 위쪽으로 이어진다. 사람이나 시선은 없다.",
        "built_space": "붕괴 개구부 하나, 그 뒤 에스컬레이터 하나와 양옆 유리 난간이 보여 참조 장소의 고정 시설을 잘 연결한다. 구멍 둘레의 노출 철근과 깨진 슬래브, 하부 기둥들이 보인다. 왼쪽 전경과 중앙 및 오른쪽에 진열대가 있고, 오른쪽에 펼친 파라솔 하나가 있다. 하단의 넓은 바닥과 중앙 진열대가 공간을 크게 차지한다. 오른쪽에는 켜진 랜턴이 추가되어 있다.",
        "entities": "굵은 밧줄 하나와 매듭 아래 풀어진 끝, 와인병, 통조림, 포장 식품, 펼친 파라솔이 보인다. 인물과 얼굴은 없다. 물방울과 젖은 잔해 바닥도 있다. 침구와 모든 출구의 차단 여부는 확인되지 않는다. 오른쪽의 불 켜진 랜턴은 프롬프트나 참조에 없는 소품이다.",
        "hard_violations": [
         "프롬프트와 참조에 없는 켜진 랜턴을 오른쪽에 추가하여 새로운 소품과 실내 광원을 만들었다."
        ],
        "physics": "밧줄 상단은 개구부 가장자리 너머로 이어지며 위에서 지지되는 매달림으로 읽힌다. 고정 장치 자체는 가려져 있고, 자유단은 바닥 위에 남아 있다. 큰 콘크리트 파편들은 노출 철근에 연결되어 있다. 진열대와 잔해는 바닥에, 병과 캔은 선반에 놓여 있으며 파라솔에는 지주가 있다. 랜턴도 물체 위에 놓여 있어 부유하는 소품은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.857
   },
   "violations": {
    "A": [
     "[gpt-high] 프롬프트와 참조에 없는 켜진 랜턴을 오른쪽에 추가하여 새로운 소품과 실내 광원을 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1417,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "레퍼런스의 건축적 특징(상단 에스컬레이터)을 훌륭하게 반영하였으며, 요구된 카메라 구도와 소품(와인, 우산, 밧줄)을 완벽히 구현함.  ★위반: [gpt-high] 프롬프트와 참조에 없는 켜진 랜턴을 오른쪽에 추가하여 새로운 소품과 실내 광원을 만들었다."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "지시된 카메라 각도와 공간의 분위기는 잘 연출되었으나, 상층부 구멍 너머의 레퍼런스 디테일이 A에 비해 다소 생략됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L108B02.png",
    "asset_id": "6666d6da-9c59-433c-b0e6-a6cd393ec81e",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cea-3dd8-7ab9-9b3d-400530b9f817",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh46::signage": {
  "fp": "70e86e388599f7eb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh46": {
  "input_fingerprint": "93176718ba916d09",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 날카로운 눈빛과 짧은 머리를 한 정한수가 구멍 앞 바닥에 꼿꼿하게 서 있는 위압적인 전신.\n\nLOCATION (lock): On the surviving upper-floor edge beside the collapse hole inside the ruined department store. Limited daylight reaches the rubble-strewn interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the pan beside the upper-level opening, maintaining the low lens position near the floor and the upward three-quarter angle from beside 이현우's emergence point. Frame 정한수's complete figure just right of center, his short hair and sharp eyes readable above a firmly planted body; he turns his attention toward 수빈 outside the left frame edge, not toward the lens. Keep a narrow broken edge of the opening in the lower foreground, using his sustained, tense stillness to make the rescuer's identity feel imposing without introducing a new action.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Upper edge of the floor opening (Broken and open after the collapse) — Its upper surface and near edge are visible across a narrow lower foreground strip; used as Maintains continuity with the ascent and separates the camera position from 정한수; Collapsed interior walls (Broken within the dim department store) — Partial wall faces sit behind 정한수 at oblique angles; used as Subordinate architectural depth around his full-body silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim department-store ambience and controlled contrast, retaining detail in 정한수's eyes and full figure without adding a dramatic new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rescue rope remains hanging through the collapsed floor opening. The stocked but blocked-off food section and open parasol remain below. 정한수: He stands at the upper opening with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 날카로운 눈빛과 짧은 머리를 한 정한수가 구멍 앞 바닥에 꼿꼿하게 서 있는 위압적인 전신.\n\nLOCATION (lock): On the surviving upper-floor edge beside the collapse hole inside the ruined department store. Limited daylight reaches the rubble-strewn interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the pan beside the upper-level opening, maintaining the low lens position near the floor and the upward three-quarter angle from beside 이현우's emergence point. Frame 정한수's complete figure just right of center, his short hair and sharp eyes readable above a firmly planted body; he turns his attention toward 수빈 outside the left frame edge, not toward the lens. Keep a narrow broken edge of the opening in the lower foreground, using his sustained, tense stillness to make the rescuer's identity feel imposing without introducing a new action.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Upper edge of the floor opening (Broken and open after the collapse) — Its upper surface and near edge are visible across a narrow lower foreground strip; used as Maintains continuity with the ascent and separates the camera position from 정한수; Collapsed interior walls (Broken within the dim department store) — Partial wall faces sit behind 정한수 at oblique angles; used as Subordinate architectural depth around his full-body silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim department-store ambience and controlled contrast, retaining detail in 정한수's eyes and full figure without adding a dramatic new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rescue rope remains hanging through the collapsed floor opening. The stocked but blocked-off food section and open parasol remain below. 정한수: He stands at the upper opening with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day, rain later.\n\nSHOT TEXT (authoritative, Korean): 날카로운 눈빛과 짧은 머리를 한 정한수가 구멍 앞 바닥에 꼿꼿하게 서 있는 위압적인 전신.\n\nLOCATION (lock): On the surviving upper-floor edge beside the collapse hole inside the ruined department store. Limited daylight reaches the rubble-strewn interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the pan beside the upper-level opening, maintaining the low lens position near the floor and the upward three-quarter angle from beside 이현우's emergence point. Frame 정한수's complete figure just right of center, his short hair and sharp eyes readable above a firmly planted body; he turns his attention toward 수빈 outside the left frame edge, not toward the lens. Keep a narrow broken edge of the opening in the lower foreground, using his sustained, tense stillness to make the rescuer's identity feel imposing without introducing a new action.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Upper edge of the floor opening (Broken and open after the collapse) — Its upper surface and near edge are visible across a narrow lower foreground strip; used as Maintains continuity with the ascent and separates the camera position from 정한수; Collapsed interior walls (Broken within the dim department store) — Partial wall faces sit behind 정한수 at oblique angles; used as Subordinate architectural depth around his full-body silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim department-store ambience and controlled contrast, retaining detail in 정한수's eyes and full figure without adding a dramatic new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rescue rope remains hanging through the collapsed floor opening. The stocked but blocked-off food section and open parasol remain below. 정한수: He stands at the upper opening with short hair and a sharp gaze.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 정한수 (북한 출신 남성, 성인의 얼굴, 짧은 검은 머리) — wearing: 흙먼지로 뒤덮인 두꺼운 생존용 갈색 점퍼와 튼튼한 카고 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "정한수는 화면 왼쪽 바깥을 향해 뚜렷하게 시선을 두고 있음.",
    "built_space": "무너진 백화점 내부. 카메라 앞 하단에 바닥 구멍의 좁은 가장자리가 배치되고, 뒤로는 에스컬레이터와 파괴된 기둥들이 보임.",
    "entities": "정한수의 복장(점퍼, 바지)은 참조와 일치하나 비니를 계속 쓰고 있어 '짧은 머리'가 직접 드러나지 않으며, 구출용 밧줄은 보이지 않음.",
    "hard_violations": [],
    "physics": "정한수는 구멍 앞 바닥에 두 발을 단단히 딛고 체중을 지탱하며 꼿꼿하게 서 있음."
   },
   {
    "label": "B",
    "direction": "정한수는 프레임 왼쪽 밖을 향해 시선을 두고 있음.",
    "built_space": "무너진 백화점 내부. 하단 전경에 구멍의 단면이 있고, 배경에 잔해와 에스컬레이터가 위치함.",
    "entities": "정한수의 재킷에 참조 이미지에 존재하지 않는 검은색 배낭 어깨끈이 추가되었으며, 구출용 밧줄은 없음.",
    "hard_violations": [
     "[gemini-pro] 참조 이미지와 프롬프트에 없는 배낭 어깨끈(스트랩)이 캐릭터 가슴에 임의로 추가됨 (invented objects)"
    ],
    "physics": "부서진 바닥 가장자리에 두 발을 딛고 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 로우 앵글, 프레임 구성, 캐릭터의 시선을 훌륭하게 구현했으나 프롬프트에 명시된 밧줄이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "구도와 배경은 양호하나, 참조 이미지에 없는 배낭 어깨끈이 임의로 추가되는 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "정한수는 화면 왼쪽 바깥을 향해 뚜렷하게 시선을 두고 있음.",
        "built_space": "무너진 백화점 내부. 카메라 앞 하단에 바닥 구멍의 좁은 가장자리가 배치되고, 뒤로는 에스컬레이터와 파괴된 기둥들이 보임.",
        "entities": "정한수의 복장(점퍼, 바지)은 참조와 일치하나 비니를 계속 쓰고 있어 '짧은 머리'가 직접 드러나지 않으며, 구출용 밧줄은 보이지 않음.",
        "hard_violations": [],
        "physics": "정한수는 구멍 앞 바닥에 두 발을 단단히 딛고 체중을 지탱하며 꼿꼿하게 서 있음."
       },
       {
        "label": "B",
        "direction": "정한수는 프레임 왼쪽 밖을 향해 시선을 두고 있음.",
        "built_space": "무너진 백화점 내부. 하단 전경에 구멍의 단면이 있고, 배경에 잔해와 에스컬레이터가 위치함.",
        "entities": "정한수의 재킷에 참조 이미지에 존재하지 않는 검은색 배낭 어깨끈이 추가되었으며, 구출용 밧줄은 없음.",
        "hard_violations": [
         "참조 이미지와 프롬프트에 없는 배낭 어깨끈(스트랩)이 캐릭터 가슴에 임의로 추가됨 (invented objects)"
        ],
        "physics": "부서진 바닥 가장자리에 두 발을 딛고 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 로우 앵글, 프레임 구성, 캐릭터의 시선을 훌륭하게 구현했으나 프롬프트에 명시된 밧줄이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "구도와 배경은 양호하나, 참조 이미지에 없는 배낭 어깨끈이 임의로 추가되는 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "정한수는 화면 왼쪽 바깥을 향해 뚜렷하게 시선을 두고 있음.",
        "built_space": "무너진 백화점 내부. 카메라 앞 하단에 바닥 구멍의 좁은 가장자리가 배치되고, 뒤로는 에스컬레이터와 파괴된 기둥들이 보임.",
        "entities": "정한수의 복장(점퍼, 바지)은 참조와 일치하나 비니를 계속 쓰고 있어 '짧은 머리'가 직접 드러나지 않으며, 구출용 밧줄은 보이지 않음.",
        "hard_violations": [],
        "physics": "정한수는 구멍 앞 바닥에 두 발을 단단히 딛고 체중을 지탱하며 꼿꼿하게 서 있음."
       },
       {
        "label": "B",
        "direction": "정한수는 프레임 왼쪽 밖을 향해 시선을 두고 있음.",
        "built_space": "무너진 백화점 내부. 하단 전경에 구멍의 단면이 있고, 배경에 잔해와 에스컬레이터가 위치함.",
        "entities": "정한수의 재킷에 참조 이미지에 존재하지 않는 검은색 배낭 어깨끈이 추가되었으며, 구출용 밧줄은 없음.",
        "hard_violations": [
         "참조 이미지와 프롬프트에 없는 배낭 어깨끈(스트랩)이 캐릭터 가슴에 임의로 추가됨 (invented objects)"
        ],
        "physics": "부서진 바닥 가장자리에 두 발을 딛고 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "낮은 앙각의 전신과 왼쪽 시선은 맞지만, 넓고 밝게 드러난 아트리움이 요구된 어두운 배경을 약화하며 모자가 짧은 머리를 가린다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "중앙 오른쪽의 꼿꼿한 전신, 화면 밖 왼쪽 시선, 파손된 전경 가장자리와 어두운 실내가 더 충실하지만, 짧은 머리를 보여야 하는 지시는 모자 때문에 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈이 화면 왼쪽을 향하며 렌즈를 바라보지 않는다. 화면 밖 수빈을 보는 지시와 부합하는 방향이나, 대상은 화면에 없어 직접 확인할 수 없다. 몸은 카메라 쪽으로 비스듬히 열려 있고 이동 동작이나 손에 든 지향성 물체는 없다.",
        "built_space": "인물은 하단의 깨진 바닥 개구부 너머 잔존 슬래브 위에 선다. 바닥 윗면과 철근이 드러난 절단면이 전경을 가로지른다. 굵은 기둥 두 개가 인물 좌우에 있고, 왼쪽 뒤에 대각선 에스컬레이터 한 구간과 여러 층의 난간·슬래브가 보인다. 낮은 카메라에서 전신을 올려다보는 구도는 맞지만, 배경은 비스듬한 파손 벽면보다 밝고 넓은 아트리움을 강조한다. 구조 로프는 철근·케이블과 구분되어 확인되지 않으며 아래층 식품 구역과 파라솔은 보이지 않는다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며 얼굴 윤곽과 체격은 인물 참고와 대체로 유사하다. 북한 출신 여부는 외관만으로 판별할 수 없다. 먼지 묻은 두꺼운 갈색 후드 점퍼, 카고 바지와 부츠는 부합한다. 참고 이미지의 검은 모자를 유지해 이번 샷이 요구한 짧은 검은 머리는 읽히지 않는다. 어깨에는 참고보다 두드러진 배낭형 끈이 보인다. 다른 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "양쪽 부츠가 잔존 콘크리트 바닥에 닿아 있고 벌어진 다리가 체중을 지지한다. 몸통은 수직으로 유지되고 빈손은 옆으로 내려와 있어 긴장한 정지 자세로 가능하다. 전경 잔해는 슬래브 위에 놓이고 돌출 철근은 파손 콘크리트에 연결되어 있다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 밖 왼쪽을 향하고 렌즈와 눈을 맞추지 않는다. 수빈이 있어야 할 방향과 일치한다. 몸은 정면에 가까운 삼사분면 방향으로 서 있으며, 손은 비어 있고 새 행동이나 이동은 나타나지 않는다.",
        "built_space": "중앙 오른쪽의 인물이 개구부 뒤 잔존 바닥에 서 있고, 하단 전경에 바닥 윗면과 부서진 가장자리가 띠처럼 이어진다. 왼쪽 중경과 오른쪽에 큰 기둥이 하나씩 있으며, 왼쪽 끝에 에스컬레이터 한 구간, 뒤로 여러 층의 슬래브·난간과 어두운 벽면이 보인다. 바닥 가까운 앙각과 인물 뒤의 어두운 건축 깊이가 요구에 더 가깝다. 참고의 파손 콘크리트와 노출 철근 재질도 이어진다. 구조 로프는 명확히 식별되지 않고 아래층 식품 구역과 파라솔은 프레임에 드러나지 않는다.",
        "entities": "성인 동아시아계 남성 한 명이며 참고와 유사한 얼굴, 수염 흔적과 체격을 보인다. 출신 국가는 이미지로 확인할 수 없다. 먼지로 더러워진 갈색 생존용 점퍼, 카고 바지와 부츠가 일치한다. 다만 검은 모자가 머리를 덮어 짧은 머리를 읽히게 하라는 명시적 요구는 두 후보 모두 충족하지 못한다. 추가 인물이나 삽입된 문자는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 개구부 바로 뒤의 바닥 윗면에 확실히 놓여 인물을 지지한다. 다리를 적당히 벌리고 상체를 세운 자세는 꼿꼿하고 긴장된 정지 상태로 자연스럽다. 양손은 비어 있으며 몸 옆에 내려와 있다. 잔해는 바닥에 쌓여 있고 철근은 슬래브와 기둥에 연결되어 있어 무지지 부유나 불가능한 신체 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "낮은 앙각의 전신과 왼쪽 시선은 맞지만, 넓고 밝게 드러난 아트리움이 요구된 어두운 배경을 약화하며 모자가 짧은 머리를 가린다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "중앙 오른쪽의 꼿꼿한 전신, 화면 밖 왼쪽 시선, 파손된 전경 가장자리와 어두운 실내가 더 충실하지만, 짧은 머리를 보여야 하는 지시는 모자 때문에 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈이 화면 왼쪽을 향하며 렌즈를 바라보지 않는다. 화면 밖 수빈을 보는 지시와 부합하는 방향이나, 대상은 화면에 없어 직접 확인할 수 없다. 몸은 카메라 쪽으로 비스듬히 열려 있고 이동 동작이나 손에 든 지향성 물체는 없다.",
        "built_space": "인물은 하단의 깨진 바닥 개구부 너머 잔존 슬래브 위에 선다. 바닥 윗면과 철근이 드러난 절단면이 전경을 가로지른다. 굵은 기둥 두 개가 인물 좌우에 있고, 왼쪽 뒤에 대각선 에스컬레이터 한 구간과 여러 층의 난간·슬래브가 보인다. 낮은 카메라에서 전신을 올려다보는 구도는 맞지만, 배경은 비스듬한 파손 벽면보다 밝고 넓은 아트리움을 강조한다. 구조 로프는 철근·케이블과 구분되어 확인되지 않으며 아래층 식품 구역과 파라솔은 보이지 않는다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며 얼굴 윤곽과 체격은 인물 참고와 대체로 유사하다. 북한 출신 여부는 외관만으로 판별할 수 없다. 먼지 묻은 두꺼운 갈색 후드 점퍼, 카고 바지와 부츠는 부합한다. 참고 이미지의 검은 모자를 유지해 이번 샷이 요구한 짧은 검은 머리는 읽히지 않는다. 어깨에는 참고보다 두드러진 배낭형 끈이 보인다. 다른 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "양쪽 부츠가 잔존 콘크리트 바닥에 닿아 있고 벌어진 다리가 체중을 지지한다. 몸통은 수직으로 유지되고 빈손은 옆으로 내려와 있어 긴장한 정지 자세로 가능하다. 전경 잔해는 슬래브 위에 놓이고 돌출 철근은 파손 콘크리트에 연결되어 있다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 밖 왼쪽을 향하고 렌즈와 눈을 맞추지 않는다. 수빈이 있어야 할 방향과 일치한다. 몸은 정면에 가까운 삼사분면 방향으로 서 있으며, 손은 비어 있고 새 행동이나 이동은 나타나지 않는다.",
        "built_space": "중앙 오른쪽의 인물이 개구부 뒤 잔존 바닥에 서 있고, 하단 전경에 바닥 윗면과 부서진 가장자리가 띠처럼 이어진다. 왼쪽 중경과 오른쪽에 큰 기둥이 하나씩 있으며, 왼쪽 끝에 에스컬레이터 한 구간, 뒤로 여러 층의 슬래브·난간과 어두운 벽면이 보인다. 바닥 가까운 앙각과 인물 뒤의 어두운 건축 깊이가 요구에 더 가깝다. 참고의 파손 콘크리트와 노출 철근 재질도 이어진다. 구조 로프는 명확히 식별되지 않고 아래층 식품 구역과 파라솔은 프레임에 드러나지 않는다.",
        "entities": "성인 동아시아계 남성 한 명이며 참고와 유사한 얼굴, 수염 흔적과 체격을 보인다. 출신 국가는 이미지로 확인할 수 없다. 먼지로 더러워진 갈색 생존용 점퍼, 카고 바지와 부츠가 일치한다. 다만 검은 모자가 머리를 덮어 짧은 머리를 읽히게 하라는 명시적 요구는 두 후보 모두 충족하지 못한다. 추가 인물이나 삽입된 문자는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 개구부 바로 뒤의 바닥 윗면에 확실히 놓여 인물을 지지한다. 다리를 적당히 벌리고 상체를 세운 자세는 꼿꼿하고 긴장된 정지 상태로 자연스럽다. 양손은 비어 있으며 몸 옆에 내려와 있다. 잔해는 바닥에 쌓여 있고 철근은 슬래브와 기둥에 연결되어 있어 무지지 부유나 불가능한 신체 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.304
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.054
   },
   "violations": {
    "B": [
     "[gemini-pro] 참조 이미지와 프롬프트에 없는 배낭 어깨끈(스트랩)이 캐릭터 가슴에 임의로 추가됨 (invented objects)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1054
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 로우 앵글, 프레임 구성, 캐릭터의 시선을 훌륭하게 구현했으나 프롬프트에 명시된 밧줄이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 1054,
    "verdict_ko": "구도와 배경은 양호하나, 참조 이미지에 없는 배낭 어깨끈이 임의로 추가되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 참조 이미지와 프롬프트에 없는 배낭 어깨끈(스트랩)이 캐릭터 가슴에 임의로 추가됨 (invented objects)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S66sh13_sel.png",
    "asset_id": "3c6757cd-f8e6-4275-9c55-d21432b7569c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 정한수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:836640>",
    "asset_id": "be570610-37f0-4bda-b6e5-2e4fdaac528b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cef-e69d-766a-9f06-c690397f23d5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S66sh13"
  }
 },
 "S67sh83::signage": {
  "fp": "29889a4684a22e86",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::village_escape_road": {
  "input_fingerprint": "613c8c80b8f673fb",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "village_escape_road",
    "tags": [
     "S67sh119",
     "S67sh83",
     "S67sh95"
    ]
   },
   "context_sig": "b5c03c63d1730c01"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 빠르게 멀어지는 민병대 트럭!\n- 트럭 앞에서 대기하고 있는...B-200.\n- 그 부품을 조용히 줍는 찰리.\n\nTIME OF DAY (lock): sunset to night, bright full moon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n익산 한옥마을 거리·중앙 공터, 물탱크 탑 공사 현장, 물탱크 탑 앞 공터, 과거 배급 장소, 외곽도로·트럭 전투 현장, 골목, 성당 앞 격투장·관중석·단상, 성당 외부·입구, 폐창고 앞: 기괴하게 변형된 한옥과 거대한 철제 물탱크가 어우러진 중앙 광장. 과거부터 현재까지 여러 군중 행사가 벌어진다. (특징: 한옥 지붕과 현대식 철제 구조물이 기괴하게 얽힌 건축 형태; 우뚝 솟은 거대한 원통형 금속 물탱크 탑; 얼굴 피부가 녹아내리고 짓무른 흉터를 가진 주민들; 가죽 하회탈을 쓴 궁사들과 횃불이 꽂힌 성당 앞 흙바닥 격투장; 격투장 철제 케이지에서 튀어나오는 육중한 B-200 (개조된 가슴 부품과 전개식 고사포 팔 장착); 최신 방진복을 입고 박격포 포탄을 쏘며 난입하는 유빅 용병대; B-200의 고사포 연사에 허리가 끊어져 무너지며 거대한 물줄기를 쏟아내는 물탱크 탑; 물에 젖어 반쯤 깨진 황금 하회탈 아래로 드러난 백산의 녹아내린 흉터 얼굴; 노을 아래 흙바닥에서 공을 차는 아이들과 방사능 피복 흉터가 남은 어른들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 빠르게 멀어지는 민병대 트럭!\n- 트럭 앞에서 대기하고 있는...B-200.\n- 그 부품을 조용히 줍는 찰리.\n\nTIME OF DAY (lock): sunset to night, bright full moon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_escape_road_a30396.png",
  "asset_id": "1725495a-ae3a-42c2-8ab7-53829be1f26d",
  "input_asset_ids": [
   "23096701-ec62-46d1-94a6-654a2505d911"
  ],
  "origin_tag": "S67sh83",
  "place_text": "In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.",
  "origin_inputs": {
   "place_text": "In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.",
   "time_of_day_en": "sunset to night, bright full moon",
   "conti_asset_id": "23096701-ec62-46d1-94a6-654a2505d911"
  }
 },
 "S67sh83::bgfirst_bg": {
  "input_fingerprint": "57d4c1af0d3d8153",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 캄캄한 도로 한가운데 우뚝 서서 트럭의 진로를 막아선 B-200의 거대한 실루엣 전경.\n\nLOCATION (lock): In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.\n\nTIME OF DAY (lock): sunset to night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the track's initial distant roadside position with a low lens, viewing B-200 diagonally across the approaching trucks' route rather than from his frontal axis. Place his entire figure near center with visible road continuing on either side and behind him, his body firmly braced across the route and cannon aligning toward the approaching truck beyond the near frame edge. Keep his attention fixed on that off-screen truck and his silhouette visually imposing through low-angle placement rather than enlarging him unnaturally, establishing the roadblock before the advance begins.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck route (Blocked by B-200 before the trucks overturn) — The road recedes diagonally past B-200, making his position across the approaching route readable; used as Open foreground and lateral space establish the obstruction and preserve the roadside camera axis; B-200's cannon (Aligning toward 박철진's approaching truck before firing) — Seen from the side at an oblique angle, aimed beyond the near frame edge rather than into the lens; used as Small but legible directional extension of the roadblock silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the road dark and B-200 predominantly silhouetted, preserving only enough tonal separation to read his braced body and cannon without introducing an unsupported backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 캄캄한 도로 한가운데 우뚝 서서 트럭의 진로를 막아선 B-200의 거대한 실루엣 전경.\n\nLOCATION (lock): In the middle of the dark road leading out of the village, directly in the departing truck convoy's path.\n\nTIME OF DAY (lock): sunset to night, bright full moon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the track's initial distant roadside position with a low lens, viewing B-200 diagonally across the approaching trucks' route rather than from his frontal axis. Place his entire figure near center with visible road continuing on either side and behind him, his body firmly braced across the route and cannon aligning toward the approaching truck beyond the near frame edge. Keep his attention fixed on that off-screen truck and his silhouette visually imposing through low-angle placement rather than enlarging him unnaturally, establishing the roadblock before the advance begins.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck route (Blocked by B-200 before the trucks overturn) — The road recedes diagonally past B-200, making his position across the approaching route readable; used as Open foreground and lateral space establish the obstruction and preserve the roadside camera axis; B-200's cannon (Aligning toward 박철진's approaching truck before firing) — Seen from the side at an oblique angle, aimed beyond the near frame edge rather than into the lens; used as Small but legible directional extension of the roadblock silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the road dark and B-200 predominantly silhouetted, preserving only enough tonal separation to read his braced body and cannon without introducing an unsupported backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S67sh83__bgfirst_bg.png",
  "asset_id": "e237972c-5b15-4c67-b588-2a2ba4abd2b6",
  "input_asset_ids": [
   "23096701-ec62-46d1-94a6-654a2505d911",
   "1725495a-ae3a-42c2-8ab7-53829be1f26d"
  ]
 },
 "S67sh83": {
  "input_fingerprint": "e2eb10c2037b45c7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 캄캄한 도로 한가운데 우뚝 서서 트럭의 진로를 막아선 B-200의 거대한 실루엣 전경.\n\nLOCATION (lock): In the middle of the dark road leading out of the village, directly in the departing truck convoy's path. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the track's initial distant roadside position with a low lens, viewing B-200 diagonally across the approaching trucks' route rather than from his frontal axis. Place his entire figure near center with visible road continuing on either side and behind him, his body firmly braced across the route and cannon aligning toward the approaching truck beyond the near frame edge. Keep his attention fixed on that off-screen truck and his silhouette visually imposing through low-angle placement rather than enlarging him unnaturally, establishing the roadblock before the advance begins.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck route (Blocked by B-200 before the trucks overturn) — The road recedes diagonally past B-200, making his position across the approaching route readable; used as Open foreground and lateral space establish the obstruction and preserve the roadside camera axis; B-200's cannon (Aligning toward 박철진's approaching truck before firing) — Seen from the side at an oblique angle, aimed beyond the near frame edge rather than into the lens; used as Small but legible directional extension of the roadblock silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the road dark and B-200 predominantly silhouetted, preserving only enough tonal separation to read his braced body and cannon without introducing an unsupported backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia trucks are still upright on the moonlit road; Charlie is confined in a transport cage with heavy restraints around his neck and wrists, his earlier battle damage still present. B-200 blocks the route with its gun-hands intact and aimed, before the ensuing gunfire damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 캄캄한 도로 한가운데 우뚝 서서 트럭의 진로를 막아선 B-200의 거대한 실루엣 전경.\n\nLOCATION (lock): In the middle of the dark road leading out of the village, directly in the departing truck convoy's path. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the track's initial distant roadside position with a low lens, viewing B-200 diagonally across the approaching trucks' route rather than from his frontal axis. Place his entire figure near center with visible road continuing on either side and behind him, his body firmly braced across the route and cannon aligning toward the approaching truck beyond the near frame edge. Keep his attention fixed on that off-screen truck and his silhouette visually imposing through low-angle placement rather than enlarging him unnaturally, establishing the roadblock before the advance begins.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck route (Blocked by B-200 before the trucks overturn) — The road recedes diagonally past B-200, making his position across the approaching route readable; used as Open foreground and lateral space establish the obstruction and preserve the roadside camera axis; B-200's cannon (Aligning toward 박철진's approaching truck before firing) — Seen from the side at an oblique angle, aimed beyond the near frame edge rather than into the lens; used as Small but legible directional extension of the roadblock silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the road dark and B-200 predominantly silhouetted, preserving only enough tonal separation to read his braced body and cannon without introducing an unsupported backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia trucks are still upright on the moonlit road; Charlie is confined in a transport cage with heavy restraints around his neck and wrists, his earlier battle damage still present. B-200 blocks the route with its gun-hands intact and aimed, before the ensuing gunfire damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 캄캄한 도로 한가운데 우뚝 서서 트럭의 진로를 막아선 B-200의 거대한 실루엣 전경.\n\nLOCATION (lock): In the middle of the dark road leading out of the village, directly in the departing truck convoy's path. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the track's initial distant roadside position with a low lens, viewing B-200 diagonally across the approaching trucks' route rather than from his frontal axis. Place his entire figure near center with visible road continuing on either side and behind him, his body firmly braced across the route and cannon aligning toward the approaching truck beyond the near frame edge. Keep his attention fixed on that off-screen truck and his silhouette visually imposing through low-angle placement rather than enlarging him unnaturally, establishing the roadblock before the advance begins.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck route (Blocked by B-200 before the trucks overturn) — The road recedes diagonally past B-200, making his position across the approaching route readable; used as Open foreground and lateral space establish the obstruction and preserve the roadside camera axis; B-200's cannon (Aligning toward 박철진's approaching truck before firing) — Seen from the side at an oblique angle, aimed beyond the near frame edge rather than into the lens; used as Small but legible directional extension of the roadblock silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the road dark and B-200 predominantly silhouetted, preserving only enough tonal separation to read his braced body and cannon without introducing an unsupported backlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The militia trucks are still upright on the moonlit road; Charlie is confined in a transport cage with heavy restraints around his neck and wrists, his earlier battle damage still present. B-200 blocks the route with its gun-hands intact and aimed, before the ensuing gunfire damage.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — wearing: 거대하고 육중한 중장비 형태 메카닉 실루엣, 암전 다크 그레이 철제 파츠 장갑판으로 덮임, 양손은 투박하고 거대한 고사포 포신 화기, 표면에 총탄 자국과 탄흔 질감이 가득함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S67sh83__bgfirst_bg.png",
     "asset_id": "e237972c-5b15-4c67-b588-2a2ba4abd2b6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S67sh83.png",
     "asset_id": "23096701-ec62-46d1-94a6-654a2505d911",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176962>",
     "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_escape_road_a30396.png",
     "asset_id": "1725495a-ae3a-42c2-8ab7-53829be1f26d",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176962>",
     "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "B-200은 화면 우측 밖을 주시하며 대포를 겨누고 있으나, 실제 트럭은 그의 뒤편(좌측 배경)에서 오고 있어 행동과 상황이 불일치함.",
    "built_space": "참조 사진의 교회, 급수탑, 한옥 지붕 등 배경 구조물이 올바른 위치에 배치됨.",
    "entities": "B-200의 외형은 참조와 일치함. 그러나 프롬프트에 명시되지 않은 인물들(트럭 운전수 및 철창 안 사람)이 추가됨.",
    "hard_violations": [
     "[gemini-pro] invented people (트럭 내부 및 철창 안의 사람들)",
     "[gpt-high] B-200만 등장해야 하는 장면에 앞 트럭 운전석의 두 사람과 적재함의 한 사람 등 허용되지 않은 인물들이 추가되었습니다."
    ],
    "physics": "B-200이 도로 위에 두 발로 안정적으로 지탱하고 서 있음."
   },
   {
    "label": "B",
    "direction": "B-200의 시선과 무기가 지시된 대로 화면 우측 밖(근경 너머의 트럭 위치)을 향해 정확히 겨눠짐.",
    "built_space": "도로 중앙에 위치하며 뒤로 길이 이어짐. 교회, 급수탑 등 참조 사진의 구조물과 배경이 정확함.",
    "entities": "B-200은 참조와 정확히 일치하며 지시되지 않은 다른 인물이나 사물은 없음.",
    "hard_violations": [],
    "physics": "B-200이 지면(도로)을 딛고 견고하게 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 오프스크린 트럭을 향한 조준 자세와 카메라 구도를 정확히 구현했으며, 불필요한 추가 요소 없이 훌륭한 샷을 완성함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오프스크린에 있어야 할 트럭을 배경에 배치해 시선 처리와 어긋났으며, 프롬프트가 금지한 추가 인물을 포함하는 치명적인 오류를 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "B-200은 화면 우측 밖을 주시하며 대포를 겨누고 있으나, 실제 트럭은 그의 뒤편(좌측 배경)에서 오고 있어 행동과 상황이 불일치함.",
        "built_space": "참조 사진의 교회, 급수탑, 한옥 지붕 등 배경 구조물이 올바른 위치에 배치됨.",
        "entities": "B-200의 외형은 참조와 일치함. 그러나 프롬프트에 명시되지 않은 인물들(트럭 운전수 및 철창 안 사람)이 추가됨.",
        "hard_violations": [
         "invented people (트럭 내부 및 철창 안의 사람들)"
        ],
        "physics": "B-200이 도로 위에 두 발로 안정적으로 지탱하고 서 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 시선과 무기가 지시된 대로 화면 우측 밖(근경 너머의 트럭 위치)을 향해 정확히 겨눠짐.",
        "built_space": "도로 중앙에 위치하며 뒤로 길이 이어짐. 교회, 급수탑 등 참조 사진의 구조물과 배경이 정확함.",
        "entities": "B-200은 참조와 정확히 일치하며 지시되지 않은 다른 인물이나 사물은 없음.",
        "hard_violations": [],
        "physics": "B-200이 지면(도로)을 딛고 견고하게 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 오프스크린 트럭을 향한 조준 자세와 카메라 구도를 정확히 구현했으며, 불필요한 추가 요소 없이 훌륭한 샷을 완성함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오프스크린에 있어야 할 트럭을 배경에 배치해 시선 처리와 어긋났으며, 프롬프트가 금지한 추가 인물을 포함하는 치명적인 오류를 범함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "B-200은 화면 우측 밖을 주시하며 대포를 겨누고 있으나, 실제 트럭은 그의 뒤편(좌측 배경)에서 오고 있어 행동과 상황이 불일치함.",
        "built_space": "참조 사진의 교회, 급수탑, 한옥 지붕 등 배경 구조물이 올바른 위치에 배치됨.",
        "entities": "B-200의 외형은 참조와 일치함. 그러나 프롬프트에 명시되지 않은 인물들(트럭 운전수 및 철창 안 사람)이 추가됨.",
        "hard_violations": [
         "invented people (트럭 내부 및 철창 안의 사람들)"
        ],
        "physics": "B-200이 도로 위에 두 발로 안정적으로 지탱하고 서 있음."
       },
       {
        "label": "B",
        "direction": "B-200의 시선과 무기가 지시된 대로 화면 우측 밖(근경 너머의 트럭 위치)을 향해 정확히 겨눠짐.",
        "built_space": "도로 중앙에 위치하며 뒤로 길이 이어짐. 교회, 급수탑 등 참조 사진의 구조물과 배경이 정확함.",
        "entities": "B-200은 참조와 정확히 일치하며 지시되지 않은 다른 인물이나 사물은 없음.",
        "hard_violations": [],
        "physics": "B-200이 지면(도로)을 딛고 견고하게 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "추가 인물 없이 도로를 막고 선 전신과 장소는 보존했지만, 정면에 가까운 큰 구도와 아래로 갈라진 포신 방향은 원거리 사선 실루엣 및 화면 밖 트럭 조준 지시와 다릅니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 트럭 탑승자들이 등장하며, 뒤쪽에서 접근하는 트럭과 오른쪽을 겨눈 포신의 방향도 어긋나 장면의 핵심 관계를 충족하지 못합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리는 화면 오른쪽으로 조금 돌아가 있습니다. 화면 오른쪽 포신은 오른쪽 앞쪽의 낮은 지점을 향하고, 왼쪽 포신은 카메라 쪽 아래를 향해 두 총구의 표적이 갈립니다. 트럭은 보이지 않으며, 두 무기가 같은 화면 밖 접근 트럭을 조준한다고 읽히지 않습니다. 특히 왼쪽 총구는 요구된 측면 조준보다 렌즈를 향한 방향에 가깝습니다.",
        "built_space": "오른쪽 뒤에 철골 지지대 위 원통형 물탱크 하나, 왼쪽 뒤에 주탑과 작은 탑이 있는 교회 하나, 그 사이와 오른쪽에 기와 건물들이 보입니다. 왼쪽에는 전주와 가로등 열, 오른쪽에는 가드레일 한 줄이 있고 도로는 왼쪽 후경으로 이어져 장소의 주요 구성을 보존합니다. 기계는 도로 중앙에 발을 벌리고 서 있으며 양옆과 앞뒤 도로가 보입니다. 다만 낮은 카메라에서도 몸의 정면이 크게 드러나, 지시된 먼 길가의 사선 관찰 구도보다 가까운 정면 전신 구도입니다.",
        "entities": "등장 개체는 B-200 한 대뿐입니다. 짙은 회색 장갑, 육중한 기계 체형, 주황색 기계 눈, 양팔의 다연장 포신, 탄흔과 가슴 표식이 참조와 대체로 일치하며 인간 피부나 얼굴은 없습니다. 보름달도 보입니다. 트럭과 찰리는 화면 밖이므로 그 상태는 확인할 수 없습니다. 장갑의 전면 세부가 상당히 밝게 드러나 주로 실루엣이어야 한다는 요구에는 못 미칩니다.",
        "hard_violations": [],
        "physics": "두 발바닥이 아스팔트에 닿아 넓은 지지면을 만들고, 벌어진 다리가 몸통의 무게를 받칩니다. 포신은 팔 관절에 연결되어 있으며 공중에 따로 떠 있는 물체는 없습니다. 발사나 전진 전 정지 자세로 성립합니다."
       },
       {
        "label": "B",
        "direction": "머리와 수평으로 든 화면 오른쪽 포신은 오른쪽 화면 밖을 향합니다. 반대쪽 포신은 왼쪽 전경의 노면 쪽으로 내려가 있습니다. 실제 보이는 트럭 두 대는 기계의 왼쪽 뒤에서 카메라 쪽으로 접근하므로, 머리와 주된 포신은 그 트럭들을 겨누지 않습니다. 트럭의 이동 방향과 조준 방향이 서로 맞지 않습니다.",
        "built_space": "원통형 물탱크 하나와 철골 구조가 오른쪽 뒤에, 교회 하나가 왼쪽 뒤에 있고, 기와 건물들과 왼쪽 전주 열 및 오른쪽 가드레일이 참조의 배치를 대체로 유지합니다. 기계는 중앙 도로에 서고 트럭 두 대는 왼쪽 후경 도로에 있습니다. 양옆 노면과 전경은 확보했지만 기계의 가슴 정면이 크게 노출되어 요구된 원거리 사선 구도보다 정면 영웅 구도에 가깝습니다.",
        "entities": "B-200의 회색 금속 장갑, 주황색 기계 눈, 양팔 포신과 탄흔은 참조의 기계 정체성을 대체로 따릅니다. 보름달과 똑바로 선 트럭 두 대가 보입니다. 그러나 앞 트럭 운전석 안에 사람 형상 두 명, 적재함에 한 명이 추가되어 B-200 이외 인물 금지 조건을 어깁니다. 이들의 나이와 민족적 외양은 식별하기 어렵고, 적재함 인물을 구속된 찰리로 확인할 목·손목 구속구도 보이지 않습니다. 기계 전면이 밝게 드러나 암전 실루엣 요구도 약합니다.",
        "hard_violations": [
         "B-200만 등장해야 하는 장면에 앞 트럭 운전석의 두 사람과 적재함의 한 사람 등 허용되지 않은 인물들이 추가되었습니다."
        ],
        "physics": "기계의 두 발이 노면에 접촉하고 벌어진 다리가 몸통을 지탱합니다. 수평 포신과 내려간 포신 모두 팔에 연결되어 지지됩니다. 트럭은 바퀴로 도로 위에 서 있으며, 탑승자들은 운전석 또는 적재함 안에 있어 지지 없는 부유는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "추가 인물 없이 도로를 막고 선 전신과 장소는 보존했지만, 정면에 가까운 큰 구도와 아래로 갈라진 포신 방향은 원거리 사선 실루엣 및 화면 밖 트럭 조준 지시와 다릅니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 트럭 탑승자들이 등장하며, 뒤쪽에서 접근하는 트럭과 오른쪽을 겨눈 포신의 방향도 어긋나 장면의 핵심 관계를 충족하지 못합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "머리는 화면 오른쪽으로 조금 돌아가 있습니다. 화면 오른쪽 포신은 오른쪽 앞쪽의 낮은 지점을 향하고, 왼쪽 포신은 카메라 쪽 아래를 향해 두 총구의 표적이 갈립니다. 트럭은 보이지 않으며, 두 무기가 같은 화면 밖 접근 트럭을 조준한다고 읽히지 않습니다. 특히 왼쪽 총구는 요구된 측면 조준보다 렌즈를 향한 방향에 가깝습니다.",
        "built_space": "오른쪽 뒤에 철골 지지대 위 원통형 물탱크 하나, 왼쪽 뒤에 주탑과 작은 탑이 있는 교회 하나, 그 사이와 오른쪽에 기와 건물들이 보입니다. 왼쪽에는 전주와 가로등 열, 오른쪽에는 가드레일 한 줄이 있고 도로는 왼쪽 후경으로 이어져 장소의 주요 구성을 보존합니다. 기계는 도로 중앙에 발을 벌리고 서 있으며 양옆과 앞뒤 도로가 보입니다. 다만 낮은 카메라에서도 몸의 정면이 크게 드러나, 지시된 먼 길가의 사선 관찰 구도보다 가까운 정면 전신 구도입니다.",
        "entities": "등장 개체는 B-200 한 대뿐입니다. 짙은 회색 장갑, 육중한 기계 체형, 주황색 기계 눈, 양팔의 다연장 포신, 탄흔과 가슴 표식이 참조와 대체로 일치하며 인간 피부나 얼굴은 없습니다. 보름달도 보입니다. 트럭과 찰리는 화면 밖이므로 그 상태는 확인할 수 없습니다. 장갑의 전면 세부가 상당히 밝게 드러나 주로 실루엣이어야 한다는 요구에는 못 미칩니다.",
        "hard_violations": [],
        "physics": "두 발바닥이 아스팔트에 닿아 넓은 지지면을 만들고, 벌어진 다리가 몸통의 무게를 받칩니다. 포신은 팔 관절에 연결되어 있으며 공중에 따로 떠 있는 물체는 없습니다. 발사나 전진 전 정지 자세로 성립합니다."
       },
       {
        "label": "A",
        "direction": "머리와 수평으로 든 화면 오른쪽 포신은 오른쪽 화면 밖을 향합니다. 반대쪽 포신은 왼쪽 전경의 노면 쪽으로 내려가 있습니다. 실제 보이는 트럭 두 대는 기계의 왼쪽 뒤에서 카메라 쪽으로 접근하므로, 머리와 주된 포신은 그 트럭들을 겨누지 않습니다. 트럭의 이동 방향과 조준 방향이 서로 맞지 않습니다.",
        "built_space": "원통형 물탱크 하나와 철골 구조가 오른쪽 뒤에, 교회 하나가 왼쪽 뒤에 있고, 기와 건물들과 왼쪽 전주 열 및 오른쪽 가드레일이 참조의 배치를 대체로 유지합니다. 기계는 중앙 도로에 서고 트럭 두 대는 왼쪽 후경 도로에 있습니다. 양옆 노면과 전경은 확보했지만 기계의 가슴 정면이 크게 노출되어 요구된 원거리 사선 구도보다 정면 영웅 구도에 가깝습니다.",
        "entities": "B-200의 회색 금속 장갑, 주황색 기계 눈, 양팔 포신과 탄흔은 참조의 기계 정체성을 대체로 따릅니다. 보름달과 똑바로 선 트럭 두 대가 보입니다. 그러나 앞 트럭 운전석 안에 사람 형상 두 명, 적재함에 한 명이 추가되어 B-200 이외 인물 금지 조건을 어깁니다. 이들의 나이와 민족적 외양은 식별하기 어렵고, 적재함 인물을 구속된 찰리로 확인할 목·손목 구속구도 보이지 않습니다. 기계 전면이 밝게 드러나 암전 실루엣 요구도 약합니다.",
        "hard_violations": [
         "B-200만 등장해야 하는 장면에 앞 트럭 운전석의 두 사람과 적재함의 한 사람 등 허용되지 않은 인물들이 추가되었습니다."
        ],
        "physics": "기계의 두 발이 노면에 접촉하고 벌어진 다리가 몸통을 지탱합니다. 수평 포신과 내려간 포신 모두 팔에 연결되어 지지됩니다. 트럭은 바퀴로 도로 위에 서 있으며, 탑승자들은 운전석 또는 적재함 안에 있어 지지 없는 부유는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (트럭 내부 및 철창 안의 사람들)",
     "[gpt-high] B-200만 등장해야 하는 장면에 앞 트럭 운전석의 두 사람과 적재함의 한 사람 등 허용되지 않은 인물들이 추가되었습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 579
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 오프스크린 트럭을 향한 조준 자세와 카메라 구도를 정확히 구현했으며, 불필요한 추가 요소 없이 훌륭한 샷을 완성함."
   },
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "오프스크린에 있어야 할 트럭을 배경에 배치해 시선 처리와 어긋났으며, 프롬프트가 금지한 추가 인물을 포함하는 치명적인 오류를 범함.  ★위반: [gemini-pro] invented people (트럭 내부 및 철창 안의 사람들) / [gpt-high] B-200만 등장해야 하는 장면에 앞 트럭 운전석의 두 사람과 적재함의 한 사람 등 허용되지 않은 인물들이 추가되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_village_escape_road_a30396.png",
    "asset_id": "1725495a-ae3a-42c2-8ab7-53829be1f26d",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1176962>",
    "asset_id": "6648e5c4-e9ff-4c31-93cf-f1d51bf069e3",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cf5-6d6c-7248-bd3d-1a50993e5bae",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S67sh83__bgfirst_bg.png",
   "bg_asset_id": "e237972c-5b15-4c67-b588-2a2ba4abd2b6",
   "bg_record_key": "S67sh83::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "village_escape_road",
   "groupbg_asset_id": "1725495a-ae3a-42c2-8ab7-53829be1f26d"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S67sh95::signage": {
  "fp": "260421120e70f30c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S67sh95": {
  "input_fingerprint": "a8995a49c1598504",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 거대한 기계 몸을 날려 B-200의 앞을 가로막은 채, 쏟아지는 총알들이 자신의 등에 부딪혀 스파크를 일으키는 찰리의 역동적인 찰나.\n\nLOCATION (lock): On the nighttime road beside the overturned convoy trucks, in the exposed firefight area outside the village. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established roadside, hold a low, rear three-quarter view and match the final lateral beat of 찰리's interception, emphasizing his changed position rather than a new camera axis. His diagonally extended body occupies the left foreground, with bullet impacts sparking across his back; leave an opening on the right through which B-200's damaged torso and missing gun-hand remain readable. 찰리 inclines his head toward the friend he is shielding, while B-200 recoils and looks up at him; retain enough road around their bodies to make the protective separation unmistakable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Road at the interception point (Visible beneath and between the two robots) — The road recedes diagonally behind the protecting body; used as Preserves the established roadside axis and provides spatial separation between the figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain readable nighttime ambient illumination with brief localized impact sparks, controlled highlights, and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks, including the transport truck, are overturned; its cage has been opened and Charlie's restraints have been broken. Charlie retains his battered, perforated body and is taking further gunfire, while B-200's torso is broken and its gun-hand has been shot off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 거대한 기계 몸을 날려 B-200의 앞을 가로막은 채, 쏟아지는 총알들이 자신의 등에 부딪혀 스파크를 일으키는 찰리의 역동적인 찰나.\n\nLOCATION (lock): On the nighttime road beside the overturned convoy trucks, in the exposed firefight area outside the village. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established roadside, hold a low, rear three-quarter view and match the final lateral beat of 찰리's interception, emphasizing his changed position rather than a new camera axis. His diagonally extended body occupies the left foreground, with bullet impacts sparking across his back; leave an opening on the right through which B-200's damaged torso and missing gun-hand remain readable. 찰리 inclines his head toward the friend he is shielding, while B-200 recoils and looks up at him; retain enough road around their bodies to make the protective separation unmistakable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Road at the interception point (Visible beneath and between the two robots) — The road recedes diagonally behind the protecting body; used as Preserves the established roadside axis and provides spatial separation between the figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain readable nighttime ambient illumination with brief localized impact sparks, controlled highlights, and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks, including the transport truck, are overturned; its cage has been opened and Charlie's restraints have been broken. Charlie retains his battered, perforated body and is taking further gunfire, while B-200's torso is broken and its gun-hand has been shot off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): 거대한 기계 몸을 날려 B-200의 앞을 가로막은 채, 쏟아지는 총알들이 자신의 등에 부딪혀 스파크를 일으키는 찰리의 역동적인 찰나.\n\nLOCATION (lock): On the nighttime road beside the overturned convoy trucks, in the exposed firefight area outside the village. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established roadside, hold a low, rear three-quarter view and match the final lateral beat of 찰리's interception, emphasizing his changed position rather than a new camera axis. His diagonally extended body occupies the left foreground, with bullet impacts sparking across his back; leave an opening on the right through which B-200's damaged torso and missing gun-hand remain readable. 찰리 inclines his head toward the friend he is shielding, while B-200 recoils and looks up at him; retain enough road around their bodies to make the protective separation unmistakable.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Road at the interception point (Visible beneath and between the two robots) — The road recedes diagonally behind the protecting body; used as Preserves the established roadside axis and provides spatial separation between the figures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain readable nighttime ambient illumination with brief localized impact sparks, controlled highlights, and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리, B-200: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several militia trucks, including the transport truck, are overturned; its cage has been opened and Charlie's restraints have been broken. Charlie retains his battered, perforated body and is taking further gunfire, while B-200's torso is broken and its gun-hand has been shot off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; B-200 (중장비형 기계 실루엣, 짙은 회색 철제 외장, 전신 장갑판, 양손의 굵은 고사포 포신, 찰리과(와) 견주어 찰리보다 크고 육중하다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "총알이 왼쪽에서 날아와 찰리의 가슴과 정면에 명중함. 찰리는 카메라를 향해 정면으로 서 있고, B-200은 그 뒤에 위치함.",
    "built_space": "젖은 아스팔트 도로 위에 위치하며, 왼쪽의 전복된 트럭과 배경의 교회, 급수탑 등 레퍼런스의 주요 지형지물이 배치됨.",
    "entities": "찰리는 고릴라 형태를 띠고 있으나 프롬프트와 달리 가슴 원자로가 뚜렷이 보이는 정면 모습임. B-200은 오른팔의 고사포가 사라지고 뭉툭한 형태로 왜곡됨.",
    "hard_violations": [
     "[gemini-pro] 카메라 시점 위반 (후면 3/4 뷰 대신 정면 뷰를 렌더링함)",
     "[gemini-pro] 프롬프트 액션 위반 (등에 부딪히는 스파크 대신 정면에 명중함)",
     "[gemini-pro] 캐릭터 디자인 위반 (B-200의 오른팔 고사포가 무단으로 삭제됨)"
    ],
    "physics": "찰리와 B-200 모두 도로 지면에 양발을 딛고 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "총알이 왼쪽에서 날아와 찰리의 등 쪽에 명중하여 스파크를 냄. 찰리는 B-200을 향해 시선을 돌리고, B-200은 물러서며 찰리를 올려다봄.",
    "built_space": "젖은 도로 위에 위치하며, 왼쪽에는 전복된 차량이 있고 배경의 교회와 급수탑이 올바른 방향과 스케일로 자리잡고 있음.",
    "entities": "찰리는 후면이 강조된 샌드 베이지 장갑판 묘사를 따름. B-200은 레퍼런스에 맞춰 오른팔 고사포, 파손된 왼팔, 부서진 몸통을 정확히 유지함.",
    "hard_violations": [],
    "physics": "찰리는 왼쪽 지면에 발을 단단히 딛고 대각선으로 몸을 뻗어 방어 자세를 취하며, B-200은 지면에 선 채로 뒤로 물러나는 무게중심을 보여줌."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 후면 3/4 뷰를 무시하고 찰리의 정면을 렌더링했으며, B-200의 핵심인 오른팔 고사포 무기를 누락하는 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 후면 3/4 카메라 앵글과 찰리의 등 뒤에서 튀는 스파크, B-200과의 공간적 분리 및 레퍼런스 디자인을 완벽하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "총알이 왼쪽에서 날아와 찰리의 가슴과 정면에 명중함. 찰리는 카메라를 향해 정면으로 서 있고, B-200은 그 뒤에 위치함.",
        "built_space": "젖은 아스팔트 도로 위에 위치하며, 왼쪽의 전복된 트럭과 배경의 교회, 급수탑 등 레퍼런스의 주요 지형지물이 배치됨.",
        "entities": "찰리는 고릴라 형태를 띠고 있으나 프롬프트와 달리 가슴 원자로가 뚜렷이 보이는 정면 모습임. B-200은 오른팔의 고사포가 사라지고 뭉툭한 형태로 왜곡됨.",
        "hard_violations": [
         "카메라 시점 위반 (후면 3/4 뷰 대신 정면 뷰를 렌더링함)",
         "프롬프트 액션 위반 (등에 부딪히는 스파크 대신 정면에 명중함)",
         "캐릭터 디자인 위반 (B-200의 오른팔 고사포가 무단으로 삭제됨)"
        ],
        "physics": "찰리와 B-200 모두 도로 지면에 양발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "총알이 왼쪽에서 날아와 찰리의 등 쪽에 명중하여 스파크를 냄. 찰리는 B-200을 향해 시선을 돌리고, B-200은 물러서며 찰리를 올려다봄.",
        "built_space": "젖은 도로 위에 위치하며, 왼쪽에는 전복된 차량이 있고 배경의 교회와 급수탑이 올바른 방향과 스케일로 자리잡고 있음.",
        "entities": "찰리는 후면이 강조된 샌드 베이지 장갑판 묘사를 따름. B-200은 레퍼런스에 맞춰 오른팔 고사포, 파손된 왼팔, 부서진 몸통을 정확히 유지함.",
        "hard_violations": [],
        "physics": "찰리는 왼쪽 지면에 발을 단단히 딛고 대각선으로 몸을 뻗어 방어 자세를 취하며, B-200은 지면에 선 채로 뒤로 물러나는 무게중심을 보여줌."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 후면 3/4 뷰를 무시하고 찰리의 정면을 렌더링했으며, B-200의 핵심인 오른팔 고사포 무기를 누락하는 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 후면 3/4 카메라 앵글과 찰리의 등 뒤에서 튀는 스파크, B-200과의 공간적 분리 및 레퍼런스 디자인을 완벽하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "총알이 왼쪽에서 날아와 찰리의 가슴과 정면에 명중함. 찰리는 카메라를 향해 정면으로 서 있고, B-200은 그 뒤에 위치함.",
        "built_space": "젖은 아스팔트 도로 위에 위치하며, 왼쪽의 전복된 트럭과 배경의 교회, 급수탑 등 레퍼런스의 주요 지형지물이 배치됨.",
        "entities": "찰리는 고릴라 형태를 띠고 있으나 프롬프트와 달리 가슴 원자로가 뚜렷이 보이는 정면 모습임. B-200은 오른팔의 고사포가 사라지고 뭉툭한 형태로 왜곡됨.",
        "hard_violations": [
         "카메라 시점 위반 (후면 3/4 뷰 대신 정면 뷰를 렌더링함)",
         "프롬프트 액션 위반 (등에 부딪히는 스파크 대신 정면에 명중함)",
         "캐릭터 디자인 위반 (B-200의 오른팔 고사포가 무단으로 삭제됨)"
        ],
        "physics": "찰리와 B-200 모두 도로 지면에 양발을 딛고 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "총알이 왼쪽에서 날아와 찰리의 등 쪽에 명중하여 스파크를 냄. 찰리는 B-200을 향해 시선을 돌리고, B-200은 물러서며 찰리를 올려다봄.",
        "built_space": "젖은 도로 위에 위치하며, 왼쪽에는 전복된 차량이 있고 배경의 교회와 급수탑이 올바른 방향과 스케일로 자리잡고 있음.",
        "entities": "찰리는 후면이 강조된 샌드 베이지 장갑판 묘사를 따름. B-200은 레퍼런스에 맞춰 오른팔 고사포, 파손된 왼팔, 부서진 몸통을 정확히 유지함.",
        "hard_violations": [],
        "physics": "찰리는 왼쪽 지면에 발을 단단히 딛고 대각선으로 몸을 뻗어 방어 자세를 취하며, B-200은 지면에 선 채로 뒤로 물러나는 무게중심을 보여줌."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 후방 사선 와이드 구도에서 찰리의 등 피탄, 오른쪽 B-200의 손상, 두 몸 사이 도로와 방어 관계가 명확하지만 몸을 날린 찰나보다는 착지 후 버티는 자세에 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "찰리의 횡방향 동세는 강하지만 B-200까지 직접 피탄하고 두 몸이 겹쳐 보호 간격이 약하며, 가슴 원자로를 등으로 옮긴 외형 오류가 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 화면 오른쪽의 B-200 쪽으로 머리를 돌리고, B-200은 왼쪽의 찰리를 바라본다. 다만 B-200이 뚜렷하게 올려다보는 각도는 약하다. 왼쪽 밖에서 들어오는 탄도는 찰리의 등과 어깨에 닿아 불꽃을 만든다. B-200의 남은 포신은 왼쪽 아래 노면 방향이며 찰리를 직접 겨누지 않는다.",
        "built_space": "왼쪽 전경에 찰리, 오른쪽 중경에 B-200이 있고 둘 사이에 젖은 도로가 넓게 드러난다. 도로는 두 몸 사이에서 배경으로 사선 후퇴한다. 왼쪽 가장자리에 전복된 트럭이 최소 한 대 보이고, 배경에는 원통형 고가 물탱크 한 기, 교회 건물 한 채와 그 종탑, 기와지붕 건물들, 전신주와 울타리가 있다. 이전 장면의 시설 종류와 야간 도로 재질을 유지한다. 노면의 빛 반사도 가능한 형태다.",
        "entities": "등장 개체는 기계인 찰리와 B-200 둘뿐이다. 찰리는 샌드 베이지 장갑, 긴 팔과 짧은 다리, 흰 마스크의 옆면과 주황색 눈, 다수의 탄흔을 보인다. 가슴의 푸른 빛은 몸 옆으로 일부만 보인다. B-200은 짙은 회색 장갑, 주황색 눈, 손상된 흉복부, 남은 포신 하나와 전선이 노출된 반대쪽 팔을 갖췄다. 원근 차이로 두 기계의 실제 크기 비교는 제한적이다. 밝은 보름달과 야간 환경이 보이며 추가 인물이나 화면 위 문자는 없다. 열린 수송 케이지와 끊어진 구속구는 명확히 식별되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 벌린 두 발을 노면에 붙이고 무릎을 굽혀 체중을 받는다. B-200도 두 발로 도로를 딛고 있다. 지지 없이 떠 있는 몸은 없다. 찰리의 자세는 횡방향으로 끼어든 뒤 제동하는 순간으로는 가능하지만, 공중으로 몸을 날리는 순간 자체는 아니다. 등 표면의 충돌점에서 불꽃이 퍼지고, B-200의 파손 전선은 팔에서 아래로 늘어진다."
       },
       {
        "label": "B",
        "direction": "찰리는 머리를 오른쪽 B-200 쪽으로 돌리고 왼팔을 반대편으로 길게 뻗는다. B-200은 찰리 쪽으로 고개를 기울이지만 올려다보는 시선은 분명하지 않다. 왼쪽에서 오른쪽으로 향하는 탄도와 불꽃이 찰리의 등뿐 아니라 B-200의 어깨와 흉부에도 이어져, 총탄을 대신 막는 관계가 약해진다. B-200의 남은 포신은 찰리 뒤에 상당 부분 가려져 조준 대상을 확정하기 어렵다.",
        "built_space": "찰리가 왼쪽부터 중앙 전경을 크게 차지하고 B-200은 오른쪽 뒤에서 찰리의 팔과 몸에 일부 겹친다. 도로는 발밑과 배경에 보이지만 두 몸 사이의 열린 간격은 A보다 좁다. 왼쪽에 전복 트럭이 최소 한 대, 뒤쪽에 교회 한 채, 오른쪽에 고가 원통형 물탱크 한 기와 기와지붕 건물들이 있다. 전신주, 비계, 가드레일과 젖은 도로도 이전 장소의 주요 요소에 부합한다.",
        "entities": "인물은 두 기계뿐이며 찰리의 베이지 장갑, 긴 팔, 짧은 다리, 흰 얼굴 옆면과 B-200의 회색 중장갑, 주황색 눈, 파손된 몸통과 노출 전선은 식별된다. 그러나 찰리의 카메라를 향한 등 중앙에 푸른 원자로가 달려 있어 가슴 중앙이라는 지정과 다르다. B-200의 포신을 잃은 쪽 팔은 오른쪽에 읽히지만 남은 무기는 일부 가려진다. 밝은 보름달은 유지된다. 열린 케이지와 끊어진 구속구는 명확하게 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 양발을 도로에 붙인 채 다리를 벌리고 상체를 오른쪽으로 기울여 버틴다. 뻗은 팔과 기울어진 몸은 횡이동을 멈추는 동작으로 가능하지만 공중 도약은 아니다. B-200은 화면 오른쪽의 발로 노면을 딛고 뒤로 기울며, 다른 다리는 찰리에게 일부 가려진다. 보이는 지지점으로 후퇴 동작을 설명할 수 있고 지지 없이 떠 있는 몸은 없다. 불꽃은 두 로봇의 표면 충돌점에서 발생하며 전선은 파손 팔에 매달려 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 후방 사선 와이드 구도에서 찰리의 등 피탄, 오른쪽 B-200의 손상, 두 몸 사이 도로와 방어 관계가 명확하지만 몸을 날린 찰나보다는 착지 후 버티는 자세에 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "찰리의 횡방향 동세는 강하지만 B-200까지 직접 피탄하고 두 몸이 겹쳐 보호 간격이 약하며, 가슴 원자로를 등으로 옮긴 외형 오류가 있다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 화면 오른쪽의 B-200 쪽으로 머리를 돌리고, B-200은 왼쪽의 찰리를 바라본다. 다만 B-200이 뚜렷하게 올려다보는 각도는 약하다. 왼쪽 밖에서 들어오는 탄도는 찰리의 등과 어깨에 닿아 불꽃을 만든다. B-200의 남은 포신은 왼쪽 아래 노면 방향이며 찰리를 직접 겨누지 않는다.",
        "built_space": "왼쪽 전경에 찰리, 오른쪽 중경에 B-200이 있고 둘 사이에 젖은 도로가 넓게 드러난다. 도로는 두 몸 사이에서 배경으로 사선 후퇴한다. 왼쪽 가장자리에 전복된 트럭이 최소 한 대 보이고, 배경에는 원통형 고가 물탱크 한 기, 교회 건물 한 채와 그 종탑, 기와지붕 건물들, 전신주와 울타리가 있다. 이전 장면의 시설 종류와 야간 도로 재질을 유지한다. 노면의 빛 반사도 가능한 형태다.",
        "entities": "등장 개체는 기계인 찰리와 B-200 둘뿐이다. 찰리는 샌드 베이지 장갑, 긴 팔과 짧은 다리, 흰 마스크의 옆면과 주황색 눈, 다수의 탄흔을 보인다. 가슴의 푸른 빛은 몸 옆으로 일부만 보인다. B-200은 짙은 회색 장갑, 주황색 눈, 손상된 흉복부, 남은 포신 하나와 전선이 노출된 반대쪽 팔을 갖췄다. 원근 차이로 두 기계의 실제 크기 비교는 제한적이다. 밝은 보름달과 야간 환경이 보이며 추가 인물이나 화면 위 문자는 없다. 열린 수송 케이지와 끊어진 구속구는 명확히 식별되지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 벌린 두 발을 노면에 붙이고 무릎을 굽혀 체중을 받는다. B-200도 두 발로 도로를 딛고 있다. 지지 없이 떠 있는 몸은 없다. 찰리의 자세는 횡방향으로 끼어든 뒤 제동하는 순간으로는 가능하지만, 공중으로 몸을 날리는 순간 자체는 아니다. 등 표면의 충돌점에서 불꽃이 퍼지고, B-200의 파손 전선은 팔에서 아래로 늘어진다."
       },
       {
        "label": "A",
        "direction": "찰리는 머리를 오른쪽 B-200 쪽으로 돌리고 왼팔을 반대편으로 길게 뻗는다. B-200은 찰리 쪽으로 고개를 기울이지만 올려다보는 시선은 분명하지 않다. 왼쪽에서 오른쪽으로 향하는 탄도와 불꽃이 찰리의 등뿐 아니라 B-200의 어깨와 흉부에도 이어져, 총탄을 대신 막는 관계가 약해진다. B-200의 남은 포신은 찰리 뒤에 상당 부분 가려져 조준 대상을 확정하기 어렵다.",
        "built_space": "찰리가 왼쪽부터 중앙 전경을 크게 차지하고 B-200은 오른쪽 뒤에서 찰리의 팔과 몸에 일부 겹친다. 도로는 발밑과 배경에 보이지만 두 몸 사이의 열린 간격은 A보다 좁다. 왼쪽에 전복 트럭이 최소 한 대, 뒤쪽에 교회 한 채, 오른쪽에 고가 원통형 물탱크 한 기와 기와지붕 건물들이 있다. 전신주, 비계, 가드레일과 젖은 도로도 이전 장소의 주요 요소에 부합한다.",
        "entities": "인물은 두 기계뿐이며 찰리의 베이지 장갑, 긴 팔, 짧은 다리, 흰 얼굴 옆면과 B-200의 회색 중장갑, 주황색 눈, 파손된 몸통과 노출 전선은 식별된다. 그러나 찰리의 카메라를 향한 등 중앙에 푸른 원자로가 달려 있어 가슴 중앙이라는 지정과 다르다. B-200의 포신을 잃은 쪽 팔은 오른쪽에 읽히지만 남은 무기는 일부 가려진다. 밝은 보름달은 유지된다. 열린 케이지와 끊어진 구속구는 명확하게 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 양발을 도로에 붙인 채 다리를 벌리고 상체를 오른쪽으로 기울여 버틴다. 뻗은 팔과 기울어진 몸은 횡이동을 멈추는 동작으로 가능하지만 공중 도약은 아니다. B-200은 화면 오른쪽의 발로 노면을 딛고 뒤로 기울며, 다른 다리는 찰리에게 일부 가려진다. 보이는 지지점으로 후퇴 동작을 설명할 수 있고 지지 없이 떠 있는 몸은 없다. 불꽃은 두 로봇의 표면 충돌점에서 발생하며 전선은 파손 팔에 매달려 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.083,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.833,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 카메라 시점 위반 (후면 3/4 뷰 대신 정면 뷰를 렌더링함)",
     "[gemini-pro] 프롬프트 액션 위반 (등에 부딪히는 스파크 대신 정면에 명중함)",
     "[gemini-pro] 캐릭터 디자인 위반 (B-200의 오른팔 고사포가 무단으로 삭제됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 833,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "지정된 후면 3/4 뷰를 무시하고 찰리의 정면을 렌더링했으며, B-200의 핵심인 오른팔 고사포 무기를 누락하는 치명적인 위반이 있습니다.  ★위반: [gemini-pro] 카메라 시점 위반 (후면 3/4 뷰 대신 정면 뷰를 렌더링함) / [gemini-pro] 프롬프트 액션 위반 (등에 부딪히는 스파크 대신 정면에 명중함) / [gemini-pro] 캐릭터 디자인 위반 (B-200의 오른팔 고사포가 무단으로 삭제됨)"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 후면 3/4 카메라 앵글과 찰리의 등 뒤에서 튀는 스파크, B-200과의 공간적 분리 및 레퍼런스 디자인을 완벽하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of B-200 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S67sh83_sel.png",
    "asset_id": "5b5b8d79-3d2b-46e2-8c42-2395827aa7e9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — B-200: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1463373>",
    "asset_id": "634d476f-22db-4067-82af-2c4e510cff32",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0cfe-5d9b-7802-bb62-673e431ea998",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S67sh83"
  },
  "staged_characters_added": [
   "C31"
  ]
 },
 "S67sh119::signage": {
  "fp": "8112e046d9d449a7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S67sh119": {
  "input_fingerprint": "def5e8399831bf3b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200의 부품을 양손으로 소중히 꽉 움켜쥔 채 고개를 푹 숙인 찰리의 슬프고 굽은 상체.\n\nLOCATION (lock): At the blast site on the village's outer road, among destroyed trucks and scattered mechanical remains at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane at its close, elevated side position on the established side of the road, looking obliquely down at 찰리's bowed upper body rather than beginning the withdrawal. Place his bowed head in the upper-left portion and his two tightly cupped hands below center, with B-200's surviving chest component small within their grasp and a narrow margin of ground beyond. His attention stays downward on the component; the enclosing curve of his shoulders and forearms carries the grief without introducing another visible figure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's surviving chest component (Recovered and held tightly in both hands) — Partly enclosed by the fingers, with its exposed portion angled toward the elevated camera; used as A small emotional focal point connecting the bowed head and enclosing hands; Ground at the explosion site (Visible around the cropped upper body after the smoke has cleared); used as Provides an unobtrusive spatial margin around the isolated act of mourning.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime ambient illumination subdued and continuous, preserving readable hand and body detail without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): B-200 has been destroyed, with only a component from his chest remaining, held in Charlie's hands. No intact head, torso, arms, or legs remain to establish a whole-body pose or facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned trucks and the aftermath of the explosion remain at the nighttime battle site; B-200 has been destroyed, leaving its chest component. Charlie's already damaged body has sustained additional gunfire, and he now holds the recovered chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200의 부품을 양손으로 소중히 꽉 움켜쥔 채 고개를 푹 숙인 찰리의 슬프고 굽은 상체.\n\nLOCATION (lock): At the blast site on the village's outer road, among destroyed trucks and scattered mechanical remains at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane at its close, elevated side position on the established side of the road, looking obliquely down at 찰리's bowed upper body rather than beginning the withdrawal. Place his bowed head in the upper-left portion and his two tightly cupped hands below center, with B-200's surviving chest component small within their grasp and a narrow margin of ground beyond. His attention stays downward on the component; the enclosing curve of his shoulders and forearms carries the grief without introducing another visible figure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's surviving chest component (Recovered and held tightly in both hands) — Partly enclosed by the fingers, with its exposed portion angled toward the elevated camera; used as A small emotional focal point connecting the bowed head and enclosing hands; Ground at the explosion site (Visible around the cropped upper body after the smoke has cleared); used as Provides an unobtrusive spatial margin around the isolated act of mourning.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime ambient illumination subdued and continuous, preserving readable hand and body detail without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): B-200 has been destroyed, with only a component from his chest remaining, held in Charlie's hands. No intact head, torso, arms, or legs remain to establish a whole-body pose or facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned trucks and the aftermath of the explosion remain at the nighttime battle site; B-200 has been destroyed, leaving its chest component. Charlie's already damaged body has sustained additional gunfire, and he now holds the recovered chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): sunset to night, bright full moon.\n\nSHOT TEXT (authoritative, Korean): B-200의 부품을 양손으로 소중히 꽉 움켜쥔 채 고개를 푹 숙인 찰리의 슬프고 굽은 상체.\n\nLOCATION (lock): At the blast site on the village's outer road, among destroyed trucks and scattered mechanical remains at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane at its close, elevated side position on the established side of the road, looking obliquely down at 찰리's bowed upper body rather than beginning the withdrawal. Place his bowed head in the upper-left portion and his two tightly cupped hands below center, with B-200's surviving chest component small within their grasp and a narrow margin of ground beyond. His attention stays downward on the component; the enclosing curve of his shoulders and forearms carries the grief without introducing another visible figure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: B-200's surviving chest component (Recovered and held tightly in both hands) — Partly enclosed by the fingers, with its exposed portion angled toward the elevated camera; used as A small emotional focal point connecting the bowed head and enclosing hands; Ground at the explosion site (Visible around the cropped upper body after the smoke has cleared); used as Provides an unobtrusive spatial margin around the isolated act of mourning.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime ambient illumination subdued and continuous, preserving readable hand and body detail without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): B-200 has been destroyed, with only a component from his chest remaining, held in Charlie's hands. No intact head, torso, arms, or legs remain to establish a whole-body pose or facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned trucks and the aftermath of the explosion remain at the nighttime battle site; B-200 has been destroyed, leaving its chest component. Charlie's already damaged body has sustained additional gunfire, and he now holds the recovered chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 고개가 숙여져 있고 시선이 손에 들린 부품에 고정되어 있습니다.",
    "built_space": "물기가 있는 도로 파편들이 보이며 우측 상단에 차량 잔해 일부가 있습니다.",
    "entities": "찰리의 외형은 레퍼런스에 부합하나, 손에 든 B-200 부품의 형태와 조명 디테일이 레퍼런스와 조금 다르게 변형되었습니다.",
    "hard_violations": [],
    "physics": "부품은 찰리의 양손에 의해 지탱되고 있으며, 자세의 무게 중심이 안정적입니다."
   },
   {
    "label": "B",
    "direction": "찰리의 고개가 아래로 숙여져 있으며 시선은 양손에 쥔 부품을 향하고 있습니다.",
    "built_space": "젖은 아스팔트 바닥과 파편들, 배경에 파괴된 차량 잔해가 보이며 지정된 폭발 현장의 야간 도로 느낌을 줍니다.",
    "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크)이 레퍼런스와 일치하며, 손에 든 B-200 부품의 내부 톱니바퀴와 푸른색 회로 패턴이 레퍼런스와 매우 흡사합니다.",
    "hard_violations": [],
    "physics": "찰리의 양손이 부품을 단단히 받치고 쥐고 있으며, 굽힌 상체는 자연스럽게 중력을 따릅니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 구도를 정확히 따랐으며, 레퍼런스와 부품의 디자인 일치도가 매우 높고 배경의 환경적 분위기도 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 찰리의 외형은 잘 구현되었으나, 손에 든 부품의 디테일이 레퍼런스와 비교해 다소 변형되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 고개가 아래로 숙여져 있으며 시선은 양손에 쥔 부품을 향하고 있습니다.",
        "built_space": "젖은 아스팔트 바닥과 파편들, 배경에 파괴된 차량 잔해가 보이며 지정된 폭발 현장의 야간 도로 느낌을 줍니다.",
        "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크)이 레퍼런스와 일치하며, 손에 든 B-200 부품의 내부 톱니바퀴와 푸른색 회로 패턴이 레퍼런스와 매우 흡사합니다.",
        "hard_violations": [],
        "physics": "찰리의 양손이 부품을 단단히 받치고 쥐고 있으며, 굽힌 상체는 자연스럽게 중력을 따릅니다."
       },
       {
        "label": "A",
        "direction": "찰리의 고개가 숙여져 있고 시선이 손에 들린 부품에 고정되어 있습니다.",
        "built_space": "물기가 있는 도로 파편들이 보이며 우측 상단에 차량 잔해 일부가 있습니다.",
        "entities": "찰리의 외형은 레퍼런스에 부합하나, 손에 든 B-200 부품의 형태와 조명 디테일이 레퍼런스와 조금 다르게 변형되었습니다.",
        "hard_violations": [],
        "physics": "부품은 찰리의 양손에 의해 지탱되고 있으며, 자세의 무게 중심이 안정적입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구된 구도를 정확히 따랐으며, 레퍼런스와 부품의 디자인 일치도가 매우 높고 배경의 환경적 분위기도 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 찰리의 외형은 잘 구현되었으나, 손에 든 부품의 디테일이 레퍼런스와 비교해 다소 변형되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 고개가 아래로 숙여져 있으며 시선은 양손에 쥔 부품을 향하고 있습니다.",
        "built_space": "젖은 아스팔트 바닥과 파편들, 배경에 파괴된 차량 잔해가 보이며 지정된 폭발 현장의 야간 도로 느낌을 줍니다.",
        "entities": "찰리의 외형(샌드 베이지 장갑, 흰색 마스크)이 레퍼런스와 일치하며, 손에 든 B-200 부품의 내부 톱니바퀴와 푸른색 회로 패턴이 레퍼런스와 매우 흡사합니다.",
        "hard_violations": [],
        "physics": "찰리의 양손이 부품을 단단히 받치고 쥐고 있으며, 굽힌 상체는 자연스럽게 중력을 따릅니다."
       },
       {
        "label": "A",
        "direction": "찰리의 고개가 숙여져 있고 시선이 손에 들린 부품에 고정되어 있습니다.",
        "built_space": "물기가 있는 도로 파편들이 보이며 우측 상단에 차량 잔해 일부가 있습니다.",
        "entities": "찰리의 외형은 레퍼런스에 부합하나, 손에 든 B-200 부품의 형태와 조명 디테일이 레퍼런스와 조금 다르게 변형되었습니다.",
        "hard_violations": [],
        "physics": "부품은 찰리의 양손에 의해 지탱되고 있으며, 자세의 무게 중심이 안정적입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "좌상단의 숙인 머리와 중앙 아래의 감싼 양손, 작고 참조 형태에 가까운 코어가 지정 구도를 더 충실히 구현하지만, 오른쪽 지면 여백과 강한 반사는 다소 과하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "고개를 숙이고 양손으로 부품을 지탱하는 행동은 맞지만, 머리와 손이 더 오른쪽으로 치우치고 코어의 외형도 참조에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 아래쪽 양손 사이의 코어를 향한다. 두 팔이 안쪽으로 굽어 부품을 둘러싸며, 코어의 기어와 푸른 회로가 드러난 면은 높은 카메라 쪽으로 기울어 있다. 카메라를 바라보거나 다른 대상을 겨냥하는 요소는 없다.",
        "built_space": "높은 사선 시점에서 상체를 가까이 내려다본다. 머리는 좌상단, 양손은 중앙 아래에 놓인다. 우상단에는 바퀴와 옆면이 보이는 전복 차량 일부가 있고, 젖은 도로와 잔해가 주변을 채운다. 별도의 온전한 차량 수나 마을 고정 시설은 이 크롭으로 확인할 수 없다. 오른쪽 지면 여백은 요구된 좁은 여백보다 넓다. 물웅덩이에 원형의 밝은 반사가 있으나 화면 밖 광원 위치를 알 수 없어 광학적으로 불가능하다고 단정할 수 없다.",
        "entities": "등장 개체는 찰리 한 대뿐이며 다른 인물이나 온전한 B-200은 없다. 샌드 베이지 장갑, 육중한 기계 팔, 흰 마스크, 주황색 점 눈 두 개와 입 선, 탄흔과 그을음이 참조 정체성과 이어진다. 인간 피부나 치아는 없다. 손에 든 작은 코어는 은색 둥근 사각 외함, 노출 기어, 푸른 회로와 하단 원통이 보여 소품 참조에 가깝다. 다만 보이는 찰리 가슴 중앙에서 참조의 푸른 원자로는 확인되지 않고 손상 부위가 두드러진다.",
        "hard_violations": [],
        "physics": "양손의 손바닥과 굽힌 손가락이 코어의 바닥과 양옆에 접촉해 무게를 받친다. 손목과 팔꿈치 연결이 이어지고, 굽힌 팔과 숙인 상체로 부품을 소중히 감싸는 동작이 가능하다. 하체 접지는 크롭 밖이므로 확인할 수 없지만 공중에 떠 있는 몸으로 보이지 않는다. 도로 잔해는 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "찰리는 고개를 아래로 숙여 양손에 든 코어 쪽을 향한다. 양팔과 손가락이 부품을 안쪽으로 감싸며, 노출된 기어와 푸른 부분은 높은 카메라에서도 보이는 방향이다. 시선이 카메라나 다른 개체로 향하지 않는다.",
        "built_space": "상체를 높은 사선에서 내려다보는 가까운 구도다. 머리는 좌상단보다는 상단 중앙에 가깝고 정수리가 화면 위에서 잘리며, 양손은 중앙 아래보다 오른쪽 아래로 이동해 있다. 우상단에는 전복 차량 한 대의 일부로 읽히는 바퀴와 차체가 보이고, 젖은 도로와 파편이 둘레를 채운다. 마을 시설은 크롭 밖이며 중복된 고정 설비는 보이지 않는다. 오른쪽 도로 여백이 넓고, 물 표면에는 차갑고 따뜻한 빛의 반사가 있다.",
        "entities": "찰리 한 대만 보이며 추가 인물이나 B-200의 온전한 신체는 없다. 베이지 장갑, 굵은 기계 팔, 흰 얼굴판과 두 점 눈 및 입 선, 총격 손상은 정체성과 대체로 맞는다. 가슴 중앙은 검게 손상되어 참조의 푸른 원자로가 보이지 않는다. 코어에는 푸른 회로와 기어가 있지만 둥근 사각 외함 및 하단 원통 구성이 덜 분명하고, 가로로 펼쳐진 기계 조립체처럼 보여 소품 참조와 차이가 더 크다.",
        "hard_violations": [],
        "physics": "코어는 양손의 손바닥과 손가락 사이에 받쳐져 있으며 손과 분리되어 떠 있지 않다. 양쪽 전완과 손목이 연결되고 팔꿈치를 굽혀 부품을 품는 자세가 성립한다. 다리와 발의 접지는 프레임 밖이라 판단할 수 없으며, 상체가 떠 있다는 증거는 없다. 배경 차체와 파편은 도로에 기대거나 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "좌상단의 숙인 머리와 중앙 아래의 감싼 양손, 작고 참조 형태에 가까운 코어가 지정 구도를 더 충실히 구현하지만, 오른쪽 지면 여백과 강한 반사는 다소 과하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "고개를 숙이고 양손으로 부품을 지탱하는 행동은 맞지만, 머리와 손이 더 오른쪽으로 치우치고 코어의 외형도 참조에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 아래쪽 양손 사이의 코어를 향한다. 두 팔이 안쪽으로 굽어 부품을 둘러싸며, 코어의 기어와 푸른 회로가 드러난 면은 높은 카메라 쪽으로 기울어 있다. 카메라를 바라보거나 다른 대상을 겨냥하는 요소는 없다.",
        "built_space": "높은 사선 시점에서 상체를 가까이 내려다본다. 머리는 좌상단, 양손은 중앙 아래에 놓인다. 우상단에는 바퀴와 옆면이 보이는 전복 차량 일부가 있고, 젖은 도로와 잔해가 주변을 채운다. 별도의 온전한 차량 수나 마을 고정 시설은 이 크롭으로 확인할 수 없다. 오른쪽 지면 여백은 요구된 좁은 여백보다 넓다. 물웅덩이에 원형의 밝은 반사가 있으나 화면 밖 광원 위치를 알 수 없어 광학적으로 불가능하다고 단정할 수 없다.",
        "entities": "등장 개체는 찰리 한 대뿐이며 다른 인물이나 온전한 B-200은 없다. 샌드 베이지 장갑, 육중한 기계 팔, 흰 마스크, 주황색 점 눈 두 개와 입 선, 탄흔과 그을음이 참조 정체성과 이어진다. 인간 피부나 치아는 없다. 손에 든 작은 코어는 은색 둥근 사각 외함, 노출 기어, 푸른 회로와 하단 원통이 보여 소품 참조에 가깝다. 다만 보이는 찰리 가슴 중앙에서 참조의 푸른 원자로는 확인되지 않고 손상 부위가 두드러진다.",
        "hard_violations": [],
        "physics": "양손의 손바닥과 굽힌 손가락이 코어의 바닥과 양옆에 접촉해 무게를 받친다. 손목과 팔꿈치 연결이 이어지고, 굽힌 팔과 숙인 상체로 부품을 소중히 감싸는 동작이 가능하다. 하체 접지는 크롭 밖이므로 확인할 수 없지만 공중에 떠 있는 몸으로 보이지 않는다. 도로 잔해는 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "찰리는 고개를 아래로 숙여 양손에 든 코어 쪽을 향한다. 양팔과 손가락이 부품을 안쪽으로 감싸며, 노출된 기어와 푸른 부분은 높은 카메라에서도 보이는 방향이다. 시선이 카메라나 다른 개체로 향하지 않는다.",
        "built_space": "상체를 높은 사선에서 내려다보는 가까운 구도다. 머리는 좌상단보다는 상단 중앙에 가깝고 정수리가 화면 위에서 잘리며, 양손은 중앙 아래보다 오른쪽 아래로 이동해 있다. 우상단에는 전복 차량 한 대의 일부로 읽히는 바퀴와 차체가 보이고, 젖은 도로와 파편이 둘레를 채운다. 마을 시설은 크롭 밖이며 중복된 고정 설비는 보이지 않는다. 오른쪽 도로 여백이 넓고, 물 표면에는 차갑고 따뜻한 빛의 반사가 있다.",
        "entities": "찰리 한 대만 보이며 추가 인물이나 B-200의 온전한 신체는 없다. 베이지 장갑, 굵은 기계 팔, 흰 얼굴판과 두 점 눈 및 입 선, 총격 손상은 정체성과 대체로 맞는다. 가슴 중앙은 검게 손상되어 참조의 푸른 원자로가 보이지 않는다. 코어에는 푸른 회로와 기어가 있지만 둥근 사각 외함 및 하단 원통 구성이 덜 분명하고, 가로로 펼쳐진 기계 조립체처럼 보여 소품 참조와 차이가 더 크다.",
        "hard_violations": [],
        "physics": "코어는 양손의 손바닥과 손가락 사이에 받쳐져 있으며 손과 분리되어 떠 있지 않다. 양쪽 전완과 손목이 연결되고 팔꿈치를 굽혀 부품을 품는 자세가 성립한다. 다리와 발의 접지는 프레임 밖이라 판단할 수 없으며, 상체가 떠 있다는 증거는 없다. 배경 차체와 파편은 도로에 기대거나 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.732,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.732,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1732
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 구도를 정확히 따랐으며, 레퍼런스와 부품의 디자인 일치도가 매우 높고 배경의 환경적 분위기도 잘 살렸습니다."
   },
   {
    "label": "A",
    "score": 1732,
    "verdict_ko": "구도와 찰리의 외형은 잘 구현되었으나, 손에 든 부품의 디테일이 레퍼런스와 비교해 다소 변형되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S67sh95_sel.png",
    "asset_id": "44769ba0-60ea-4e49-a4a9-1abc0499141e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — B-200의 가슴 코어 부품: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1066829>",
    "asset_id": "49df2f5c-8e78-419f-a82d-962582f6b8be",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d04-60f8-7d3f-aa6d-259dc04403de",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S67sh95"
  }
 },
 "S68sh4::signage": {
  "fp": "e5dd304c80fcbe2b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::departure_road_fork": {
  "input_fingerprint": "952f6ab8f70f4b56",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "departure_road_fork",
    "tags": [
     "S68sh4"
    ]
   },
   "context_sig": "8982d04fd01599ac"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a rural road fork outside the village in daylight, where the two truck groups take separate branches.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 이내 도달한 갈림길. 그러자 현우와 수빈의 트럭이 양 갈래로 갈라지고.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a rural road fork outside the village in daylight, where the two truck groups take separate branches.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 이내 도달한 갈림길. 그러자 현우와 수빈의 트럭이 양 갈래로 갈라지고.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_departure_road_fork_bcc32e.png",
  "asset_id": "ba0541fa-bfcd-471b-a4bf-ec26a45d05cf",
  "input_asset_ids": [
   "cbc7fcb1-4bf4-4a5c-99a7-b07b039e1614"
  ],
  "origin_tag": "S68sh4",
  "place_text": "At a rural road fork outside the village in daylight, where the two truck groups take separate branches.",
  "origin_inputs": {
   "place_text": "At a rural road fork outside the village in daylight, where the two truck groups take separate branches.",
   "time_of_day_en": "day",
   "conti_asset_id": "cbc7fcb1-4bf4-4a5c-99a7-b07b039e1614"
  }
 },
 "S68sh4::bgfirst_bg": {
  "input_fingerprint": "074d32b604ef02dc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갈림길에서 이현우의 트럭과 수빈의 트럭이 양 갈래로 벌어지는 찰나.\n\nLOCATION (lock): At a rural road fork outside the village in daylight, where the two truck groups take separate branches.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rising crane move, remain behind and oblique to the roofless trucks, looking diagonally down across the fork as their separation becomes the principal change. Place 이현우's truck on the left branch and 수빈's on the right, each occupying less than a quarter of the frame, with following vehicles receding behind 수빈 rather than arranged at identical intervals. Both drivers are seen from behind, leaning into their respective turns and attending to the road ahead; leave the leftward route open for the subsequent continuous descent toward 찰리's truck.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road branch taken by 이현우's truck in the upper-left of the frame, background; Road branch taken by 수빈's truck and following vehicles in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Forked road (Dividing into the two routes taken by the trucks) — The shared approach enters from below and separates toward the upper-left and upper-right; used as Makes the farewell readable through diverging travel paths; 이현우's truck (Old, roofless, moving onto one branch) — Rear and near side visible from above; the open driving area reveals 이현우; used as Leftward moving anchor and destination of the next camera move; 수빈's truck and following vehicles (Moving onto the other branch) — Rear quarters remain visible, with following vehicles staggered in depth along the route; used as Carries the departing group away without making the convoy look mechanically repeated.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight with controlled contrast and clear separation between the vehicles and roadway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 갈림길에서 이현우의 트럭과 수빈의 트럭이 양 갈래로 벌어지는 찰나.\n\nLOCATION (lock): At a rural road fork outside the village in daylight, where the two truck groups take separate branches.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rising crane move, remain behind and oblique to the roofless trucks, looking diagonally down across the fork as their separation becomes the principal change. Place 이현우's truck on the left branch and 수빈's on the right, each occupying less than a quarter of the frame, with following vehicles receding behind 수빈 rather than arranged at identical intervals. Both drivers are seen from behind, leaning into their respective turns and attending to the road ahead; leave the leftward route open for the subsequent continuous descent toward 찰리's truck.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road branch taken by 이현우's truck in the upper-left of the frame, background; Road branch taken by 수빈's truck and following vehicles in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Forked road (Dividing into the two routes taken by the trucks) — The shared approach enters from below and separates toward the upper-left and upper-right; used as Makes the farewell readable through diverging travel paths; 이현우's truck (Old, roofless, moving onto one branch) — Rear and near side visible from above; the open driving area reveals 이현우; used as Leftward moving anchor and destination of the next camera move; 수빈's truck and following vehicles (Moving onto the other branch) — Rear quarters remain visible, with following vehicles staggered in depth along the route; used as Carries the departing group away without making the convoy look mechanically repeated.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight with controlled contrast and clear separation between the vehicles and roadway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh4__bgfirst_bg.png",
  "asset_id": "944ba677-25c7-4539-9f9a-eb8676d9142a",
  "input_asset_ids": [
   "cbc7fcb1-4bf4-4a5c-99a7-b07b039e1614",
   "ba0541fa-bfcd-471b-a4bf-ec26a45d05cf"
  ]
 },
 "S68sh4": {
  "input_fingerprint": "a99b32c628f8ac99",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 이현우의 트럭과 수빈의 트럭이 양 갈래로 벌어지는 찰나.\n\nLOCATION (lock): At a rural road fork outside the village in daylight, where the two truck groups take separate branches. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rising crane move, remain behind and oblique to the roofless trucks, looking diagonally down across the fork as their separation becomes the principal change. Place 이현우's truck on the left branch and 수빈's on the right, each occupying less than a quarter of the frame, with following vehicles receding behind 수빈 rather than arranged at identical intervals. Both drivers are seen from behind, leaning into their respective turns and attending to the road ahead; leave the leftward route open for the subsequent continuous descent toward 찰리's truck.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road branch taken by 이현우's truck in the upper-left of the frame, background; Road branch taken by 수빈's truck and following vehicles in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Forked road (Dividing into the two routes taken by the trucks) — The shared approach enters from below and separates toward the upper-left and upper-right; used as Makes the farewell readable through diverging travel paths; 이현우's truck (Old, roofless, moving onto one branch) — Rear and near side visible from above; the open driving area reveals 이현우; used as Leftward moving anchor and destination of the next camera move; 수빈's truck and following vehicles (Moving onto the other branch) — Rear quarters remain visible, with following vehicles staggered in depth along the route; used as Carries the departing group away without making the convoy look mechanically repeated.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight with controlled contrast and clear separation between the vehicles and roadway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at a daytime fork, with additional vehicles following one branch. Charlie rides in the cargo bed, retaining his battle-damaged body and the recovered B-200 chest component. 이현우: He is driving a roofless truck and still bears his battle injuries. 수빈: She is driving the other truck; her battle wounds and pre-existing radiation damage remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 이현우의 트럭과 수빈의 트럭이 양 갈래로 벌어지는 찰나.\n\nLOCATION (lock): At a rural road fork outside the village in daylight, where the two truck groups take separate branches. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rising crane move, remain behind and oblique to the roofless trucks, looking diagonally down across the fork as their separation becomes the principal change. Place 이현우's truck on the left branch and 수빈's on the right, each occupying less than a quarter of the frame, with following vehicles receding behind 수빈 rather than arranged at identical intervals. Both drivers are seen from behind, leaning into their respective turns and attending to the road ahead; leave the leftward route open for the subsequent continuous descent toward 찰리's truck.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road branch taken by 이현우's truck in the upper-left of the frame, background; Road branch taken by 수빈's truck and following vehicles in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Forked road (Dividing into the two routes taken by the trucks) — The shared approach enters from below and separates toward the upper-left and upper-right; used as Makes the farewell readable through diverging travel paths; 이현우's truck (Old, roofless, moving onto one branch) — Rear and near side visible from above; the open driving area reveals 이현우; used as Leftward moving anchor and destination of the next camera move; 수빈's truck and following vehicles (Moving onto the other branch) — Rear quarters remain visible, with following vehicles staggered in depth along the route; used as Carries the departing group away without making the convoy look mechanically repeated.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight with controlled contrast and clear separation between the vehicles and roadway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at a daytime fork, with additional vehicles following one branch. Charlie rides in the cargo bed, retaining his battle-damaged body and the recovered B-200 chest component. 이현우: He is driving a roofless truck and still bears his battle injuries. 수빈: She is driving the other truck; her battle wounds and pre-existing radiation damage remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 갈림길에서 이현우의 트럭과 수빈의 트럭이 양 갈래로 벌어지는 찰나.\n\nLOCATION (lock): At a rural road fork outside the village in daylight, where the two truck groups take separate branches. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the rising crane move, remain behind and oblique to the roofless trucks, looking diagonally down across the fork as their separation becomes the principal change. Place 이현우's truck on the left branch and 수빈's on the right, each occupying less than a quarter of the frame, with following vehicles receding behind 수빈 rather than arranged at identical intervals. Both drivers are seen from behind, leaning into their respective turns and attending to the road ahead; leave the leftward route open for the subsequent continuous descent toward 찰리's truck.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Road branch taken by 이현우's truck in the upper-left of the frame, background; Road branch taken by 수빈's truck and following vehicles in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Forked road (Dividing into the two routes taken by the trucks) — The shared approach enters from below and separates toward the upper-left and upper-right; used as Makes the farewell readable through diverging travel paths; 이현우's truck (Old, roofless, moving onto one branch) — Rear and near side visible from above; the open driving area reveals 이현우; used as Leftward moving anchor and destination of the next camera move; 수빈's truck and following vehicles (Moving onto the other branch) — Rear quarters remain visible, with following vehicles staggered in depth along the route; used as Carries the departing group away without making the convoy look mechanically repeated.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight with controlled contrast and clear separation between the vehicles and roadway.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two old roofless trucks separate at a daytime fork, with additional vehicles following one branch. Charlie rides in the cargo bed, retaining his battle-damaged body and the recovered B-200 chest component. 이현우: He is driving a roofless truck and still bears his battle injuries. 수빈: She is driving the other truck; her battle wounds and pre-existing radiation damage remain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 수빈 (북한 출신 여성, 20세의 젊은 얼굴, 검은 단발머리, 날렵하게 자른 머리끝) — wearing: 마르고 근육질인 체격에 핏되는 낡고 실용적인 올리브색 테크웨어 상의, 카고 팬츠, 전술 부츠. 거친 피폭 흔적과 흙먼지가 있음. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh4__bgfirst_bg.png",
     "asset_id": "944ba677-25c7-4539-9f9a-eb8676d9142a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S68sh4.png",
     "asset_id": "cbc7fcb1-4bf4-4a5c-99a7-b07b039e1614",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_departure_road_fork_bcc32e.png",
     "asset_id": "ba0541fa-bfcd-471b-a4bf-ec26a45d05cf",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1001789>",
     "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 대의 트럭이 각각 왼쪽과 오른쪽 갈림길을 향해 주행 중이며, 두 운전자 모두 전방의 도로를 주시하며 나아가고 있음.",
    "built_space": "레퍼런스와 동일한 농촌의 Y자형 분기점이 묘사되었으며, 카메라는 지시된 대로 뒤쪽 상단에서 아래를 내려다보는 앵글을 취하고 있음.",
    "entities": "왼쪽 트럭 운전자는 짧은 머리와 어두운 셔츠(이현우)를, 오른쪽 트럭 운전자는 단발머리와 올리브색 상의(수빈)를 입고 있음. 두 트럭 모두 지붕이 없고 우측 경로에 후행 차량이 있으나, 화물칸의 찰리는 명확한 형상으로 보이지 않음.",
    "hard_violations": [],
    "physics": "모든 차량의 바퀴가 도로에 안정적으로 접지되어 있으며 주행 방향과 그림자가 자연스럽게 형성됨."
   },
   {
    "label": "B",
    "direction": "두 대의 트럭이 양 갈래 길을 따라 전방으로 멀어지고 있으며 운전자의 방향도 진행 방향과 일치함.",
    "built_space": "주어진 장소 레퍼런스를 정확히 반영한 시골길 분기점이며, 뒤쪽에서 부감으로 내려다보는 카메라 위치가 바르게 설정됨.",
    "entities": "왼쪽 운전자가 올리브색 상의를 입어 이현우의 설정과 다르고, 오른쪽 운전자는 단발이 아닌 짧은 머리로 묘사됨. 우측 트럭의 운전석이 오른쪽에 위치해 있으며 화물칸의 찰리는 보이지 않음.",
    "hard_violations": [],
    "physics": "차량들이 노면에 맞닿아 정상적인 주행 물리법칙을 보여주며 이질적인 부유 현상은 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 부감 카메라 구도와 배경을 정확히 구현했으며 운전자의 헤어스타일과 의상이 설정과 잘 일치하나, 화물칸에 탑승한 찰리의 디테일이 생략되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경과 분기점 구도는 훌륭하게 표현되었으나, 두 운전자의 의상 및 헤어스타일이 레퍼런스와 어긋나며 우측 트럭이 우핸들 차량으로 잘못 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 대의 트럭이 각각 왼쪽과 오른쪽 갈림길을 향해 주행 중이며, 두 운전자 모두 전방의 도로를 주시하며 나아가고 있음.",
        "built_space": "레퍼런스와 동일한 농촌의 Y자형 분기점이 묘사되었으며, 카메라는 지시된 대로 뒤쪽 상단에서 아래를 내려다보는 앵글을 취하고 있음.",
        "entities": "왼쪽 트럭 운전자는 짧은 머리와 어두운 셔츠(이현우)를, 오른쪽 트럭 운전자는 단발머리와 올리브색 상의(수빈)를 입고 있음. 두 트럭 모두 지붕이 없고 우측 경로에 후행 차량이 있으나, 화물칸의 찰리는 명확한 형상으로 보이지 않음.",
        "hard_violations": [],
        "physics": "모든 차량의 바퀴가 도로에 안정적으로 접지되어 있으며 주행 방향과 그림자가 자연스럽게 형성됨."
       },
       {
        "label": "B",
        "direction": "두 대의 트럭이 양 갈래 길을 따라 전방으로 멀어지고 있으며 운전자의 방향도 진행 방향과 일치함.",
        "built_space": "주어진 장소 레퍼런스를 정확히 반영한 시골길 분기점이며, 뒤쪽에서 부감으로 내려다보는 카메라 위치가 바르게 설정됨.",
        "entities": "왼쪽 운전자가 올리브색 상의를 입어 이현우의 설정과 다르고, 오른쪽 운전자는 단발이 아닌 짧은 머리로 묘사됨. 우측 트럭의 운전석이 오른쪽에 위치해 있으며 화물칸의 찰리는 보이지 않음.",
        "hard_violations": [],
        "physics": "차량들이 노면에 맞닿아 정상적인 주행 물리법칙을 보여주며 이질적인 부유 현상은 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 부감 카메라 구도와 배경을 정확히 구현했으며 운전자의 헤어스타일과 의상이 설정과 잘 일치하나, 화물칸에 탑승한 찰리의 디테일이 생략되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경과 분기점 구도는 훌륭하게 표현되었으나, 두 운전자의 의상 및 헤어스타일이 레퍼런스와 어긋나며 우측 트럭이 우핸들 차량으로 잘못 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 대의 트럭이 각각 왼쪽과 오른쪽 갈림길을 향해 주행 중이며, 두 운전자 모두 전방의 도로를 주시하며 나아가고 있음.",
        "built_space": "레퍼런스와 동일한 농촌의 Y자형 분기점이 묘사되었으며, 카메라는 지시된 대로 뒤쪽 상단에서 아래를 내려다보는 앵글을 취하고 있음.",
        "entities": "왼쪽 트럭 운전자는 짧은 머리와 어두운 셔츠(이현우)를, 오른쪽 트럭 운전자는 단발머리와 올리브색 상의(수빈)를 입고 있음. 두 트럭 모두 지붕이 없고 우측 경로에 후행 차량이 있으나, 화물칸의 찰리는 명확한 형상으로 보이지 않음.",
        "hard_violations": [],
        "physics": "모든 차량의 바퀴가 도로에 안정적으로 접지되어 있으며 주행 방향과 그림자가 자연스럽게 형성됨."
       },
       {
        "label": "B",
        "direction": "두 대의 트럭이 양 갈래 길을 따라 전방으로 멀어지고 있으며 운전자의 방향도 진행 방향과 일치함.",
        "built_space": "주어진 장소 레퍼런스를 정확히 반영한 시골길 분기점이며, 뒤쪽에서 부감으로 내려다보는 카메라 위치가 바르게 설정됨.",
        "entities": "왼쪽 운전자가 올리브색 상의를 입어 이현우의 설정과 다르고, 오른쪽 운전자는 단발이 아닌 짧은 머리로 묘사됨. 우측 트럭의 운전석이 오른쪽에 위치해 있으며 화물칸의 찰리는 보이지 않음.",
        "hard_violations": [],
        "physics": "차량들이 노면에 맞닿아 정상적인 주행 물리법칙을 보여주며 이질적인 부유 현상은 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "더 높은 후방 시점에서 적재함과 갈라지는 경로를 내려다봐 크레인 종료 구도에 더 가깝지만, 후속 차량들이 수빈 뒤가 아니라 앞에 있다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "수빈의 단발머리와 장소는 잘 맞지만, 카메라가 기준 사진의 낮은 시점에 더 가깝고 후속 차량도 수빈보다 앞서 있어 지정된 구도와 행렬 관계가 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 트럭은 좌상단 도로로, 오른쪽 트럭은 우상단 도로로 향하며 두 운전자도 각각 전방을 보는 뒷모습이다. 다만 오른쪽 추가 차량들은 수빈의 트럭보다 도로 진행 방향 앞쪽에 있어 '뒤따르는 차량' 관계가 반대다. 회전 방향으로 몸을 기울이는 동작은 뚜렷하지 않다.",
        "built_space": "하단 진입로 하나가 중앙의 식생 삼각지대를 두고 두 갈래로 나뉜다. 왼쪽 가드레일, 중앙 바위 하나, 우측 배경 비닐하우스 하나, 논밭과 주택군 및 양쪽 전신주가 기준 장소와 대응한다. 두 트럭은 각각 화면의 4분의 1보다 작고 좌우 지선에 놓였다. 적재함 내부를 볼 수 있는 높은 후방 시점이며 B보다 하향 관찰이 강하지만, 좌우 갈림길을 거의 정면으로 보는 축은 남아 있다.",
        "entities": "낡고 지붕 없는 주역 트럭 두 대와 오른쪽 지선의 추가 차량들이 보인다. 왼쪽 운전자는 짧은 검은 머리와 어두운 셔츠가 이현우 설정에 부합한다. 오른쪽 운전자는 올리브색 상의를 입었지만 머리가 짧고 둥글게 보여 수빈의 선명한 단발 윤곽과는 차이가 있다. 얼굴이 돌아서 있고 작아 정확한 나이·얼굴·민족적 외형, 상처와 인이어는 확인할 수 없다. 적재함에는 짐이 보이나 찰리나 B-200 부품으로 확정할 대상은 없다. 별도의 추가 인물이나 화면 위 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "트럭들은 타이어로 도로에 지지되고, 운전자들의 상체는 운전석 위치에서 올라온다. 하체와 손의 접촉은 차체에 가려져 정밀하게 확인되지 않지만 공중에 뜬 신체는 없다. 적재물은 적재함 바닥과 측벽 안에 놓여 있으며, 회전 중 차량 자세도 물리적으로 가능하다."
       },
       {
        "label": "B",
        "direction": "왼쪽 트럭의 앞부분은 좌상단 지선을, 오른쪽 트럭은 우상단 지선을 향한다. 두 운전자의 머리도 각 도로 전방을 향하며 카메라를 돌아보지 않는다. 오른쪽의 추가 차량 두 대는 수빈보다 앞서 멀어지고 있어 후속 행렬의 진행 순서가 맞지 않는다. 두 운전자는 비교적 곧게 앉아 있어 회전에 몸을 싣는 동작이 약하다.",
        "built_space": "공통 진입로 하나와 좌우 지선 두 개, 중앙 식생 지대와 바위 하나, 왼쪽 가드레일, 우측 비닐하우스 하나 및 논밭·주택·전신주가 기준 장소를 잘 재현한다. 두 주역 트럭은 화면의 4분의 1 미만이며 왼쪽 이동 경로도 열려 있다. 그러나 도로와 운전자를 보는 높이가 기준 사진과 비슷해, 상승 크레인 끝에서 대각선 아래로 내려다보는 구도는 A보다 약하다.",
        "entities": "지붕 없는 낡은 트럭 두 대, 오른쪽의 추가 차량 두 대, 운전자 두 명이 보인다. 왼쪽 인물의 짧은 검은 머리와 어두운 셔츠는 이현우에 대응하고, 오른쪽 인물의 검은 단발머리와 올리브색 상의는 수빈에 대응한다. 뒷모습과 작은 인물 크기 때문에 정확한 얼굴·나이·민족적 외형 및 전투 상처와 피폭 흔적은 판별할 수 없다. 적재함의 가방과 장비 중 찰리나 B-200 부품으로 식별되는 것은 없다. 추가 인물이나 문자 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "차량 바퀴는 도로에 닿아 있고 적재물은 적재함에 실려 있다. 두 운전자는 운전석에 앉은 상체로 읽히며 좌석 아래 신체는 가려져 있다. 손과 조향장치의 정확한 접촉은 작아서 확인하기 어렵지만, 지지 없이 떠 있는 사람이나 물건, 불가능한 회전 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "더 높은 후방 시점에서 적재함과 갈라지는 경로를 내려다봐 크레인 종료 구도에 더 가깝지만, 후속 차량들이 수빈 뒤가 아니라 앞에 있다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "수빈의 단발머리와 장소는 잘 맞지만, 카메라가 기준 사진의 낮은 시점에 더 가깝고 후속 차량도 수빈보다 앞서 있어 지정된 구도와 행렬 관계가 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 트럭은 좌상단 도로로, 오른쪽 트럭은 우상단 도로로 향하며 두 운전자도 각각 전방을 보는 뒷모습이다. 다만 오른쪽 추가 차량들은 수빈의 트럭보다 도로 진행 방향 앞쪽에 있어 '뒤따르는 차량' 관계가 반대다. 회전 방향으로 몸을 기울이는 동작은 뚜렷하지 않다.",
        "built_space": "하단 진입로 하나가 중앙의 식생 삼각지대를 두고 두 갈래로 나뉜다. 왼쪽 가드레일, 중앙 바위 하나, 우측 배경 비닐하우스 하나, 논밭과 주택군 및 양쪽 전신주가 기준 장소와 대응한다. 두 트럭은 각각 화면의 4분의 1보다 작고 좌우 지선에 놓였다. 적재함 내부를 볼 수 있는 높은 후방 시점이며 B보다 하향 관찰이 강하지만, 좌우 갈림길을 거의 정면으로 보는 축은 남아 있다.",
        "entities": "낡고 지붕 없는 주역 트럭 두 대와 오른쪽 지선의 추가 차량들이 보인다. 왼쪽 운전자는 짧은 검은 머리와 어두운 셔츠가 이현우 설정에 부합한다. 오른쪽 운전자는 올리브색 상의를 입었지만 머리가 짧고 둥글게 보여 수빈의 선명한 단발 윤곽과는 차이가 있다. 얼굴이 돌아서 있고 작아 정확한 나이·얼굴·민족적 외형, 상처와 인이어는 확인할 수 없다. 적재함에는 짐이 보이나 찰리나 B-200 부품으로 확정할 대상은 없다. 별도의 추가 인물이나 화면 위 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "트럭들은 타이어로 도로에 지지되고, 운전자들의 상체는 운전석 위치에서 올라온다. 하체와 손의 접촉은 차체에 가려져 정밀하게 확인되지 않지만 공중에 뜬 신체는 없다. 적재물은 적재함 바닥과 측벽 안에 놓여 있으며, 회전 중 차량 자세도 물리적으로 가능하다."
       },
       {
        "label": "A",
        "direction": "왼쪽 트럭의 앞부분은 좌상단 지선을, 오른쪽 트럭은 우상단 지선을 향한다. 두 운전자의 머리도 각 도로 전방을 향하며 카메라를 돌아보지 않는다. 오른쪽의 추가 차량 두 대는 수빈보다 앞서 멀어지고 있어 후속 행렬의 진행 순서가 맞지 않는다. 두 운전자는 비교적 곧게 앉아 있어 회전에 몸을 싣는 동작이 약하다.",
        "built_space": "공통 진입로 하나와 좌우 지선 두 개, 중앙 식생 지대와 바위 하나, 왼쪽 가드레일, 우측 비닐하우스 하나 및 논밭·주택·전신주가 기준 장소를 잘 재현한다. 두 주역 트럭은 화면의 4분의 1 미만이며 왼쪽 이동 경로도 열려 있다. 그러나 도로와 운전자를 보는 높이가 기준 사진과 비슷해, 상승 크레인 끝에서 대각선 아래로 내려다보는 구도는 A보다 약하다.",
        "entities": "지붕 없는 낡은 트럭 두 대, 오른쪽의 추가 차량 두 대, 운전자 두 명이 보인다. 왼쪽 인물의 짧은 검은 머리와 어두운 셔츠는 이현우에 대응하고, 오른쪽 인물의 검은 단발머리와 올리브색 상의는 수빈에 대응한다. 뒷모습과 작은 인물 크기 때문에 정확한 얼굴·나이·민족적 외형 및 전투 상처와 피폭 흔적은 판별할 수 없다. 적재함의 가방과 장비 중 찰리나 B-200 부품으로 식별되는 것은 없다. 추가 인물이나 문자 오버레이는 보이지 않는다.",
        "hard_violations": [],
        "physics": "차량 바퀴는 도로에 닿아 있고 적재물은 적재함에 실려 있다. 두 운전자는 운전석에 앉은 상체로 읽히며 좌석 아래 신체는 가려져 있다. 손과 조향장치의 정확한 접촉은 작아서 확인하기 어렵지만, 지지 없이 떠 있는 사람이나 물건, 불가능한 회전 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "요구된 부감 카메라 구도와 배경을 정확히 구현했으며 운전자의 헤어스타일과 의상이 설정과 잘 일치하나, 화물칸에 탑승한 찰리의 디테일이 생략되었습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "배경과 분기점 구도는 훌륭하게 표현되었으나, 두 운전자의 의상 및 헤어스타일이 레퍼런스와 어긋나며 우측 트럭이 우핸들 차량으로 잘못 묘사되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_departure_road_fork_bcc32e.png",
    "asset_id": "ba0541fa-bfcd-471b-a4bf-ec26a45d05cf",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 수빈: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1001789>",
    "asset_id": "8618e5e5-c604-4da4-9033-6858f6203370",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d0a-1df1-7583-bda6-0a3dcdf71895",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh4__bgfirst_bg.png",
   "bg_asset_id": "944ba677-25c7-4539-9f9a-eb8676d9142a",
   "bg_record_key": "S68sh4::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "departure_road_fork",
   "groupbg_asset_id": "ba0541fa-bfcd-471b-a4bf-ec26a45d05cf"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S68sh8::signage": {
  "fp": "4f391b678780ebdd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::travel_truck_bed": {
  "input_fingerprint": "795e823d9a856a4e",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "travel_truck_bed",
    "tags": [
     "S68sh8",
     "S69sh2",
     "S69sh5",
     "S71sh6"
    ]
   },
   "context_sig": "a87ee34f780c0e65"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘) / 이현우의 트럭 적재함: 짐을 싣는 뒷공간으로 사람들이 이불을 덮고 탈 수 있다. (특징: 짐칸 모서리에 기대앉은 찰리와 배낭 더미) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 짐칸에 앉은 찰리 B-200의 가슴 부품을 꺼내본다.\n- 바위틈을 그늘 삼아 현우 차가 세워져있다. 뒤 짐칸에 현우와 앰버 라울 셋이 머릴 맞대고 자고 있다.\n- (짐칸에 앉은 찰리에게) 앞으로 30분 안에 해남에 도착해.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘) / 이현우의 트럭 적재함: 짐을 싣는 뒷공간으로 사람들이 이불을 덮고 탈 수 있다. (특징: 짐칸 모서리에 기대앉은 찰리와 배낭 더미) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 짐칸에 앉은 찰리 B-200의 가슴 부품을 꺼내본다.\n- 바위틈을 그늘 삼아 현우 차가 세워져있다. 뒤 짐칸에 현우와 앰버 라울 셋이 머릴 맞대고 자고 있다.\n- (짐칸에 앉은 찰리에게) 앞으로 30분 안에 해남에 도착해.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_travel_truck_bed_0f596c.png",
  "asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67",
  "input_asset_ids": [
   "6ed020ac-24c2-4481-bbcd-027c83f6b30b"
  ],
  "origin_tag": "S68sh8",
  "place_text": "On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.",
  "origin_inputs": {
   "place_text": "On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.",
   "time_of_day_en": "day",
   "conti_asset_id": "6ed020ac-24c2-4481-bbcd-027c83f6b30b"
  }
 },
 "S68sh8::bgfirst_bg": {
  "input_fingerprint": "f9e40fba49c0c46c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공에서 날아온 작은 아기 새가 찰리의 굵은 금속 손가락 끝에 사뿐히 내려앉은 근접 찰나.\n\nLOCATION (lock): On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 찰리 inside the cargo bed, holding the established close lateral position slightly above the raised finger and looking obliquely down, with open arrival space above it. Catch 아기 새 just as its feet settle and its wings begin folding near center; 찰리's metal finger enters from the lower-left, while a softly resolved portion of his seated torso remains behind as a scale reference. Keep 찰리's head outside the crop, his attention lowered toward the bird, and the bird's head inclined toward its new perch; focus the landing without enlarging the finger beyond its natural relation to his body.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck cargo bed (Carrying the seated 찰리 during travel) — A partial inner edge is visible behind the hand and torso; used as Maintains the real traveling setting beneath the shallow-focus landing detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the tiny bird and the precisely rendered metal finger gently distinct without exaggerated shine.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공에서 날아온 작은 아기 새가 찰리의 굵은 금속 손가락 끝에 사뿐히 내려앉은 근접 찰나.\n\nLOCATION (lock): On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 찰리 inside the cargo bed, holding the established close lateral position slightly above the raised finger and looking obliquely down, with open arrival space above it. Catch 아기 새 just as its feet settle and its wings begin folding near center; 찰리's metal finger enters from the lower-left, while a softly resolved portion of his seated torso remains behind as a scale reference. Keep 찰리's head outside the crop, his attention lowered toward the bird, and the bird's head inclined toward its new perch; focus the landing without enlarging the finger beyond its natural relation to his body.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck cargo bed (Carrying the seated 찰리 during travel) — A partial inner edge is visible behind the hand and torso; used as Maintains the real traveling setting beneath the shallow-focus landing detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the tiny bird and the precisely rendered metal finger gently distinct without exaggerated shine.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh8__bgfirst_bg.png",
  "asset_id": "6495be78-8a5d-4e57-8242-99d4994cf7be",
  "input_asset_ids": [
   "6ed020ac-24c2-4481-bbcd-027c83f6b30b",
   "71e106e5-8061-49a4-96c1-92ad1c3d2c67"
  ]
 },
 "S68sh8": {
  "input_fingerprint": "5226cbd9a225cf4e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 작은 아기 새가 찰리의 굵은 금속 손가락 끝에 사뿐히 내려앉은 근접 찰나.\n\nLOCATION (lock): On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 찰리 inside the cargo bed, holding the established close lateral position slightly above the raised finger and looking obliquely down, with open arrival space above it. Catch 아기 새 just as its feet settle and its wings begin folding near center; 찰리's metal finger enters from the lower-left, while a softly resolved portion of his seated torso remains behind as a scale reference. Keep 찰리's head outside the crop, his attention lowered toward the bird, and the bird's head inclined toward its new perch; focus the landing without enlarging the finger beyond its natural relation to his body.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck cargo bed (Carrying the seated 찰리 during travel) — A partial inner edge is visible behind the hand and torso; used as Maintains the real traveling setting beneath the shallow-focus landing detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the tiny bird and the precisely rendered metal finger gently distinct without exaggerated shine.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck continues along the daytime road, with battle-damaged Charlie seated in its cargo bed and retaining B-200's chest component. A baby bird has landed on Charlie's finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 작은 아기 새가 찰리의 굵은 금속 손가락 끝에 사뿐히 내려앉은 근접 찰나.\n\nLOCATION (lock): On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 찰리 inside the cargo bed, holding the established close lateral position slightly above the raised finger and looking obliquely down, with open arrival space above it. Catch 아기 새 just as its feet settle and its wings begin folding near center; 찰리's metal finger enters from the lower-left, while a softly resolved portion of his seated torso remains behind as a scale reference. Keep 찰리's head outside the crop, his attention lowered toward the bird, and the bird's head inclined toward its new perch; focus the landing without enlarging the finger beyond its natural relation to his body.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck cargo bed (Carrying the seated 찰리 during travel) — A partial inner edge is visible behind the hand and torso; used as Maintains the real traveling setting beneath the shallow-focus landing detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the tiny bird and the precisely rendered metal finger gently distinct without exaggerated shine.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck continues along the daytime road, with battle-damaged Charlie seated in its cargo bed and retaining B-200's chest component. A baby bird has landed on Charlie's finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 작은 아기 새가 찰리의 굵은 금속 손가락 끝에 사뿐히 내려앉은 근접 찰나.\n\nLOCATION (lock): On the open rear load bed of the moving truck, exposed to daylight and the surrounding countryside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track beside 찰리 inside the cargo bed, holding the established close lateral position slightly above the raised finger and looking obliquely down, with open arrival space above it. Catch 아기 새 just as its feet settle and its wings begin folding near center; 찰리's metal finger enters from the lower-left, while a softly resolved portion of his seated torso remains behind as a scale reference. Keep 찰리's head outside the crop, his attention lowered toward the bird, and the bird's head inclined toward its new perch; focus the landing without enlarging the finger beyond its natural relation to his body.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Truck cargo bed (Carrying the seated 찰리 during travel) — A partial inner edge is visible behind the hand and torso; used as Maintains the real traveling setting beneath the shallow-focus landing detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the tiny bird and the precisely rendered metal finger gently distinct without exaggerated shine.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck continues along the daytime road, with battle-damaged Charlie seated in its cargo bed and retaining B-200's chest component. A baby bird has landed on Charlie's finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh8__bgfirst_bg.png",
     "asset_id": "6495be78-8a5d-4e57-8242-99d4994cf7be",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S68sh8.png",
     "asset_id": "6ed020ac-24c2-4481-bbcd-027c83f6b30b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 아기 새: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761765>",
     "asset_id": "4f173cfe-d96c-4d9d-9779-afb9d9a8dd4b",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_travel_truck_bed_0f596c.png",
     "asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 아기 새: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761765>",
     "asset_id": "4f173cfe-d96c-4d9d-9779-afb9d9a8dd4b",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "새의 부리는 왼쪽 아래를 향하고 있으며, 찰리의 뻗은 손가락은 오른쪽을 향하고 있음.",
    "built_space": "트럭 적재함 내부 가장자리 철제 벽면이 보이고, 배경으로는 위치 레퍼런스와 일치하는 도로와 가드레일이 보임.",
    "entities": "찰리(가슴의 푸른 원자로, 베이지색 장갑판 토르소 및 팔), 아기 새(작은 부리와 펼친 날개, 둥근 몸체).",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (찰리의 뻗은 손가락 아래로 접힌 손가락이 4개 이상 노출되어 해부학적 구조가 붕괴됨)"
    ],
    "physics": "아기 새의 발이 찰리의 뻗은 손가락에 닿아 지지받고 있으며, 찰리의 팔은 프레임 밖 하단의 몸체에 의해 지탱됨."
   },
   {
    "label": "B",
    "direction": "새는 왼쪽을 향해 시선을 두며 내려앉고 있고, 찰리의 손가락은 오른쪽으로 뻗어 있음.",
    "built_space": "트럭 적재함 내부 측면이 근경에 위치하며, 뒤로는 초점이 흐려진 자연 풍경(나무와 강)이 보임.",
    "entities": "찰리(베이지색 장갑판, 푸른 원자로, 지시와 달리 턱 부분의 하얀 마스크가 살짝 보임), 아기 새(가느다란 발가락과 작은 날개).",
    "hard_violations": [],
    "physics": "새의 두 발이 찰리의 손가락 끝에 닿아 안착하고 있으며, 배경 쪽에 찰리의 다리 부품이 바닥에 닿아 자세를 지지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍 지시와 달리 찰리의 머리 하단부가 화면에 노출되었으나, 아기 새의 착지 순간과 금속 손가락의 질감을 전반적으로 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 손가락 개수가 정상적인 구조를 벗어나 6개처럼 렌더링되어 해부학적으로 불가능한 치명적 위반(Hard Violation)에 해당합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "새의 부리는 왼쪽 아래를 향하고 있으며, 찰리의 뻗은 손가락은 오른쪽을 향하고 있음.",
        "built_space": "트럭 적재함 내부 가장자리 철제 벽면이 보이고, 배경으로는 위치 레퍼런스와 일치하는 도로와 가드레일이 보임.",
        "entities": "찰리(가슴의 푸른 원자로, 베이지색 장갑판 토르소 및 팔), 아기 새(작은 부리와 펼친 날개, 둥근 몸체).",
        "hard_violations": [
         "physically impossible anatomy (찰리의 뻗은 손가락 아래로 접힌 손가락이 4개 이상 노출되어 해부학적 구조가 붕괴됨)"
        ],
        "physics": "아기 새의 발이 찰리의 뻗은 손가락에 닿아 지지받고 있으며, 찰리의 팔은 프레임 밖 하단의 몸체에 의해 지탱됨."
       },
       {
        "label": "B",
        "direction": "새는 왼쪽을 향해 시선을 두며 내려앉고 있고, 찰리의 손가락은 오른쪽으로 뻗어 있음.",
        "built_space": "트럭 적재함 내부 측면이 근경에 위치하며, 뒤로는 초점이 흐려진 자연 풍경(나무와 강)이 보임.",
        "entities": "찰리(베이지색 장갑판, 푸른 원자로, 지시와 달리 턱 부분의 하얀 마스크가 살짝 보임), 아기 새(가느다란 발가락과 작은 날개).",
        "hard_violations": [],
        "physics": "새의 두 발이 찰리의 손가락 끝에 닿아 안착하고 있으며, 배경 쪽에 찰리의 다리 부품이 바닥에 닿아 자세를 지지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레이밍 지시와 달리 찰리의 머리 하단부가 화면에 노출되었으나, 아기 새의 착지 순간과 금속 손가락의 질감을 전반적으로 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "찰리의 손가락 개수가 정상적인 구조를 벗어나 6개처럼 렌더링되어 해부학적으로 불가능한 치명적 위반(Hard Violation)에 해당합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "새의 부리는 왼쪽 아래를 향하고 있으며, 찰리의 뻗은 손가락은 오른쪽을 향하고 있음.",
        "built_space": "트럭 적재함 내부 가장자리 철제 벽면이 보이고, 배경으로는 위치 레퍼런스와 일치하는 도로와 가드레일이 보임.",
        "entities": "찰리(가슴의 푸른 원자로, 베이지색 장갑판 토르소 및 팔), 아기 새(작은 부리와 펼친 날개, 둥근 몸체).",
        "hard_violations": [
         "physically impossible anatomy (찰리의 뻗은 손가락 아래로 접힌 손가락이 4개 이상 노출되어 해부학적 구조가 붕괴됨)"
        ],
        "physics": "아기 새의 발이 찰리의 뻗은 손가락에 닿아 지지받고 있으며, 찰리의 팔은 프레임 밖 하단의 몸체에 의해 지탱됨."
       },
       {
        "label": "B",
        "direction": "새는 왼쪽을 향해 시선을 두며 내려앉고 있고, 찰리의 손가락은 오른쪽으로 뻗어 있음.",
        "built_space": "트럭 적재함 내부 측면이 근경에 위치하며, 뒤로는 초점이 흐려진 자연 풍경(나무와 강)이 보임.",
        "entities": "찰리(베이지색 장갑판, 푸른 원자로, 지시와 달리 턱 부분의 하얀 마스크가 살짝 보임), 아기 새(가느다란 발가락과 작은 날개).",
        "hard_violations": [],
        "physics": "새의 두 발이 찰리의 손가락 끝에 닿아 안착하고 있으며, 배경 쪽에 찰리의 다리 부품이 바닥에 닿아 자세를 지지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "착지에 집중한 근접 구도와 흐린 몸통은 좋지만, 새가 손가락 끝보다 안쪽 마디에 앉고 손이 과대하게 강조되며 흰 얼굴 일부까지 프레임에 들어온다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "몸통 대비 자연스러운 금속 손가락의 끝에 새가 발을 붙이는 순간과 머리 제외 구도를 더 충실히 구현하지만, 배경과 몸통은 요구보다 넓고 선명하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "새의 부리와 머리는 왼쪽 아래의 금속 손가락을 향한다. 손가락은 왼쪽 아래에서 중앙 오른쪽으로 뻗어 있으나 새의 발은 끝부분보다 안쪽 마디 위에 놓인다. 찰리의 눈은 잘려 있어 새를 내려다보는 시선은 확인할 수 없고, 흰 마스크의 입과 턱은 상단에 보인다.",
        "built_space": "오른쪽 배경에 녹슨 밝은색 적재함 내벽 한 면과 상단 테두리, 세로 보강대 두 곳이 보이며 아래에는 어두운 바닥이 드러난다. 찰리의 몸통은 손 뒤쪽 왼편, 하체 일부는 적재함 안쪽에 있다. 참고 장소의 낡은 금속 재질과 낮의 산·수면 배경에 부합한다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 대와 갈색 아기 새 한 마리가 보인다. 찰리의 마모된 샌드 베이지 장갑, 굵은 기계 손가락, 푸른 원형 가슴 원자로와 부분적으로 보이는 흰 마스크는 참고와 일치한다. 새의 둥근 솜털 몸, 짧은 부리, 가는 발가락도 일치한다. 다만 전경의 손이 몸통에 비해 크게 강조된다. 별도로 보유한 가슴 부품은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "새의 두 발과 발톱이 금속 손가락 윗면에 접촉해 몸을 지탱한다. 펼쳐진 날개와 굽힌 다리는 착지 직후 균형을 잡는 동작으로 가능하지만 날개를 접기 시작했는지는 뚜렷하지 않다. 손은 손목과 팔에 연결되어 있어 부유하지 않는다. 찰리의 엉덩이 접촉점은 가려져 있지만 적재함 안의 하체가 이어져 보이며, 공중에 떠 있다는 징후는 없다."
       },
       {
        "label": "B",
        "direction": "새의 머리와 부리는 왼쪽 아래의 새 횃대를 향해 기울어져 있다. 왼쪽 아래에서 들어온 금속 손가락이 중앙으로 뻗고 새의 발은 마지막 마디 끝부분을 붙잡는다. 찰리의 얼굴과 눈은 화면 밖이어서 시선 자체는 확인할 수 없다. 새 위에는 날아들 수 있는 빈 공간이 충분하다.",
        "built_space": "뒤쪽에 녹슨 적재함 후면 내벽 한 면, 오른쪽에 측벽 한 면, 아래에 바닥 일부가 보인다. 후면에는 세로 보강대 두 곳과 오른쪽 모서리 연결부가 드러난다. 찰리는 적재함 안 왼쪽을 차지한다. 뒤로 이어지는 도로, 오른쪽 전신주와 가드레일, 산과 수면은 참고 장소에 잘 맞는다. 다만 부분 테두리만 은근하게 남기라는 요구보다 적재함과 풍경이 넓고 선명하다.",
        "entities": "찰리 한 대와 작은 갈색 아기 새 한 마리가 있다. 찰리의 각진 베이지 장갑과 마모, 기계 관절, 푸른 가슴 원자로가 참고 정체성을 유지하며 손과 몸통의 크기 관계도 비교적 자연스럽다. 새는 짧은 부리와 둥근 어린 몸, 가는 발가락을 갖춘다. 얼굴은 크롭 밖이며 별도 가슴 부품의 보유 여부는 보이는 부분만으로 확인하기 어렵다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "새는 양발로 마지막 손가락 마디를 잡아 지지받고 있으며, 들어 올린 날개는 막 내려앉아 균형을 잡는 순간으로 성립한다. 다만 날개가 이미 접히기 시작했다는 표현은 약하다. 금속 손가락은 손바닥·손목·팔로 연결되어 있다. 아래쪽 하체 일부는 적재함 내부에 놓여 있고 정확한 착석 접촉점은 크롭에 가려진다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "착지에 집중한 근접 구도와 흐린 몸통은 좋지만, 새가 손가락 끝보다 안쪽 마디에 앉고 손이 과대하게 강조되며 흰 얼굴 일부까지 프레임에 들어온다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "몸통 대비 자연스러운 금속 손가락의 끝에 새가 발을 붙이는 순간과 머리 제외 구도를 더 충실히 구현하지만, 배경과 몸통은 요구보다 넓고 선명하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "새의 부리와 머리는 왼쪽 아래의 금속 손가락을 향한다. 손가락은 왼쪽 아래에서 중앙 오른쪽으로 뻗어 있으나 새의 발은 끝부분보다 안쪽 마디 위에 놓인다. 찰리의 눈은 잘려 있어 새를 내려다보는 시선은 확인할 수 없고, 흰 마스크의 입과 턱은 상단에 보인다.",
        "built_space": "오른쪽 배경에 녹슨 밝은색 적재함 내벽 한 면과 상단 테두리, 세로 보강대 두 곳이 보이며 아래에는 어두운 바닥이 드러난다. 찰리의 몸통은 손 뒤쪽 왼편, 하체 일부는 적재함 안쪽에 있다. 참고 장소의 낡은 금속 재질과 낮의 산·수면 배경에 부합한다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 대와 갈색 아기 새 한 마리가 보인다. 찰리의 마모된 샌드 베이지 장갑, 굵은 기계 손가락, 푸른 원형 가슴 원자로와 부분적으로 보이는 흰 마스크는 참고와 일치한다. 새의 둥근 솜털 몸, 짧은 부리, 가는 발가락도 일치한다. 다만 전경의 손이 몸통에 비해 크게 강조된다. 별도로 보유한 가슴 부품은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "새의 두 발과 발톱이 금속 손가락 윗면에 접촉해 몸을 지탱한다. 펼쳐진 날개와 굽힌 다리는 착지 직후 균형을 잡는 동작으로 가능하지만 날개를 접기 시작했는지는 뚜렷하지 않다. 손은 손목과 팔에 연결되어 있어 부유하지 않는다. 찰리의 엉덩이 접촉점은 가려져 있지만 적재함 안의 하체가 이어져 보이며, 공중에 떠 있다는 징후는 없다."
       },
       {
        "label": "A",
        "direction": "새의 머리와 부리는 왼쪽 아래의 새 횃대를 향해 기울어져 있다. 왼쪽 아래에서 들어온 금속 손가락이 중앙으로 뻗고 새의 발은 마지막 마디 끝부분을 붙잡는다. 찰리의 얼굴과 눈은 화면 밖이어서 시선 자체는 확인할 수 없다. 새 위에는 날아들 수 있는 빈 공간이 충분하다.",
        "built_space": "뒤쪽에 녹슨 적재함 후면 내벽 한 면, 오른쪽에 측벽 한 면, 아래에 바닥 일부가 보인다. 후면에는 세로 보강대 두 곳과 오른쪽 모서리 연결부가 드러난다. 찰리는 적재함 안 왼쪽을 차지한다. 뒤로 이어지는 도로, 오른쪽 전신주와 가드레일, 산과 수면은 참고 장소에 잘 맞는다. 다만 부분 테두리만 은근하게 남기라는 요구보다 적재함과 풍경이 넓고 선명하다.",
        "entities": "찰리 한 대와 작은 갈색 아기 새 한 마리가 있다. 찰리의 각진 베이지 장갑과 마모, 기계 관절, 푸른 가슴 원자로가 참고 정체성을 유지하며 손과 몸통의 크기 관계도 비교적 자연스럽다. 새는 짧은 부리와 둥근 어린 몸, 가는 발가락을 갖춘다. 얼굴은 크롭 밖이며 별도 가슴 부품의 보유 여부는 보이는 부분만으로 확인하기 어렵다. 추가 인물이나 문구는 없다.",
        "hard_violations": [],
        "physics": "새는 양발로 마지막 손가락 마디를 잡아 지지받고 있으며, 들어 올린 날개는 막 내려앉아 균형을 잡는 순간으로 성립한다. 다만 날개가 이미 접히기 시작했다는 표현은 약하다. 금속 손가락은 손바닥·손목·팔로 연결되어 있다. 아래쪽 하체 일부는 적재함 내부에 놓여 있고 정확한 착석 접촉점은 크롭에 가려진다. 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (찰리의 뻗은 손가락 아래로 접힌 손가락이 4개 이상 노출되어 해부학적 구조가 붕괴됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "프레이밍 지시와 달리 찰리의 머리 하단부가 화면에 노출되었으나, 아기 새의 착지 순간과 금속 손가락의 질감을 전반적으로 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "찰리의 손가락 개수가 정상적인 구조를 벗어나 6개처럼 렌더링되어 해부학적으로 불가능한 치명적 위반(Hard Violation)에 해당합니다.  ★위반: [gemini-pro] physically impossible anatomy (찰리의 뻗은 손가락 아래로 접힌 손가락이 4개 이상 노출되어 해부학적 구조가 붕괴됨)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_travel_truck_bed_0f596c.png",
    "asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 아기 새: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761765>",
    "asset_id": "4f173cfe-d96c-4d9d-9779-afb9d9a8dd4b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d14-76c7-7b13-9d98-bdcc47283caa",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh8__bgfirst_bg.png",
   "bg_asset_id": "6495be78-8a5d-4e57-8242-99d4994cf7be",
   "bg_record_key": "S68sh8::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "travel_truck_bed",
   "groupbg_asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S68sh20::signage": {
  "fp": "fe6117f5b09335c5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::violet_field": {
  "input_fingerprint": "5eefa25c0550b8b2",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "violet_field",
    "tags": [
     "S68sh20",
     "S71sh28"
    ]
   },
   "context_sig": "8680b8e5ac741c24"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In an imagined sunlit meadow of purple violets near a distant research building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 보라색 제비꽃이 만방에 피었고, 저 멀리 보이는 지동현의 해남 연구소.\n- 엄청나게 많은 제비꽃이 화려하게 피어있고, 차에서 내려 내려다보는 아름다운 제비꽃들의 향연. 마치 천국과도 같은. 저 멀리 우뚝 솟아있는 지동현의 연구소도 보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In an imagined sunlit meadow of purple violets near a distant research building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘) / 찰리가 홀로 걷는 지방도로·캠핑카 수색 도로, 지방도로에서 갈라지는 숲길, 익산 지방도로·갈림길: 외곽의 한산한 아스팔트 도로와 숲으로 이어지는 흙길이다. (특징: 커다란 밀짚모자와 고무장화, 알록달록한 우비를 걸치고 걷는 찰리의 실루엣; 빗방울이 떨어지는 흙길 웅덩이와 자동차 바퀴 자국; 꽃 위에 앉았다가 날아오르는 생물 나비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 보라색 제비꽃이 만방에 피었고, 저 멀리 보이는 지동현의 해남 연구소.\n- 엄청나게 많은 제비꽃이 화려하게 피어있고, 차에서 내려 내려다보는 아름다운 제비꽃들의 향연. 마치 천국과도 같은. 저 멀리 우뚝 솟아있는 지동현의 연구소도 보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_violet_field_621f9d.png",
  "asset_id": "bba38d59-2419-4fea-98d2-3d47523d7b7c",
  "input_asset_ids": [
   "ca3ec43a-0832-4cb9-bb1a-f51d83efa39b"
  ],
  "origin_tag": "S68sh20",
  "place_text": "In an imagined sunlit meadow of purple violets near a distant research building.",
  "origin_inputs": {
   "place_text": "In an imagined sunlit meadow of purple violets near a distant research building.",
   "time_of_day_en": "day",
   "conti_asset_id": "ca3ec43a-0832-4cb9-bb1a-f51d83efa39b"
  }
 },
 "S68sh20::bgfirst_bg": {
  "input_fingerprint": "9ce8c4578cf9b6ef",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 의식 속 환상, 찰리의 거대한 기계 손이 지소영의 작은 손을 맞잡은 근접 찰나.\n\nLOCATION (lock): In an imagined sunlit meadow of purple violets near a distant research building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 찰리's dream, finish the close dolly beside the space between the two bodies, looking diagonally down across their hands from just above hand height rather than along either frontal axis. 찰리's larger hand enters from the left and closes gently around 지소영's smaller hand from the right, with their forearms and cropped body edges retaining scale and purple flowers visible beyond the joined grip. Their heads remain outside the frame as their bodies begin orienting toward the off-screen caller, 지동현; hold the completed contact before the camera retreats behind their movement.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Purple violets (Blooming and moving in the breeze within the dream); used as A softly resolved surrounding field keeps the clasp situated in the imagined place of friendship.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the dream's daylight soft and restrained, allowing the established purple flowers to carry color without artificial glow or visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 의식 속 환상, 찰리의 거대한 기계 손이 지소영의 작은 손을 맞잡은 근접 찰나.\n\nLOCATION (lock): In an imagined sunlit meadow of purple violets near a distant research building.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 찰리's dream, finish the close dolly beside the space between the two bodies, looking diagonally down across their hands from just above hand height rather than along either frontal axis. 찰리's larger hand enters from the left and closes gently around 지소영's smaller hand from the right, with their forearms and cropped body edges retaining scale and purple flowers visible beyond the joined grip. Their heads remain outside the frame as their bodies begin orienting toward the off-screen caller, 지동현; hold the completed contact before the camera retreats behind their movement.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Purple violets (Blooming and moving in the breeze within the dream); used as A softly resolved surrounding field keeps the clasp situated in the imagined place of friendship.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the dream's daylight soft and restrained, allowing the established purple flowers to carry color without artificial glow or visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh20__bgfirst_bg.png",
  "asset_id": "b01ffceb-46c9-43e6-99b0-7f2e082004a2",
  "input_asset_ids": [
   "ca3ec43a-0832-4cb9-bb1a-f51d83efa39b",
   "bba38d59-2419-4fea-98d2-3d47523d7b7c"
  ]
 },
 "S68sh20": {
  "input_fingerprint": "0f21168ba9bacbd6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의식 속 환상, 찰리의 거대한 기계 손이 지소영의 작은 손을 맞잡은 근접 찰나.\n\nLOCATION (lock): In an imagined sunlit meadow of purple violets near a distant research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 찰리's dream, finish the close dolly beside the space between the two bodies, looking diagonally down across their hands from just above hand height rather than along either frontal axis. 찰리's larger hand enters from the left and closes gently around 지소영's smaller hand from the right, with their forearms and cropped body edges retaining scale and purple flowers visible beyond the joined grip. Their heads remain outside the frame as their bodies begin orienting toward the off-screen caller, 지동현; hold the completed contact before the camera retreats behind their movement.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Purple violets (Blooming and moving in the breeze within the dream); used as A softly resolved surrounding field keeps the clasp situated in the imagined place of friendship.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the dream's daylight soft and restrained, allowing the established purple flowers to carry color without artificial glow or visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This imagined daytime Haenam landscape is filled with purple violets moving in a gentle spring breeze, with the research institute visible in the distance. B-200 is present in the imagined scene, without the destruction carried by the real-world remains. 지소영: She appears as her childhood self in the imagined flower field, with a hand extended for the departure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의식 속 환상, 찰리의 거대한 기계 손이 지소영의 작은 손을 맞잡은 근접 찰나.\n\nLOCATION (lock): In an imagined sunlit meadow of purple violets near a distant research building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 찰리's dream, finish the close dolly beside the space between the two bodies, looking diagonally down across their hands from just above hand height rather than along either frontal axis. 찰리's larger hand enters from the left and closes gently around 지소영's smaller hand from the right, with their forearms and cropped body edges retaining scale and purple flowers visible beyond the joined grip. Their heads remain outside the frame as their bodies begin orienting toward the off-screen caller, 지동현; hold the completed contact before the camera retreats behind their movement.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Purple violets (Blooming and moving in the breeze within the dream); used as A softly resolved surrounding field keeps the clasp situated in the imagined place of friendship.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the dream's daylight soft and restrained, allowing the established purple flowers to carry color without artificial glow or visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This imagined daytime Haenam landscape is filled with purple violets moving in a gentle spring breeze, with the research institute visible in the distance. B-200 is present in the imagined scene, without the destruction carried by the real-world remains. 지소영: She appears as her childhood self in the imagined flower field, with a hand extended for the departure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의식 속 환상, 찰리의 거대한 기계 손이 지소영의 작은 손을 맞잡은 근접 찰나.\n\nLOCATION (lock): In an imagined sunlit meadow of purple violets near a distant research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Within 찰리's dream, finish the close dolly beside the space between the two bodies, looking diagonally down across their hands from just above hand height rather than along either frontal axis. 찰리's larger hand enters from the left and closes gently around 지소영's smaller hand from the right, with their forearms and cropped body edges retaining scale and purple flowers visible beyond the joined grip. Their heads remain outside the frame as their bodies begin orienting toward the off-screen caller, 지동현; hold the completed contact before the camera retreats behind their movement.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Purple violets (Blooming and moving in the breeze within the dream); used as A softly resolved surrounding field keeps the clasp situated in the imagined place of friendship.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the dream's daylight soft and restrained, allowing the established purple flowers to carry color without artificial glow or visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): This imagined daytime Haenam landscape is filled with purple violets moving in a gentle spring breeze, with the research institute visible in the distance. B-200 is present in the imagined scene, without the destruction carried by the real-world remains. 지소영: She appears as her childhood self in the imagined flower field, with a hand extended for the departure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh20__bgfirst_bg.png",
     "asset_id": "b01ffceb-46c9-43e6-99b0-7f2e082004a2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S68sh20.png",
     "asset_id": "ca3ec43a-0832-4cb9-bb1a-f51d83efa39b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1237799>",
     "asset_id": "3674e762-acc5-4d71-a06e-9b4c0f9b740c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_violet_field_621f9d.png",
     "asset_id": "bba38d59-2419-4fea-98d2-3d47523d7b7c",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1237799>",
     "asset_id": "3674e762-acc5-4d71-a06e-9b4c0f9b740c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 두 인물의 손 위에서 대각선 아래를 향해 내려다보고 있으며, 왼쪽에서 나타난 거대한 기계 손이 오른쪽에서 나타난 사람의 손을 감싸 쥐고 있습니다.",
    "built_space": "프롬프트의 배경인 보라색 제비꽃이 만발한 들판이 펼쳐져 있고, 화면 우측 원경에 위치 레퍼런스와 동일한 연구소 건물이 보입니다.",
    "entities": "왼쪽의 팔은 샌드 베이지색 장갑판을 두른 찰리의 기계 팔이며, 오른쪽의 팔은 짙은 남색 소매를 입은 지소영의 팔입니다. 두 인물의 머리는 화면 밖에 위치합니다.",
    "hard_violations": [],
    "physics": "기계 손의 손가락과 엄지가 사람의 손을 물리적으로 자연스럽게 쥐고 있으며, 팔의 무게와 자세가 안정적으로 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 손을 향해 대각선 아래로 향하고 있으며, 왼쪽의 기계 팔과 오른쪽의 사람 팔이 화면 중앙에서 만나 손을 잡고 있습니다.",
    "built_space": "보라색 꽃밭이 넓게 펼쳐져 있으며, 우측 멀리 레퍼런스와 일치하는 연구소 건물이 배치되어 있습니다.",
    "entities": "왼쪽에는 푸른색 원자로가 빛나는 찰리의 몸체 일부와 팔이 있고, 오른쪽에는 짙은 남색 스웨터를 입은 지소영의 팔이 있습니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 구조 (기계 손의 구조가 엄지와 손바닥의 구분이 없이 위아래 양쪽에서 관절이 뻗어나와 비정상적인 집게 형태를 띰)"
    ],
    "physics": "오른쪽 사람의 팔과 자세는 안정적이나, 왼쪽 기계 손의 파지 형태가 구조적으로 성립하지 않는 기형적인 관절 배치를 보입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 카메라 앵글과 인물의 신체 일부분, 보라색 꽃밭과 멀리 보이는 연구소 배경까지 프롬프트의 지시를 충실하고 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 의상과 배경은 참조 이미지와 일치하나, 로봇 손이 엄지와 손바닥 없이 위아래로 둥글게 이어지는 불가능한 구조로 렌더링되어 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 두 인물의 손 위에서 대각선 아래를 향해 내려다보고 있으며, 왼쪽에서 나타난 거대한 기계 손이 오른쪽에서 나타난 사람의 손을 감싸 쥐고 있습니다.",
        "built_space": "프롬프트의 배경인 보라색 제비꽃이 만발한 들판이 펼쳐져 있고, 화면 우측 원경에 위치 레퍼런스와 동일한 연구소 건물이 보입니다.",
        "entities": "왼쪽의 팔은 샌드 베이지색 장갑판을 두른 찰리의 기계 팔이며, 오른쪽의 팔은 짙은 남색 소매를 입은 지소영의 팔입니다. 두 인물의 머리는 화면 밖에 위치합니다.",
        "hard_violations": [],
        "physics": "기계 손의 손가락과 엄지가 사람의 손을 물리적으로 자연스럽게 쥐고 있으며, 팔의 무게와 자세가 안정적으로 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 손을 향해 대각선 아래로 향하고 있으며, 왼쪽의 기계 팔과 오른쪽의 사람 팔이 화면 중앙에서 만나 손을 잡고 있습니다.",
        "built_space": "보라색 꽃밭이 넓게 펼쳐져 있으며, 우측 멀리 레퍼런스와 일치하는 연구소 건물이 배치되어 있습니다.",
        "entities": "왼쪽에는 푸른색 원자로가 빛나는 찰리의 몸체 일부와 팔이 있고, 오른쪽에는 짙은 남색 스웨터를 입은 지소영의 팔이 있습니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 구조 (기계 손의 구조가 엄지와 손바닥의 구분이 없이 위아래 양쪽에서 관절이 뻗어나와 비정상적인 집게 형태를 띰)"
        ],
        "physics": "오른쪽 사람의 팔과 자세는 안정적이나, 왼쪽 기계 손의 파지 형태가 구조적으로 성립하지 않는 기형적인 관절 배치를 보입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 카메라 앵글과 인물의 신체 일부분, 보라색 꽃밭과 멀리 보이는 연구소 배경까지 프롬프트의 지시를 충실하고 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 의상과 배경은 참조 이미지와 일치하나, 로봇 손이 엄지와 손바닥 없이 위아래로 둥글게 이어지는 불가능한 구조로 렌더링되어 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 두 인물의 손 위에서 대각선 아래를 향해 내려다보고 있으며, 왼쪽에서 나타난 거대한 기계 손이 오른쪽에서 나타난 사람의 손을 감싸 쥐고 있습니다.",
        "built_space": "프롬프트의 배경인 보라색 제비꽃이 만발한 들판이 펼쳐져 있고, 화면 우측 원경에 위치 레퍼런스와 동일한 연구소 건물이 보입니다.",
        "entities": "왼쪽의 팔은 샌드 베이지색 장갑판을 두른 찰리의 기계 팔이며, 오른쪽의 팔은 짙은 남색 소매를 입은 지소영의 팔입니다. 두 인물의 머리는 화면 밖에 위치합니다.",
        "hard_violations": [],
        "physics": "기계 손의 손가락과 엄지가 사람의 손을 물리적으로 자연스럽게 쥐고 있으며, 팔의 무게와 자세가 안정적으로 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 손을 향해 대각선 아래로 향하고 있으며, 왼쪽의 기계 팔과 오른쪽의 사람 팔이 화면 중앙에서 만나 손을 잡고 있습니다.",
        "built_space": "보라색 꽃밭이 넓게 펼쳐져 있으며, 우측 멀리 레퍼런스와 일치하는 연구소 건물이 배치되어 있습니다.",
        "entities": "왼쪽에는 푸른색 원자로가 빛나는 찰리의 몸체 일부와 팔이 있고, 오른쪽에는 짙은 남색 스웨터를 입은 지소영의 팔이 있습니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 구조 (기계 손의 구조가 엄지와 손바닥의 구분이 없이 위아래 양쪽에서 관절이 뻗어나와 비정상적인 집게 형태를 띰)"
        ],
        "physics": "오른쪽 사람의 팔과 자세는 안정적이나, 왼쪽 기계 손의 파지 형태가 구조적으로 성립하지 않는 기형적인 관절 배치를 보입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "맞잡은 손을 중심으로 비스듬히 내려다보는 근접 구도가 더 충실하지만, 지소영의 손은 지정된 어린 시절보다 성인의 손으로 보인다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "좌우에서 이어지는 손잡기는 맞지만, 하늘이 크게 보이는 낮은 시점이 지정된 내려다보기와 다르고 지소영의 손도 성인으로 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 팔이 왼쪽 위에서 중앙으로 내려오고, 지소영의 손이 오른쪽에서 왼쪽으로 뻗어 기계 손 안에 들어간다. 기계 손가락은 실제로 그 손을 둘러싸며 접촉이 완성되어 있다. 머리는 모두 화면 밖이므로 시선과 지동현 쪽으로 돌아서는 방향은 확인할 수 없다.",
        "built_space": "야외 보라색 꽃밭 너머 오른쪽 위에 낮은 연구동 한 덩어리와 수직 탑 하나가 보이고, 뒤로 산줄기가 이어져 장소 참고와 부합한다. 건물은 작고 흐리게 유지된다. 양쪽 몸통 가장자리와 전완 사이에 맞잡은 손이 놓이며, 손 위쪽에서 비스듬히 보는 구도에 가깝다. 반사상이나 중복 시설은 보이지 않는다.",
        "entities": "보이는 인물 부분은 왼쪽의 찰리와 오른쪽의 지소영뿐이다. 찰리의 샌드 베이지 장갑판, 노출 금속 관절, 거대한 손과 가장자리에 일부 보이는 푸른 원자로가 참고 외형에 맞는다. 지소영의 남색 니트는 참고와 맞지만, 손등과 손목은 어린아이보다 성인의 것으로 보이므로 어린 시절 설정에는 맞지 않는다. 얼굴이 없어 한국인 여성이라는 신원과 얼굴의 일치는 판정할 수 없다. 보라색 꽃밭과 먼 연구소는 확인되며 별도의 B-200은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 전완에 연결되고, 기계 손가락이 사람 손의 아래와 옆을 받치며 감싼다. 팔도 양쪽 몸체로 이어져 떠 있는 부위가 없다. 발과 지면 접촉은 프레임 밖이지만, 보이는 부분에 불가능한 지지나 관절 배치는 없다."
       },
       {
        "label": "B",
        "direction": "찰리의 큰 손이 왼쪽 위에서 내려와 오른쪽에서 뻗은 지소영의 손을 잡는다. 기계 손가락은 사람 손을 감싸 접촉 대상을 정확히 향한다. 머리는 화면 밖이고, 잘린 몸통만으로는 화면 밖 지동현을 향한 회전 여부를 확인할 수 없다.",
        "built_space": "보라색 꽃밭과 산줄기, 오른쪽 원경의 낮은 연구동 및 탑 하나가 보인다. 연구소의 크기와 배치는 장소 참고에 대체로 부합한다. 두 몸의 가장자리 사이 아래쪽에 손잡기가 놓이지만, 넓게 드러난 하늘과 손 뒤로 보이는 낮은 지평선 때문에 손보다 조금 높은 위치에서 내려다보는 지정 시점보다 낮고 수평적인 구도로 읽힌다. 중복 시설이나 반사상은 없다.",
        "entities": "찰리의 베이지 장갑판과 금속 관절, 사람보다 큰 기계 손은 참고의 재질과 정체성에 맞는다. 오른쪽 인물은 남색 니트 소매와 밝은색 하의를 입었으며, 드러난 손등의 힘줄과 전완은 어린아이보다 성인으로 보인다. 따라서 두 후보 모두 어린 시절 지소영이라는 명시 조건을 충족하지 못한다. 얼굴이 잘려 성별·한국인 정체성과 얼굴 일치는 확인할 수 없다. 보라색 꽃과 먼 연구소는 있으며 별도의 B-200은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "기계 손과 사람 손이 각각 전완에 연속적으로 연결되고, 굽힌 기계 손가락과 사람 손가락이 서로 맞물린다. 손과 팔에 보이는 지지가 있으며 공중에 분리되어 떠 있는 물체는 없다. 손목의 굽힘도 손을 잡는 동작으로 가능한 범위다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "맞잡은 손을 중심으로 비스듬히 내려다보는 근접 구도가 더 충실하지만, 지소영의 손은 지정된 어린 시절보다 성인의 손으로 보인다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "좌우에서 이어지는 손잡기는 맞지만, 하늘이 크게 보이는 낮은 시점이 지정된 내려다보기와 다르고 지소영의 손도 성인으로 보인다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 팔이 왼쪽 위에서 중앙으로 내려오고, 지소영의 손이 오른쪽에서 왼쪽으로 뻗어 기계 손 안에 들어간다. 기계 손가락은 실제로 그 손을 둘러싸며 접촉이 완성되어 있다. 머리는 모두 화면 밖이므로 시선과 지동현 쪽으로 돌아서는 방향은 확인할 수 없다.",
        "built_space": "야외 보라색 꽃밭 너머 오른쪽 위에 낮은 연구동 한 덩어리와 수직 탑 하나가 보이고, 뒤로 산줄기가 이어져 장소 참고와 부합한다. 건물은 작고 흐리게 유지된다. 양쪽 몸통 가장자리와 전완 사이에 맞잡은 손이 놓이며, 손 위쪽에서 비스듬히 보는 구도에 가깝다. 반사상이나 중복 시설은 보이지 않는다.",
        "entities": "보이는 인물 부분은 왼쪽의 찰리와 오른쪽의 지소영뿐이다. 찰리의 샌드 베이지 장갑판, 노출 금속 관절, 거대한 손과 가장자리에 일부 보이는 푸른 원자로가 참고 외형에 맞는다. 지소영의 남색 니트는 참고와 맞지만, 손등과 손목은 어린아이보다 성인의 것으로 보이므로 어린 시절 설정에는 맞지 않는다. 얼굴이 없어 한국인 여성이라는 신원과 얼굴의 일치는 판정할 수 없다. 보라색 꽃밭과 먼 연구소는 확인되며 별도의 B-200은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "두 손은 각각 손목과 전완에 연결되고, 기계 손가락이 사람 손의 아래와 옆을 받치며 감싼다. 팔도 양쪽 몸체로 이어져 떠 있는 부위가 없다. 발과 지면 접촉은 프레임 밖이지만, 보이는 부분에 불가능한 지지나 관절 배치는 없다."
       },
       {
        "label": "A",
        "direction": "찰리의 큰 손이 왼쪽 위에서 내려와 오른쪽에서 뻗은 지소영의 손을 잡는다. 기계 손가락은 사람 손을 감싸 접촉 대상을 정확히 향한다. 머리는 화면 밖이고, 잘린 몸통만으로는 화면 밖 지동현을 향한 회전 여부를 확인할 수 없다.",
        "built_space": "보라색 꽃밭과 산줄기, 오른쪽 원경의 낮은 연구동 및 탑 하나가 보인다. 연구소의 크기와 배치는 장소 참고에 대체로 부합한다. 두 몸의 가장자리 사이 아래쪽에 손잡기가 놓이지만, 넓게 드러난 하늘과 손 뒤로 보이는 낮은 지평선 때문에 손보다 조금 높은 위치에서 내려다보는 지정 시점보다 낮고 수평적인 구도로 읽힌다. 중복 시설이나 반사상은 없다.",
        "entities": "찰리의 베이지 장갑판과 금속 관절, 사람보다 큰 기계 손은 참고의 재질과 정체성에 맞는다. 오른쪽 인물은 남색 니트 소매와 밝은색 하의를 입었으며, 드러난 손등의 힘줄과 전완은 어린아이보다 성인으로 보인다. 따라서 두 후보 모두 어린 시절 지소영이라는 명시 조건을 충족하지 못한다. 얼굴이 잘려 성별·한국인 정체성과 얼굴 일치는 확인할 수 없다. 보라색 꽃과 먼 연구소는 있으며 별도의 B-200은 이 크롭에서 식별되지 않는다.",
        "hard_violations": [],
        "physics": "기계 손과 사람 손이 각각 전완에 연속적으로 연결되고, 굽힌 기계 손가락과 사람 손가락이 서로 맞물린다. 손과 팔에 보이는 지지가 있으며 공중에 분리되어 떠 있는 물체는 없다. 손목의 굽힘도 손을 잡는 동작으로 가능한 범위다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.179
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학 구조 (기계 손의 구조가 엄지와 손바닥의 구분이 없이 위아래 양쪽에서 관절이 뻗어나와 비정상적인 집게 형태를 띰)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1179
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 카메라 앵글과 인물의 신체 일부분, 보라색 꽃밭과 멀리 보이는 연구소 배경까지 프롬프트의 지시를 충실하고 자연스럽게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "캐릭터의 의상과 배경은 참조 이미지와 일치하나, 로봇 손이 엄지와 손바닥 없이 위아래로 둥글게 이어지는 불가능한 구조로 렌더링되어 감점되었습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 구조 (기계 손의 구조가 엄지와 손바닥의 구분이 없이 위아래 양쪽에서 관절이 뻗어나와 비정상적인 집게 형태를 띰)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_violet_field_621f9d.png",
    "asset_id": "bba38d59-2419-4fea-98d2-3d47523d7b7c",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1237799>",
    "asset_id": "3674e762-acc5-4d71-a06e-9b4c0f9b740c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d1d-7eab-7e59-9b88-de448b291d5e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S68sh20__bgfirst_bg.png",
   "bg_asset_id": "b01ffceb-46c9-43e6-99b0-7f2e082004a2",
   "bg_record_key": "S68sh20::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "violet_field",
   "groupbg_asset_id": "bba38d59-2419-4fea-98d2-3d47523d7b7c"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C32"
  ]
 },
 "S69sh2::signage": {
  "fp": "b3526b0d9f0e9aa0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::868cb6b91854fad0": {
  "subjects": [],
  "subject_text": "이현우의 트럭 적재함\n낡은 금속 바닥과 낮은 측벽으로 둘러싸인 개방형 적재 공간. 고정 지붕이 없어 위가 트여 있고, 이동식 그늘막을 걸칠 수 있다.",
  "identity": "canonical",
  "scope_id": "L98",
  "scope_role": "location_exterior",
  "scope_sha": "b416fffdedb8e64e"
 },
 "S69sh2::bgfirst_bg": {
  "input_fingerprint": "72a948837180179a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 트럭 짐칸 바닥에 이현우, 앰버, 라울이 머리를 맞댄 채 곤히 잠들어 있는 구도.\n\nLOCATION (lock): On the open rear bed of a parked truck sheltered beside a rock outcrop, beneath a makeshift shade cover.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless just outside the rear corner of the cargo bed, looking diagonally downward from above the sleepers and retaining their full resting arrangement. Cluster 이현우, 앰버, and 라울's heads near center while their bodies extend in different directions: 이현우 lies slightly turned, 앰버 curls toward the shared center, and 라울 rests with one shoulder rolled back. All three have their eyes closed in sleep, with naturally different arm and knee positions; leave the near bed edge readable for the ensuing move along the same side.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-bed floor and rear edge (Stationary, supporting the three sleeping companions) — The floor is seen obliquely from the rear corner, with the near edge below the sleepers; used as Defines their shared resting space and the camera's continuous path; Sheltering rocks (The truck is parked in the space between them) — Only peripheral portions beyond the cargo bed are included; used as Grounds the resting place without competing with the clustered heads.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the sheltering rocks creates a quiet shaded resting place with gentle, readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 트럭 짐칸 바닥에 이현우, 앰버, 라울이 머리를 맞댄 채 곤히 잠들어 있는 구도.\n\nLOCATION (lock): On the open rear bed of a parked truck sheltered beside a rock outcrop, beneath a makeshift shade cover.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless just outside the rear corner of the cargo bed, looking diagonally downward from above the sleepers and retaining their full resting arrangement. Cluster 이현우, 앰버, and 라울's heads near center while their bodies extend in different directions: 이현우 lies slightly turned, 앰버 curls toward the shared center, and 라울 rests with one shoulder rolled back. All three have their eyes closed in sleep, with naturally different arm and knee positions; leave the near bed edge readable for the ensuing move along the same side.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-bed floor and rear edge (Stationary, supporting the three sleeping companions) — The floor is seen obliquely from the rear corner, with the near edge below the sleepers; used as Defines their shared resting space and the camera's continuous path; Sheltering rocks (The truck is parked in the space between them) — Only peripheral portions beyond the cargo bed are included; used as Grounds the resting place without competing with the clustered heads.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the sheltering rocks creates a quiet shaded resting place with gentle, readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S69sh2__bgfirst_bg.png",
  "asset_id": "2213f06c-53ef-4c05-9f78-fb802e861cf9",
  "input_asset_ids": [
   "bca8100b-96f9-4c74-ba97-6d1492d43b65",
   "2e3db7e5-7f1f-47b6-8d3a-2d3a8a41eaaf"
  ]
 },
 "S69sh2": {
  "input_fingerprint": "43d59730a7494c30",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 바닥에 이현우, 앰버, 라울이 머리를 맞댄 채 곤히 잠들어 있는 구도.\n\nLOCATION (lock): On the open rear bed of a parked truck sheltered beside a rock outcrop, beneath a makeshift shade cover. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless just outside the rear corner of the cargo bed, looking diagonally downward from above the sleepers and retaining their full resting arrangement. Cluster 이현우, 앰버, and 라울's heads near center while their bodies extend in different directions: 이현우 lies slightly turned, 앰버 curls toward the shared center, and 라울 rests with one shoulder rolled back. All three have their eyes closed in sleep, with naturally different arm and knee positions; leave the near bed edge readable for the ensuing move along the same side.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-bed floor and rear edge (Stationary, supporting the three sleeping companions) — The floor is seen obliquely from the rear corner, with the near edge below the sleepers; used as Defines their shared resting space and the camera's continuous path; Sheltering rocks (The truck is parked in the space between them) — Only peripheral portions beyond the cargo bed are included; used as Grounds the resting place without competing with the clustered heads.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the sheltering rocks creates a quiet shaded resting place with gentle, readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is parked in shade beside rocks during the day, with a shade screen available over the cargo area. Charlie retains his battle damage and possession of B-200's recovered chest component. 이현우: He is asleep in the cargo bed with his existing battle wounds. 앰버: She is asleep in the cargo bed; the injury to the back of her head remains unresolved. 라울: He is asleep in the cargo bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 바닥에 이현우, 앰버, 라울이 머리를 맞댄 채 곤히 잠들어 있는 구도.\n\nLOCATION (lock): On the open rear bed of a parked truck sheltered beside a rock outcrop, beneath a makeshift shade cover. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless just outside the rear corner of the cargo bed, looking diagonally downward from above the sleepers and retaining their full resting arrangement. Cluster 이현우, 앰버, and 라울's heads near center while their bodies extend in different directions: 이현우 lies slightly turned, 앰버 curls toward the shared center, and 라울 rests with one shoulder rolled back. All three have their eyes closed in sleep, with naturally different arm and knee positions; leave the near bed edge readable for the ensuing move along the same side.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-bed floor and rear edge (Stationary, supporting the three sleeping companions) — The floor is seen obliquely from the rear corner, with the near edge below the sleepers; used as Defines their shared resting space and the camera's continuous path; Sheltering rocks (The truck is parked in the space between them) — Only peripheral portions beyond the cargo bed are included; used as Grounds the resting place without competing with the clustered heads.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the sheltering rocks creates a quiet shaded resting place with gentle, readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is parked in shade beside rocks during the day, with a shade screen available over the cargo area. Charlie retains his battle damage and possession of B-200's recovered chest component. 이현우: He is asleep in the cargo bed with his existing battle wounds. 앰버: She is asleep in the cargo bed; the injury to the back of her head remains unresolved. 라울: He is asleep in the cargo bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 트럭 짐칸 바닥에 이현우, 앰버, 라울이 머리를 맞댄 채 곤히 잠들어 있는 구도.\n\nLOCATION (lock): On the open rear bed of a parked truck sheltered beside a rock outcrop, beneath a makeshift shade cover. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless just outside the rear corner of the cargo bed, looking diagonally downward from above the sleepers and retaining their full resting arrangement. Cluster 이현우, 앰버, and 라울's heads near center while their bodies extend in different directions: 이현우 lies slightly turned, 앰버 curls toward the shared center, and 라울 rests with one shoulder rolled back. All three have their eyes closed in sleep, with naturally different arm and knee positions; leave the near bed edge readable for the ensuing move along the same side.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Truck cargo-bed floor and rear edge (Stationary, supporting the three sleeping companions) — The floor is seen obliquely from the rear corner, with the near edge below the sleepers; used as Defines their shared resting space and the camera's continuous path; Sheltering rocks (The truck is parked in the space between them) — Only peripheral portions beyond the cargo bed are included; used as Grounds the resting place without competing with the clustered heads.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight moderated by the sheltering rocks creates a quiet shaded resting place with gentle, readable contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is parked in shade beside rocks during the day, with a shade screen available over the cargo area. Charlie retains his battle damage and possession of B-200's recovered chest component. 이현우: He is asleep in the cargo bed with his existing battle wounds. 앰버: She is asleep in the cargo bed; the injury to the back of her head remains unresolved. 라울: He is asleep in the cargo bed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S69sh2__bgfirst_bg.png",
     "asset_id": "2213f06c-53ef-4c05-9f78-fb802e861cf9",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S69sh2.png",
     "asset_id": "bca8100b-96f9-4c74-ba97-6d1492d43b65",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_open_travel_truck_sel.png",
     "asset_id": "2e3db7e5-7f1f-47b6-8d3a-2d3a8a41eaaf",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 잠든 세 사람을 대각선 아래로 향함.",
    "built_space": "암석 사이 주차된 트럭 짐칸으로, 기둥으로 지지된 차광막이 정확히 배치됨.",
    "entities": "이현우(무전기 누락), 앰버(마스크, 공구벨트), 라울(꽁지머리 누락). 모두 지시된 복장으로 짐칸에 모여 잠듦.",
    "hard_violations": [],
    "physics": "모든 인물의 신체가 짐칸 바닥에 중력에 맞게 완전히 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 세 사람을 내려다봄.",
    "built_space": "암석 사이 트럭 짐칸과 그물망 형태의 차광막이 존재함.",
    "entities": "이현우(무전기 착용), 앰버(마스크 착용), 라울(꽁지머리). 지시되지 않은 담요와 배낭, 상자들이 다수 생성됨.",
    "hard_violations": [
     "[gemini-pro] 지시되지 않은 물건(담요, 대형 배낭, 철제 상자) 임의 생성",
     "[gpt-high] 지시나 참조에 없는 배낭 여러 개, 보관 상자 및 병·물통을 적재함에 추가해 상당한 면적을 차지하게 했다."
    ],
    "physics": "인물들이 짐칸 바닥과 임의로 생성된 담요 위에 자연스럽게 지지되어 누워있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 환경과 구도를 정확히 구현했으며, 불필요한 임의의 사물을 추가하지 않아 프롬프트에 가장 충실함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 세부 외형은 잘 묘사되었으나, 프롬프트에 명시되지 않은 담요와 짐들을 대거 추가하여 지침을 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 잠든 세 사람을 대각선 아래로 향함.",
        "built_space": "암석 사이 주차된 트럭 짐칸으로, 기둥으로 지지된 차광막이 정확히 배치됨.",
        "entities": "이현우(무전기 누락), 앰버(마스크, 공구벨트), 라울(꽁지머리 누락). 모두 지시된 복장으로 짐칸에 모여 잠듦.",
        "hard_violations": [],
        "physics": "모든 인물의 신체가 짐칸 바닥에 중력에 맞게 완전히 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 세 사람을 내려다봄.",
        "built_space": "암석 사이 트럭 짐칸과 그물망 형태의 차광막이 존재함.",
        "entities": "이현우(무전기 착용), 앰버(마스크 착용), 라울(꽁지머리). 지시되지 않은 담요와 배낭, 상자들이 다수 생성됨.",
        "hard_violations": [
         "지시되지 않은 물건(담요, 대형 배낭, 철제 상자) 임의 생성"
        ],
        "physics": "인물들이 짐칸 바닥과 임의로 생성된 담요 위에 자연스럽게 지지되어 누워있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 환경과 구도를 정확히 구현했으며, 불필요한 임의의 사물을 추가하지 않아 프롬프트에 가장 충실함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 세부 외형은 잘 묘사되었으나, 프롬프트에 명시되지 않은 담요와 짐들을 대거 추가하여 지침을 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 잠든 세 사람을 대각선 아래로 향함.",
        "built_space": "암석 사이 주차된 트럭 짐칸으로, 기둥으로 지지된 차광막이 정확히 배치됨.",
        "entities": "이현우(무전기 누락), 앰버(마스크, 공구벨트), 라울(꽁지머리 누락). 모두 지시된 복장으로 짐칸에 모여 잠듦.",
        "hard_violations": [],
        "physics": "모든 인물의 신체가 짐칸 바닥에 중력에 맞게 완전히 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 트럭 짐칸 뒤쪽 모서리에서 세 사람을 내려다봄.",
        "built_space": "암석 사이 트럭 짐칸과 그물망 형태의 차광막이 존재함.",
        "entities": "이현우(무전기 착용), 앰버(마스크 착용), 라울(꽁지머리). 지시되지 않은 담요와 배낭, 상자들이 다수 생성됨.",
        "hard_violations": [
         "지시되지 않은 물건(담요, 대형 배낭, 철제 상자) 임의 생성"
        ],
        "physics": "인물들이 짐칸 바닥과 임의로 생성된 담요 위에 자연스럽게 지지되어 누워있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "세 사람의 수면과 머리 모음은 구현했지만, 지시되지 않은 다수의 짐과 용기를 추가했고 하체를 잘라 전체 휴식 배치를 보존하라는 구도에 미달한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "뒤 모서리의 사선 하향 와이드 숏, 읽히는 뒤 적재함 가장자리, 서로 다른 수면 자세를 더 충실히 구현했으나 지붕 없는 트럭 조건과 주변부로 제한해야 할 배경 비중은 아쉽다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 눈을 감고 있으며 시선의 대상은 없다. 이현우의 몸은 중앙의 머리에서 왼쪽 아래로, 앰버는 오른쪽 위로, 라울은 오른쪽 아래로 이어진다. 앰버는 중앙을 향해 웅크렸지만 라울 역시 옆으로 웅크려, 한쪽 어깨를 뒤로 젖힌 자세는 뚜렷하지 않다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "적재함 바닥 하나와 금속 벽 세 면의 일부, 위쪽 차광망 하나 및 이를 받치는 기둥과 끈이 보인다. 가까운 측벽은 왼쪽 아래를 대각선으로 지나지만 뒤쪽 끝과 전체 휴식 배치는 프레임 밖으로 잘린다. 세 사람은 모두 적재함 안에 있다. 양쪽의 층상 암벽과 틈 사이 바다는 장소 참조의 재질과 지형에 부합한다. 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 정확히 세 명이다. 이현우는 짧고 흐트러진 검은 머리의 동아시아계 후기 청소년 남성으로 보이며, 어두운 낡은 옷과 인이어 장치, 얼굴의 상처가 보인다. 앰버는 밝은 피부와 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 벨트와 목 부근 방진 마스크가 보인다. 라울은 갈색 피부와 묶은 곱슬머리의 어린 남자아이이며 빛바랜 티셔츠와 반바지를 입었다. 두 아이의 외형은 참조와 대체로 맞지만 혼혈 배경 자체는 외관만으로 확정할 수 없다. 앰버의 후두부 부상은 가려져 확인할 수 없다. 배낭 여러 개, 상자, 병과 물통, 담요가 추가되어 있다.",
        "hard_violations": [
         "지시나 참조에 없는 배낭 여러 개, 보관 상자 및 병·물통을 적재함에 추가해 상당한 면적을 차지하게 했다."
        ],
        "physics": "세 사람의 몸통과 다리는 적재함 바닥 및 깔린 천에 놓여 있다. 머리는 접은 팔과 손, 천에 받쳐지고 손들은 바닥이나 몸에 내려앉아 있다. 떠 있거나 스스로 들어 올린 사지는 보이지 않는다. 추가된 짐은 바닥에 놓여 있고 차광망은 기둥과 끈으로 지지된다."
       },
       {
        "label": "B",
        "direction": "세 사람 모두 눈을 감았다. 모인 머리에서 이현우의 몸은 왼쪽 아래로, 앰버는 중앙 아래로, 라울은 오른쪽 아래로 뻗는다. 앰버는 머리 모음 쪽으로 몸을 말았고 이현우는 옆으로 돌아누웠으며, 라울은 얼굴을 위로 향하고 어깨를 뒤로 연 자세다. 이동이나 조준 대상은 없다.",
        "built_space": "뒤 모서리 바깥에서 적재함을 비스듬히 내려다보며, 바닥 하나와 앞벽 하나, 좌우 측벽 두 개, 가까운 뒤판 하나가 구분된다. 세 사람의 휴식 배치와 가까운 뒤 가장자리가 함께 읽힌다. 차광막 하나와 지지대 세 개가 보인다. 앞쪽에는 유리창과 지붕 테두리를 갖춘 운전실이 있어 지붕 없는 트럭이라는 조건과 차이가 있다. 양쪽 층상 암벽과 바다는 장소 참조에 부합하지만, 바위와 바다가 주변부 이상으로 눈에 들어온다. 창문에 보이는 반사는 불가능하다고 판단할 근거가 없다.",
        "entities": "추가 인물 없이 이현우, 앰버, 라울에 해당하는 세 사람만 보인다. 이현우의 동아시아계 후기 청소년 남성 외형, 짧은 검은 머리, 마른 체격과 더럽혀진 어두운 옷, 얼굴 상처가 맞는다. 인이어 장치는 명확히 식별되지 않는다. 앰버는 금발의 어린 여자아이이며 카키 작업복과 가죽 공구 벨트, 얼굴 아래쪽의 방진 마스크가 보인다. 라울은 갈색 피부의 어린 남자아이로 묶은 곱슬머리, 낡은 티셔츠와 반바지가 참조에 가깝다. 혼혈 배경은 외형만으로 확정할 수 없다. 앰버의 후두부는 머리카락에 가려 부상 지속 여부를 확인할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 옆구리와 굽힌 다리, 앰버의 몸통과 접은 다리, 라울의 등과 다리가 바닥에 지지된다. 이현우와 앰버의 머리는 내려놓은 팔과 손에 받쳐진다. 라울의 한 손은 배 위에, 다른 팔과 손은 머리 뒤쪽 바닥에 놓인 것으로 읽히며 공중에 유지하는 자세는 아니다. 차광막은 지지대와 연결 끈에 매달려 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "세 사람의 수면과 머리 모음은 구현했지만, 지시되지 않은 다수의 짐과 용기를 추가했고 하체를 잘라 전체 휴식 배치를 보존하라는 구도에 미달한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "뒤 모서리의 사선 하향 와이드 숏, 읽히는 뒤 적재함 가장자리, 서로 다른 수면 자세를 더 충실히 구현했으나 지붕 없는 트럭 조건과 주변부로 제한해야 할 배경 비중은 아쉽다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 눈을 감고 있으며 시선의 대상은 없다. 이현우의 몸은 중앙의 머리에서 왼쪽 아래로, 앰버는 오른쪽 위로, 라울은 오른쪽 아래로 이어진다. 앰버는 중앙을 향해 웅크렸지만 라울 역시 옆으로 웅크려, 한쪽 어깨를 뒤로 젖힌 자세는 뚜렷하지 않다. 조준하거나 이동하는 물체는 없다.",
        "built_space": "적재함 바닥 하나와 금속 벽 세 면의 일부, 위쪽 차광망 하나 및 이를 받치는 기둥과 끈이 보인다. 가까운 측벽은 왼쪽 아래를 대각선으로 지나지만 뒤쪽 끝과 전체 휴식 배치는 프레임 밖으로 잘린다. 세 사람은 모두 적재함 안에 있다. 양쪽의 층상 암벽과 틈 사이 바다는 장소 참조의 재질과 지형에 부합한다. 불가능한 반사는 보이지 않는다.",
        "entities": "인물은 정확히 세 명이다. 이현우는 짧고 흐트러진 검은 머리의 동아시아계 후기 청소년 남성으로 보이며, 어두운 낡은 옷과 인이어 장치, 얼굴의 상처가 보인다. 앰버는 밝은 피부와 금발의 어린 여자아이이며 카키 작업복, 가죽 공구 벨트와 목 부근 방진 마스크가 보인다. 라울은 갈색 피부와 묶은 곱슬머리의 어린 남자아이이며 빛바랜 티셔츠와 반바지를 입었다. 두 아이의 외형은 참조와 대체로 맞지만 혼혈 배경 자체는 외관만으로 확정할 수 없다. 앰버의 후두부 부상은 가려져 확인할 수 없다. 배낭 여러 개, 상자, 병과 물통, 담요가 추가되어 있다.",
        "hard_violations": [
         "지시나 참조에 없는 배낭 여러 개, 보관 상자 및 병·물통을 적재함에 추가해 상당한 면적을 차지하게 했다."
        ],
        "physics": "세 사람의 몸통과 다리는 적재함 바닥 및 깔린 천에 놓여 있다. 머리는 접은 팔과 손, 천에 받쳐지고 손들은 바닥이나 몸에 내려앉아 있다. 떠 있거나 스스로 들어 올린 사지는 보이지 않는다. 추가된 짐은 바닥에 놓여 있고 차광망은 기둥과 끈으로 지지된다."
       },
       {
        "label": "A",
        "direction": "세 사람 모두 눈을 감았다. 모인 머리에서 이현우의 몸은 왼쪽 아래로, 앰버는 중앙 아래로, 라울은 오른쪽 아래로 뻗는다. 앰버는 머리 모음 쪽으로 몸을 말았고 이현우는 옆으로 돌아누웠으며, 라울은 얼굴을 위로 향하고 어깨를 뒤로 연 자세다. 이동이나 조준 대상은 없다.",
        "built_space": "뒤 모서리 바깥에서 적재함을 비스듬히 내려다보며, 바닥 하나와 앞벽 하나, 좌우 측벽 두 개, 가까운 뒤판 하나가 구분된다. 세 사람의 휴식 배치와 가까운 뒤 가장자리가 함께 읽힌다. 차광막 하나와 지지대 세 개가 보인다. 앞쪽에는 유리창과 지붕 테두리를 갖춘 운전실이 있어 지붕 없는 트럭이라는 조건과 차이가 있다. 양쪽 층상 암벽과 바다는 장소 참조에 부합하지만, 바위와 바다가 주변부 이상으로 눈에 들어온다. 창문에 보이는 반사는 불가능하다고 판단할 근거가 없다.",
        "entities": "추가 인물 없이 이현우, 앰버, 라울에 해당하는 세 사람만 보인다. 이현우의 동아시아계 후기 청소년 남성 외형, 짧은 검은 머리, 마른 체격과 더럽혀진 어두운 옷, 얼굴 상처가 맞는다. 인이어 장치는 명확히 식별되지 않는다. 앰버는 금발의 어린 여자아이이며 카키 작업복과 가죽 공구 벨트, 얼굴 아래쪽의 방진 마스크가 보인다. 라울은 갈색 피부의 어린 남자아이로 묶은 곱슬머리, 낡은 티셔츠와 반바지가 참조에 가깝다. 혼혈 배경은 외형만으로 확정할 수 없다. 앰버의 후두부는 머리카락에 가려 부상 지속 여부를 확인할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 옆구리와 굽힌 다리, 앰버의 몸통과 접은 다리, 라울의 등과 다리가 바닥에 지지된다. 이현우와 앰버의 머리는 내려놓은 팔과 손에 받쳐진다. 라울의 한 손은 배 위에, 다른 팔과 손은 머리 뒤쪽 바닥에 놓인 것으로 읽히며 공중에 유지하는 자세는 아니다. 차광막은 지지대와 연결 끈에 매달려 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시되지 않은 물건(담요, 대형 배낭, 철제 상자) 임의 생성",
     "[gpt-high] 지시나 참조에 없는 배낭 여러 개, 보관 상자 및 병·물통을 적재함에 추가해 상당한 면적을 차지하게 했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 환경과 구도를 정확히 구현했으며, 불필요한 임의의 사물을 추가하지 않아 프롬프트에 가장 충실함."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "캐릭터의 세부 외형은 잘 묘사되었으나, 프롬프트에 명시되지 않은 담요와 짐들을 대거 추가하여 지침을 위반함.  ★위반: [gemini-pro] 지시되지 않은 물건(담요, 대형 배낭, 철제 상자) 임의 생성 / [gpt-high] 지시나 참조에 없는 배낭 여러 개, 보관 상자 및 병·물통을 적재함에 추가해 상당한 면적을 차지하게 했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_open_travel_truck_sel.png",
    "asset_id": "2e3db7e5-7f1f-47b6-8d3a-2d3a8a41eaaf",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d27-1260-7cae-b698-c3be357fd84f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S69sh2__bgfirst_bg.png",
   "bg_asset_id": "2213f06c-53ef-4c05-9f78-fb802e861cf9",
   "bg_record_key": "S69sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S69sh5::signage": {
  "fp": "9806c4a0d113af53",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S69sh5": {
  "input_fingerprint": "b82a6698a677c87a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막을 쥔 채 곤히 자는 일행을 내려다보며 환한 미소 이모티콘을 띄운 찰리의 상체.\n\nLOCATION (lock): On the parked truck's open rear load bed beside the rock shelter, at the edge of the movable shade cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the upward tilt settle at the same low bedside endpoint, viewing 찰리's upper body from a three-quarter side angle outside the line of his downward gaze. Place his smiling face above center and retain both hands holding the shade cover along the lower edge, with only a modest portion of the cover entering the frame. His head stays inclined toward the sleeping companions immediately below the crop, making the smile a response to their rest rather than an address to the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shade cover (Repositioned and held to shield the sleeping companions) — An oblique portion extends from 찰리's hands toward the sleepers below frame; used as Links the visible smile to the protective action without obscuring the upper body; Cargo-bed edge (Stationary beside the sleeping group) — A narrow near-side section remains at the bottom of the composition; used as Preserves the bedside endpoint established by the continuous move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use gentle daylight with the held cover interrupting the light toward the sleepers, keeping 찰리's caring expression clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the truck-bed floor, surrounding rock shade, and daylight appearance from the reference. Exclude the moving-road setting and objects from the imagined flower field.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks, and the shade screen is now repositioned to block sunlight from the sleeping area. Charlie retains his damaged body and B-200's chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막을 쥔 채 곤히 자는 일행을 내려다보며 환한 미소 이모티콘을 띄운 찰리의 상체.\n\nLOCATION (lock): On the parked truck's open rear load bed beside the rock shelter, at the edge of the movable shade cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the upward tilt settle at the same low bedside endpoint, viewing 찰리's upper body from a three-quarter side angle outside the line of his downward gaze. Place his smiling face above center and retain both hands holding the shade cover along the lower edge, with only a modest portion of the cover entering the frame. His head stays inclined toward the sleeping companions immediately below the crop, making the smile a response to their rest rather than an address to the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shade cover (Repositioned and held to shield the sleeping companions) — An oblique portion extends from 찰리's hands toward the sleepers below frame; used as Links the visible smile to the protective action without obscuring the upper body; Cargo-bed edge (Stationary beside the sleeping group) — A narrow near-side section remains at the bottom of the composition; used as Preserves the bedside endpoint established by the continuous move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use gentle daylight with the held cover interrupting the light toward the sleepers, keeping 찰리's caring expression clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the truck-bed floor, surrounding rock shade, and daylight appearance from the reference. Exclude the moving-road setting and objects from the imagined flower field.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks, and the shade screen is now repositioned to block sunlight from the sleeping area. Charlie retains his damaged body and B-200's chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 그늘막을 쥔 채 곤히 자는 일행을 내려다보며 환한 미소 이모티콘을 띄운 찰리의 상체.\n\nLOCATION (lock): On the parked truck's open rear load bed beside the rock shelter, at the edge of the movable shade cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the upward tilt settle at the same low bedside endpoint, viewing 찰리's upper body from a three-quarter side angle outside the line of his downward gaze. Place his smiling face above center and retain both hands holding the shade cover along the lower edge, with only a modest portion of the cover entering the frame. His head stays inclined toward the sleeping companions immediately below the crop, making the smile a response to their rest rather than an address to the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Shade cover (Repositioned and held to shield the sleeping companions) — An oblique portion extends from 찰리's hands toward the sleepers below frame; used as Links the visible smile to the protective action without obscuring the upper body; Cargo-bed edge (Stationary beside the sleeping group) — A narrow near-side section remains at the bottom of the composition; used as Preserves the bedside endpoint established by the continuous move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use gentle daylight with the held cover interrupting the light toward the sleepers, keeping 찰리's caring expression clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the truck-bed floor, surrounding rock shade, and daylight appearance from the reference. Exclude the moving-road setting and objects from the imagined flower field.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep in the truck's rear cargo bed, with his head gathered together with Amber's and Raul's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is asleep in the truck's rear cargo bed, with her head gathered together with Hyunwoo's and Raul's heads. The scene text does not specify her torso's orientation, her head's supporting surface, or the placement of her arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Raul is asleep in the truck's rear cargo bed, with his head gathered together with Hyunwoo's and Amber's heads. The scene text does not specify his torso's orientation, his head's supporting surface, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck remains parked beside the rocks, and the shade screen is now repositioned to block sunlight from the sleeping area. Charlie retains his damaged body and B-200's chest component.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 찰리 right now, so 찰리's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 찰리: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선이 프레임 아래쪽(일행이 있는 위치)을 향하고 있음.",
    "built_space": "화면 하단에 트럭 짐칸의 모서리가 보이며, 배경으로 바위 절벽과 바다가 올바르게 배치됨.",
    "entities": "찰리의 고릴라형 기계 몸체, 마스크 등 외형이 레퍼런스와 일치함 (표정은 기본 입술 선 유지).",
    "hard_violations": [],
    "physics": "두 손으로 그늘막 천을 안정적으로 쥐고 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "찰리의 시선이 프레임 아래쪽을 향하고 있음.",
    "built_space": "하단에 트럭 가장자리와 배경이 보이나, 우측 상단에 지시되지 않은 커다란 천막 지붕이 프레임 안으로 크게 침범함.",
    "entities": "찰리의 외형이 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "손가락이 천의 질감과 융합되어 물리적인 쥐기 형태가 무너짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "화면 하단에만 그늘막을 배치하라는 구도 지시를 정확히 따랐으며, 두 손의 물리적 상호작용이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "구도 지시와 달리 화면 우측 상단에 큰 천막이 침범하였고, 천을 쥔 손가락의 묘사가 부자연스럽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 프레임 아래쪽(일행이 있는 위치)을 향하고 있음.",
        "built_space": "화면 하단에 트럭 짐칸의 모서리가 보이며, 배경으로 바위 절벽과 바다가 올바르게 배치됨.",
        "entities": "찰리의 고릴라형 기계 몸체, 마스크 등 외형이 레퍼런스와 일치함 (표정은 기본 입술 선 유지).",
        "hard_violations": [],
        "physics": "두 손으로 그늘막 천을 안정적으로 쥐고 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 프레임 아래쪽을 향하고 있음.",
        "built_space": "하단에 트럭 가장자리와 배경이 보이나, 우측 상단에 지시되지 않은 커다란 천막 지붕이 프레임 안으로 크게 침범함.",
        "entities": "찰리의 외형이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "손가락이 천의 질감과 융합되어 물리적인 쥐기 형태가 무너짐."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "화면 하단에만 그늘막을 배치하라는 구도 지시를 정확히 따랐으며, 두 손의 물리적 상호작용이 자연스럽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "구도 지시와 달리 화면 우측 상단에 큰 천막이 침범하였고, 천을 쥔 손가락의 묘사가 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선이 프레임 아래쪽(일행이 있는 위치)을 향하고 있음.",
        "built_space": "화면 하단에 트럭 짐칸의 모서리가 보이며, 배경으로 바위 절벽과 바다가 올바르게 배치됨.",
        "entities": "찰리의 고릴라형 기계 몸체, 마스크 등 외형이 레퍼런스와 일치함 (표정은 기본 입술 선 유지).",
        "hard_violations": [],
        "physics": "두 손으로 그늘막 천을 안정적으로 쥐고 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선이 프레임 아래쪽을 향하고 있음.",
        "built_space": "하단에 트럭 가장자리와 배경이 보이나, 우측 상단에 지시되지 않은 커다란 천막 지붕이 프레임 안으로 크게 침범함.",
        "entities": "찰리의 외형이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "손가락이 천의 질감과 융합되어 물리적인 쥐기 형태가 무너짐."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 사선 시점과 아래쪽으로 기울인 얼굴이 잠든 일행을 돌보는 상체 숏에 더 가깝지만, 그늘막과 적재함 측판의 비중이 크고 미소는 다소 은은하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "양손으로 그늘막을 잡는 동작과 기계 정체성은 맞지만, 정면에 가까운 얼굴이 렌즈를 향하는 인상을 주어 요구된 측면 시점과 일행을 향한 미소의 관계가 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 머리는 화면 오른쪽 아래로 기울어 있고 얼굴 면도 그쪽을 향한다. 발광하는 점 형태의 눈만으로 시선을 확정하기는 어렵지만, 렌즈 정면보다는 그늘막 너머 화면 아래의 잠자리 쪽을 보는 자세에 가깝다. 양손은 아래쪽 그늘막 가장자리를 잡고 있으며 천은 손에서 오른쪽 아래로 이어진다. 잠든 일행 자체는 보이지 않는다.",
        "built_space": "왼쪽에 긴 금속 지주 하나, 왼손 아래 천 고리 부근에 짧게 드러난 금속 지지부 하나가 보인다. 녹슨 적재함 측판 한 구간이 하단을 비스듬히 가로지르고, 층상 암벽과 바다가 배경에 있다. 화면 위쪽에도 그늘막 일부가 남아 있으나 아래 천과 별개의 중복 설비인지는 확인되지 않는다. 찰리는 측판 뒤에 있으며 낮은 사선 카메라 위치가 읽힌다. 다만 측판이 요구된 좁은 가장자리보다 넓게 보인다.",
        "entities": "보이는 인물은 찰리 하나뿐이다. 샌드 베이지색의 긁히고 마모된 장갑판, 육중한 팔, 흰 마스크, 주황색 점 눈 두 개와 검은 입 선, 푸르게 빛나는 원형 가슴 부품이 캐릭터 참조와 부합한다. 인간 피부나 치아는 없다. 입 선은 부드러운 미소이지만 환한 미소 이모티콘이라는 지시보다는 절제되어 있다. 녹색 망사 그늘막과 낡은 흰색 트럭, 해안 암벽은 장소 참조와 일치한다. 잠든 세 사람과 다리는 프레임 밖이므로 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "양쪽 기계 손가락이 천의 윗단을 감싸 실제로 붙잡고 있다. 그늘막에는 손 사이의 장력과 아래로 처지는 주름이 보이며, 상단 천은 지주가 있는 차양 구조와 양립한다. 팔은 어깨와 팔꿈치 관절에 연결되어 있다. 하체의 접지점은 측판과 화면 경계에 가려져 있으나 상체가 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "B",
        "direction": "머리는 아래로 숙여져 있지만 얼굴과 두 눈은 카메라 쪽에 거의 정면으로 제시된다. 아래쪽 일행을 비껴서 관찰하는 측면 시점보다는 관객을 내려다보는 인상이 강하다. 양팔은 좌우로 벌어져 각각 천의 가장자리를 잡으며, 천은 손에서 화면 아래와 오른쪽 지주로 이어진다. 시선의 실제 목표가 잠든 일행이라고 확인할 단서는 부족하다.",
        "built_space": "오른쪽에 금속 지주 하나와 천을 고정하는 체결부가 보인다. 전경의 녹슨 측판 한 구간과 오른쪽으로 이어지는 측판 한 구간이 적재함 모서리를 구성한다. 찰리는 그 뒤에 있고, 양옆의 층상 암벽 사이로 바다가 보인다. 낮은 카메라 위치는 맞지만 찰리를 거의 정면으로 바라본다. 그늘막이 화면 하부를 넓게 채우고 측판도 두껍게 들어와, 소량의 천과 좁은 적재함 가장자리라는 구도에서 벗어난다.",
        "entities": "찰리 한 기만 보이며 추가 인물은 없다. 흰 마스크와 점 눈 두 개, 검은 입 선, 마모된 베이지 장갑판, 긴 기계 팔과 푸른 원형 가슴 부품은 참조의 정체성을 유지한다. 미소는 보이지만 환하게 웃는 표정보다는 작은 미소에 가깝다. 녹색 망사 천과 부식된 트럭 측판, 암벽과 바다는 지정된 장소의 재질과 일치한다. 가려진 잠든 일행의 외모나 자세는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 천의 가장자리를 움켜쥐고 있으며, 오른쪽 끝은 지주 체결부에도 고정되어 있다. 손과 고정부 사이에 생긴 당김과 천의 처짐이 자연스럽다. 상체의 전방 기울기와 굽힌 팔 관절도 그늘막을 당겨 유지하는 동작으로 가능하다. 발은 보이지 않지만 몸이 지지 없이 떠 있다고 볼 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 사선 시점과 아래쪽으로 기울인 얼굴이 잠든 일행을 돌보는 상체 숏에 더 가깝지만, 그늘막과 적재함 측판의 비중이 크고 미소는 다소 은은하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "양손으로 그늘막을 잡는 동작과 기계 정체성은 맞지만, 정면에 가까운 얼굴이 렌즈를 향하는 인상을 주어 요구된 측면 시점과 일행을 향한 미소의 관계가 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 머리는 화면 오른쪽 아래로 기울어 있고 얼굴 면도 그쪽을 향한다. 발광하는 점 형태의 눈만으로 시선을 확정하기는 어렵지만, 렌즈 정면보다는 그늘막 너머 화면 아래의 잠자리 쪽을 보는 자세에 가깝다. 양손은 아래쪽 그늘막 가장자리를 잡고 있으며 천은 손에서 오른쪽 아래로 이어진다. 잠든 일행 자체는 보이지 않는다.",
        "built_space": "왼쪽에 긴 금속 지주 하나, 왼손 아래 천 고리 부근에 짧게 드러난 금속 지지부 하나가 보인다. 녹슨 적재함 측판 한 구간이 하단을 비스듬히 가로지르고, 층상 암벽과 바다가 배경에 있다. 화면 위쪽에도 그늘막 일부가 남아 있으나 아래 천과 별개의 중복 설비인지는 확인되지 않는다. 찰리는 측판 뒤에 있으며 낮은 사선 카메라 위치가 읽힌다. 다만 측판이 요구된 좁은 가장자리보다 넓게 보인다.",
        "entities": "보이는 인물은 찰리 하나뿐이다. 샌드 베이지색의 긁히고 마모된 장갑판, 육중한 팔, 흰 마스크, 주황색 점 눈 두 개와 검은 입 선, 푸르게 빛나는 원형 가슴 부품이 캐릭터 참조와 부합한다. 인간 피부나 치아는 없다. 입 선은 부드러운 미소이지만 환한 미소 이모티콘이라는 지시보다는 절제되어 있다. 녹색 망사 그늘막과 낡은 흰색 트럭, 해안 암벽은 장소 참조와 일치한다. 잠든 세 사람과 다리는 프레임 밖이므로 평가 대상이 아니다.",
        "hard_violations": [],
        "physics": "양쪽 기계 손가락이 천의 윗단을 감싸 실제로 붙잡고 있다. 그늘막에는 손 사이의 장력과 아래로 처지는 주름이 보이며, 상단 천은 지주가 있는 차양 구조와 양립한다. 팔은 어깨와 팔꿈치 관절에 연결되어 있다. 하체의 접지점은 측판과 화면 경계에 가려져 있으나 상체가 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "A",
        "direction": "머리는 아래로 숙여져 있지만 얼굴과 두 눈은 카메라 쪽에 거의 정면으로 제시된다. 아래쪽 일행을 비껴서 관찰하는 측면 시점보다는 관객을 내려다보는 인상이 강하다. 양팔은 좌우로 벌어져 각각 천의 가장자리를 잡으며, 천은 손에서 화면 아래와 오른쪽 지주로 이어진다. 시선의 실제 목표가 잠든 일행이라고 확인할 단서는 부족하다.",
        "built_space": "오른쪽에 금속 지주 하나와 천을 고정하는 체결부가 보인다. 전경의 녹슨 측판 한 구간과 오른쪽으로 이어지는 측판 한 구간이 적재함 모서리를 구성한다. 찰리는 그 뒤에 있고, 양옆의 층상 암벽 사이로 바다가 보인다. 낮은 카메라 위치는 맞지만 찰리를 거의 정면으로 바라본다. 그늘막이 화면 하부를 넓게 채우고 측판도 두껍게 들어와, 소량의 천과 좁은 적재함 가장자리라는 구도에서 벗어난다.",
        "entities": "찰리 한 기만 보이며 추가 인물은 없다. 흰 마스크와 점 눈 두 개, 검은 입 선, 마모된 베이지 장갑판, 긴 기계 팔과 푸른 원형 가슴 부품은 참조의 정체성을 유지한다. 미소는 보이지만 환하게 웃는 표정보다는 작은 미소에 가깝다. 녹색 망사 천과 부식된 트럭 측판, 암벽과 바다는 지정된 장소의 재질과 일치한다. 가려진 잠든 일행의 외모나 자세는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 천의 가장자리를 움켜쥐고 있으며, 오른쪽 끝은 지주 체결부에도 고정되어 있다. 손과 고정부 사이에 생긴 당김과 천의 처짐이 자연스럽다. 상체의 전방 기울기와 굽힌 팔 관절도 그늘막을 당겨 유지하는 동작으로 가능하다. 발은 보이지 않지만 몸이 지지 없이 떠 있다고 볼 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.571
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "화면 하단에만 그늘막을 배치하라는 구도 지시를 정확히 따랐으며, 두 손의 물리적 상호작용이 자연스럽습니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "구도 지시와 달리 화면 우측 상단에 큰 천막이 침범하였고, 천을 쥔 손가락의 묘사가 부자연스럽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S69sh2_sel.png",
    "asset_id": "ecdea7fe-9859-47e7-b6ba-478040a8e48a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d2f-415c-767d-b85f-1165bd013117",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S69sh2"
  }
 },
 "S70sh8::signage": {
  "fp": "e9c95b677ae18d34",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::aea9422aa337aaa2": {
  "subjects": [],
  "subject_text": "군 병원 복도·병실 문 앞\n군사 시설 내의 정돈된 의료 통로 공간이다.",
  "identity": "canonical",
  "scope_id": "L113",
  "scope_role": "location_interior",
  "scope_sha": "89847c12bfb2a23e"
 },
 "S70sh8::bgfirst_bg": {
  "input_fingerprint": "6d16ea264ec251a9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 침대 위, 한쪽 눈과 팔다리에 두꺼운 깁스를 감은 채 꼼짝 못하고 누워있는 박철진의 처참한 전신.\n\nLOCATION (lock): At a patient's bed inside a military-hospital ward, illuminated for nighttime care.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane reveal beyond and to one side of the bed's foot, looking diagonally downward along 박철진's entire immobilized body. Let the body run from the lower-right feet toward the upper-left head, clearly separating the thick arm and leg casts and the covered injured eye, while keeping the bed subordinate to the patient. His remaining eye strains upward toward the minister outside the frame; the rigid posture reads as the consequence of severe injury, not deliberate stillness.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by the severely injured 박철진) — Foot and one long side visible beneath the oblique full-body view; used as Establishes the patient's helpless horizontal position and the side used for the next move; Arm and leg casts (Thick casts immobilizing the injured limbs) — Their outer contours remain separately visible rather than overlapping in projection; used as Makes the extent of immobilization legible in the full-body reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior ambient illumination appropriate to the nighttime hospital, with controlled contrast and no stylized color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 침대 위, 한쪽 눈과 팔다리에 두꺼운 깁스를 감은 채 꼼짝 못하고 누워있는 박철진의 처참한 전신.\n\nLOCATION (lock): At a patient's bed inside a military-hospital ward, illuminated for nighttime care.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane reveal beyond and to one side of the bed's foot, looking diagonally downward along 박철진's entire immobilized body. Let the body run from the lower-right feet toward the upper-left head, clearly separating the thick arm and leg casts and the covered injured eye, while keeping the bed subordinate to the patient. His remaining eye strains upward toward the minister outside the frame; the rigid posture reads as the consequence of severe injury, not deliberate stillness.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by the severely injured 박철진) — Foot and one long side visible beneath the oblique full-body view; used as Establishes the patient's helpless horizontal position and the side used for the next move; Arm and leg casts (Thick casts immobilizing the injured limbs) — Their outer contours remain separately visible rather than overlapping in projection; used as Makes the extent of immobilization legible in the full-body reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior ambient illumination appropriate to the nighttime hospital, with controlled contrast and no stylized color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S70sh8__bgfirst_bg.png",
  "asset_id": "6efdab49-d170-4857-8312-56183482dce5",
  "input_asset_ids": [
   "959e7c4a-cff6-4e0d-bfd9-7f270432ea80",
   "85a5a8f3-ccce-4681-8741-a11b7aba26d3"
  ]
 },
 "S70sh8": {
  "input_fingerprint": "cb329e9b98c57413",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위, 한쪽 눈과 팔다리에 두꺼운 깁스를 감은 채 꼼짝 못하고 누워있는 박철진의 처참한 전신.\n\nLOCATION (lock): At a patient's bed inside a military-hospital ward, illuminated for nighttime care. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane reveal beyond and to one side of the bed's foot, looking diagonally downward along 박철진's entire immobilized body. Let the body run from the lower-right feet toward the upper-left head, clearly separating the thick arm and leg casts and the covered injured eye, while keeping the bed subordinate to the patient. His remaining eye strains upward toward the minister outside the frame; the rigid posture reads as the consequence of severe injury, not deliberate stillness.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by the severely injured 박철진) — Foot and one long side visible beneath the oblique full-body view; used as Establishes the patient's helpless horizontal position and the side used for the next move; Arm and leg casts (Thick casts immobilizing the injured limbs) — Their outer contours remain separately visible rather than overlapping in projection; used as Makes the extent of immobilization legible in the full-body reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior ambient illumination appropriate to the nighttime hospital, with controlled contrast and no stylized color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting remains a military hospital at night, with a bed in the ward. 박철진: He lies in bed under treatment, with one eye covered and an arm and leg immobilized in casts. His consciousness is faint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위, 한쪽 눈과 팔다리에 두꺼운 깁스를 감은 채 꼼짝 못하고 누워있는 박철진의 처참한 전신.\n\nLOCATION (lock): At a patient's bed inside a military-hospital ward, illuminated for nighttime care. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane reveal beyond and to one side of the bed's foot, looking diagonally downward along 박철진's entire immobilized body. Let the body run from the lower-right feet toward the upper-left head, clearly separating the thick arm and leg casts and the covered injured eye, while keeping the bed subordinate to the patient. His remaining eye strains upward toward the minister outside the frame; the rigid posture reads as the consequence of severe injury, not deliberate stillness.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by the severely injured 박철진) — Foot and one long side visible beneath the oblique full-body view; used as Establishes the patient's helpless horizontal position and the side used for the next move; Arm and leg casts (Thick casts immobilizing the injured limbs) — Their outer contours remain separately visible rather than overlapping in projection; used as Makes the extent of immobilization legible in the full-body reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior ambient illumination appropriate to the nighttime hospital, with controlled contrast and no stylized color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting remains a military hospital at night, with a bed in the ward. 박철진: He lies in bed under treatment, with one eye covered and an arm and leg immobilized in casts. His consciousness is faint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위, 한쪽 눈과 팔다리에 두꺼운 깁스를 감은 채 꼼짝 못하고 누워있는 박철진의 처참한 전신.\n\nLOCATION (lock): At a patient's bed inside a military-hospital ward, illuminated for nighttime care. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane reveal beyond and to one side of the bed's foot, looking diagonally downward along 박철진's entire immobilized body. Let the body run from the lower-right feet toward the upper-left head, clearly separating the thick arm and leg casts and the covered injured eye, while keeping the bed subordinate to the patient. His remaining eye strains upward toward the minister outside the frame; the rigid posture reads as the consequence of severe injury, not deliberate stillness.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hospital bed (Occupied by the severely injured 박철진) — Foot and one long side visible beneath the oblique full-body view; used as Establishes the patient's helpless horizontal position and the side used for the next move; Arm and leg casts (Thick casts immobilizing the injured limbs) — Their outer contours remain separately visible rather than overlapping in projection; used as Makes the extent of immobilization legible in the full-body reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained interior ambient illumination appropriate to the nighttime hospital, with controlled contrast and no stylized color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting remains a military hospital at night, with a bed in the ward. 박철진: He lies in bed under treatment, with one eye covered and an arm and leg immobilized in casts. His consciousness is faint.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S70sh8__bgfirst_bg.png",
     "asset_id": "6efdab49-d170-4857-8312-56183482dce5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S70sh8.png",
     "asset_id": "959e7c4a-cff6-4e0d-bfd9-7f270432ea80",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1411720>",
     "asset_id": "b9c6497f-2ecb-48b6-9a70-a0073dfdd650",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L113B01.png",
     "asset_id": "85a5a8f3-ccce-4681-8741-a11b7aba26d3",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1411720>",
     "asset_id": "b9c6497f-2ecb-48b6-9a70-a0073dfdd650",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "환자의 남은 한쪽 눈 시선이 침대 옆에 서 있는 정장 차림의 남성을 향하고 있음.",
    "built_space": "야간 병실 내부. 환자는 우측 병상에 누워 있으며, 좌측 전경에 지시문에 어긋나는 인물이 큰 비중으로 배치됨.",
    "entities": "박철진(머리 붕대, 팔과 다리 깁스), 프레임 안에 임의로 생성된 남성(장관).",
    "hard_violations": [
     "[gemini-pro] invented people (프레임 밖에 있어야 할 장관을 화면 안에 추가함)",
     "[gpt-high] 화면 밖에 있어야 하고 이 샷의 등장인물로 허용되지 않은 장관으로 보이는 남성을 왼쪽 전경에 추가했다."
    ],
    "physics": "환자의 신체가 병상 위에 안정적으로 뉘어져 있음."
   },
   {
    "label": "B",
    "direction": "환자의 시선이 화면 좌측 상단 프레임 밖을 향해 긴장감 있게 올라가 있음.",
    "built_space": "야간 병실 내부. 카메라는 침대 발치에서 머리 쪽을 향해 하향 대각선으로 바라보며, 참조 이미지와 일치하는 병상과 벽면 구조가 나타남.",
    "entities": "단독으로 배치된 박철진(한쪽 눈 안대, 우측 팔 깁스와 슬링, 좌측 다리 두꺼운 깁스).",
    "hard_violations": [
     "[gpt-high] 참조와 본문에 없는 환자 감시장치 및 주입 펌프를 추가했으며, 감시장치 화면에 허용되지 않은 새 숫자·문자 표시도 생성했다."
    ],
    "physics": "팔과 다리의 무거운 깁스가 각각 몸통과 침대 매트리스 위에 중력에 맞게 자연스럽게 얹혀 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레임 밖에 있어야 할 장관을 화면 내에 추가하여 지정 인물 외 등장 금지 조건을 명백히 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임 밖을 향하는 시선, 대각선 카메라 구도, 단독 인물 조건 및 깁스 묘사를 모두 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "환자의 남은 한쪽 눈 시선이 침대 옆에 서 있는 정장 차림의 남성을 향하고 있음.",
        "built_space": "야간 병실 내부. 환자는 우측 병상에 누워 있으며, 좌측 전경에 지시문에 어긋나는 인물이 큰 비중으로 배치됨.",
        "entities": "박철진(머리 붕대, 팔과 다리 깁스), 프레임 안에 임의로 생성된 남성(장관).",
        "hard_violations": [
         "invented people (프레임 밖에 있어야 할 장관을 화면 안에 추가함)"
        ],
        "physics": "환자의 신체가 병상 위에 안정적으로 뉘어져 있음."
       },
       {
        "label": "B",
        "direction": "환자의 시선이 화면 좌측 상단 프레임 밖을 향해 긴장감 있게 올라가 있음.",
        "built_space": "야간 병실 내부. 카메라는 침대 발치에서 머리 쪽을 향해 하향 대각선으로 바라보며, 참조 이미지와 일치하는 병상과 벽면 구조가 나타남.",
        "entities": "단독으로 배치된 박철진(한쪽 눈 안대, 우측 팔 깁스와 슬링, 좌측 다리 두꺼운 깁스).",
        "hard_violations": [],
        "physics": "팔과 다리의 무거운 깁스가 각각 몸통과 침대 매트리스 위에 중력에 맞게 자연스럽게 얹혀 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레임 밖에 있어야 할 장관을 화면 내에 추가하여 지정 인물 외 등장 금지 조건을 명백히 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임 밖을 향하는 시선, 대각선 카메라 구도, 단독 인물 조건 및 깁스 묘사를 모두 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "환자의 남은 한쪽 눈 시선이 침대 옆에 서 있는 정장 차림의 남성을 향하고 있음.",
        "built_space": "야간 병실 내부. 환자는 우측 병상에 누워 있으며, 좌측 전경에 지시문에 어긋나는 인물이 큰 비중으로 배치됨.",
        "entities": "박철진(머리 붕대, 팔과 다리 깁스), 프레임 안에 임의로 생성된 남성(장관).",
        "hard_violations": [
         "invented people (프레임 밖에 있어야 할 장관을 화면 안에 추가함)"
        ],
        "physics": "환자의 신체가 병상 위에 안정적으로 뉘어져 있음."
       },
       {
        "label": "B",
        "direction": "환자의 시선이 화면 좌측 상단 프레임 밖을 향해 긴장감 있게 올라가 있음.",
        "built_space": "야간 병실 내부. 카메라는 침대 발치에서 머리 쪽을 향해 하향 대각선으로 바라보며, 참조 이미지와 일치하는 병상과 벽면 구조가 나타남.",
        "entities": "단독으로 배치된 박철진(한쪽 눈 안대, 우측 팔 깁스와 슬링, 좌측 다리 두꺼운 깁스).",
        "hard_violations": [],
        "physics": "팔과 다리의 무거운 깁스가 각각 몸통과 침대 매트리스 위에 중력에 맞게 자연스럽게 얹혀 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "발치 옆의 하향 사선 구도와 전신·깁스 분리는 가장 충실하지만, 참조에 없는 감시장치와 표시 문자를 추가해 엄격한 조건에서는 부적격이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "화면 밖에 있어야 할 장관을 전경에 추가했고, 한쪽 눈도 제대로 가리지 않아 환자 단독 전신 공개라는 핵심 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "환자의 머리는 화면 왼쪽 위, 발은 오른쪽 아래로 놓였다. 노출된 눈은 화면 위쪽 오른편의 화면 밖을 향해 올라가 있어, 침상 옆 장관을 올려다보라는 지시와 부합한다. 무기나 이동 동작은 없다.",
        "built_space": "침대 한 개의 머리판·발판과 긴 측면 난간이 보인다. 머리맡 벽면 설비대와 조명 각 한 개, 왼쪽 창과 협탁 각 한 개, 오른쪽 칸막이 커튼이 있어 참조 병실의 주요 재료와 배치를 따른다. 다만 머리 오른편 감시장치와 왼편 주입 펌프가 추가되었다. 카메라는 발치 한쪽 위에서 환자 전체를 비스듬히 내려다보며, 환자가 프레임의 중심을 차지한다.",
        "entities": "짧은 검은 머리와 얼굴 상처가 있는 중년 동아시아계 남성 한 명으로, 박철진의 외형에 대체로 맞는다. 한쪽 눈에는 실제 천 안대와 붕대가 있고, 양팔과 양다리에 두꺼운 고정물이 보이며 각각의 외곽이 구분된다. 참조의 어두운 작업복 대신 무늬 있는 환자복을 입었다. 장관은 보이지 않는다. 감시장치에는 참조가 정하지 않은 숫자와 표시가 보인다.",
        "hard_violations": [
         "참조와 본문에 없는 환자 감시장치 및 주입 펌프를 추가했으며, 감시장치 화면에 허용되지 않은 새 숫자·문자 표시도 생성했다."
        ],
        "physics": "머리와 목은 베개, 몸통과 다리는 매트리스에 받쳐져 있다. 몸 위로 접힌 깁스 팔은 슬링과 몸통에 지지되고, 반대쪽 팔과 손은 침구 위에 놓였다. 다리 깁스와 발뒤꿈치도 침상에 닿아 있으며, 지지 없이 떠 있는 신체 부위는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "환자는 화면 왼쪽 전경의 정장 남성을 올려다보고, 정장 남성은 환자 쪽으로 고개를 숙인다. 시선 관계 자체는 읽히지만, 목표 인물을 화면 밖에 두라는 지시와 반대다. 몸의 방향은 왼쪽 위 머리에서 오른쪽 아래 발로 이어진다.",
        "built_space": "침대 한 개, 머리판과 발판, 양쪽 난간, 벽면 설비대와 조명 각 한 개, 협탁·창·수액대 각 한 개 및 오른쪽 커튼이 보인다. 참조 병실의 기본 설비는 대체로 유지한다. 그러나 카메라가 더 낮고 전경 남성이 왼쪽 화면을 크게 차지해, 환자 단독의 하향 전신 공개보다 두 사람의 대면 장면에 가깝다.",
        "entities": "침대에는 검은 머리의 중년 동아시아계 남성이 있으며, 얼굴 상처와 어두운 작업복, 붉은 완장은 인물 참조에 가깝다. 한 팔과 한 다리의 두꺼운 깁스는 구분되지만, 반대 다리는 담요 아래에 가려져 있다. 머리 붕대는 이마와 관자놀이를 감싸고 두 눈은 노출되어 보여, 한쪽 눈을 가리라는 조건을 충족하지 못한다. 허용된 환자 외에 정장 차림 성인 남성 한 명이 추가되었다.",
        "hard_violations": [
         "화면 밖에 있어야 하고 이 샷의 등장인물로 허용되지 않은 장관으로 보이는 남성을 왼쪽 전경에 추가했다."
        ],
        "physics": "환자의 머리는 베개에, 몸통과 다리는 침상에 놓여 있다. 깁스 팔은 복부와 담요 위에 받쳐지고 맨손도 담요 위에 쉬고 있어 중력에 맞는다. 전경 남성은 서 있는 자세이며 발은 프레임 밖이지만, 몸이 공중에 떠 있다고 볼 단서는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "발치 옆의 하향 사선 구도와 전신·깁스 분리는 가장 충실하지만, 참조에 없는 감시장치와 표시 문자를 추가해 엄격한 조건에서는 부적격이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "화면 밖에 있어야 할 장관을 전경에 추가했고, 한쪽 눈도 제대로 가리지 않아 환자 단독 전신 공개라는 핵심 지시를 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "환자의 머리는 화면 왼쪽 위, 발은 오른쪽 아래로 놓였다. 노출된 눈은 화면 위쪽 오른편의 화면 밖을 향해 올라가 있어, 침상 옆 장관을 올려다보라는 지시와 부합한다. 무기나 이동 동작은 없다.",
        "built_space": "침대 한 개의 머리판·발판과 긴 측면 난간이 보인다. 머리맡 벽면 설비대와 조명 각 한 개, 왼쪽 창과 협탁 각 한 개, 오른쪽 칸막이 커튼이 있어 참조 병실의 주요 재료와 배치를 따른다. 다만 머리 오른편 감시장치와 왼편 주입 펌프가 추가되었다. 카메라는 발치 한쪽 위에서 환자 전체를 비스듬히 내려다보며, 환자가 프레임의 중심을 차지한다.",
        "entities": "짧은 검은 머리와 얼굴 상처가 있는 중년 동아시아계 남성 한 명으로, 박철진의 외형에 대체로 맞는다. 한쪽 눈에는 실제 천 안대와 붕대가 있고, 양팔과 양다리에 두꺼운 고정물이 보이며 각각의 외곽이 구분된다. 참조의 어두운 작업복 대신 무늬 있는 환자복을 입었다. 장관은 보이지 않는다. 감시장치에는 참조가 정하지 않은 숫자와 표시가 보인다.",
        "hard_violations": [
         "참조와 본문에 없는 환자 감시장치 및 주입 펌프를 추가했으며, 감시장치 화면에 허용되지 않은 새 숫자·문자 표시도 생성했다."
        ],
        "physics": "머리와 목은 베개, 몸통과 다리는 매트리스에 받쳐져 있다. 몸 위로 접힌 깁스 팔은 슬링과 몸통에 지지되고, 반대쪽 팔과 손은 침구 위에 놓였다. 다리 깁스와 발뒤꿈치도 침상에 닿아 있으며, 지지 없이 떠 있는 신체 부위는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "환자는 화면 왼쪽 전경의 정장 남성을 올려다보고, 정장 남성은 환자 쪽으로 고개를 숙인다. 시선 관계 자체는 읽히지만, 목표 인물을 화면 밖에 두라는 지시와 반대다. 몸의 방향은 왼쪽 위 머리에서 오른쪽 아래 발로 이어진다.",
        "built_space": "침대 한 개, 머리판과 발판, 양쪽 난간, 벽면 설비대와 조명 각 한 개, 협탁·창·수액대 각 한 개 및 오른쪽 커튼이 보인다. 참조 병실의 기본 설비는 대체로 유지한다. 그러나 카메라가 더 낮고 전경 남성이 왼쪽 화면을 크게 차지해, 환자 단독의 하향 전신 공개보다 두 사람의 대면 장면에 가깝다.",
        "entities": "침대에는 검은 머리의 중년 동아시아계 남성이 있으며, 얼굴 상처와 어두운 작업복, 붉은 완장은 인물 참조에 가깝다. 한 팔과 한 다리의 두꺼운 깁스는 구분되지만, 반대 다리는 담요 아래에 가려져 있다. 머리 붕대는 이마와 관자놀이를 감싸고 두 눈은 노출되어 보여, 한쪽 눈을 가리라는 조건을 충족하지 못한다. 허용된 환자 외에 정장 차림 성인 남성 한 명이 추가되었다.",
        "hard_violations": [
         "화면 밖에 있어야 하고 이 샷의 등장인물로 허용되지 않은 장관으로 보이는 남성을 왼쪽 전경에 추가했다."
        ],
        "physics": "환자의 머리는 베개에, 몸통과 다리는 침상에 놓여 있다. 깁스 팔은 복부와 담요 위에 받쳐지고 맨손도 담요 위에 쉬고 있어 중력에 맞는다. 전경 남성은 서 있는 자세이며 발은 프레임 밖이지만, 몸이 공중에 떠 있다고 볼 단서는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.829,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.579,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (프레임 밖에 있어야 할 장관을 화면 안에 추가함)",
     "[gpt-high] 화면 밖에 있어야 하고 이 샷의 등장인물로 허용되지 않은 장관으로 보이는 남성을 왼쪽 전경에 추가했다."
    ],
    "B": [
     "[gpt-high] 참조와 본문에 없는 환자 감시장치 및 주입 펌프를 추가했으며, 감시장치 화면에 허용되지 않은 새 숫자·문자 표시도 생성했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 579,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 579,
    "verdict_ko": "프레임 밖에 있어야 할 장관을 화면 내에 추가하여 지정 인물 외 등장 금지 조건을 명백히 위반했습니다.  ★위반: [gemini-pro] invented people (프레임 밖에 있어야 할 장관을 화면 안에 추가함) / [gpt-high] 화면 밖에 있어야 하고 이 샷의 등장인물로 허용되지 않은 장관으로 보이는 남성을 왼쪽 전경에 추가했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "프레임 밖을 향하는 시선, 대각선 카메라 구도, 단독 인물 조건 및 깁스 묘사를 모두 정확하게 구현했습니다.  ★위반: [gpt-high] 참조와 본문에 없는 환자 감시장치 및 주입 펌프를 추가했으며, 감시장치 화면에 허용되지 않은 새 숫자·문자 표시도 생성했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L113B01.png",
    "asset_id": "85a5a8f3-ccce-4681-8741-a11b7aba26d3",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1411720>",
    "asset_id": "b9c6497f-2ecb-48b6-9a70-a0073dfdd650",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d34-ebda-733b-b101-6e31c2839297",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S70sh8__bgfirst_bg.png",
   "bg_asset_id": "6efdab49-d170-4857-8312-56183482dce5",
   "bg_record_key": "S70sh8::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S70sh10::signage": {
  "fp": "5d8c4bb2f8ea5987",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S70sh10": {
  "input_fingerprint": "928215c8f67df000",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진을 향해 손가락을 뻗은 채 차갑게 호통치듯 입을 벌리고 있는 국방장관의 매정한 상체.\n\nLOCATION (lock): At the military-hospital ward doorway facing the injured patient's bed, under nighttime hospital lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the established bedside endpoint beside 박철진's upper torso, settle just above mattress height and look obliquely upward toward 국방장관 at the doorway, staying offset from his line of address. Frame the minister's upper body on the right with his pointing hand extending diagonally down toward the patient's cropped shoulder at the lower-left edge, keeping the hand naturally proportioned. 국방장관 looks past the lens toward the patient's face as he delivers the dismissal, while 박철진's face remains outside the crop and his body stays pinned to the bed by injury.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital room doorway (Visible behind the minister as the route of departure) — Seen obliquely from the bedside rather than square-on; used as Places the minister between the patient and his impending abandonment; Bed beside the patient's shoulder (Supporting the immobilized patient) — A narrow mattress edge crosses the lower foreground; used as Anchors the low viewpoint without implying the patient's literal optical POV.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained hospital illumination, allowing expression and the pointing gesture rather than a lighting change to convey cruelty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital ward and its bed remain unchanged at night. 박철진: He remains in bed with one eye covered and an arm and leg in casts, barely conscious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진을 향해 손가락을 뻗은 채 차갑게 호통치듯 입을 벌리고 있는 국방장관의 매정한 상체.\n\nLOCATION (lock): At the military-hospital ward doorway facing the injured patient's bed, under nighttime hospital lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the established bedside endpoint beside 박철진's upper torso, settle just above mattress height and look obliquely upward toward 국방장관 at the doorway, staying offset from his line of address. Frame the minister's upper body on the right with his pointing hand extending diagonally down toward the patient's cropped shoulder at the lower-left edge, keeping the hand naturally proportioned. 국방장관 looks past the lens toward the patient's face as he delivers the dismissal, while 박철진's face remains outside the crop and his body stays pinned to the bed by injury.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital room doorway (Visible behind the minister as the route of departure) — Seen obliquely from the bedside rather than square-on; used as Places the minister between the patient and his impending abandonment; Bed beside the patient's shoulder (Supporting the immobilized patient) — A narrow mattress edge crosses the lower foreground; used as Anchors the low viewpoint without implying the patient's literal optical POV.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained hospital illumination, allowing expression and the pointing gesture rather than a lighting change to convey cruelty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital ward and its bed remain unchanged at night. 박철진: He remains in bed with one eye covered and an arm and leg in casts, barely conscious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박철진을 향해 손가락을 뻗은 채 차갑게 호통치듯 입을 벌리고 있는 국방장관의 매정한 상체.\n\nLOCATION (lock): At the military-hospital ward doorway facing the injured patient's bed, under nighttime hospital lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the established bedside endpoint beside 박철진's upper torso, settle just above mattress height and look obliquely upward toward 국방장관 at the doorway, staying offset from his line of address. Frame the minister's upper body on the right with his pointing hand extending diagonally down toward the patient's cropped shoulder at the lower-left edge, keeping the hand naturally proportioned. 국방장관 looks past the lens toward the patient's face as he delivers the dismissal, while 박철진's face remains outside the crop and his body stays pinned to the bed by injury.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Hospital room doorway (Visible behind the minister as the route of departure) — Seen obliquely from the bedside rather than square-on; used as Places the minister between the patient and his impending abandonment; Bed beside the patient's shoulder (Supporting the immobilized patient) — A narrow mattress edge crosses the lower foreground; used as Anchors the low viewpoint without implying the patient's literal optical POV.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained hospital illumination, allowing expression and the pointing gesture rather than a lighting change to convey cruelty.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital ward and its bed remain unchanged at night. 박철진: He remains in bed with one eye covered and an arm and leg in casts, barely conscious.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 국방장관 (한국인 남성, 성숙한 얼굴, 정돈된 짧은 검은 머리) — wearing: 권위적인 느낌을 주는 짙은 네이비 색상의 고급 양복 정장과 화이트 셔츠.; 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "국방장관은 왼쪽 아래의 박철진을 향해 시선을 두고 오른손 검지로 정확히 가리킴.",
    "built_space": "전경 하단에 침대 난간이 가로지르며, 배경 우측에 열린 병실 문이 위치함.",
    "entities": "국방장관은 지정된 네이비 정장과 외모를 따름. 박철진은 안대를 착용했으나 지시와 달리 얼굴이 화면에 포함됨.",
    "hard_violations": [],
    "physics": "국방장관은 안정적으로 서서 팔을 뻗고 있으며, 환자는 침대에 누워 지탱됨."
   },
   {
    "label": "B",
    "direction": "국방장관은 왼쪽 아래를 향해 팔을 뻗었으나, 가리키는 손가락의 방향과 형태가 불분명함.",
    "built_space": "전경에 침대 난간, 좌측에 커튼, 배경에 열린 병실 문이 배치됨.",
    "entities": "국방장관은 정장을 착용함. 박철진은 지시를 어기고 얼굴 일부가 프레임에 노출됨.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (마디가 잘린 듯 뭉개진 검지 손가락)"
    ],
    "physics": "환자는 침대에 누워 있으나, 장관의 뻗은 손은 인체공학적으로 불가능한 형태를 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "박철진의 얼굴을 프레임에서 배제하라는 지시를 어겼으나, 침대 난간의 레퍼런스 연속성과 손가락 형태를 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "박철진의 얼굴이 프레임에 노출되었으며, 가리키는 손가락이 기형적으로 묘사되어 해부학적 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "국방장관은 왼쪽 아래의 박철진을 향해 시선을 두고 오른손 검지로 정확히 가리킴.",
        "built_space": "전경 하단에 침대 난간이 가로지르며, 배경 우측에 열린 병실 문이 위치함.",
        "entities": "국방장관은 지정된 네이비 정장과 외모를 따름. 박철진은 안대를 착용했으나 지시와 달리 얼굴이 화면에 포함됨.",
        "hard_violations": [],
        "physics": "국방장관은 안정적으로 서서 팔을 뻗고 있으며, 환자는 침대에 누워 지탱됨."
       },
       {
        "label": "B",
        "direction": "국방장관은 왼쪽 아래를 향해 팔을 뻗었으나, 가리키는 손가락의 방향과 형태가 불분명함.",
        "built_space": "전경에 침대 난간, 좌측에 커튼, 배경에 열린 병실 문이 배치됨.",
        "entities": "국방장관은 정장을 착용함. 박철진은 지시를 어기고 얼굴 일부가 프레임에 노출됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (마디가 잘린 듯 뭉개진 검지 손가락)"
        ],
        "physics": "환자는 침대에 누워 있으나, 장관의 뻗은 손은 인체공학적으로 불가능한 형태를 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "박철진의 얼굴을 프레임에서 배제하라는 지시를 어겼으나, 침대 난간의 레퍼런스 연속성과 손가락 형태를 사실적으로 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "박철진의 얼굴이 프레임에 노출되었으며, 가리키는 손가락이 기형적으로 묘사되어 해부학적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "국방장관은 왼쪽 아래의 박철진을 향해 시선을 두고 오른손 검지로 정확히 가리킴.",
        "built_space": "전경 하단에 침대 난간이 가로지르며, 배경 우측에 열린 병실 문이 위치함.",
        "entities": "국방장관은 지정된 네이비 정장과 외모를 따름. 박철진은 안대를 착용했으나 지시와 달리 얼굴이 화면에 포함됨.",
        "hard_violations": [],
        "physics": "국방장관은 안정적으로 서서 팔을 뻗고 있으며, 환자는 침대에 누워 지탱됨."
       },
       {
        "label": "B",
        "direction": "국방장관은 왼쪽 아래를 향해 팔을 뻗었으나, 가리키는 손가락의 방향과 형태가 불분명함.",
        "built_space": "전경에 침대 난간, 좌측에 커튼, 배경에 열린 병실 문이 배치됨.",
        "entities": "국방장관은 정장을 착용함. 박철진은 지시를 어기고 얼굴 일부가 프레임에 노출됨.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (마디가 잘린 듯 뭉개진 검지 손가락)"
        ],
        "physics": "환자는 침대에 누워 있으나, 장관의 뻗은 손은 인체공학적으로 불가능한 형태를 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "환자를 향한 시선과 낮게 뻗은 손, 오른쪽 장관 상체는 더 충실하지만, 환자의 머리와 얼굴까지 보여 어깨만 남기라는 핵심 크롭을 어겼다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "장관의 외형과 호통치는 표정은 맞지만, 손가락이 환자 어깨보다 렌즈 쪽을 향하고 환자의 머리·몸통·침구를 넓게 보여 지정 구도에서 더 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장관의 눈은 렌즈 왼쪽에 있는 환자의 머리를 향한다. 오른팔은 왼쪽 아래로 뻗어 있으며 손가락은 단축되어 보이지만 환자의 상체 쪽을 지목하는 것으로 읽힌다. 다만 손끝이 왼쪽 아래 가장자리의 어깨로 명확히 이어지는 지정 대각선은 약하다. 환자는 장관 쪽으로 얼굴을 둔 채 누워 있다.",
        "built_space": "장관 뒤로 출입구 하나와 좁은 유리창이 있는 열린 문짝 하나가 보이며, 문은 비스듬히 관찰된다. 왼쪽에는 줄무늬 커튼 하나와 벽면 제어기 하나, 오른쪽에는 제어기 하나와 벽걸이 용기 하나가 보인다. 하단에는 침대 양쪽 난간 일부가 보인다. 커튼과 벽의 색감은 이전 병실과 유사하며, 참조에 출입구가 보이지 않아 문 주변 설비의 정확한 연속성은 확인할 수 없다. 낮은 침상 옆 시점과 오른쪽 장관 배치는 맞지만, 좁은 매트리스 가장자리 대신 환자의 머리와 큰 상체, 난간이 전경을 차지한다.",
        "entities": "장관 한 명과 침대에 누운 환자 한 명만 보인다. 장관은 성숙한 한국인 남성으로 보이고, 정돈된 검은 머리와 얼굴 윤곽이 인물 참조에 가깝다. 짙은 네이비 정장, 흰 셔츠, 푸른 넥타이와 타이바도 일치하며 입을 벌려 꾸짖고 있다. 환자는 검은 머리, 눈 주변 붕대, 무늬 있는 환자복을 유지한다. 얼굴 일부가 드러나지만 신원 전체를 확인할 만큼 보이지는 않는다. 다리와 석고붕대의 전체 상태는 크롭 밖이므로 평가하지 않는다. 추가 인물이나 사진 위에 얹힌 문구는 없다.",
        "hard_violations": [],
        "physics": "환자의 머리는 베개에, 몸통은 등받이가 올라간 침대에 지지되어 있다. 장관의 발은 프레임 밖이지만 몸통은 서 있는 사람의 자세로 이어지며 공중에 뜬 징후는 없다. 뻗은 팔과 손은 어깨·팔꿈치·손목으로 자연스럽게 연결되고 손의 크기도 과장되지 않았다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "장관의 시선은 화면 왼쪽으로 약간 벗어나 있으나 환자의 얼굴보다 렌즈 가까운 지점을 향하는 인상이 강하다. 검지는 강하게 단축되어 거의 카메라 정면을 가리키며, 왼쪽 아래 환자 어깨로 내려가는 방향이 아니다. 환자는 장관을 향해 얼굴을 둔 채 누워 있다.",
        "built_space": "장관 뒤에는 출입구 하나, 오른쪽으로 열린 유리창 달린 문짝 하나가 있다. 왼쪽 벽에는 벽걸이 용기 하나, 화면이 있는 제어기 하나, 스위치 판 하나가 보이고, 하단에는 침대 양쪽 난간 일부가 보인다. 문과 장관의 위치 관계는 가능한 구조이며 중복된 출입구나 불가능한 반사는 없다. 다만 이전 병실의 커튼은 이 크롭에 없고, 출입구 설비의 일치 여부는 참조로 확인할 수 없다. 환자의 머리와 넓은 몸통·담요가 전경을 채워, 어깨 일부와 좁은 매트리스 가장자리만 두라는 구도를 충족하지 못한다.",
        "entities": "장관 한 명과 환자 한 명이 보인다. 장관의 성숙한 남성 외모, 검은 머리, 네이비 정장과 흰 셔츠, 넥타이·타이바는 참조에 가깝고 입도 열려 있다. 환자의 검은 머리, 눈 주변 붕대, 환자복과 회색 담요는 이전 장면과 대체로 맞는다. 환자의 얼굴 일부가 불필요하게 프레임 안에 들어오며, 얼굴 전체의 동일성은 확인하기 어렵다. 가려진 팔과 다리의 석고붕대는 평가할 수 없다. 추가 인물이나 오버레이 문구는 없다.",
        "hard_violations": [],
        "physics": "환자는 베개와 기울어진 침대에 머리와 몸통을 기대고 있고, 담요는 몸과 매트리스 위에 놓여 있다. 장관은 앞으로 팔을 뻗은 정상적인 기립 자세이며, 보이는 팔과 손의 연결도 물리적으로 가능하다. 발이 크롭 밖이라는 이유로 부유 상태로 볼 근거는 없다. 지지 없는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "환자를 향한 시선과 낮게 뻗은 손, 오른쪽 장관 상체는 더 충실하지만, 환자의 머리와 얼굴까지 보여 어깨만 남기라는 핵심 크롭을 어겼다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "장관의 외형과 호통치는 표정은 맞지만, 손가락이 환자 어깨보다 렌즈 쪽을 향하고 환자의 머리·몸통·침구를 넓게 보여 지정 구도에서 더 벗어났다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "장관의 눈은 렌즈 왼쪽에 있는 환자의 머리를 향한다. 오른팔은 왼쪽 아래로 뻗어 있으며 손가락은 단축되어 보이지만 환자의 상체 쪽을 지목하는 것으로 읽힌다. 다만 손끝이 왼쪽 아래 가장자리의 어깨로 명확히 이어지는 지정 대각선은 약하다. 환자는 장관 쪽으로 얼굴을 둔 채 누워 있다.",
        "built_space": "장관 뒤로 출입구 하나와 좁은 유리창이 있는 열린 문짝 하나가 보이며, 문은 비스듬히 관찰된다. 왼쪽에는 줄무늬 커튼 하나와 벽면 제어기 하나, 오른쪽에는 제어기 하나와 벽걸이 용기 하나가 보인다. 하단에는 침대 양쪽 난간 일부가 보인다. 커튼과 벽의 색감은 이전 병실과 유사하며, 참조에 출입구가 보이지 않아 문 주변 설비의 정확한 연속성은 확인할 수 없다. 낮은 침상 옆 시점과 오른쪽 장관 배치는 맞지만, 좁은 매트리스 가장자리 대신 환자의 머리와 큰 상체, 난간이 전경을 차지한다.",
        "entities": "장관 한 명과 침대에 누운 환자 한 명만 보인다. 장관은 성숙한 한국인 남성으로 보이고, 정돈된 검은 머리와 얼굴 윤곽이 인물 참조에 가깝다. 짙은 네이비 정장, 흰 셔츠, 푸른 넥타이와 타이바도 일치하며 입을 벌려 꾸짖고 있다. 환자는 검은 머리, 눈 주변 붕대, 무늬 있는 환자복을 유지한다. 얼굴 일부가 드러나지만 신원 전체를 확인할 만큼 보이지는 않는다. 다리와 석고붕대의 전체 상태는 크롭 밖이므로 평가하지 않는다. 추가 인물이나 사진 위에 얹힌 문구는 없다.",
        "hard_violations": [],
        "physics": "환자의 머리는 베개에, 몸통은 등받이가 올라간 침대에 지지되어 있다. 장관의 발은 프레임 밖이지만 몸통은 서 있는 사람의 자세로 이어지며 공중에 뜬 징후는 없다. 뻗은 팔과 손은 어깨·팔꿈치·손목으로 자연스럽게 연결되고 손의 크기도 과장되지 않았다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "장관의 시선은 화면 왼쪽으로 약간 벗어나 있으나 환자의 얼굴보다 렌즈 가까운 지점을 향하는 인상이 강하다. 검지는 강하게 단축되어 거의 카메라 정면을 가리키며, 왼쪽 아래 환자 어깨로 내려가는 방향이 아니다. 환자는 장관을 향해 얼굴을 둔 채 누워 있다.",
        "built_space": "장관 뒤에는 출입구 하나, 오른쪽으로 열린 유리창 달린 문짝 하나가 있다. 왼쪽 벽에는 벽걸이 용기 하나, 화면이 있는 제어기 하나, 스위치 판 하나가 보이고, 하단에는 침대 양쪽 난간 일부가 보인다. 문과 장관의 위치 관계는 가능한 구조이며 중복된 출입구나 불가능한 반사는 없다. 다만 이전 병실의 커튼은 이 크롭에 없고, 출입구 설비의 일치 여부는 참조로 확인할 수 없다. 환자의 머리와 넓은 몸통·담요가 전경을 채워, 어깨 일부와 좁은 매트리스 가장자리만 두라는 구도를 충족하지 못한다.",
        "entities": "장관 한 명과 환자 한 명이 보인다. 장관의 성숙한 남성 외모, 검은 머리, 네이비 정장과 흰 셔츠, 넥타이·타이바는 참조에 가깝고 입도 열려 있다. 환자의 검은 머리, 눈 주변 붕대, 환자복과 회색 담요는 이전 장면과 대체로 맞는다. 환자의 얼굴 일부가 불필요하게 프레임 안에 들어오며, 얼굴 전체의 동일성은 확인하기 어렵다. 가려진 팔과 다리의 석고붕대는 평가할 수 없다. 추가 인물이나 오버레이 문구는 없다.",
        "hard_violations": [],
        "physics": "환자는 베개와 기울어진 침대에 머리와 몸통을 기대고 있고, 담요는 몸과 매트리스 위에 놓여 있다. 장관은 앞으로 팔을 뻗은 정상적인 기립 자세이며, 보이는 팔과 손의 연결도 물리적으로 가능하다. 발이 크롭 밖이라는 이유로 부유 상태로 볼 근거는 없다. 지지 없는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (마디가 잘린 듯 뭉개진 검지 손가락)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1833,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "박철진의 얼굴을 프레임에서 배제하라는 지시를 어겼으나, 침대 난간의 레퍼런스 연속성과 손가락 형태를 사실적으로 구현함."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "박철진의 얼굴이 프레임에 노출되었으며, 가리키는 손가락이 기형적으로 묘사되어 해부학적 오류가 발생함.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (마디가 잘린 듯 뭉개진 검지 손가락)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S70sh8_sel.png",
    "asset_id": "d3d6dfe0-d6e4-4e51-b440-c796f1abc923",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 국방장관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:805558>",
    "asset_id": "2ee9d902-fac0-42dc-986b-7bd16edba439",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박철진: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1411720>",
    "asset_id": "b9c6497f-2ecb-48b6-9a70-a0073dfdd650",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d3c-9030-7e4d-8a64-ebd4960a1e14",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S70sh8"
  },
  "staged_characters_added": [
   "C15"
  ]
 },
 "S70sh14::signage": {
  "fp": "346ecfab746059d8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S70sh14": {
  "input_fingerprint": "24914c0958ac4c58",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 눈물이 고인 채 분노로 일그러져 얼굴 근육이 잔뜩 경직된 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the injured patient's bed inside the military-hospital ward, under the continuing nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from the same bedside position, keeping the lens slightly above 박철진's face and angled down in an oblique three-quarter view; proximity, rather than another spatial reset, supplies the final emphasis. His face occupies most of the composition, with the covered eye retained at one side and the remaining bloodshot eye visibly holding tears. His gaze stays directed toward the off-screen doorway through which the men have left, while a tightened jaw and contracted cheek muscles turn helplessness into anger.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hospital bed beneath the head (Supporting the motionless patient) — Only a narrow, softly resolved portion remains around the face; used as Keeps the close-up grounded in physical helplessness while excluding the departed visitors.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established neutral hospital illumination, retaining subtle tear highlights and facial tension without a new dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, bedding, nearby room surfaces, and nighttime interior lighting from the reference. Exclude the minister and accompanying officer, who have left.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital bed remains in the nighttime ward. 박철진: He remains bedridden with one eye covered and an arm and leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 눈물이 고인 채 분노로 일그러져 얼굴 근육이 잔뜩 경직된 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the injured patient's bed inside the military-hospital ward, under the continuing nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from the same bedside position, keeping the lens slightly above 박철진's face and angled down in an oblique three-quarter view; proximity, rather than another spatial reset, supplies the final emphasis. His face occupies most of the composition, with the covered eye retained at one side and the remaining bloodshot eye visibly holding tears. His gaze stays directed toward the off-screen doorway through which the men have left, while a tightened jaw and contracted cheek muscles turn helplessness into anger.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hospital bed beneath the head (Supporting the motionless patient) — Only a narrow, softly resolved portion remains around the face; used as Keeps the close-up grounded in physical helplessness while excluding the departed visitors.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established neutral hospital illumination, retaining subtle tear highlights and facial tension without a new dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, bedding, nearby room surfaces, and nighttime interior lighting from the reference. Exclude the minister and accompanying officer, who have left.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital bed remains in the nighttime ward. 박철진: He remains bedridden with one eye covered and an arm and leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 눈물이 고인 채 분노로 일그러져 얼굴 근육이 잔뜩 경직된 박철진의 얼굴 클로즈업.\n\nLOCATION (lock): At the injured patient's bed inside the military-hospital ward, under the continuing nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the approach from the same bedside position, keeping the lens slightly above 박철진's face and angled down in an oblique three-quarter view; proximity, rather than another spatial reset, supplies the final emphasis. His face occupies most of the composition, with the covered eye retained at one side and the remaining bloodshot eye visibly holding tears. His gaze stays directed toward the off-screen doorway through which the men have left, while a tightened jaw and contracted cheek muscles turn helplessness into anger.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Hospital bed beneath the head (Supporting the motionless patient) — Only a narrow, softly resolved portion remains around the face; used as Keeps the close-up grounded in physical helplessness while excluding the departed visitors.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established neutral hospital illumination, retaining subtle tear highlights and facial tension without a new dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, bedding, nearby room surfaces, and nighttime interior lighting from the reference. Exclude the minister and accompanying officer, who have left.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Park Cheoljin is recumbent in a hospital bed, with one eye covered and his injured arm and leg immobilized in casts, looking up at the defense minister with his remaining eye. The scene text does not specify which side is injured, the angle of his torso and head, or the placement of his other arm and leg.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The military hospital bed remains in the nighttime ward. 박철진: He remains bedridden with one eye covered and an arm and leg in casts. His uncovered eye is bloodshot and filling with tears.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박철진 (한국인 남성, 40대 중반의 얼굴, 짧은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 밖 우측 상단(방문 쪽)을 향하고 있습니다.",
    "built_space": "머리를 받치고 있는 흰색 베개와 흐릿한 배경의 금속 침대 난간이 보입니다.",
    "entities": "박철진(40대 중반 한국인 남성). 우측 눈은 붕대로 덮여 있고, 좌측 눈은 핏발이 서고 눈물이 고여 있습니다. 턱과 뺨 근육이 극도로 긴장된 분노의 표정입니다. 환자복과 남색 팔걸이를 착용 중이나 가슴의 호스가 흰색 끈으로 변형되었습니다.",
    "hard_violations": [],
    "physics": "머리는 중력에 의해 베개에 자연스럽게 기대어 있습니다."
   },
   {
    "label": "B",
    "direction": "시선은 화면 밖 우측 상단을 향하고 있습니다.",
    "built_space": "머리를 받치고 있는 흰색 베개와 배경의 병원 침대 난간이 올바르게 위치합니다.",
    "entities": "박철진(40대 중반 한국인 남성). 우측 눈에 붕대, 좌측 눈에 눈물이 고여 있으나 표정은 근육 긴장 없이 이완되어 있습니다. 환자복, 남색 팔걸이, 가슴을 가로지르는 투명 호스가 레퍼런스와 일치합니다.",
    "hard_violations": [],
    "physics": "머리는 베개에 안정적으로 눕혀져 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 강력하게 요구한 분노로 일그러진 표정과 경직된 얼굴 근육을 매우 훌륭하게 연출했으나, 가슴을 지나는 투명 호스가 엉뚱한 캔버스 스트랩으로 대체된 점은 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스의 투명 호스와 의상 디테일은 정확히 재현했지만, 핵심 지시 사항인 '분노로 턱을 앙다물고 근육이 경직된 표정'을 전혀 반영하지 못하고 이완된 슬픈 표정에 머물렀습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 밖 우측 상단(방문 쪽)을 향하고 있습니다.",
        "built_space": "머리를 받치고 있는 흰색 베개와 흐릿한 배경의 금속 침대 난간이 보입니다.",
        "entities": "박철진(40대 중반 한국인 남성). 우측 눈은 붕대로 덮여 있고, 좌측 눈은 핏발이 서고 눈물이 고여 있습니다. 턱과 뺨 근육이 극도로 긴장된 분노의 표정입니다. 환자복과 남색 팔걸이를 착용 중이나 가슴의 호스가 흰색 끈으로 변형되었습니다.",
        "hard_violations": [],
        "physics": "머리는 중력에 의해 베개에 자연스럽게 기대어 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 밖 우측 상단을 향하고 있습니다.",
        "built_space": "머리를 받치고 있는 흰색 베개와 배경의 병원 침대 난간이 올바르게 위치합니다.",
        "entities": "박철진(40대 중반 한국인 남성). 우측 눈에 붕대, 좌측 눈에 눈물이 고여 있으나 표정은 근육 긴장 없이 이완되어 있습니다. 환자복, 남색 팔걸이, 가슴을 가로지르는 투명 호스가 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "머리는 베개에 안정적으로 눕혀져 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 강력하게 요구한 분노로 일그러진 표정과 경직된 얼굴 근육을 매우 훌륭하게 연출했으나, 가슴을 지나는 투명 호스가 엉뚱한 캔버스 스트랩으로 대체된 점은 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "레퍼런스의 투명 호스와 의상 디테일은 정확히 재현했지만, 핵심 지시 사항인 '분노로 턱을 앙다물고 근육이 경직된 표정'을 전혀 반영하지 못하고 이완된 슬픈 표정에 머물렀습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 밖 우측 상단(방문 쪽)을 향하고 있습니다.",
        "built_space": "머리를 받치고 있는 흰색 베개와 흐릿한 배경의 금속 침대 난간이 보입니다.",
        "entities": "박철진(40대 중반 한국인 남성). 우측 눈은 붕대로 덮여 있고, 좌측 눈은 핏발이 서고 눈물이 고여 있습니다. 턱과 뺨 근육이 극도로 긴장된 분노의 표정입니다. 환자복과 남색 팔걸이를 착용 중이나 가슴의 호스가 흰색 끈으로 변형되었습니다.",
        "hard_violations": [],
        "physics": "머리는 중력에 의해 베개에 자연스럽게 기대어 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 화면 밖 우측 상단을 향하고 있습니다.",
        "built_space": "머리를 받치고 있는 흰색 베개와 배경의 병원 침대 난간이 올바르게 위치합니다.",
        "entities": "박철진(40대 중반 한국인 남성). 우측 눈에 붕대, 좌측 눈에 눈물이 고여 있으나 표정은 근육 긴장 없이 이완되어 있습니다. 환자복, 남색 팔걸이, 가슴을 가로지르는 투명 호스가 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "머리는 베개에 안정적으로 눕혀져 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "약간 높은 사선 얼굴 클로즈업에서 눈물 고인 눈, 굳게 다문 입과 수축한 볼이 드러나며, 참조의 시선과 병상 자세도 더 충실하게 이어진다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "병상 얼굴 클로즈업과 눈물은 충실하지만, 시선이 반대쪽으로 이동하고 입이 벌어져 악문 턱의 분노보다 고통과 불안이 더 강하게 읽힌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "드러난 눈은 화면 오른쪽 위의 프레임 밖 지점을 바라본다. 참조에서도 시선은 대체로 같은 방향이다. 문 자체는 보이지 않아 그 지점이 출입문인지는 확인할 수 없지만, 렌즈를 응시하지 않고 화면 밖 대상을 보는 요구에는 부합한다. 무기나 방향을 평가할 이동 물체는 없다.",
        "built_space": "얼굴 바로 아래 베개 하나, 왼쪽 위 머리판 일부, 오른쪽 침대 측면 난간 하나가 보인다. 얼굴과 목이 화면 대부분을 차지하고 병상 구조는 주변에 흐리게 남는다. 침대 옆에서 얼굴보다 조금 높은 사선 시점이며, 참조의 밝은 침구와 회색 계열 침대 구조가 이어진다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "참조와 닮은 40대 중반의 동아시아계 남성 한 명만 보이며, 짧고 흐트러진 검은 머리, 수염 자국, 얼굴 상처, 화면 왼쪽 눈을 덮은 흰 거즈가 유지된다. 반대쪽 눈은 충혈되고 아래 눈꺼풀에 눈물이 고여 있다. 청색 무늬 환자복과 짙은 지지대 끈도 이어진다. 미간과 볼에 긴장이 있고 입술을 닫아 턱을 굳힌 표정이다. 팔과 다리의 깁스는 올바른 클로즈업 범위 밖이라 평가하지 않는다. 다른 인물이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리 옆면이 눌린 베개에 받쳐져 있고 목과 어깨는 침대에 누운 몸으로 이어진다. 거즈는 얼굴에 밀착되어 고정되고, 옷과 지지대 끈은 몸 위에 놓여 있다. 눈물은 아래 눈꺼풀에 고여 반사된다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "드러난 눈의 동공은 화면 왼쪽 위, 코와 가린 눈 쪽의 프레임 밖 지점을 향한다. A와 참조의 오른쪽 위 시선과는 다르다. 출입문은 보이지 않아 실제 목표가 문인지 확정할 수 없으며, 떠난 사람들을 향한 시선의 연속성은 A보다 약하다. 무기나 이동 물체는 없다.",
        "built_space": "베개 하나가 머리를 받치며, 왼쪽 위 머리판 일부와 오른쪽 측면 난간 하나가 보인다. 오른쪽 위에는 의료기기 일부가 작게 잘려 있다. 얼굴 중심의 약간 높은 사선 클로즈업으로, 참조의 병상 재료와 실내 조명을 대체로 유지한다. 중복된 난간이나 불가능한 반사는 없다.",
        "entities": "참조와 닮은 중년 동아시아계 남성 한 명이며, 검은 머리, 상처와 수염 자국, 화면 왼쪽 눈의 거즈, 청색 무늬 환자복과 지지대 끈이 보인다. 드러난 눈은 충혈되고 눈물이 고여 있으며 볼에도 젖은 흔적이 있다. 미간은 수축했지만 입술이 벌어지고 치아가 조금 보여, 요구된 악문 턱의 분노보다 괴로움이 두드러진다. 깁스가 있는 팔다리는 프레임 밖이다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "머리는 베개에 깊이 기대고 목과 어깨는 누운 몸에 연결되어 안정적으로 받쳐진다. 가슴의 끈과 버클은 환자복 위에 놓여 있으며, 안대는 붕대로 고정되어 있다. 눈물의 고임과 볼을 따라 내려온 흔적도 중력에 맞는다. 지지 없는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "약간 높은 사선 얼굴 클로즈업에서 눈물 고인 눈, 굳게 다문 입과 수축한 볼이 드러나며, 참조의 시선과 병상 자세도 더 충실하게 이어진다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "병상 얼굴 클로즈업과 눈물은 충실하지만, 시선이 반대쪽으로 이동하고 입이 벌어져 악문 턱의 분노보다 고통과 불안이 더 강하게 읽힌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "드러난 눈은 화면 오른쪽 위의 프레임 밖 지점을 바라본다. 참조에서도 시선은 대체로 같은 방향이다. 문 자체는 보이지 않아 그 지점이 출입문인지는 확인할 수 없지만, 렌즈를 응시하지 않고 화면 밖 대상을 보는 요구에는 부합한다. 무기나 방향을 평가할 이동 물체는 없다.",
        "built_space": "얼굴 바로 아래 베개 하나, 왼쪽 위 머리판 일부, 오른쪽 침대 측면 난간 하나가 보인다. 얼굴과 목이 화면 대부분을 차지하고 병상 구조는 주변에 흐리게 남는다. 침대 옆에서 얼굴보다 조금 높은 사선 시점이며, 참조의 밝은 침구와 회색 계열 침대 구조가 이어진다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "참조와 닮은 40대 중반의 동아시아계 남성 한 명만 보이며, 짧고 흐트러진 검은 머리, 수염 자국, 얼굴 상처, 화면 왼쪽 눈을 덮은 흰 거즈가 유지된다. 반대쪽 눈은 충혈되고 아래 눈꺼풀에 눈물이 고여 있다. 청색 무늬 환자복과 짙은 지지대 끈도 이어진다. 미간과 볼에 긴장이 있고 입술을 닫아 턱을 굳힌 표정이다. 팔과 다리의 깁스는 올바른 클로즈업 범위 밖이라 평가하지 않는다. 다른 인물이나 덧씌운 문자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리 옆면이 눌린 베개에 받쳐져 있고 목과 어깨는 침대에 누운 몸으로 이어진다. 거즈는 얼굴에 밀착되어 고정되고, 옷과 지지대 끈은 몸 위에 놓여 있다. 눈물은 아래 눈꺼풀에 고여 반사된다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "드러난 눈의 동공은 화면 왼쪽 위, 코와 가린 눈 쪽의 프레임 밖 지점을 향한다. A와 참조의 오른쪽 위 시선과는 다르다. 출입문은 보이지 않아 실제 목표가 문인지 확정할 수 없으며, 떠난 사람들을 향한 시선의 연속성은 A보다 약하다. 무기나 이동 물체는 없다.",
        "built_space": "베개 하나가 머리를 받치며, 왼쪽 위 머리판 일부와 오른쪽 측면 난간 하나가 보인다. 오른쪽 위에는 의료기기 일부가 작게 잘려 있다. 얼굴 중심의 약간 높은 사선 클로즈업으로, 참조의 병상 재료와 실내 조명을 대체로 유지한다. 중복된 난간이나 불가능한 반사는 없다.",
        "entities": "참조와 닮은 중년 동아시아계 남성 한 명이며, 검은 머리, 상처와 수염 자국, 화면 왼쪽 눈의 거즈, 청색 무늬 환자복과 지지대 끈이 보인다. 드러난 눈은 충혈되고 눈물이 고여 있으며 볼에도 젖은 흔적이 있다. 미간은 수축했지만 입술이 벌어지고 치아가 조금 보여, 요구된 악문 턱의 분노보다 괴로움이 두드러진다. 깁스가 있는 팔다리는 프레임 밖이다. 추가 인물이나 문자 오버레이는 없다.",
        "hard_violations": [],
        "physics": "머리는 베개에 깊이 기대고 목과 어깨는 누운 몸에 연결되어 안정적으로 받쳐진다. 가슴의 끈과 버클은 환자복 위에 놓여 있으며, 안대는 붕대로 고정되어 있다. 눈물의 고임과 볼을 따라 내려온 흔적도 중력에 맞는다. 지지 없는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.778,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.778,
    "B": 1.571
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1778,
   "B": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1778,
    "verdict_ko": "프롬프트가 강력하게 요구한 분노로 일그러진 표정과 경직된 얼굴 근육을 매우 훌륭하게 연출했으나, 가슴을 지나는 투명 호스가 엉뚱한 캔버스 스트랩으로 대체된 점은 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "레퍼런스의 투명 호스와 의상 디테일은 정확히 재현했지만, 핵심 지시 사항인 '분노로 턱을 앙다물고 근육이 경직된 표정'을 전혀 반영하지 못하고 이완된 슬픈 표정에 머물렀습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박철진 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S70sh8_sel.png",
    "asset_id": "d3d6dfe0-d6e4-4e51-b440-c796f1abc923",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d41-f372-77cb-9ab1-188dc24c1547",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S70sh8"
  },
  "locked_char_refs_excluded": [
   "박철진(C15)"
  ]
 },
 "S71sh6::signage": {
  "fp": "a9d12fd4d6c06226",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::2797d89bf0931272": {
  "subjects": [],
  "subject_text": "이현우의 트럭 운전실\n낡은 운전석과 조수석이 나란히 놓인 간소한 운전 공간. 운전대와 계기판, 라디오가 있으며 전면 유리와 측면 창문으로 시야가 열린다.",
  "identity": "canonical",
  "scope_id": "L97",
  "scope_role": "location_exterior",
  "scope_sha": "42660a2880e69c4c"
 },
 "S71sh6::bgfirst_bg": {
  "input_fingerprint": "025d2127f84cf507",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리가 그려낸 보라색 꽃 그림을 빤히 내려다보는 앰버의 어깨너머 시점 구도.\n\nLOCATION (lock): At the rear passenger and cargo area of the old truck, where the drawing is passed from the robot riding in the load bed. Early-dawn light reaches the open vehicle.\n\nTIME OF DAY (lock): dawn to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward approach just behind 앰버's outer shoulder, slightly above her seated shoulder height and tilted down toward the drawing. Her shoulder and bowed head border the left foreground while her hands hold the paper obliquely across the lower center, occupying roughly one third of the image; she studies the purple flower, with 찰리 outside the crop. Keep the truck-bed conversation axis unchanged, making camera distance the sole emphasis as the movement settles before the daytime cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 꽃 그림이 그려진 종이 (Received by 앰버, with the purple flower drawing visible) — The reverse side bearing the flower drawing tilts toward the camera and 앰버; used as Primary focal detail, bounded by her hands and shoulder rather than isolated from context; 트럭 짐칸 (Occupied during the journey) — A limited downward view of the bed remains around the seated figure; used as Peripheral spatial context beneath the drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination preserves the drawing's purple color and readable hands without lifting the surrounding truck bed excessively.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리가 그려낸 보라색 꽃 그림을 빤히 내려다보는 앰버의 어깨너머 시점 구도.\n\nLOCATION (lock): At the rear passenger and cargo area of the old truck, where the drawing is passed from the robot riding in the load bed. Early-dawn light reaches the open vehicle.\n\nTIME OF DAY (lock): dawn to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward approach just behind 앰버's outer shoulder, slightly above her seated shoulder height and tilted down toward the drawing. Her shoulder and bowed head border the left foreground while her hands hold the paper obliquely across the lower center, occupying roughly one third of the image; she studies the purple flower, with 찰리 outside the crop. Keep the truck-bed conversation axis unchanged, making camera distance the sole emphasis as the movement settles before the daytime cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 꽃 그림이 그려진 종이 (Received by 앰버, with the purple flower drawing visible) — The reverse side bearing the flower drawing tilts toward the camera and 앰버; used as Primary focal detail, bounded by her hands and shoulder rather than isolated from context; 트럭 짐칸 (Occupied during the journey) — A limited downward view of the bed remains around the seated figure; used as Peripheral spatial context beneath the drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination preserves the drawing's purple color and readable hands without lifting the surrounding truck bed excessively.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh6__bgfirst_bg.png",
  "asset_id": "d3583e8e-1f98-413a-b44e-e70f975cc0a1",
  "input_asset_ids": [
   "ce3545ce-59ac-4ae4-ae13-d92cf3a890d6",
   "71e106e5-8061-49a4-96c1-92ad1c3d2c67"
  ]
 },
 "S71sh6": {
  "input_fingerprint": "ae5b1323ebfb4038",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 그려낸 보라색 꽃 그림을 빤히 내려다보는 앰버의 어깨너머 시점 구도.\n\nLOCATION (lock): At the rear passenger and cargo area of the old truck, where the drawing is passed from the robot riding in the load bed. Early-dawn light reaches the open vehicle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward approach just behind 앰버's outer shoulder, slightly above her seated shoulder height and tilted down toward the drawing. Her shoulder and bowed head border the left foreground while her hands hold the paper obliquely across the lower center, occupying roughly one third of the image; she studies the purple flower, with 찰리 outside the crop. Keep the truck-bed conversation axis unchanged, making camera distance the sole emphasis as the movement settles before the daytime cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 꽃 그림이 그려진 종이 (Received by 앰버, with the purple flower drawing visible) — The reverse side bearing the flower drawing tilts toward the camera and 앰버; used as Primary focal detail, bounded by her hands and shoulder rather than isolated from context; 트럭 짐칸 (Occupied during the journey) — A limited downward view of the bed remains around the seated figure; used as Peripheral spatial context beneath the drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination preserves the drawing's purple color and readable hands without lifting the surrounding truck bed excessively.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is traveling south at dawn, and a flower sketch has been drawn on the reverse of a sheet of paper; the text identifies the flower as a purple violet but does not establish colored drawing media. Charlie remains battle-damaged in the cargo bed and retains B-200's chest component. 앰버: She holds the sheet bearing the flower drawing; her prior head injury has no stated treatment yet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 그려낸 보라색 꽃 그림을 빤히 내려다보는 앰버의 어깨너머 시점 구도.\n\nLOCATION (lock): At the rear passenger and cargo area of the old truck, where the drawing is passed from the robot riding in the load bed. Early-dawn light reaches the open vehicle. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward approach just behind 앰버's outer shoulder, slightly above her seated shoulder height and tilted down toward the drawing. Her shoulder and bowed head border the left foreground while her hands hold the paper obliquely across the lower center, occupying roughly one third of the image; she studies the purple flower, with 찰리 outside the crop. Keep the truck-bed conversation axis unchanged, making camera distance the sole emphasis as the movement settles before the daytime cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 꽃 그림이 그려진 종이 (Received by 앰버, with the purple flower drawing visible) — The reverse side bearing the flower drawing tilts toward the camera and 앰버; used as Primary focal detail, bounded by her hands and shoulder rather than isolated from context; 트럭 짐칸 (Occupied during the journey) — A limited downward view of the bed remains around the seated figure; used as Peripheral spatial context beneath the drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination preserves the drawing's purple color and readable hands without lifting the surrounding truck bed excessively.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is traveling south at dawn, and a flower sketch has been drawn on the reverse of a sheet of paper; the text identifies the flower as a purple violet but does not establish colored drawing media. Charlie remains battle-damaged in the cargo bed and retains B-200's chest component. 앰버: She holds the sheet bearing the flower drawing; her prior head injury has no stated treatment yet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리가 그려낸 보라색 꽃 그림을 빤히 내려다보는 앰버의 어깨너머 시점 구도.\n\nLOCATION (lock): At the rear passenger and cargo area of the old truck, where the drawing is passed from the robot riding in the load bed. Early-dawn light reaches the open vehicle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the inward approach just behind 앰버's outer shoulder, slightly above her seated shoulder height and tilted down toward the drawing. Her shoulder and bowed head border the left foreground while her hands hold the paper obliquely across the lower center, occupying roughly one third of the image; she studies the purple flower, with 찰리 outside the crop. Keep the truck-bed conversation axis unchanged, making camera distance the sole emphasis as the movement settles before the daytime cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 꽃 그림이 그려진 종이 (Received by 앰버, with the purple flower drawing visible) — The reverse side bearing the flower drawing tilts toward the camera and 앰버; used as Primary focal detail, bounded by her hands and shoulder rather than isolated from context; 트럭 짐칸 (Occupied during the journey) — A limited downward view of the bed remains around the seated figure; used as Peripheral spatial context beneath the drawing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued dawn illumination preserves the drawing's purple color and readable hands without lifting the surrounding truck bed excessively.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old roofless truck is traveling south at dawn, and a flower sketch has been drawn on the reverse of a sheet of paper; the text identifies the flower as a purple violet but does not establish colored drawing media. Charlie remains battle-damaged in the cargo bed and retains B-200's chest component. 앰버: She holds the sheet bearing the flower drawing; her prior head injury has no stated treatment yet.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 앰버 right now, so 앰버's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 앰버: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh6__bgfirst_bg.png",
     "asset_id": "d3583e8e-1f98-413a-b44e-e70f975cc0a1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S71sh6.png",
     "asset_id": "ce3545ce-59ac-4ae4-ae13-d92cf3a890d6",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 제비꽃 그림 종이: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1220530>",
     "asset_id": "6bd68403-8866-4247-8992-f87863a2ea65",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_travel_truck_bed_0f596c.png",
     "asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 찰리의 제비꽃 그림 종이: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1220530>",
     "asset_id": "6bd68403-8866-4247-8992-f87863a2ea65",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "소녀의 얼굴 방향이 손에 든 종이를 향하고 있음.",
    "built_space": "트럭 짐칸 내부. 카메라가 수평을 향해 먼 산과 도로가 배경으로 보이며, 프롬프트가 요구한 하향 시점의 짐칸 바닥 뷰가 아님.",
    "entities": "금발 머리의 인물, 매끄러운 종이에 그려진 보라색 꽃. 인물의 목에 걸린 장비가 방진 마스크라기보다 헤드폰에 가깝게 묘사됨.",
    "hard_violations": [],
    "physics": "두 손으로 종이를 쥐고 몸을 구부린 채 트럭 짐칸 바닥에 앉아 지지됨."
   },
   {
    "label": "B",
    "direction": "인물의 고개가 손에 든 종이를 향해 숙여져 시선이 꽃 그림을 향함.",
    "built_space": "트럭 짐칸 내부. 카메라가 아래를 향해 종이 아래로 짐칸 바닥 판자가 배경으로 펼쳐지며 지시된 공간적 맥락을 형성함.",
    "entities": "금발 머리의 인물(앰버), 낡고 거친 질감의 종이에 그려진 보라색 꽃 그림. 카키색 작업복과 목에 걸친 방진 마스크가 레퍼런스와 부합함.",
    "hard_violations": [],
    "physics": "두 손으로 종이를 안정적으로 쥐고 있으며, 인물은 트럭 바닥에 앉아 무게를 지지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "어깨 너머로 종이를 내려다보는 하향 카메라 각도와 짐칸 바닥을 주변 배경으로 삼는 구도 지시를 정확하게 구현했으며, 소품의 질감도 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라가 아래를 향해 짐칸 바닥을 비춰야 한다는 명확한 프레이밍 지시를 어기고 수평선과 먼 풍경을 배경으로 연출하여 샷의 의도를 크게 훼손함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "인물의 고개가 손에 든 종이를 향해 숙여져 시선이 꽃 그림을 향함.",
        "built_space": "트럭 짐칸 내부. 카메라가 아래를 향해 종이 아래로 짐칸 바닥 판자가 배경으로 펼쳐지며 지시된 공간적 맥락을 형성함.",
        "entities": "금발 머리의 인물(앰버), 낡고 거친 질감의 종이에 그려진 보라색 꽃 그림. 카키색 작업복과 목에 걸친 방진 마스크가 레퍼런스와 부합함.",
        "hard_violations": [],
        "physics": "두 손으로 종이를 안정적으로 쥐고 있으며, 인물은 트럭 바닥에 앉아 무게를 지지하고 있음."
       },
       {
        "label": "A",
        "direction": "소녀의 얼굴 방향이 손에 든 종이를 향하고 있음.",
        "built_space": "트럭 짐칸 내부. 카메라가 수평을 향해 먼 산과 도로가 배경으로 보이며, 프롬프트가 요구한 하향 시점의 짐칸 바닥 뷰가 아님.",
        "entities": "금발 머리의 인물, 매끄러운 종이에 그려진 보라색 꽃. 인물의 목에 걸린 장비가 방진 마스크라기보다 헤드폰에 가깝게 묘사됨.",
        "hard_violations": [],
        "physics": "두 손으로 종이를 쥐고 몸을 구부린 채 트럭 짐칸 바닥에 앉아 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "어깨 너머로 종이를 내려다보는 하향 카메라 각도와 짐칸 바닥을 주변 배경으로 삼는 구도 지시를 정확하게 구현했으며, 소품의 질감도 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라가 아래를 향해 짐칸 바닥을 비춰야 한다는 명확한 프레이밍 지시를 어기고 수평선과 먼 풍경을 배경으로 연출하여 샷의 의도를 크게 훼손함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 고개가 손에 든 종이를 향해 숙여져 시선이 꽃 그림을 향함.",
        "built_space": "트럭 짐칸 내부. 카메라가 아래를 향해 종이 아래로 짐칸 바닥 판자가 배경으로 펼쳐지며 지시된 공간적 맥락을 형성함.",
        "entities": "금발 머리의 인물(앰버), 낡고 거친 질감의 종이에 그려진 보라색 꽃 그림. 카키색 작업복과 목에 걸친 방진 마스크가 레퍼런스와 부합함.",
        "hard_violations": [],
        "physics": "두 손으로 종이를 안정적으로 쥐고 있으며, 인물은 트럭 바닥에 앉아 무게를 지지하고 있음."
       },
       {
        "label": "A",
        "direction": "소녀의 얼굴 방향이 손에 든 종이를 향하고 있음.",
        "built_space": "트럭 짐칸 내부. 카메라가 수평을 향해 먼 산과 도로가 배경으로 보이며, 프롬프트가 요구한 하향 시점의 짐칸 바닥 뷰가 아님.",
        "entities": "금발 머리의 인물, 매끄러운 종이에 그려진 보라색 꽃. 인물의 목에 걸린 장비가 방진 마스크라기보다 헤드폰에 가깝게 묘사됨.",
        "hard_violations": [],
        "physics": "두 손으로 종이를 쥐고 몸을 구부린 채 트럭 짐칸 바닥에 앉아 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "왼쪽 어깨너머에서 종이를 크게 내려다보는 근접 구도와 제한된 짐칸 배경이 지시에 충실하며, 종이의 낡은 질감과 제비꽃 형태도 참조에 더 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "종이를 보는 행동은 맞지만 도로와 하늘까지 넓게 드러내 종이 중심의 하향 클로즈업을 약화했고, 종이 가장자리와 꽃 도안도 참조에서 더 멀어졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버의 고개가 아래 중앙의 꽃 그림을 향해 숙여져 있다. 눈 자체는 가려져 시선을 직접 확인할 수 없지만 머리 방향은 종이를 보는 행동과 맞는다. 그림 면은 앰버 쪽으로 기울어 있고, 뒤쪽 어깨너머 카메라에서도 자연스럽게 보인다.",
        "built_space": "녹슨 밝은 금속 적재함의 서로 만나는 벽 두 면과 어두운 바닥 일부가 보인다. 벽에는 수직 보강대와 체결부가 있으며 지붕이나 별도 좌석은 보이지 않는다. 앰버의 굽힌 무릎이 아래쪽에 있고, 주변 바닥이 제한적으로 남아 있어 적재함 안에 앉은 배치로 읽힌다. 참조의 낡은 금속 재질과 평평한 바닥에 가깝고 외부 풍경은 상단의 좁은 띠로 제한된다.",
        "entities": "보이는 사람은 앰버 한 명뿐이며 찰리는 지시대로 화면 밖이다. 금발, 밝은 피부, 작은 체구, 때 묻은 카키 작업복과 목 주변의 기계식 방진 마스크가 보인다. 얼굴 대부분이 가려져 정확한 나이와 혼혈 정체성은 확인하기 어렵다. 양손과 소매는 같은 인물의 것으로 자연스럽게 이어진다. 종이는 거칠고 낡은 가장자리를 가진 한 장이며, 꽃잎 다섯 장과 잎 두 장, 아래로 처진 꽃봉오리의 배치가 소품 참조에 가깝다. 다만 참조의 무채색 윤곽선보다 보라색 음영이 많이 들어갔다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "왼손과 오른손이 종이의 양쪽 가장자리를 실제로 잡고 있어 종이가 지지된다. 종이의 기울기와 약한 휨도 양손으로 든 상태에 맞는다. 굽힌 무릎과 몸통은 앉아서 무릎 위에 종이를 든 자세로 연결되며, 엉덩이 접촉면은 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "앰버는 고개를 아래로 숙여 두 손 사이의 보라색 꽃 그림을 향한다. 눈은 보이지 않지만 머리 방향은 그림을 살펴보는 행동과 맞는다. 종이의 그림 면이 앰버와 어깨 뒤 카메라 양쪽에서 보일 수 있는 방향으로 놓였다.",
        "built_space": "적재함 뒤판 한 면과 좌우 측판 일부, 넓은 바닥이 보인다. 뒤판 양쪽에 수직 체결부가 있고 오른쪽에는 손잡이 하나가 보인다. 바닥은 참조보다 골이 뚜렷하며 뒤판 아래에 둥근 돌출부가 보인다. 앰버는 적재함 안에 무릎을 굽히고 앉아 있다. 녹슨 밝은 금속 벽은 장소 참조와 유사하지만 카메라가 도로, 전신주, 산과 하늘을 넓게 담아 요구된 제한적 하향 배경보다 외부 공간이 훨씬 커졌다.",
        "entities": "앰버 한 명의 금발 뒷머리, 밝은 피부, 카키 작업복, 목 주변 방진 마스크와 두 손이 보이며 찰리는 없다. 얼굴이 숨겨져 정확한 나이와 참조 얼굴의 일치 여부는 확인할 수 없다. 종이는 한 장이지만 가장자리가 참조보다 반듯하고 깨끗하다. 보라색 꽃과 초록색 잎이 그려져 있으나 참조의 처진 꽃봉오리가 없고 꽃잎 및 잎 형태도 달라졌다. 도구 벨트와 머리 부상은 이 구도에서 판별하기 어렵다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "양손이 종이 좌우 가장자리를 잡고 있으며 손목과 팔은 작업복 소매로 이어진다. 종이는 손에 의해 지지되고 무릎 위쪽에 자연스럽게 놓여 있다. 몸통과 굽힌 다리는 앉은 자세로 읽히며 엉덩이의 직접 접촉부는 가려져 있다. 공중에 뜨거나 물리적으로 불가능한 동작은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "왼쪽 어깨너머에서 종이를 크게 내려다보는 근접 구도와 제한된 짐칸 배경이 지시에 충실하며, 종이의 낡은 질감과 제비꽃 형태도 참조에 더 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "종이를 보는 행동은 맞지만 도로와 하늘까지 넓게 드러내 종이 중심의 하향 클로즈업을 약화했고, 종이 가장자리와 꽃 도안도 참조에서 더 멀어졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버의 고개가 아래 중앙의 꽃 그림을 향해 숙여져 있다. 눈 자체는 가려져 시선을 직접 확인할 수 없지만 머리 방향은 종이를 보는 행동과 맞는다. 그림 면은 앰버 쪽으로 기울어 있고, 뒤쪽 어깨너머 카메라에서도 자연스럽게 보인다.",
        "built_space": "녹슨 밝은 금속 적재함의 서로 만나는 벽 두 면과 어두운 바닥 일부가 보인다. 벽에는 수직 보강대와 체결부가 있으며 지붕이나 별도 좌석은 보이지 않는다. 앰버의 굽힌 무릎이 아래쪽에 있고, 주변 바닥이 제한적으로 남아 있어 적재함 안에 앉은 배치로 읽힌다. 참조의 낡은 금속 재질과 평평한 바닥에 가깝고 외부 풍경은 상단의 좁은 띠로 제한된다.",
        "entities": "보이는 사람은 앰버 한 명뿐이며 찰리는 지시대로 화면 밖이다. 금발, 밝은 피부, 작은 체구, 때 묻은 카키 작업복과 목 주변의 기계식 방진 마스크가 보인다. 얼굴 대부분이 가려져 정확한 나이와 혼혈 정체성은 확인하기 어렵다. 양손과 소매는 같은 인물의 것으로 자연스럽게 이어진다. 종이는 거칠고 낡은 가장자리를 가진 한 장이며, 꽃잎 다섯 장과 잎 두 장, 아래로 처진 꽃봉오리의 배치가 소품 참조에 가깝다. 다만 참조의 무채색 윤곽선보다 보라색 음영이 많이 들어갔다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "왼손과 오른손이 종이의 양쪽 가장자리를 실제로 잡고 있어 종이가 지지된다. 종이의 기울기와 약한 휨도 양손으로 든 상태에 맞는다. 굽힌 무릎과 몸통은 앉아서 무릎 위에 종이를 든 자세로 연결되며, 엉덩이 접촉면은 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "앰버는 고개를 아래로 숙여 두 손 사이의 보라색 꽃 그림을 향한다. 눈은 보이지 않지만 머리 방향은 그림을 살펴보는 행동과 맞는다. 종이의 그림 면이 앰버와 어깨 뒤 카메라 양쪽에서 보일 수 있는 방향으로 놓였다.",
        "built_space": "적재함 뒤판 한 면과 좌우 측판 일부, 넓은 바닥이 보인다. 뒤판 양쪽에 수직 체결부가 있고 오른쪽에는 손잡이 하나가 보인다. 바닥은 참조보다 골이 뚜렷하며 뒤판 아래에 둥근 돌출부가 보인다. 앰버는 적재함 안에 무릎을 굽히고 앉아 있다. 녹슨 밝은 금속 벽은 장소 참조와 유사하지만 카메라가 도로, 전신주, 산과 하늘을 넓게 담아 요구된 제한적 하향 배경보다 외부 공간이 훨씬 커졌다.",
        "entities": "앰버 한 명의 금발 뒷머리, 밝은 피부, 카키 작업복, 목 주변 방진 마스크와 두 손이 보이며 찰리는 없다. 얼굴이 숨겨져 정확한 나이와 참조 얼굴의 일치 여부는 확인할 수 없다. 종이는 한 장이지만 가장자리가 참조보다 반듯하고 깨끗하다. 보라색 꽃과 초록색 잎이 그려져 있으나 참조의 처진 꽃봉오리가 없고 꽃잎 및 잎 형태도 달라졌다. 도구 벨트와 머리 부상은 이 구도에서 판별하기 어렵다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "양손이 종이 좌우 가장자리를 잡고 있으며 손목과 팔은 작업복 소매로 이어진다. 종이는 손에 의해 지지되고 무릎 위쪽에 자연스럽게 놓여 있다. 몸통과 굽힌 다리는 앉은 자세로 읽히며 엉덩이의 직접 접촉부는 가려져 있다. 공중에 뜨거나 물리적으로 불가능한 동작은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.238,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.238,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1238
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "어깨 너머로 종이를 내려다보는 하향 카메라 각도와 짐칸 바닥을 주변 배경으로 삼는 구도 지시를 정확하게 구현했으며, 소품의 질감도 우수함."
   },
   {
    "label": "A",
    "score": 1238,
    "verdict_ko": "카메라가 아래를 향해 짐칸 바닥을 비춰야 한다는 명확한 프레이밍 지시를 어기고 수평선과 먼 풍경을 배경으로 연출하여 샷의 의도를 크게 훼손함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_travel_truck_bed_0f596c.png",
    "asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 제비꽃 그림 종이: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1220530>",
    "asset_id": "6bd68403-8866-4247-8992-f87863a2ea65",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d46-73a7-7cd0-b33e-179c3e02c514",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh6__bgfirst_bg.png",
   "bg_asset_id": "d3583e8e-1f98-413a-b44e-e70f975cc0a1",
   "bg_record_key": "S71sh6::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "travel_truck_bed",
   "groupbg_asset_id": "71e106e5-8061-49a4-96c1-92ad1c3d2c67"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S71sh15::signage": {
  "fp": "ac830ef2509ab0a2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::coastal_search_hill": {
  "input_fingerprint": "a7d946b63c4ba996",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "coastal_search_hill",
    "tags": [
     "S71sh15"
    ]
   },
   "context_sig": "2f9ccc7451af914a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 언덕 위에 서서 능선을 쳐다보는 현우. 저 멀리 능선 사이사이로 보이는 바다윤슬. 그러나 어디에도 제비꽃은 보이지 않고.\n\nTIME OF DAY (lock): dawn to day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n이현우의 트럭 운전실: 지붕 없이 개방되거나 앞유리가 낡은 형태의 노후 트럭 운전석이다. (특징: 먼지 낀 낡은 조향 장치와 빈 연료 게이지 바늘)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 언덕 위에 서서 능선을 쳐다보는 현우. 저 멀리 능선 사이사이로 보이는 바다윤슬. 그러나 어디에도 제비꽃은 보이지 않고.\n\nTIME OF DAY (lock): dawn to day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
  "asset_id": "f1390459-bf60-4198-ac9c-83c7241d7ec7",
  "input_asset_ids": [
   "21067bef-99f9-4b59-8cd5-cc610f5127fa"
  ],
  "origin_tag": "S71sh15",
  "place_text": "On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.",
  "origin_inputs": {
   "place_text": "On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.",
   "time_of_day_en": "dawn to day",
   "conti_asset_id": "21067bef-99f9-4b59-8cd5-cc610f5127fa"
  }
 },
 "S71sh15::bgfirst_bg": {
  "input_fingerprint": "9be2ff5244f68d71",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 표정을 살피며 슬픔이 차오른 눈빛으로 아랫입술을 꾹 깨문 앰버의 측면.\n\nLOCATION (lock): On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.\n\nTIME OF DAY (lock): dawn to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at 앰버's seated eye height, perpendicular to her gaze, with her right-facing profile occupying the left half of the close frame. She lightly bites her lower lip and studies 찰리 just beyond the right edge, leaving open space ahead of her face rather than turning her toward the lens. Hold the established distance and lighting, concentrating the beat on her gaze as sadness gathers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 언덕의 지면 (The ground where the group has been playing with soil); used as A softly resolved background below her profile, keeping the reaction rooted in the hill location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the moisture in her eyes legible without turning the sadness into a heightened lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 표정을 살피며 슬픔이 차오른 눈빛으로 아랫입술을 꾹 깨문 앰버의 측면.\n\nLOCATION (lock): On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing.\n\nTIME OF DAY (lock): dawn to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at 앰버's seated eye height, perpendicular to her gaze, with her right-facing profile occupying the left half of the close frame. She lightly bites her lower lip and studies 찰리 just beyond the right edge, leaving open space ahead of her face rather than turning her toward the lens. Hold the established distance and lighting, concentrating the beat on her gaze as sadness gathers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 언덕의 지면 (The ground where the group has been playing with soil); used as A softly resolved background below her profile, keeping the reaction rooted in the hill location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the moisture in her eyes legible without turning the sadness into a heightened lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh15__bgfirst_bg.png",
  "asset_id": "f5de7de7-f0ce-4577-8bdc-6c35baa1e0e6",
  "input_asset_ids": [
   "21067bef-99f9-4b59-8cd5-cc610f5127fa",
   "f1390459-bf60-4198-ac9c-83c7241d7ec7"
  ]
 },
 "S71sh15": {
  "input_fingerprint": "2301db152eeb8703",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 표정을 살피며 슬픔이 차오른 눈빛으로 아랫입술을 꾹 깨문 앰버의 측면.\n\nLOCATION (lock): On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at 앰버's seated eye height, perpendicular to her gaze, with her right-facing profile occupying the left half of the close frame. She lightly bites her lower lip and studies 찰리 just beyond the right edge, leaving open space ahead of her face rather than turning her toward the lens. Hold the established distance and lighting, concentrating the beat on her gaze as sadness gathers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 언덕의 지면 (The ground where the group has been playing with soil); used as A softly resolved background below her profile, keeping the reaction rooted in the hill location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the moisture in her eyes legible without turning the sadness into a heightened lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight reveals coastal ridges and glinting seawater, with no violets found at this stopping place; the truck and the earlier flower drawing remain available. Charlie's battle damage and possession of B-200's chest component continue. 앰버: She is outside the truck after playing with dirt, now visibly sad and tearful. Her head injury remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 표정을 살피며 슬픔이 차오른 눈빛으로 아랫입술을 꾹 깨문 앰버의 측면.\n\nLOCATION (lock): On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at 앰버's seated eye height, perpendicular to her gaze, with her right-facing profile occupying the left half of the close frame. She lightly bites her lower lip and studies 찰리 just beyond the right edge, leaving open space ahead of her face rather than turning her toward the lens. Hold the established distance and lighting, concentrating the beat on her gaze as sadness gathers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 언덕의 지면 (The ground where the group has been playing with soil); used as A softly resolved background below her profile, keeping the reaction rooted in the hill location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the moisture in her eyes legible without turning the sadness into a heightened lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight reveals coastal ridges and glinting seawater, with no violets found at this stopping place; the truck and the earlier flower drawing remain available. Charlie's battle damage and possession of B-200's chest component continue. 앰버: She is outside the truck after playing with dirt, now visibly sad and tearful. Her head injury remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 표정을 살피며 슬픔이 차오른 눈빛으로 아랫입술을 꾹 깨문 앰버의 측면.\n\nLOCATION (lock): On the roadside hill where the truck has stopped, near the patch of earth where the children and robot are playing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Briefly settle the lateral track at 앰버's seated eye height, perpendicular to her gaze, with her right-facing profile occupying the left half of the close frame. She lightly bites her lower lip and studies 찰리 just beyond the right edge, leaving open space ahead of her face rather than turning her toward the lens. Hold the established distance and lighting, concentrating the beat on her gaze as sadness gathers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 언덕의 지면 (The ground where the group has been playing with soil); used as A softly resolved background below her profile, keeping the reaction rooted in the hill location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight keeps the moisture in her eyes legible without turning the sadness into a heightened lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight reveals coastal ridges and glinting seawater, with no violets found at this stopping place; the truck and the earlier flower drawing remain available. Charlie's battle damage and possession of B-200's chest component continue. 앰버: She is outside the truck after playing with dirt, now visibly sad and tearful. Her head injury remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh15__bgfirst_bg.png",
     "asset_id": "f5de7de7-f0ce-4577-8bdc-6c35baa1e0e6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S71sh15.png",
     "asset_id": "21067bef-99f9-4b59-8cd5-cc610f5127fa",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
     "asset_id": "f1390459-bf60-4198-ac9c-83c7241d7ec7",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
    "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측에 트럭, 우측에 둥근 로봇 부품이 배치됨.",
    "entities": "금발의 여자아이(앰버)가 카키색 작업복과 목에 건 마스크를 착용하고 있으며, 이마에 붉은 상처가 있고 눈물이 흐르고 있음.",
    "hard_violations": [],
    "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
   },
   {
    "label": "B",
    "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
    "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측 끝에 트럭 일부, 우측에 둥근 로봇 부품이 배치됨.",
    "entities": "금발의 여자아이가 카키색 작업복과 목에 건 마스크를 착용하고 눈물이 맺혀 있으나, 요구된 머리 상처는 뚜렷하게 보이지 않음.",
    "hard_violations": [],
    "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 우측면 클로즈업 구도와 슬픔을 참으며 입술을 깨문 표정을 잘 담아냈으며, 이전 장면에서 이어진 이마의 상처까지 정확히 반영함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 표정은 지시사항을 따랐으나, 필수 유지 상태인 이마의 상처가 명확히 묘사되지 않음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
        "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측에 트럭, 우측에 둥근 로봇 부품이 배치됨.",
        "entities": "금발의 여자아이(앰버)가 카키색 작업복과 목에 건 마스크를 착용하고 있으며, 이마에 붉은 상처가 있고 눈물이 흐르고 있음.",
        "hard_violations": [],
        "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
       },
       {
        "label": "B",
        "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
        "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측 끝에 트럭 일부, 우측에 둥근 로봇 부품이 배치됨.",
        "entities": "금발의 여자아이가 카키색 작업복과 목에 건 마스크를 착용하고 눈물이 맺혀 있으나, 요구된 머리 상처는 뚜렷하게 보이지 않음.",
        "hard_violations": [],
        "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 우측면 클로즈업 구도와 슬픔을 참으며 입술을 깨문 표정을 잘 담아냈으며, 이전 장면에서 이어진 이마의 상처까지 정확히 반영함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 표정은 지시사항을 따랐으나, 필수 유지 상태인 이마의 상처가 명확히 묘사되지 않음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
        "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측에 트럭, 우측에 둥근 로봇 부품이 배치됨.",
        "entities": "금발의 여자아이(앰버)가 카키색 작업복과 목에 건 마스크를 착용하고 있으며, 이마에 붉은 상처가 있고 눈물이 흐르고 있음.",
        "hard_violations": [],
        "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
       },
       {
        "label": "B",
        "direction": "인물은 오른쪽 화면 밖을 향해 시선을 두고 있다.",
        "built_space": "야외 언덕 지면에 인물이 위치하며, 배경 좌측 끝에 트럭 일부, 우측에 둥근 로봇 부품이 배치됨.",
        "entities": "금발의 여자아이가 카키색 작업복과 목에 건 마스크를 착용하고 눈물이 맺혀 있으나, 요구된 머리 상처는 뚜렷하게 보이지 않음.",
        "hard_violations": [],
        "physics": "인물의 자세는 지면의 중력을 자연스럽게 받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "완전한 정측면보다 카메라 쪽으로 조금 돌아왔지만, 왼쪽 얼굴 클로즈업과 오른쪽 여백, 젖은 눈빛과 눌러 문 입술이 핵심 반응 숏에 더 충실하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오른쪽 정측면과 이마의 부상은 더 정확하지만, 무릎과 주변 풍경까지 담도록 넓어진 구도가 지정된 얼굴 클로즈업을 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 오른쪽 바깥을 향하며 렌즈를 보지 않는다. 찰리는 보이지 않으므로 실제 시선의 도착점은 확인할 수 없지만 지정된 화면 밖 위치와 맞는다. 양쪽 눈이 보여 시선에 정확히 수직인 정측면보다는 약간 앞쪽에서 촬영한 각도다.",
        "built_space": "왼쪽 끝에 트럭 한 대의 일부, 오른쪽 중경 지면에 구형 로봇 부품 하나가 보인다. 돌과 마른 풀의 언덕 아래로 해안 능선과 바다가 이어져 장소 참고와 부합한다. 얼굴 아래의 흙과 돌은 부드럽게 흐려져 있다. 얼굴이 왼쪽 절반을 크게 차지하고 오른쪽에 여백이 남는 근접 구도다.",
        "entities": "보이는 사람은 앰버 한 명뿐이다. 금발, 밝은 피부, 큰 눈과 어린 여자아이의 얼굴은 참고 인물에 대체로 부합하며, 혼혈 배경 자체는 외양만으로 확정할 수 없다. 더러워진 카키 작업복과 목에 내려 걸린 기계식 방진 마스크가 보인다. 눈물과 안으로 눌린 아랫입술이 슬픔을 드러낸다. 노출된 이마에서는 부상 흔적이 뚜렷하지 않다. 허리 공구 벨트와 찰리의 소지품은 프레임 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 마스크는 목의 끈에 매달려 가슴에 놓여 있다. 하체와 좌면은 잘려 있어 앉은 자세의 접촉점은 확인할 수 없지만 공중에 뜬 신체는 보이지 않는다. 배경 트럭과 구형 부품은 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "코와 입의 윤곽이 오른쪽을 향한 거의 정확한 정측면이며, 눈도 오른쪽 화면 밖을 응시한다. 보이지 않는 찰리의 지정 위치와 방향이 일치하고 카메라를 향한 시선은 아니다.",
        "built_space": "왼쪽 배경에 트럭 한 대, 오른쪽 흙과 돌 사이에 구형 로봇 부품 하나가 보인다. 해안 능선과 섬, 바다 및 메마른 언덕의 배치는 장소 참고와 부합한다. 다만 머리 전체와 상체, 두 무릎까지 포함하며 지면과 바다가 넓고 비교적 선명하게 보여 지정된 클로즈업보다 넓다.",
        "entities": "앰버로 읽히는 금발의 어린 여자아이 한 명만 보인다. 밝은 피부와 큰 눈, 더러워진 카키 작업복, 가슴에 내려진 기계식 방진 마스크가 참고의 주요 특징과 맞는다. 정확한 혼혈 배경은 외양만으로 판별할 수 없다. 이마의 붉은 상처와 눈 아래의 눈물이 보이고 아랫입술은 안쪽으로 눌려 있다. 공구 벨트와 찰리의 흉부 부품은 보이지 않는다.",
        "hard_violations": [],
        "physics": "무릎을 몸 앞으로 굽혀 모은 자세이며 몸통과 다리의 연결은 자연스럽다. 엉덩이의 지면 접촉은 화면 아래로 가려져 있지만 부유를 시사하는 모습은 없다. 마스크는 목끈과 가슴에 의해 지지되며 트럭과 구형 부품도 땅에 닿아 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "완전한 정측면보다 카메라 쪽으로 조금 돌아왔지만, 왼쪽 얼굴 클로즈업과 오른쪽 여백, 젖은 눈빛과 눌러 문 입술이 핵심 반응 숏에 더 충실하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오른쪽 정측면과 이마의 부상은 더 정확하지만, 무릎과 주변 풍경까지 담도록 넓어진 구도가 지정된 얼굴 클로즈업을 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 오른쪽 바깥을 향하며 렌즈를 보지 않는다. 찰리는 보이지 않으므로 실제 시선의 도착점은 확인할 수 없지만 지정된 화면 밖 위치와 맞는다. 양쪽 눈이 보여 시선에 정확히 수직인 정측면보다는 약간 앞쪽에서 촬영한 각도다.",
        "built_space": "왼쪽 끝에 트럭 한 대의 일부, 오른쪽 중경 지면에 구형 로봇 부품 하나가 보인다. 돌과 마른 풀의 언덕 아래로 해안 능선과 바다가 이어져 장소 참고와 부합한다. 얼굴 아래의 흙과 돌은 부드럽게 흐려져 있다. 얼굴이 왼쪽 절반을 크게 차지하고 오른쪽에 여백이 남는 근접 구도다.",
        "entities": "보이는 사람은 앰버 한 명뿐이다. 금발, 밝은 피부, 큰 눈과 어린 여자아이의 얼굴은 참고 인물에 대체로 부합하며, 혼혈 배경 자체는 외양만으로 확정할 수 없다. 더러워진 카키 작업복과 목에 내려 걸린 기계식 방진 마스크가 보인다. 눈물과 안으로 눌린 아랫입술이 슬픔을 드러낸다. 노출된 이마에서는 부상 흔적이 뚜렷하지 않다. 허리 공구 벨트와 찰리의 소지품은 프레임 밖이라 평가할 수 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 마스크는 목의 끈에 매달려 가슴에 놓여 있다. 하체와 좌면은 잘려 있어 앉은 자세의 접촉점은 확인할 수 없지만 공중에 뜬 신체는 보이지 않는다. 배경 트럭과 구형 부품은 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "코와 입의 윤곽이 오른쪽을 향한 거의 정확한 정측면이며, 눈도 오른쪽 화면 밖을 응시한다. 보이지 않는 찰리의 지정 위치와 방향이 일치하고 카메라를 향한 시선은 아니다.",
        "built_space": "왼쪽 배경에 트럭 한 대, 오른쪽 흙과 돌 사이에 구형 로봇 부품 하나가 보인다. 해안 능선과 섬, 바다 및 메마른 언덕의 배치는 장소 참고와 부합한다. 다만 머리 전체와 상체, 두 무릎까지 포함하며 지면과 바다가 넓고 비교적 선명하게 보여 지정된 클로즈업보다 넓다.",
        "entities": "앰버로 읽히는 금발의 어린 여자아이 한 명만 보인다. 밝은 피부와 큰 눈, 더러워진 카키 작업복, 가슴에 내려진 기계식 방진 마스크가 참고의 주요 특징과 맞는다. 정확한 혼혈 배경은 외양만으로 판별할 수 없다. 이마의 붉은 상처와 눈 아래의 눈물이 보이고 아랫입술은 안쪽으로 눌려 있다. 공구 벨트와 찰리의 흉부 부품은 보이지 않는다.",
        "hard_violations": [],
        "physics": "무릎을 몸 앞으로 굽혀 모은 자세이며 몸통과 다리의 연결은 자연스럽다. 엉덩이의 지면 접촉은 화면 아래로 가려져 있지만 부유를 시사하는 모습은 없다. 마스크는 목끈과 가슴에 의해 지지되며 트럭과 구형 부품도 땅에 닿아 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지시된 우측면 클로즈업 구도와 슬픔을 참으며 입술을 깨문 표정을 잘 담아냈으며, 이전 장면에서 이어진 이마의 상처까지 정확히 반영함."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "전반적인 구도와 표정은 지시사항을 따랐으나, 필수 유지 상태인 이마의 상처가 명확히 묘사되지 않음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_coastal_search_hill_7092b9.png",
    "asset_id": "f1390459-bf60-4198-ac9c-83c7241d7ec7",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d4e-61a9-7457-89f2-3d5c33ab359e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S71sh15__bgfirst_bg.png",
   "bg_asset_id": "f5de7de7-f0ce-4577-8bdc-6c35baa1e0e6",
   "bg_record_key": "S71sh15::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "coastal_search_hill",
   "groupbg_asset_id": "f1390459-bf60-4198-ac9c-83c7241d7ec7"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S71sh28::signage": {
  "fp": "27c5187e86c440c7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S71sh28": {
  "input_fingerprint": "9243fda3ada19e47",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 만개한 보라색 제비꽃밭 너머로 우뚝 솟아 있는 거대한 연구소 건물의 웅장한 전경.\n\nLOCATION (lock): At the edge of an extensive violet field beside the road, overlooking a tall research building beyond the flowers. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the crane at its elevated roadside endpoint, beyond the group's previous framing position, looking gently downward across the violet field. Blooming flowers spread through the foreground and middle distance toward the complete laboratory building in the upper center, with the building occupying less than one third of the image and no people visible. Let the increased viewing distance carry the reveal without adding a simultaneous lighting change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: blooming violet field in the lower-center of the frame, foreground; distant laboratory beyond the flowers in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 제비꽃밭 (Abundantly blooming with purple violets); used as Continuous foreground-to-background depth leading toward the laboratory; 지동현의 연구소 (A large building rising in the distance beyond the flowers) — The complete field-facing elevation is visible from the elevated roadside view; used as Distant destination and visual endpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast retain the purple blossoms while keeping the distant laboratory clearly separated from its surroundings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A large field of blooming purple violets now fills the daylight landscape, with the Haenam research institute visible beyond it and the truck stopped nearby. Charlie still has his unrepaired battle damage and B-200's chest component.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 만개한 보라색 제비꽃밭 너머로 우뚝 솟아 있는 거대한 연구소 건물의 웅장한 전경.\n\nLOCATION (lock): At the edge of an extensive violet field beside the road, overlooking a tall research building beyond the flowers. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the crane at its elevated roadside endpoint, beyond the group's previous framing position, looking gently downward across the violet field. Blooming flowers spread through the foreground and middle distance toward the complete laboratory building in the upper center, with the building occupying less than one third of the image and no people visible. Let the increased viewing distance carry the reveal without adding a simultaneous lighting change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: blooming violet field in the lower-center of the frame, foreground; distant laboratory beyond the flowers in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 제비꽃밭 (Abundantly blooming with purple violets); used as Continuous foreground-to-background depth leading toward the laboratory; 지동현의 연구소 (A large building rising in the distance beyond the flowers) — The complete field-facing elevation is visible from the elevated roadside view; used as Distant destination and visual endpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast retain the purple blossoms while keeping the distant laboratory clearly separated from its surroundings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A large field of blooming purple violets now fills the daylight landscape, with the Haenam research institute visible beyond it and the truck stopped nearby. Charlie still has his unrepaired battle damage and B-200's chest component.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dawn to day.\n\nSHOT TEXT (authoritative, Korean): 만개한 보라색 제비꽃밭 너머로 우뚝 솟아 있는 거대한 연구소 건물의 웅장한 전경.\n\nLOCATION (lock): At the edge of an extensive violet field beside the road, overlooking a tall research building beyond the flowers. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the crane at its elevated roadside endpoint, beyond the group's previous framing position, looking gently downward across the violet field. Blooming flowers spread through the foreground and middle distance toward the complete laboratory building in the upper center, with the building occupying less than one third of the image and no people visible. Let the increased viewing distance carry the reveal without adding a simultaneous lighting change.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: blooming violet field in the lower-center of the frame, foreground; distant laboratory beyond the flowers in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 제비꽃밭 (Abundantly blooming with purple violets); used as Continuous foreground-to-background depth leading toward the laboratory; 지동현의 연구소 (A large building rising in the distance beyond the flowers) — The complete field-facing elevation is visible from the elevated roadside view; used as Distant destination and visual endpoint.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight and controlled contrast retain the purple blossoms while keeping the distant laboratory clearly separated from its surroundings.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A large field of blooming purple violets now fills the daylight landscape, with the Haenam research institute visible beyond it and the truck stopped nearby. Charlie still has his unrepaired battle damage and B-200's chest component.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라 시선은 도로변에서 언덕을 내려다보며, 보라색 제비꽃밭을 지나 원경의 연구소 건물을 향하고 있음.",
    "built_space": "원경 우측 중앙에 거대한 연구소 건물이 자리하고, 좌측 전경에 가드레일이 있는 도로가 있으며, 좌측 중경에는 레퍼런스와 일치하는 층리 암벽이 해안과 함께 배치됨.",
    "entities": "만개한 보라색 제비꽃밭, 거대한 연구소, 가드레일, 바다, 층리 형태의 해안 암벽 모두 프롬프트 및 레퍼런스와 일치하게 나타남.",
    "hard_violations": [],
    "physics": "건물, 가드레일, 암벽, 식물 등 모든 객체가 지면에 안정적으로 고정되어 있으며 물리적 결함이 없음."
   },
   {
    "label": "B",
    "direction": "카메라 시선은 도로변에서 제비꽃밭을 가로질러 멀리 떨어진 연구소를 향함.",
    "built_space": "원경 중앙에 연구소 건물이 위치하고, 좌측 전경에 도로와 가드레일, 좌측 해안에 층리 암벽이 배치됨.",
    "entities": "보라색 제비꽃밭, 연구소 건물, 도로, 암벽 등이 모두 식별되나, 하늘에 강한 태양이 떠 있음.",
    "hard_violations": [],
    "physics": "모든 객체가 지면과 암벽 위에 자연스럽게 고정되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "'차분한 일광과 제어된 대비'라는 조명 지시를 정확히 따랐으며, 레퍼런스의 암벽 지형과 프롬프트의 제비꽃밭 및 연구소를 완벽하게 조화시켰습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구성과 배경 요소는 잘 구현되었으나, 우측 상단에 강한 태양광이 비치어 '차분한 일광과 제어된 대비'라는 핵심 조명 지시를 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 시선은 도로변에서 언덕을 내려다보며, 보라색 제비꽃밭을 지나 원경의 연구소 건물을 향하고 있음.",
        "built_space": "원경 우측 중앙에 거대한 연구소 건물이 자리하고, 좌측 전경에 가드레일이 있는 도로가 있으며, 좌측 중경에는 레퍼런스와 일치하는 층리 암벽이 해안과 함께 배치됨.",
        "entities": "만개한 보라색 제비꽃밭, 거대한 연구소, 가드레일, 바다, 층리 형태의 해안 암벽 모두 프롬프트 및 레퍼런스와 일치하게 나타남.",
        "hard_violations": [],
        "physics": "건물, 가드레일, 암벽, 식물 등 모든 객체가 지면에 안정적으로 고정되어 있으며 물리적 결함이 없음."
       },
       {
        "label": "B",
        "direction": "카메라 시선은 도로변에서 제비꽃밭을 가로질러 멀리 떨어진 연구소를 향함.",
        "built_space": "원경 중앙에 연구소 건물이 위치하고, 좌측 전경에 도로와 가드레일, 좌측 해안에 층리 암벽이 배치됨.",
        "entities": "보라색 제비꽃밭, 연구소 건물, 도로, 암벽 등이 모두 식별되나, 하늘에 강한 태양이 떠 있음.",
        "hard_violations": [],
        "physics": "모든 객체가 지면과 암벽 위에 자연스럽게 고정되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "'차분한 일광과 제어된 대비'라는 조명 지시를 정확히 따랐으며, 레퍼런스의 암벽 지형과 프롬프트의 제비꽃밭 및 연구소를 완벽하게 조화시켰습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구성과 배경 요소는 잘 구현되었으나, 우측 상단에 강한 태양광이 비치어 '차분한 일광과 제어된 대비'라는 핵심 조명 지시를 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 시선은 도로변에서 언덕을 내려다보며, 보라색 제비꽃밭을 지나 원경의 연구소 건물을 향하고 있음.",
        "built_space": "원경 우측 중앙에 거대한 연구소 건물이 자리하고, 좌측 전경에 가드레일이 있는 도로가 있으며, 좌측 중경에는 레퍼런스와 일치하는 층리 암벽이 해안과 함께 배치됨.",
        "entities": "만개한 보라색 제비꽃밭, 거대한 연구소, 가드레일, 바다, 층리 형태의 해안 암벽 모두 프롬프트 및 레퍼런스와 일치하게 나타남.",
        "hard_violations": [],
        "physics": "건물, 가드레일, 암벽, 식물 등 모든 객체가 지면에 안정적으로 고정되어 있으며 물리적 결함이 없음."
       },
       {
        "label": "B",
        "direction": "카메라 시선은 도로변에서 제비꽃밭을 가로질러 멀리 떨어진 연구소를 향함.",
        "built_space": "원경 중앙에 연구소 건물이 위치하고, 좌측 전경에 도로와 가드레일, 좌측 해안에 층리 암벽이 배치됨.",
        "entities": "보라색 제비꽃밭, 연구소 건물, 도로, 암벽 등이 모두 식별되나, 하늘에 강한 태양이 떠 있음.",
        "hard_violations": [],
        "physics": "모든 객체가 지면과 암벽 위에 자연스럽게 고정되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "꽃밭 너머 연구소를 상단 중앙에 둔 완만한 하향 와이드 구도와 절제된 주광이 더 충실하지만, 참조 장소의 고정 지형은 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "꽃밭과 연구소 전경은 구현했으나 연구소가 오른쪽으로 치우치고 주광 대비가 더 강하며, 참조 장소의 고정 지형 역시 확인되지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 도로 가장자리의 높은 위치에서 꽃밭을 완만하게 내려다보며 상단 중앙의 연구소를 향한다. 전경의 꽃에서 중경의 꽃밭을 지나 건물로 시선이 이어진다. 사람의 시선, 무기, 이동 중인 몸체는 없다.",
        "built_space": "전경 왼쪽부터 하단으로 가드레일 한 줄이 지나간다. 연구소에는 높은 직사각형 주탑 하나와 여러 낮은 연결동, 오른쪽 연결교 하나가 보이며 꽃밭을 향한 건물 전면이 대체로 온전히 드러난다. 건물의 화면 점유 면적은 삼분의 일보다 작다. 참조는 양쪽 층리 암벽 사이의 좁은 바위 통로와 그 끝의 바다를 보여주지만, 여기서는 넓은 해안 부지로 재구성되어 그 통로와 암벽 배치를 특정할 수 없다. 참조에는 연구소 자체가 없어 건물 외형의 일치 여부는 검증할 수 없다.",
        "entities": "보라색 다섯 꽃잎과 넓은 잎을 가진 제비꽃들이 전경부터 중경까지 풍성하게 보이고, 그 너머 콘크리트와 유리로 된 대형 연구시설이 있다. 바다와 수평 층리의 해안 암석도 보인다. 사람이나 얼굴은 없고, 추가 자막이나 읽을 수 있는 표지도 없다. 정차한 트럭은 확실히 식별되지 않는다. 찰리와 흉부 부품은 인물 없는 이 구도에서 보이지 않는 것이 적절하다.",
        "hard_violations": [],
        "physics": "꽃과 관목은 지면에서 자라며 바위와 연구소는 지형 위에 놓여 있다. 가드레일은 도로변을 따라 설치되어 있고 떠 있는 물체는 보이지 않는다. 건물 연결부도 양끝 구조물에 이어져 있다. 바다의 밝은 반사는 오른쪽 위 낮은 태양 방향과 부합한다."
       },
       {
        "label": "B",
        "direction": "카메라는 도로 위쪽에서 꽃밭을 내려다보며 상단 중앙보다 오른쪽에 있는 연구소를 향한다. 중경의 진입로도 그 건물 쪽으로 이어진다. 사람의 시선이나 조준 대상, 이동 중인 몸체는 없다.",
        "built_space": "왼쪽 아래에 아스팔트 도로와 기둥으로 지지된 가드레일 한 줄이 보인다. 연구소에는 꼭대기가 두 갈래로 솟은 중앙 탑 하나, 오른쪽 돔 하나, 양옆 낮은 연결동과 오른쪽의 작은 부속동이 있다. 건물 전면 전체가 보이고 화면 점유 면적은 삼분의 일 미만이지만 중심은 오른쪽으로 치우친다. 왼쪽 해안에 층리 암벽과 틈이 있으나 참조의 양쪽 암벽, 중앙 통로, 끝의 바위턱과 대응하는 고정 배치는 확인되지 않는다. 참조에 없는 연구소 외형 자체는 비교할 근거가 없다.",
        "entities": "보라색 제비꽃밭과 큰 연구시설이 있으며, 전경의 꽃잎과 잎도 제비꽃의 형태로 읽힌다. 해안 암석과 바다가 배경을 이룬다. 인물과 얼굴, 자막은 없다. 정차한 트럭은 확실히 식별되지 않는다. 찰리와 부품은 이 무인 장소 전경에 등장하지 않는다. 밝은 바다와 선명한 그림자 때문에 요구된 절제된 주광보다 대비가 강하다.",
        "hard_violations": [],
        "physics": "가드레일은 보이는 도로변 기둥들이 받치고, 꽃과 바위는 지면에 붙어 있다. 주탑과 돔, 연결동은 연속된 건물 구조 위에 서 있다. 공중에 지지 없이 떠 있는 물체나 불가능한 동작은 없다. 오른쪽 바다의 반짝임은 오른쪽에서 오는 햇빛과 양립한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "꽃밭 너머 연구소를 상단 중앙에 둔 완만한 하향 와이드 구도와 절제된 주광이 더 충실하지만, 참조 장소의 고정 지형은 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "꽃밭과 연구소 전경은 구현했으나 연구소가 오른쪽으로 치우치고 주광 대비가 더 강하며, 참조 장소의 고정 지형 역시 확인되지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 도로 가장자리의 높은 위치에서 꽃밭을 완만하게 내려다보며 상단 중앙의 연구소를 향한다. 전경의 꽃에서 중경의 꽃밭을 지나 건물로 시선이 이어진다. 사람의 시선, 무기, 이동 중인 몸체는 없다.",
        "built_space": "전경 왼쪽부터 하단으로 가드레일 한 줄이 지나간다. 연구소에는 높은 직사각형 주탑 하나와 여러 낮은 연결동, 오른쪽 연결교 하나가 보이며 꽃밭을 향한 건물 전면이 대체로 온전히 드러난다. 건물의 화면 점유 면적은 삼분의 일보다 작다. 참조는 양쪽 층리 암벽 사이의 좁은 바위 통로와 그 끝의 바다를 보여주지만, 여기서는 넓은 해안 부지로 재구성되어 그 통로와 암벽 배치를 특정할 수 없다. 참조에는 연구소 자체가 없어 건물 외형의 일치 여부는 검증할 수 없다.",
        "entities": "보라색 다섯 꽃잎과 넓은 잎을 가진 제비꽃들이 전경부터 중경까지 풍성하게 보이고, 그 너머 콘크리트와 유리로 된 대형 연구시설이 있다. 바다와 수평 층리의 해안 암석도 보인다. 사람이나 얼굴은 없고, 추가 자막이나 읽을 수 있는 표지도 없다. 정차한 트럭은 확실히 식별되지 않는다. 찰리와 흉부 부품은 인물 없는 이 구도에서 보이지 않는 것이 적절하다.",
        "hard_violations": [],
        "physics": "꽃과 관목은 지면에서 자라며 바위와 연구소는 지형 위에 놓여 있다. 가드레일은 도로변을 따라 설치되어 있고 떠 있는 물체는 보이지 않는다. 건물 연결부도 양끝 구조물에 이어져 있다. 바다의 밝은 반사는 오른쪽 위 낮은 태양 방향과 부합한다."
       },
       {
        "label": "A",
        "direction": "카메라는 도로 위쪽에서 꽃밭을 내려다보며 상단 중앙보다 오른쪽에 있는 연구소를 향한다. 중경의 진입로도 그 건물 쪽으로 이어진다. 사람의 시선이나 조준 대상, 이동 중인 몸체는 없다.",
        "built_space": "왼쪽 아래에 아스팔트 도로와 기둥으로 지지된 가드레일 한 줄이 보인다. 연구소에는 꼭대기가 두 갈래로 솟은 중앙 탑 하나, 오른쪽 돔 하나, 양옆 낮은 연결동과 오른쪽의 작은 부속동이 있다. 건물 전면 전체가 보이고 화면 점유 면적은 삼분의 일 미만이지만 중심은 오른쪽으로 치우친다. 왼쪽 해안에 층리 암벽과 틈이 있으나 참조의 양쪽 암벽, 중앙 통로, 끝의 바위턱과 대응하는 고정 배치는 확인되지 않는다. 참조에 없는 연구소 외형 자체는 비교할 근거가 없다.",
        "entities": "보라색 제비꽃밭과 큰 연구시설이 있으며, 전경의 꽃잎과 잎도 제비꽃의 형태로 읽힌다. 해안 암석과 바다가 배경을 이룬다. 인물과 얼굴, 자막은 없다. 정차한 트럭은 확실히 식별되지 않는다. 찰리와 부품은 이 무인 장소 전경에 등장하지 않는다. 밝은 바다와 선명한 그림자 때문에 요구된 절제된 주광보다 대비가 강하다.",
        "hard_violations": [],
        "physics": "가드레일은 보이는 도로변 기둥들이 받치고, 꽃과 바위는 지면에 붙어 있다. 주탑과 돔, 연결동은 연속된 건물 구조 위에 서 있다. 공중에 지지 없이 떠 있는 물체나 불가능한 동작은 없다. 오른쪽 바다의 반짝임은 오른쪽에서 오는 햇빛과 양립한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "'차분한 일광과 제어된 대비'라는 조명 지시를 정확히 따랐으며, 레퍼런스의 암벽 지형과 프롬프트의 제비꽃밭 및 연구소를 완벽하게 조화시켰습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "구성과 배경 요소는 잘 구현되었으나, 우측 상단에 강한 태양광이 비치어 '차분한 일광과 제어된 대비'라는 핵심 조명 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_open_travel_truck_sel.png",
    "asset_id": "2e3db7e5-7f1f-47b6-8d3a-2d3a8a41eaaf",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d57-8807-78a6-9ad7-4a005ac4821b",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S72sh48::signage": {
  "fp": "2332191eedd8eebf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::414b27c857429341": {
  "subjects": [],
  "subject_text": "지동현의 연구소 외부·앞마당·뒷문, 인근 절벽·구출 지점\n오래 방치되어 낡은 대형 연구 시설 외부와 바다로 이어진 깎아지른 절벽이다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L172",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::flower_field_cliff": {
  "input_fingerprint": "4fb936086bd3a61a",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "flower_field_cliff",
    "tags": [
     "S72sh48"
    ]
   },
   "context_sig": "7a34966c0e0f2aaf"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n지동현의 연구소 외부·앞마당·뒷문, 인근 절벽·구출 지점: 오래 방치되어 낡은 대형 연구 시설 외부와 바다로 이어진 깎아지른 절벽이다. (특징: 녹슨 거대한 낡은 철문과 버려진 연구소 건물 벽면; 줄줄이 주차되는 유빅사의 검은 트럭들; 관자놀이에 소형 스위치가 달린 인간형 인공지능 전투 병사들 (유빅 유니폼 착용); 총구를 겨누며 다가오는 병사들과 절벽 끝으로 물러서는 찰리; 병사들을 들이받으며 난입하는 현우의 트럭 범퍼)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그러다 막다른 절벽에 도착한 찰리.\n- 재빨리 트럭 문을 여는 현우. 찰리에게 손을 내밀며-! ‘어서! 타!’\n\nTIME OF DAY (lock): day through sunset to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n지동현의 연구소 외부·앞마당·뒷문, 인근 절벽·구출 지점: 오래 방치되어 낡은 대형 연구 시설 외부와 바다로 이어진 깎아지른 절벽이다. (특징: 녹슨 거대한 낡은 철문과 버려진 연구소 건물 벽면; 줄줄이 주차되는 유빅사의 검은 트럭들; 관자놀이에 소형 스위치가 달린 인간형 인공지능 전투 병사들 (유빅 유니폼 착용); 총구를 겨누며 다가오는 병사들과 절벽 끝으로 물러서는 찰리; 병사들을 들이받으며 난입하는 현우의 트럭 범퍼)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그러다 막다른 절벽에 도착한 찰리.\n- 재빨리 트럭 문을 여는 현우. 찰리에게 손을 내밀며-! ‘어서! 타!’\n\nTIME OF DAY (lock): day through sunset to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flower_field_cliff_5989be.png",
  "asset_id": "61e06d8f-77ab-4fb9-82eb-f3a235fa4221",
  "input_asset_ids": [
   "39bcb0b3-d802-4810-8f20-04ee1d5af087"
  ],
  "origin_tag": "S72sh48",
  "place_text": "Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.",
  "origin_inputs": {
   "place_text": "Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.",
   "time_of_day_en": "day through sunset to night",
   "conti_asset_id": "39bcb0b3-d802-4810-8f20-04ee1d5af087"
  }
 },
 "S72sh48::bgfirst_bg": {
  "input_fingerprint": "69efe77efae4082d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 기계 손이 이현우의 손을 덥석 부여잡은 채 바닥을 박차고 허공에 뜬 mid-action 순간.\n\nLOCATION (lock): Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.\n\nTIME OF DAY (lock): day through sunset to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the doorway endpoint from the established low, rear-side tracking position and tilt slightly upward with 찰리's leap, keeping his full airborne body at left and 이현우 leaning out from the doorway at right. Their clasped hands meet near the center at natural scale, with visible ground beneath 찰리's lifted feet; 찰리 looks toward the opening while 이현우 looks down toward their grip. Emphasize the upward change in character position, retaining the same oblique action axis and enough doorway context to make the rescue physically readable.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open truck doorway receiving the leap in the middle-right of the frame, midground; ground visibly separated from Charlie's airborne feet in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 트럭 출입문 (Open for 찰리 to board) — The opening is seen obliquely from outside, with the threshold visible behind the joined hands; used as Destination of the leap and structural frame around 이현우; 제비꽃밭의 지면 (Below the boarding action during the pursuit); used as A visible gap beneath 찰리's feet establishes the airborne instant.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established deep sunset illumination retains readable human expressions and precise mechanical contours without an exaggerated action-lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리의 거대한 기계 손이 이현우의 손을 덥석 부여잡은 채 바닥을 박차고 허공에 뜬 mid-action 순간.\n\nLOCATION (lock): Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building.\n\nTIME OF DAY (lock): day through sunset to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the doorway endpoint from the established low, rear-side tracking position and tilt slightly upward with 찰리's leap, keeping his full airborne body at left and 이현우 leaning out from the doorway at right. Their clasped hands meet near the center at natural scale, with visible ground beneath 찰리's lifted feet; 찰리 looks toward the opening while 이현우 looks down toward their grip. Emphasize the upward change in character position, retaining the same oblique action axis and enough doorway context to make the rescue physically readable.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open truck doorway receiving the leap in the middle-right of the frame, midground; ground visibly separated from Charlie's airborne feet in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 트럭 출입문 (Open for 찰리 to board) — The opening is seen obliquely from outside, with the threshold visible behind the joined hands; used as Destination of the leap and structural frame around 이현우; 제비꽃밭의 지면 (Below the boarding action during the pursuit); used as A visible gap beneath 찰리's feet establishes the airborne instant.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established deep sunset illumination retains readable human expressions and precise mechanical contours without an exaggerated action-lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh48__bgfirst_bg.png",
  "asset_id": "f9fb961c-f191-418f-a0c7-5dadabb654aa",
  "input_asset_ids": [
   "39bcb0b3-d802-4810-8f20-04ee1d5af087",
   "61e06d8f-77ab-4fb9-82eb-f3a235fa4221"
  ]
 },
 "S72sh48": {
  "input_fingerprint": "232f4c0250a1c164",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 기계 손이 이현우의 손을 덥석 부여잡은 채 바닥을 박차고 허공에 뜬 mid-action 순간.\n\nLOCATION (lock): Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the doorway endpoint from the established low, rear-side tracking position and tilt slightly upward with 찰리's leap, keeping his full airborne body at left and 이현우 leaning out from the doorway at right. Their clasped hands meet near the center at natural scale, with visible ground beneath 찰리's lifted feet; 찰리 looks toward the opening while 이현우 looks down toward their grip. Emphasize the upward change in character position, retaining the same oblique action axis and enough doorway context to make the rescue physically readable.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open truck doorway receiving the leap in the middle-right of the frame, midground; ground visibly separated from Charlie's airborne feet in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 트럭 출입문 (Open for 찰리 to board) — The opening is seen obliquely from outside, with the threshold visible behind the joined hands; used as Destination of the leap and structural frame around 이현우; 제비꽃밭의 지면 (Below the boarding action during the pursuit); used as A visible gap beneath 찰리's feet establishes the airborne instant.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established deep sunset illumination retains readable human expressions and precise mechanical contours without an exaggerated action-lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field at sunset with its door open; armed humanoid robots wear Ubik uniforms and have small switches at their temples. Charlie retains his battered body, including the open chest damage exposing its interior, and possession of B-200's chest component. 이현우: He is at the open truck doorway with an arm extended, retaining his earlier wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 기계 손이 이현우의 손을 덥석 부여잡은 채 바닥을 박차고 허공에 뜬 mid-action 순간.\n\nLOCATION (lock): Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the doorway endpoint from the established low, rear-side tracking position and tilt slightly upward with 찰리's leap, keeping his full airborne body at left and 이현우 leaning out from the doorway at right. Their clasped hands meet near the center at natural scale, with visible ground beneath 찰리's lifted feet; 찰리 looks toward the opening while 이현우 looks down toward their grip. Emphasize the upward change in character position, retaining the same oblique action axis and enough doorway context to make the rescue physically readable.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open truck doorway receiving the leap in the middle-right of the frame, midground; ground visibly separated from Charlie's airborne feet in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 트럭 출입문 (Open for 찰리 to board) — The opening is seen obliquely from outside, with the threshold visible behind the joined hands; used as Destination of the leap and structural frame around 이현우; 제비꽃밭의 지면 (Below the boarding action during the pursuit); used as A visible gap beneath 찰리's feet establishes the airborne instant.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established deep sunset illumination retains readable human expressions and precise mechanical contours without an exaggerated action-lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field at sunset with its door open; armed humanoid robots wear Ubik uniforms and have small switches at their temples. Charlie retains his battered body, including the open chest damage exposing its interior, and possession of B-200's chest component. 이현우: He is at the open truck doorway with an arm extended, retaining his earlier wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 기계 손이 이현우의 손을 덥석 부여잡은 채 바닥을 박차고 허공에 뜬 mid-action 순간.\n\nLOCATION (lock): Beside the rescue truck's open doorway at the cliff-edge end of the violet field near the abandoned research building. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Reach the doorway endpoint from the established low, rear-side tracking position and tilt slightly upward with 찰리's leap, keeping his full airborne body at left and 이현우 leaning out from the doorway at right. Their clasped hands meet near the center at natural scale, with visible ground beneath 찰리's lifted feet; 찰리 looks toward the opening while 이현우 looks down toward their grip. Emphasize the upward change in character position, retaining the same oblique action axis and enough doorway context to make the rescue physically readable.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open truck doorway receiving the leap in the middle-right of the frame, midground; ground visibly separated from Charlie's airborne feet in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 트럭 출입문 (Open for 찰리 to board) — The opening is seen obliquely from outside, with the threshold visible behind the joined hands; used as Destination of the leap and structural frame around 이현우; 제비꽃밭의 지면 (Below the boarding action during the pursuit); used as A visible gap beneath 찰리's feet establishes the airborne instant.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established deep sunset illumination retains readable human expressions and precise mechanical contours without an exaggerated action-lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The truck has reached the cliff-edge violet field at sunset with its door open; armed humanoid robots wear Ubik uniforms and have small switches at their temples. Charlie retains his battered body, including the open chest damage exposing its interior, and possession of B-200's chest component. 이현우: He is at the open truck doorway with an arm extended, retaining his earlier wounds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh48__bgfirst_bg.png",
     "asset_id": "f9fb961c-f191-418f-a0c7-5dadabb654aa",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S72sh48.png",
     "asset_id": "39bcb0b3-d802-4810-8f20-04ee1d5af087",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flower_field_cliff_5989be.png",
     "asset_id": "61e06d8f-77ab-4fb9-82eb-f3a235fa4221",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 출입문 쪽을 바라보고, 이현우는 아래쪽의 맞잡은 손을 향해 시선을 두고 있습니다.",
    "built_space": "우측의 트럭 출입문, 배경의 폐건물 및 절벽 가장자리 공간이 프롬프트와 레퍼런스의 공간 구조에 부합합니다.",
    "entities": "이현우의 외형과 복장, 인이어 무전기는 일치합니다. 찰리의 외형도 레퍼런스와 같으나, 프롬프트에서 요구한 파손된 흉부와 내부 노출 상태가 묘사되지 않고 온전한 장갑판으로 표현되었습니다.",
    "hard_violations": [],
    "physics": "찰리가 바닥을 박차고 뛰어오른 공중 도약 상태(바닥의 흙먼지)와 이현우가 트럭 내부에 다리와 한쪽 팔을 의지해 몸을 내민 구조적 지지가 잘 표현되었습니다."
   },
   {
    "label": "B",
    "direction": "찰리는 트럭의 출입문을 향해 시선을 두고 있으며, 이현우는 맞잡은 손을 내려다보고 있습니다.",
    "built_space": "우측에 위치한 트럭의 열린 출입문, 절벽 가장자리의 제비꽃밭, 배경의 폐연구소 건물이 레퍼런스와 정확히 일치하는 위치에 배치되어 있습니다.",
    "entities": "이현우의 인체적 특징, 복장, 귀에 꽂힌 무전기 및 부상 흔적이 일치합니다. 찰리의 고릴라형 외형과 얼굴, 특히 프롬프트에서 명시한 흉부 장갑판 파손 및 내부 노출이 정확히 반영되었습니다.",
    "hard_violations": [],
    "physics": "찰리는 지면을 박차고 도약하여 공중에 떠 있으며(지면의 흙먼지가 발사점임을 보여줌), 이현우는 트럭 내부 프레임에 하체를 지지한 채 밖으로 몸을 기울여 물리적으로 자연스러운 자세를 취하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 찰리의 흉부 파손 및 내부 노출 디테일(Carried State)을 충실히 구현하여 프롬프트 일치도가 더 높습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 인물의 동작은 훌륭하나, 찰리의 기계 몸체가 파손되지 않은 깨끗한 상태로 묘사되어 유지되어야 할 상태(Carried State) 조건을 누락했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 트럭의 출입문을 향해 시선을 두고 있으며, 이현우는 맞잡은 손을 내려다보고 있습니다.",
        "built_space": "우측에 위치한 트럭의 열린 출입문, 절벽 가장자리의 제비꽃밭, 배경의 폐연구소 건물이 레퍼런스와 정확히 일치하는 위치에 배치되어 있습니다.",
        "entities": "이현우의 인체적 특징, 복장, 귀에 꽂힌 무전기 및 부상 흔적이 일치합니다. 찰리의 고릴라형 외형과 얼굴, 특히 프롬프트에서 명시한 흉부 장갑판 파손 및 내부 노출이 정확히 반영되었습니다.",
        "hard_violations": [],
        "physics": "찰리는 지면을 박차고 도약하여 공중에 떠 있으며(지면의 흙먼지가 발사점임을 보여줌), 이현우는 트럭 내부 프레임에 하체를 지지한 채 밖으로 몸을 기울여 물리적으로 자연스러운 자세를 취하고 있습니다."
       },
       {
        "label": "A",
        "direction": "찰리는 출입문 쪽을 바라보고, 이현우는 아래쪽의 맞잡은 손을 향해 시선을 두고 있습니다.",
        "built_space": "우측의 트럭 출입문, 배경의 폐건물 및 절벽 가장자리 공간이 프롬프트와 레퍼런스의 공간 구조에 부합합니다.",
        "entities": "이현우의 외형과 복장, 인이어 무전기는 일치합니다. 찰리의 외형도 레퍼런스와 같으나, 프롬프트에서 요구한 파손된 흉부와 내부 노출 상태가 묘사되지 않고 온전한 장갑판으로 표현되었습니다.",
        "hard_violations": [],
        "physics": "찰리가 바닥을 박차고 뛰어오른 공중 도약 상태(바닥의 흙먼지)와 이현우가 트럭 내부에 다리와 한쪽 팔을 의지해 몸을 내민 구조적 지지가 잘 표현되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 찰리의 흉부 파손 및 내부 노출 디테일(Carried State)을 충실히 구현하여 프롬프트 일치도가 더 높습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "구도와 인물의 동작은 훌륭하나, 찰리의 기계 몸체가 파손되지 않은 깨끗한 상태로 묘사되어 유지되어야 할 상태(Carried State) 조건을 누락했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 트럭의 출입문을 향해 시선을 두고 있으며, 이현우는 맞잡은 손을 내려다보고 있습니다.",
        "built_space": "우측에 위치한 트럭의 열린 출입문, 절벽 가장자리의 제비꽃밭, 배경의 폐연구소 건물이 레퍼런스와 정확히 일치하는 위치에 배치되어 있습니다.",
        "entities": "이현우의 인체적 특징, 복장, 귀에 꽂힌 무전기 및 부상 흔적이 일치합니다. 찰리의 고릴라형 외형과 얼굴, 특히 프롬프트에서 명시한 흉부 장갑판 파손 및 내부 노출이 정확히 반영되었습니다.",
        "hard_violations": [],
        "physics": "찰리는 지면을 박차고 도약하여 공중에 떠 있으며(지면의 흙먼지가 발사점임을 보여줌), 이현우는 트럭 내부 프레임에 하체를 지지한 채 밖으로 몸을 기울여 물리적으로 자연스러운 자세를 취하고 있습니다."
       },
       {
        "label": "A",
        "direction": "찰리는 출입문 쪽을 바라보고, 이현우는 아래쪽의 맞잡은 손을 향해 시선을 두고 있습니다.",
        "built_space": "우측의 트럭 출입문, 배경의 폐건물 및 절벽 가장자리 공간이 프롬프트와 레퍼런스의 공간 구조에 부합합니다.",
        "entities": "이현우의 외형과 복장, 인이어 무전기는 일치합니다. 찰리의 외형도 레퍼런스와 같으나, 프롬프트에서 요구한 파손된 흉부와 내부 노출 상태가 묘사되지 않고 온전한 장갑판으로 표현되었습니다.",
        "hard_violations": [],
        "physics": "찰리가 바닥을 박차고 뛰어오른 공중 도약 상태(바닥의 흙먼지)와 이현우가 트럭 내부에 다리와 한쪽 팔을 의지해 몸을 내민 구조적 지지가 잘 표현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "가슴 파손과 짧고 육중한 체형은 잘 맞지만, 찰리를 정면에 가깝게 보여 주고 이현우의 시선도 맞잡은 손보다 찰리 얼굴에 가까워 지정된 사선 도약 구도에서는 B보다 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 지면에서 오른쪽 출입구로 상승하는 전신 도약과 손을 내려다보는 이현우가 더 정확하지만, 찰리의 늘씬해진 비율과 드러나지 않은 가슴 파손은 불일치한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 뻗은 팔은 오른쪽 출입구와 이현우를 향하고, 두 손은 화면 중앙 부근에서 맞잡혀 있다. 이현우는 왼쪽 아래를 보지만 시선이 손보다는 찰리 얼굴 쪽에 가깝다. 찰리의 몸은 출입구 쪽으로 기울었으나 두 발이 몸 아래로 처져 상승 진행 방향은 B보다 덜 선명하다.",
        "built_space": "오른쪽에 열린 출입구 하나, 왼쪽으로 젖혀진 문짝 하나, 문턱 하나, 외부 발판 하나와 오른쪽 세로 손잡이 하나가 보인다. 이현우는 출입구 안에서 발을 문턱 부근에 대고 상체를 내밀었다. 맞잡은 손 뒤쪽으로 문턱과 실내가 보인다. 바퀴 두 개가 일부 보이며, 바다·절벽·제비꽃밭·돔 연구동의 관계도 장소 참조와 맞는다. 낮은 외부 사선 와이드 구도지만 찰리의 가슴 정면을 많이 보여 후방 측면 시점에는 덜 가깝다.",
        "entities": "찰리와 이현우만 등장한다. 찰리는 샌드 베이지 장갑, 긴 팔과 짧고 육중한 다리, 흰 마스크와 주황색 눈, 선형 입, 푸른 원자로를 갖추고 가슴에는 내부가 노출된 큰 파손이 있다. 별도의 B-200 가슴 부품은 확인되지 않는다. 이현우는 참조와 유사한 젊은 동아시아계 남성 외형에 짧고 헝클어진 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠·바지, 인이어 장치를 갖췄다. 노을과 금속·천·흙의 질감도 요구에 맞는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발과 지면 사이에 분명한 틈이 있으며 아래에 흩날리는 흙과 돌이 보여 박차고 오른 순간으로 읽힌다. 무릎이 굽고 몸이 출입구 쪽으로 기울어 있으며 맞잡은 손이 추가 접점을 만든다. 이현우는 반대 손으로 고정 손잡이를 잡고 발을 문턱 부근에 지지한다. 도약의 추진력 표현은 약하지만 근거 없이 떠 있는 몸으로 볼 정도는 아니다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴, 뻗은 팔과 몸통의 상승 대각선이 모두 오른쪽 열린 출입구를 향한다. 이현우는 고개를 왼쪽 아래로 숙여 중앙 부근의 손 결합을 바라본다. 뒤로 남긴 자유 팔과 왼쪽 아래로 뻗은 다리가 지면에서 출입구로 이동하는 방향을 명확히 만든다.",
        "built_space": "오른쪽에 출입구 하나, 바깥으로 열린 문짝 하나, 문턱 하나, 그 아래 발판 하나, 문 오른쪽의 세로 손잡이 하나가 보인다. 이현우는 출입구 안쪽에서 낮게 몸을 지지한 채 밖으로 팔을 뻗고 있어 출입구가 구조 프레임으로 기능한다. 손 뒤에 문턱이 보이고, 왼쪽 아래에는 찰리의 발과 분리된 꽃밭 지면이 충분히 남는다. 절벽·바다·돔 연구동과 트럭의 배치는 참조 장소에 부합한다. 낮은 외부 사선 와이드이며 A보다 측면성이 강하지만 완전한 후방 측면은 아니다.",
        "entities": "등장 인물은 찰리와 이현우뿐이다. 찰리의 베이지 장갑, 흰 기계 마스크, 주황색 눈과 푸른 가슴 원자로는 일치하지만 다리와 전체 몸이 참조보다 길고 날렵하게 보인다. 가슴은 원자로 주변 장갑이 대체로 온전해 요구된 개방 파손이 확인되지 않으며, 별도의 B-200 가슴 부품도 보이지 않는다. 이현우의 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 어두운 오염 의복과 인이어 장치는 부합한다. 상처 표현은 A보다 약하다.",
        "hard_violations": [],
        "physics": "찰리의 발 아래 왼쪽 지면에서 흙과 돌이 튀고, 뒤로 뻗은 다리와 앞으로 기울어진 몸이 이륙의 추진 방향을 보여 준다. 두 발은 지면에서 떨어져 있고 도착 지점은 오른쪽 문턱으로 명확하다. 기계 손은 이현우의 손 부위를 감싸 연결되며, 이현우는 다른 손으로 고정 손잡이를 잡고 굽힌 하체를 차량 내부에 두어 당기는 힘을 받는다. 도약과 지지 관계가 물리적으로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "가슴 파손과 짧고 육중한 체형은 잘 맞지만, 찰리를 정면에 가깝게 보여 주고 이현우의 시선도 맞잡은 손보다 찰리 얼굴에 가까워 지정된 사선 도약 구도에서는 B보다 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 지면에서 오른쪽 출입구로 상승하는 전신 도약과 손을 내려다보는 이현우가 더 정확하지만, 찰리의 늘씬해진 비율과 드러나지 않은 가슴 파손은 불일치한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 뻗은 팔은 오른쪽 출입구와 이현우를 향하고, 두 손은 화면 중앙 부근에서 맞잡혀 있다. 이현우는 왼쪽 아래를 보지만 시선이 손보다는 찰리 얼굴 쪽에 가깝다. 찰리의 몸은 출입구 쪽으로 기울었으나 두 발이 몸 아래로 처져 상승 진행 방향은 B보다 덜 선명하다.",
        "built_space": "오른쪽에 열린 출입구 하나, 왼쪽으로 젖혀진 문짝 하나, 문턱 하나, 외부 발판 하나와 오른쪽 세로 손잡이 하나가 보인다. 이현우는 출입구 안에서 발을 문턱 부근에 대고 상체를 내밀었다. 맞잡은 손 뒤쪽으로 문턱과 실내가 보인다. 바퀴 두 개가 일부 보이며, 바다·절벽·제비꽃밭·돔 연구동의 관계도 장소 참조와 맞는다. 낮은 외부 사선 와이드 구도지만 찰리의 가슴 정면을 많이 보여 후방 측면 시점에는 덜 가깝다.",
        "entities": "찰리와 이현우만 등장한다. 찰리는 샌드 베이지 장갑, 긴 팔과 짧고 육중한 다리, 흰 마스크와 주황색 눈, 선형 입, 푸른 원자로를 갖추고 가슴에는 내부가 노출된 큰 파손이 있다. 별도의 B-200 가슴 부품은 확인되지 않는다. 이현우는 참조와 유사한 젊은 동아시아계 남성 외형에 짧고 헝클어진 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠·바지, 인이어 장치를 갖췄다. 노을과 금속·천·흙의 질감도 요구에 맞는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발과 지면 사이에 분명한 틈이 있으며 아래에 흩날리는 흙과 돌이 보여 박차고 오른 순간으로 읽힌다. 무릎이 굽고 몸이 출입구 쪽으로 기울어 있으며 맞잡은 손이 추가 접점을 만든다. 이현우는 반대 손으로 고정 손잡이를 잡고 발을 문턱 부근에 지지한다. 도약의 추진력 표현은 약하지만 근거 없이 떠 있는 몸으로 볼 정도는 아니다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴, 뻗은 팔과 몸통의 상승 대각선이 모두 오른쪽 열린 출입구를 향한다. 이현우는 고개를 왼쪽 아래로 숙여 중앙 부근의 손 결합을 바라본다. 뒤로 남긴 자유 팔과 왼쪽 아래로 뻗은 다리가 지면에서 출입구로 이동하는 방향을 명확히 만든다.",
        "built_space": "오른쪽에 출입구 하나, 바깥으로 열린 문짝 하나, 문턱 하나, 그 아래 발판 하나, 문 오른쪽의 세로 손잡이 하나가 보인다. 이현우는 출입구 안쪽에서 낮게 몸을 지지한 채 밖으로 팔을 뻗고 있어 출입구가 구조 프레임으로 기능한다. 손 뒤에 문턱이 보이고, 왼쪽 아래에는 찰리의 발과 분리된 꽃밭 지면이 충분히 남는다. 절벽·바다·돔 연구동과 트럭의 배치는 참조 장소에 부합한다. 낮은 외부 사선 와이드이며 A보다 측면성이 강하지만 완전한 후방 측면은 아니다.",
        "entities": "등장 인물은 찰리와 이현우뿐이다. 찰리의 베이지 장갑, 흰 기계 마스크, 주황색 눈과 푸른 가슴 원자로는 일치하지만 다리와 전체 몸이 참조보다 길고 날렵하게 보인다. 가슴은 원자로 주변 장갑이 대체로 온전해 요구된 개방 파손이 확인되지 않으며, 별도의 B-200 가슴 부품도 보이지 않는다. 이현우의 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 어두운 오염 의복과 인이어 장치는 부합한다. 상처 표현은 A보다 약하다.",
        "hard_violations": [],
        "physics": "찰리의 발 아래 왼쪽 지면에서 흙과 돌이 튀고, 뒤로 뻗은 다리와 앞으로 기울어진 몸이 이륙의 추진 방향을 보여 준다. 두 발은 지면에서 떨어져 있고 도착 지점은 오른쪽 문턱으로 명확하다. 기계 손은 이현우의 손 부위를 감싸 연결되며, 이현우는 다른 손으로 고정 손잡이를 잡고 굽힌 하체를 차량 내부에 두어 당기는 힘을 받는다. 도약과 지지 관계가 물리적으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1875,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "프롬프트가 요구한 찰리의 흉부 파손 및 내부 노출 디테일(Carried State)을 충실히 구현하여 프롬프트 일치도가 더 높습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "구도와 인물의 동작은 훌륭하나, 찰리의 기계 몸체가 파손되지 않은 깨끗한 상태로 묘사되어 유지되어야 할 상태(Carried State) 조건을 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_flower_field_cliff_5989be.png",
    "asset_id": "61e06d8f-77ab-4fb9-82eb-f3a235fa4221",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d5c-d00e-7054-b066-b1e9153d7133",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh48__bgfirst_bg.png",
   "bg_asset_id": "f9fb961c-f191-418f-a0c7-5dadabb654aa",
   "bg_record_key": "S72sh48::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "flower_field_cliff",
   "groupbg_asset_id": "61e06d8f-77ab-4fb9-82eb-f3a235fa4221"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S72sh74::confined_fp_apt": {
  "applies": true,
  "reason_ko": "해당 샷은 트럭 캡 내부라는 밀폐되고 좌석 배치가 중요한 공간에서 진행되며, 인물들이 '조수석'에 위치한다는 명확한 지정이 있어 차량 내부의 방향과 구조가 올바르게 표현되어야 하므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "7f53d4e37d5fc596"
 },
 "S72sh74::signage": {
  "fp": "6944b5b7caed6d45",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "confinedfp::ac82c64ce3b1": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/confinedfp_base_ac82c64ce3b1.png",
  "place_text": "Inside the stopped truck's front passenger area on the nighttime road. The dark windows and rain outside enclose the dim cab.",
  "input_fingerprint": "46805e5708b5b66c"
 },
 "S72sh74::confined_fp": {
  "reads": {
   "controls": "The steering wheel is attached to the left-hand driver station, ahead of the empty driver seat. The passenger station has no steering wheel or other primary control. A center console separates the seats, but no controls are drawn on it.",
   "mirrors": "No mirror or explicitly reflective surface is marked. Side windows are shown along both outer walls; the broad outlined opening across the front appears to be the windshield. No mirror-face orientation or reflected view is specified.",
   "camera": "The camera is inside the cab at the rear, outboard corner of the passenger seat. Its arrow points diagonally forward and inward across the passenger station, toward the upper left of the plan. Camera height, downward tilt, field of view, and movement are not specified.",
   "occupants": "이현우 and 앰버 share the passenger station in an embracing arrangement. 이현우 is on its inboard side, toward the center console; 앰버 is immediately outboard of him, toward the passenger window. The driver seat is empty, and no third occupant is shown. Individual gaze directions and facial expressions are not marked."
  },
  "mismatches": [
   "The requested camera position is diagonally behind 이현우's shoulder, with his bent back forming the left foreground. The diagram instead places the camera behind the outboard edge of the passenger seat, on 앰버's side of the pair: 이현우 projects to screen-left but is farther from the camera than 앰버, rather than establishing the specified near shoulder-and-back foreground."
  ],
  "scene_description_en": "The view looks diagonally forward and inward from the rear outboard corner of the passenger station, placing the shared passenger seat in the near center-right and the driver station farther to screen-left. 이현우 occupies the inboard, screen-left side of the passenger seat, embracing 앰버 on his screen-right; she is slightly nearer the camera, and the diagram does not establish either person's precise facial direction or eye state. The passenger seat surrounds the pair, with its backrest nearest the camera and its seating direction toward the front dashboard. The empty driver seat appears farther left, facing the front, with its steering wheel ahead of it in the far-left part of the view and oriented toward the driver position. The center console lies to the left of the occupied passenger seat, between the two stations, with no controls marked on its visible upper surface. The dashboard and apparent windshield extend across the far front, while the passenger-side window lies to screen-right and the driver-side window farther to screen-left, their interior sides facing into the cab. No mirror is depicted, so there is no specified mirror face or physically supported reflected image to show.",
  "readback_fallback": {
   "first_model": "gemini-pro",
   "first_error": "LLM returned empty response for step=confined_fp_readback_S72sh74_fix, model=gemini-pro, finish_reason='content_filter'",
   "model": "gpt-high",
   "physical_model": "gpt-6-astra"
  },
  "readback_fallback_initial": {
   "first_model": "gemini-pro",
   "first_error": "LLM returned empty response for step=confined_fp_readback_S72sh74, model=gemini-pro, finish_reason='content_filter'",
   "model": "gpt-high",
   "physical_model": "gpt-6-astra"
  },
  "fixed": true,
  "input_fingerprint": "ee2b05454f3cf99d"
 },
 "S72sh74": {
  "input_fingerprint": "af1560d0fc5d17b4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 조수석에서 이현우가 축 늘어진 앰버의 상체를 자신의 품에 와락 껴안은 절망적인 구도.\n\nLOCATION (lock): Inside the stopped truck's front passenger area on the nighttime road. The dark windows and rain outside enclose the dim cab. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the short inward move from diagonally behind 이현우's shoulder inside the stopped truck, looking down across his arms toward the passenger seat. His bent back occupies the left foreground while 앰버's slack upper body lies diagonally through the center-right, her closed eyes visible above his enclosing arms; his lowered head remains directed toward her rather than the lens. Keep both upper bodies readable and make only the closing camera distance emphatic, withholding the later turn toward 찰리 and the window.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 조수석 (Supporting the embrace inside the stopped truck) — The seat is viewed diagonally from above and behind 이현우; used as A narrow visible perimeter anchors the bodies without competing with them.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low, source-unspecified nighttime illumination preserves the embrace and 앰버's face with controlled shadow detail before the later flashlight intrusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is held tightly in Hyunwoo's arms inside the stopped truck, her severely wounded body supported against him as she loses consciousness. Her head's precise angle, her body's facing direction, and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck has stopped on the road at night with its fuel completely exhausted, and rain is falling outside. Charlie retains his unrepaired body and chest damage and possession of B-200's component. 이현우: He is inside the stopped truck with his arms tightly gathered in an embrace, crying. His existing wounds and dirty appearance remain. 앰버: She has a severe bleeding wound in her side and is limp and increasingly drowsy. Her earlier head injury also remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe view looks diagonally forward and inward from the rear outboard corner of the passenger station, placing the shared passenger seat in the near center-right and the driver station farther to screen-left. 이현우 occupies the inboard, screen-left side of the passenger seat, embracing 앰버 on his screen-right; she is slightly nearer the camera, and the diagram does not establish either person's precise facial direction or eye state. The passenger seat surrounds the pair, with its backrest nearest the camera and its seating direction toward the front dashboard. The empty driver seat appears farther left, facing the front, with its steering wheel ahead of it in the far-left part of the view and oriented toward the driver position. The center console lies to the left of the occupied passenger seat, between the two stations, with no controls marked on its visible upper surface. The dashboard and apparent windshield extend across the far front, while the passenger-side window lies to screen-right and the driver-side window farther to screen-left, their interior sides facing into the cab. No mirror is depicted, so there is no specified mirror face or physically supported reflected image to show.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 조수석에서 이현우가 축 늘어진 앰버의 상체를 자신의 품에 와락 껴안은 절망적인 구도.\n\nLOCATION (lock): Inside the stopped truck's front passenger area on the nighttime road. The dark windows and rain outside enclose the dim cab. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low, source-unspecified nighttime illumination preserves the embrace and 앰버's face with controlled shadow detail before the later flashlight intrusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is held tightly in Hyunwoo's arms inside the stopped truck, her severely wounded body supported against him as she loses consciousness. Her head's precise angle, her body's facing direction, and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck has stopped on the road at night with its fuel completely exhausted, and rain is falling outside. Charlie retains his unrepaired body and chest damage and possession of B-200's component. 이현우: He is inside the stopped truck with his arms tightly gathered in an embrace, crying. His existing wounds and dirty appearance remain. 앰버: She has a severe bleeding wound in her side and is limp and increasingly drowsy. Her earlier head injury also remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe view looks diagonally forward and inward from the rear outboard corner of the passenger station, placing the shared passenger seat in the near center-right and the driver station farther to screen-left. 이현우 occupies the inboard, screen-left side of the passenger seat, embracing 앰버 on his screen-right; she is slightly nearer the camera, and the diagram does not establish either person's precise facial direction or eye state. The passenger seat surrounds the pair, with its backrest nearest the camera and its seating direction toward the front dashboard. The empty driver seat appears farther left, facing the front, with its steering wheel ahead of it in the far-left part of the view and oriented toward the driver position. The center console lies to the left of the occupied passenger seat, between the two stations, with no controls marked on its visible upper surface. The dashboard and apparent windshield extend across the far front, while the passenger-side window lies to screen-right and the driver-side window farther to screen-left, their interior sides facing into the cab. No mirror is depicted, so there is no specified mirror face or physically supported reflected image to show.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 조수석에서 이현우가 축 늘어진 앰버의 상체를 자신의 품에 와락 껴안은 절망적인 구도.\n\nLOCATION (lock): Inside the stopped truck's front passenger area on the nighttime road. The dark windows and rain outside enclose the dim cab. The shot takes place here.\n\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low, source-unspecified nighttime illumination preserves the embrace and 앰버's face with controlled shadow detail before the later flashlight intrusion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is held tightly in Hyunwoo's arms inside the stopped truck, her severely wounded body supported against him as she loses consciousness. Her head's precise angle, her body's facing direction, and the placement of her arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old truck has stopped on the road at night with its fuel completely exhausted, and rain is falling outside. Charlie retains his unrepaired body and chest damage and possession of B-200's component. 이현우: He is inside the stopped truck with his arms tightly gathered in an embrace, crying. His existing wounds and dirty appearance remain. 앰버: She has a severe bleeding wound in her side and is limp and increasingly drowsy. Her earlier head injury also remains unresolved.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh74_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "앰버",
     "path": "<bytes:1750847>",
     "asset_id": "4fd99635-8ccf-4aa4-be5a-bce26f963e04",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh74_confinedfp.png",
     "asset_id": null,
     "role": null
    },
    {
     "label": "이현우",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "앰버",
     "path": "<bytes:1750847>",
     "asset_id": "4fd99635-8ccf-4aa4-be5a-bce26f963e04",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "현우는 앰버 쪽으로 고개를 깊이 숙이지 않고 화면 왼쪽 앞을 바라본다. 앰버는 눈을 뜬 채 카메라 쪽을 본다. 요구된 현우의 하향 시선과 앰버의 감긴 눈 모두 다르다. 겨누는 무기나 이동 중인 물체는 없다.",
    "built_space": "왼쪽에 운전대 하나와 계기판, 그 뒤 운전석 방석 일부가 있고 오른쪽에 두 사람을 받치는 조수석 등받이 하나가 보인다. 변속 레버 하나, 실내 거울 하나, 빗물 맺힌 창과 문틀이 보이며 중복 좌석이나 불가능한 반사는 없다. 두 사람은 조수석에 함께 있어 공간 관계는 성립한다. 다만 카메라는 조수석 문 개구부 바깥에서 안을 보는 듯하며, 요구된 실내 어깨 뒤 하향 시점이 아니다. 좌석과 운전 장치도 좁은 테두리 이상으로 드러난다.",
    "entities": "두 사람만 보인다. 현우는 참조와 대체로 맞는 동아시아계의 젊은 남성으로 짧고 헝클어진 검은 머리, 어두운 셔츠와 인이어 장치를 갖췄다. 눈물은 보이지만 의복의 피와 기존 상처 표현은 약하다. 앰버는 금발의 어린 여자아이로 연령과 대략적인 외형은 맞지만, 참조의 얼굴 부상과 찢어진 작업복, 목의 호흡 장비가 재현되지 않았고 드러난 옆구리에도 심한 출혈이 보이지 않는다. 한국계 혼혈 여부 자체는 외관만으로 확정할 수 없다. 창밖은 회색으로 밝아 잠긴 밤의 인상도 약하다.",
    "hard_violations": [],
    "physics": "현우는 조수석에 앉아 등받이에 기대고, 앰버의 골반과 다리는 그의 무릎에 놓인다. 앰버의 머리는 현우의 가슴과 어깨에 기대며 몸통은 그의 두 팔이 받친다. 앰버의 팔과 손도 현우의 팔에 닿아 있어 지지 없는 부유는 없다. 그러나 앰버가 손가락을 펴 현우의 위팔을 감싸고 눈을 뜬 자세는 무력하게 늘어진 무의식 상태보다 능동적인 포옹으로 읽힌다."
   },
   {
    "label": "A",
    "direction": "현우는 얼굴을 앰버의 머리 쪽으로 숙여 울고 있으며 렌즈를 보지 않는다. 앰버는 눈을 감고 고개를 왼쪽으로 떨군다. 두 사람의 주의 방향은 요구에 맞지만, 카메라에는 현우의 등 대신 얼굴과 가슴이 정면에 가깝게 보인다. 허리의 작은 칼 모양 도구는 아래쪽을 향하며 누구를 겨누지는 않는다.",
    "built_space": "왼쪽 전경에 운전석 등받이 하나, 그 앞에 운전대 하나와 대시보드가 보이고 중앙 콘솔 하나가 두 좌석 사이에 있다. 앞유리 너머로 밤 도로가 펼쳐져 카메라는 객실 뒤에서 앞을 보는 위치다. 그런데 오른쪽 조수석 등받이는 현우 뒤, 즉 차량 앞유리 쪽에 놓이고 방석과 두 사람은 카메라 쪽으로 펼쳐져 조수석만 뒤를 향하는 구조다. 정상적인 전방 지향 좌석 배치와 맞지 않는다. 빗물 맺힌 창은 적절하지만 빈 운전석이 화면 왼쪽을 크게 차지하여 요구된 현우의 굽은 등과 좁은 좌석 테두리를 대체한다.",
    "entities": "현우와 앰버 두 사람만 있다. 현우의 젊은 동아시아계 남성 외형, 검은 머리, 더러운 어두운 셔츠, 인이어 장치와 우는 표정은 참조 및 지시에 대체로 맞는다. 앰버는 금발의 어린 여자아이이며 얼굴의 멍과 상처, 찢어진 카키색 작업복, 목의 호흡 장비와 허리 도구가 참조와 연결된다. 심한 출혈도 보이나 주로 몸통 앞쪽에 강조된다. 혼혈 정체성은 외관만으로 확정할 수 없다. 밤과 비는 분명하고 추가 인물이나 자막은 없다.",
    "hard_violations": [
     "앞유리와 운전대가 있는 차량 전방 쪽에 조수석 등받이를 놓고 방석은 후방 카메라 쪽으로 펼쳐, 평면도가 정한 전방 지향 조수석을 역방향으로 구성했다."
    ],
    "physics": "앰버의 몸통은 현우의 가슴과 감싼 두 팔에 기대고, 골반과 다리는 그의 무릎 위에 놓인다. 머리는 가슴 쪽으로 기울어 지지되며 왼팔은 아래로 늘어지고 오른손은 현우의 다리 위에 떨어져 있다. 호흡 장비는 목의 끈과 가슴에, 허리 도구는 벨트에 지지된다. 신체 자체의 무지지 부유는 없고 무력한 상태도 자연스럽지만, 이를 받치는 좌석의 차량 내 방향이 잘못되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "조수석 배치와 신체 지지는 성립하지만, 실내 어깨 뒤 시점 대신 문밖 정면에 가까운 구도이며 앰버가 눈을 뜨고 능동적으로 안아 핵심 순간을 놓쳤다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "울며 고개를 숙인 현우와 의식 잃은 앰버는 충실하지만, 조수석 등받이를 차량 앞쪽에 둔 역방향 구조가 치명적이며 왼쪽 전경도 현우의 등이 아닌 빈 좌석이 차지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 앰버 쪽으로 고개를 깊이 숙이지 않고 화면 왼쪽 앞을 바라본다. 앰버는 눈을 뜬 채 카메라 쪽을 본다. 요구된 현우의 하향 시선과 앰버의 감긴 눈 모두 다르다. 겨누는 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽에 운전대 하나와 계기판, 그 뒤 운전석 방석 일부가 있고 오른쪽에 두 사람을 받치는 조수석 등받이 하나가 보인다. 변속 레버 하나, 실내 거울 하나, 빗물 맺힌 창과 문틀이 보이며 중복 좌석이나 불가능한 반사는 없다. 두 사람은 조수석에 함께 있어 공간 관계는 성립한다. 다만 카메라는 조수석 문 개구부 바깥에서 안을 보는 듯하며, 요구된 실내 어깨 뒤 하향 시점이 아니다. 좌석과 운전 장치도 좁은 테두리 이상으로 드러난다.",
        "entities": "두 사람만 보인다. 현우는 참조와 대체로 맞는 동아시아계의 젊은 남성으로 짧고 헝클어진 검은 머리, 어두운 셔츠와 인이어 장치를 갖췄다. 눈물은 보이지만 의복의 피와 기존 상처 표현은 약하다. 앰버는 금발의 어린 여자아이로 연령과 대략적인 외형은 맞지만, 참조의 얼굴 부상과 찢어진 작업복, 목의 호흡 장비가 재현되지 않았고 드러난 옆구리에도 심한 출혈이 보이지 않는다. 한국계 혼혈 여부 자체는 외관만으로 확정할 수 없다. 창밖은 회색으로 밝아 잠긴 밤의 인상도 약하다.",
        "hard_violations": [],
        "physics": "현우는 조수석에 앉아 등받이에 기대고, 앰버의 골반과 다리는 그의 무릎에 놓인다. 앰버의 머리는 현우의 가슴과 어깨에 기대며 몸통은 그의 두 팔이 받친다. 앰버의 팔과 손도 현우의 팔에 닿아 있어 지지 없는 부유는 없다. 그러나 앰버가 손가락을 펴 현우의 위팔을 감싸고 눈을 뜬 자세는 무력하게 늘어진 무의식 상태보다 능동적인 포옹으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "현우는 얼굴을 앰버의 머리 쪽으로 숙여 울고 있으며 렌즈를 보지 않는다. 앰버는 눈을 감고 고개를 왼쪽으로 떨군다. 두 사람의 주의 방향은 요구에 맞지만, 카메라에는 현우의 등 대신 얼굴과 가슴이 정면에 가깝게 보인다. 허리의 작은 칼 모양 도구는 아래쪽을 향하며 누구를 겨누지는 않는다.",
        "built_space": "왼쪽 전경에 운전석 등받이 하나, 그 앞에 운전대 하나와 대시보드가 보이고 중앙 콘솔 하나가 두 좌석 사이에 있다. 앞유리 너머로 밤 도로가 펼쳐져 카메라는 객실 뒤에서 앞을 보는 위치다. 그런데 오른쪽 조수석 등받이는 현우 뒤, 즉 차량 앞유리 쪽에 놓이고 방석과 두 사람은 카메라 쪽으로 펼쳐져 조수석만 뒤를 향하는 구조다. 정상적인 전방 지향 좌석 배치와 맞지 않는다. 빗물 맺힌 창은 적절하지만 빈 운전석이 화면 왼쪽을 크게 차지하여 요구된 현우의 굽은 등과 좁은 좌석 테두리를 대체한다.",
        "entities": "현우와 앰버 두 사람만 있다. 현우의 젊은 동아시아계 남성 외형, 검은 머리, 더러운 어두운 셔츠, 인이어 장치와 우는 표정은 참조 및 지시에 대체로 맞는다. 앰버는 금발의 어린 여자아이이며 얼굴의 멍과 상처, 찢어진 카키색 작업복, 목의 호흡 장비와 허리 도구가 참조와 연결된다. 심한 출혈도 보이나 주로 몸통 앞쪽에 강조된다. 혼혈 정체성은 외관만으로 확정할 수 없다. 밤과 비는 분명하고 추가 인물이나 자막은 없다.",
        "hard_violations": [
         "앞유리와 운전대가 있는 차량 전방 쪽에 조수석 등받이를 놓고 방석은 후방 카메라 쪽으로 펼쳐, 평면도가 정한 전방 지향 조수석을 역방향으로 구성했다."
        ],
        "physics": "앰버의 몸통은 현우의 가슴과 감싼 두 팔에 기대고, 골반과 다리는 그의 무릎 위에 놓인다. 머리는 가슴 쪽으로 기울어 지지되며 왼팔은 아래로 늘어지고 오른손은 현우의 다리 위에 떨어져 있다. 호흡 장비는 목의 끈과 가슴에, 허리 도구는 벨트에 지지된다. 신체 자체의 무지지 부유는 없고 무력한 상태도 자연스럽지만, 이를 받치는 좌석의 차량 내 방향이 잘못되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "조수석 배치와 신체 지지는 성립하지만, 실내 어깨 뒤 시점 대신 문밖 정면에 가까운 구도이며 앰버가 눈을 뜨고 능동적으로 안아 핵심 순간을 놓쳤다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "울며 고개를 숙인 현우와 의식 잃은 앰버는 충실하지만, 조수석 등받이를 차량 앞쪽에 둔 역방향 구조가 치명적이며 왼쪽 전경도 현우의 등이 아닌 빈 좌석이 차지한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 앰버 쪽으로 고개를 깊이 숙이지 않고 화면 왼쪽 앞을 바라본다. 앰버는 눈을 뜬 채 카메라 쪽을 본다. 요구된 현우의 하향 시선과 앰버의 감긴 눈 모두 다르다. 겨누는 무기나 이동 중인 물체는 없다.",
        "built_space": "왼쪽에 운전대 하나와 계기판, 그 뒤 운전석 방석 일부가 있고 오른쪽에 두 사람을 받치는 조수석 등받이 하나가 보인다. 변속 레버 하나, 실내 거울 하나, 빗물 맺힌 창과 문틀이 보이며 중복 좌석이나 불가능한 반사는 없다. 두 사람은 조수석에 함께 있어 공간 관계는 성립한다. 다만 카메라는 조수석 문 개구부 바깥에서 안을 보는 듯하며, 요구된 실내 어깨 뒤 하향 시점이 아니다. 좌석과 운전 장치도 좁은 테두리 이상으로 드러난다.",
        "entities": "두 사람만 보인다. 현우는 참조와 대체로 맞는 동아시아계의 젊은 남성으로 짧고 헝클어진 검은 머리, 어두운 셔츠와 인이어 장치를 갖췄다. 눈물은 보이지만 의복의 피와 기존 상처 표현은 약하다. 앰버는 금발의 어린 여자아이로 연령과 대략적인 외형은 맞지만, 참조의 얼굴 부상과 찢어진 작업복, 목의 호흡 장비가 재현되지 않았고 드러난 옆구리에도 심한 출혈이 보이지 않는다. 한국계 혼혈 여부 자체는 외관만으로 확정할 수 없다. 창밖은 회색으로 밝아 잠긴 밤의 인상도 약하다.",
        "hard_violations": [],
        "physics": "현우는 조수석에 앉아 등받이에 기대고, 앰버의 골반과 다리는 그의 무릎에 놓인다. 앰버의 머리는 현우의 가슴과 어깨에 기대며 몸통은 그의 두 팔이 받친다. 앰버의 팔과 손도 현우의 팔에 닿아 있어 지지 없는 부유는 없다. 그러나 앰버가 손가락을 펴 현우의 위팔을 감싸고 눈을 뜬 자세는 무력하게 늘어진 무의식 상태보다 능동적인 포옹으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "현우는 얼굴을 앰버의 머리 쪽으로 숙여 울고 있으며 렌즈를 보지 않는다. 앰버는 눈을 감고 고개를 왼쪽으로 떨군다. 두 사람의 주의 방향은 요구에 맞지만, 카메라에는 현우의 등 대신 얼굴과 가슴이 정면에 가깝게 보인다. 허리의 작은 칼 모양 도구는 아래쪽을 향하며 누구를 겨누지는 않는다.",
        "built_space": "왼쪽 전경에 운전석 등받이 하나, 그 앞에 운전대 하나와 대시보드가 보이고 중앙 콘솔 하나가 두 좌석 사이에 있다. 앞유리 너머로 밤 도로가 펼쳐져 카메라는 객실 뒤에서 앞을 보는 위치다. 그런데 오른쪽 조수석 등받이는 현우 뒤, 즉 차량 앞유리 쪽에 놓이고 방석과 두 사람은 카메라 쪽으로 펼쳐져 조수석만 뒤를 향하는 구조다. 정상적인 전방 지향 좌석 배치와 맞지 않는다. 빗물 맺힌 창은 적절하지만 빈 운전석이 화면 왼쪽을 크게 차지하여 요구된 현우의 굽은 등과 좁은 좌석 테두리를 대체한다.",
        "entities": "현우와 앰버 두 사람만 있다. 현우의 젊은 동아시아계 남성 외형, 검은 머리, 더러운 어두운 셔츠, 인이어 장치와 우는 표정은 참조 및 지시에 대체로 맞는다. 앰버는 금발의 어린 여자아이이며 얼굴의 멍과 상처, 찢어진 카키색 작업복, 목의 호흡 장비와 허리 도구가 참조와 연결된다. 심한 출혈도 보이나 주로 몸통 앞쪽에 강조된다. 혼혈 정체성은 외관만으로 확정할 수 없다. 밤과 비는 분명하고 추가 인물이나 자막은 없다.",
        "hard_violations": [
         "앞유리와 운전대가 있는 차량 전방 쪽에 조수석 등받이를 놓고 방석은 후방 카메라 쪽으로 펼쳐, 평면도가 정한 전방 지향 조수석을 역방향으로 구성했다."
        ],
        "physics": "앰버의 몸통은 현우의 가슴과 감싼 두 팔에 기대고, 골반과 다리는 그의 무릎 위에 놓인다. 머리는 가슴 쪽으로 기울어 지지되며 왼팔은 아래로 늘어지고 오른손은 현우의 다리 위에 떨어져 있다. 호흡 장비는 목의 끈과 가슴에, 허리 도구는 벨트에 지지된다. 신체 자체의 무지지 부유는 없고 무력한 상태도 자연스럽지만, 이를 받치는 좌석의 차량 내 방향이 잘못되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 4,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "조수석 배치와 신체 지지는 성립하지만, 실내 어깨 뒤 시점 대신 문밖 정면에 가까운 구도이며 앰버가 눈을 뜨고 능동적으로 안아 핵심 순간을 놓쳤다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "울며 고개를 숙인 현우와 의식 잃은 앰버는 충실하지만, 조수석 등받이를 차량 앞쪽에 둔 역방향 구조가 치명적이며 왼쪽 전경도 현우의 등이 아닌 빈 좌석이 차지한다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh74_confinedfp.png",
    "asset_id": null,
    "role": null
   },
   {
    "label": "이현우",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "앰버",
    "path": "<bytes:1750847>",
    "asset_id": "4fd99635-8ccf-4aa4-be5a-bce26f963e04",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d66-57e1-7c79-9c95-55b34f752d42",
  "confined_fp": {
   "base_key": "confinedfp::ac82c64ce3b1",
   "apt_reason": "해당 샷은 트럭 캡 내부라는 밀폐되고 좌석 배치가 중요한 공간에서 진행되며, 인물들이 '조수석'에 위치한다는 명확한 지정이 있어 차량 내부의 방향과 구조가 올바르게 표현되어야 하므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": [
    "The requested camera position is diagonally behind 이현우's shoulder, with his bent back forming the left foreground. The diagram instead places the camera behind the outboard edge of the passenger seat, on 앰버's side of the pair: 이현우 projects to screen-left but is farther from the camera than 앰버, rather than establishing the specified near shoulder-and-back foreground."
   ]
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S72sh80::signage": {
  "fp": "5f7e954d665f2979",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S72sh80": {
  "input_fingerprint": "d42ea061a66a2342",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 앞의 빛 무리 속에서 사제복 차림의 신부가 손전등을 든 채 굳은 표정으로 서 있는 전신.\n\nLOCATION (lock): On the rain-soaked road immediately outside the stopped truck's open door, among the rescuers' flashlight beams. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the pan from outside and diagonally beside the passenger doorway, settling below eye height with a slight upward view of 신부's full body at center-right. In three-quarter view, he holds the flashlight and pauses with recognition tightening his expression, looking into the truck toward 이현우 at the left edge; 라울 remains beside him, also watching the interior, while the other arrivals stay beyond the crop. Keep the open door as a narrow intervening edge and emphasize the newly revealed character position rather than changing exposure or moving into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open passenger doorway between the interior and the priest in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 트럭 조수석 출입문 (Open) — Seen from outside at an oblique angle, with a slim interior portion visible at the left edge; used as Separates the priest outside from 이현우 inside while preserving their shared space; 신부의 손전등 (Held and switched on) — Directed into the truck rather than at the filming lens; used as A small, physically held object clarifying his arrival.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted strong flashlight illumination interrupts the nighttime darkness, with controlled highlights preserving the priest's expression instead of suggesting suspended luminous fog.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, and its door is now open. Charlie retains his unrepaired chest and body damage and possession of B-200's component. 이현우: He remains inside the truck, wounded and tearful, raising a hand against the glare. 라울: He has returned and is standing at the truck doorway. 신부: He stands at the open truck doorway, retaining his clerical collar. Strong flashlight beams flood the truck interior, creating a glare that obscures the figures beyond the open doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 신부 right now, so 신부's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 신부: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 앞의 빛 무리 속에서 사제복 차림의 신부가 손전등을 든 채 굳은 표정으로 서 있는 전신.\n\nLOCATION (lock): On the rain-soaked road immediately outside the stopped truck's open door, among the rescuers' flashlight beams. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the pan from outside and diagonally beside the passenger doorway, settling below eye height with a slight upward view of 신부's full body at center-right. In three-quarter view, he holds the flashlight and pauses with recognition tightening his expression, looking into the truck toward 이현우 at the left edge; 라울 remains beside him, also watching the interior, while the other arrivals stay beyond the crop. Keep the open door as a narrow intervening edge and emphasize the newly revealed character position rather than changing exposure or moving into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open passenger doorway between the interior and the priest in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 트럭 조수석 출입문 (Open) — Seen from outside at an oblique angle, with a slim interior portion visible at the left edge; used as Separates the priest outside from 이현우 inside while preserving their shared space; 신부의 손전등 (Held and switched on) — Directed into the truck rather than at the filming lens; used as A small, physically held object clarifying his arrival.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted strong flashlight illumination interrupts the nighttime darkness, with controlled highlights preserving the priest's expression instead of suggesting suspended luminous fog.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, and its door is now open. Charlie retains his unrepaired chest and body damage and possession of B-200's component. 이현우: He remains inside the truck, wounded and tearful, raising a hand against the glare. 라울: He has returned and is standing at the truck doorway. 신부: He stands at the open truck doorway, retaining his clerical collar. Strong flashlight beams flood the truck interior, creating a glare that obscures the figures beyond the open doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 신부 right now, so 신부's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 신부: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day through sunset to night.\n\nSHOT TEXT (authoritative, Korean): 열린 문 앞의 빛 무리 속에서 사제복 차림의 신부가 손전등을 든 채 굳은 표정으로 서 있는 전신.\n\nLOCATION (lock): On the rain-soaked road immediately outside the stopped truck's open door, among the rescuers' flashlight beams. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the pan from outside and diagonally beside the passenger doorway, settling below eye height with a slight upward view of 신부's full body at center-right. In three-quarter view, he holds the flashlight and pauses with recognition tightening his expression, looking into the truck toward 이현우 at the left edge; 라울 remains beside him, also watching the interior, while the other arrivals stay beyond the crop. Keep the open door as a narrow intervening edge and emphasize the newly revealed character position rather than changing exposure or moving into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open passenger doorway between the interior and the priest in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 트럭 조수석 출입문 (Open) — Seen from outside at an oblique angle, with a slim interior portion visible at the left edge; used as Separates the priest outside from 이현우 inside while preserving their shared space; 신부의 손전등 (Held and switched on) — Directed into the truck rather than at the filming lens; used as A small, physically held object clarifying his arrival.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted strong flashlight illumination interrupts the nighttime darkness, with controlled highlights preserving the priest's expression instead of suggesting suspended luminous fog.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The fuel-empty truck remains stopped on the rainy nighttime road, and its door is now open. Charlie retains his unrepaired chest and body damage and possession of B-200's component. 이현우: He remains inside the truck, wounded and tearful, raising a hand against the glare. 라울: He has returned and is standing at the truck doorway. 신부: He stands at the open truck doorway, retaining his clerical collar. Strong flashlight beams flood the truck interior, creating a glare that obscures the figures beyond the open doorway.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 신부 right now, so 신부's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 신부: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "신부의 시선과 손전등이 트럭 안을 향하고, 이현우가 손을 뻗어 빛을 막고 있음.",
    "built_space": "왼쪽에 열린 조수석 문과 내부가 보이고, 오른쪽에 비 오는 도로가 위치함.",
    "entities": "신부가 레퍼런스와 달리 라펠이 있는 일반 양복을 입음. 등장해서는 안 될 이전 샷의 금발 아이가 트럭 내부에 존재함.",
    "hard_violations": [
     "[gemini-pro] invented people or objects (프롬프트가 허용하지 않은 금발 아이 포함)",
     "[gpt-high] 이번 장면에 등장시키지 말라고 명시한 이전 장면의 금발 아이를 이현우 품에 재등장시켰다.",
     "[gpt-high] 화면 밖에 머물러야 할 다른 도착자들을 배경의 여러 인물로 노출했다.",
     "[gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
    ],
    "physics": "인물들은 지면과 좌석에 닿아 있으며 물리적 오류는 없음."
   },
   {
    "label": "B",
    "direction": "신부의 시선과 손전등 불빛이 트럭 내부의 이현우를 향하고 있으며, 이현우는 손을 들어 빛을 가리고 있음.",
    "built_space": "화면 왼쪽에 열린 트럭 문이 좁게 걸쳐 있고, 오른쪽에 비 내리는 도로와 배경의 불빛들이 배치됨.",
    "entities": "신부의 의상(차이나 칼라 스타일)과 외모가 레퍼런스와 일치함. 라울과 이현우가 명시된 대로 존재하며, 금지된 잉여 인물이 없음.",
    "hard_violations": [
     "[gpt-high] 화면 밖에 있어야 할 다른 도착자들이 배경에 적어도 두 명의 사람으로 노출된다.",
     "[gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
    ],
    "physics": "모든 인물이 지면에 서 있거나 차체에 안정적으로 앉아 있으며 손전등은 손에 쥐어져 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 대로 이전 샷의 다른 인물을 배제하고 명시된 인물만 정확히 배치했으며, 신부의 의상과 전체적인 구도 및 조명 연출이 프롬프트와 훌륭하게 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에서 명시적으로 제외할 것을 요구한 이전 샷의 금발 아이를 포함하는 심각한 위규를 범했으며, 신부의 복장도 레퍼런스와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "신부의 시선과 손전등 불빛이 트럭 내부의 이현우를 향하고 있으며, 이현우는 손을 들어 빛을 가리고 있음.",
        "built_space": "화면 왼쪽에 열린 트럭 문이 좁게 걸쳐 있고, 오른쪽에 비 내리는 도로와 배경의 불빛들이 배치됨.",
        "entities": "신부의 의상(차이나 칼라 스타일)과 외모가 레퍼런스와 일치함. 라울과 이현우가 명시된 대로 존재하며, 금지된 잉여 인물이 없음.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 서 있거나 차체에 안정적으로 앉아 있으며 손전등은 손에 쥐어져 있음."
       },
       {
        "label": "A",
        "direction": "신부의 시선과 손전등이 트럭 안을 향하고, 이현우가 손을 뻗어 빛을 막고 있음.",
        "built_space": "왼쪽에 열린 조수석 문과 내부가 보이고, 오른쪽에 비 오는 도로가 위치함.",
        "entities": "신부가 레퍼런스와 달리 라펠이 있는 일반 양복을 입음. 등장해서는 안 될 이전 샷의 금발 아이가 트럭 내부에 존재함.",
        "hard_violations": [
         "invented people or objects (프롬프트가 허용하지 않은 금발 아이 포함)"
        ],
        "physics": "인물들은 지면과 좌석에 닿아 있으며 물리적 오류는 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 대로 이전 샷의 다른 인물을 배제하고 명시된 인물만 정확히 배치했으며, 신부의 의상과 전체적인 구도 및 조명 연출이 프롬프트와 훌륭하게 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에서 명시적으로 제외할 것을 요구한 이전 샷의 금발 아이를 포함하는 심각한 위규를 범했으며, 신부의 복장도 레퍼런스와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부의 시선과 손전등 불빛이 트럭 내부의 이현우를 향하고 있으며, 이현우는 손을 들어 빛을 가리고 있음.",
        "built_space": "화면 왼쪽에 열린 트럭 문이 좁게 걸쳐 있고, 오른쪽에 비 내리는 도로와 배경의 불빛들이 배치됨.",
        "entities": "신부의 의상(차이나 칼라 스타일)과 외모가 레퍼런스와 일치함. 라울과 이현우가 명시된 대로 존재하며, 금지된 잉여 인물이 없음.",
        "hard_violations": [],
        "physics": "모든 인물이 지면에 서 있거나 차체에 안정적으로 앉아 있으며 손전등은 손에 쥐어져 있음."
       },
       {
        "label": "A",
        "direction": "신부의 시선과 손전등이 트럭 안을 향하고, 이현우가 손을 뻗어 빛을 막고 있음.",
        "built_space": "왼쪽에 열린 조수석 문과 내부가 보이고, 오른쪽에 비 오는 도로가 위치함.",
        "entities": "신부가 레퍼런스와 달리 라펠이 있는 일반 양복을 입음. 등장해서는 안 될 이전 샷의 금발 아이가 트럭 내부에 존재함.",
        "hard_violations": [
         "invented people or objects (프롬프트가 허용하지 않은 금발 아이 포함)"
        ],
        "physics": "인물들은 지면과 좌석에 닿아 있으며 물리적 오류는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금발 아이를 제외하고 신부와 라울의 내부를 향한 시선을 구현했지만, 배경 인물 추가와 실내 시점, 신부의 발이 잘린 구도로 핵심 지시를 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "신부의 전신과 지면 접촉은 구현했지만, 명시적으로 제외된 금발 아이까지 재등장시키고 구조대원을 노출했으며 카메라도 외부 사선 시점이 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부와 라울은 화면 왼쪽 트럭 안의 이현우 쪽을 보고 있다. 신부가 오른손에 쥔 손전등은 왼쪽 아래 실내 방향으로 빛을 보내며, 이현우의 얼굴보다 낮은 쪽을 비춘다. 발광면은 카메라에도 보이지만 광선의 주축은 왼쪽으로 비껴간다. 이현우는 신부 쪽으로 고개를 돌리고 손을 들어 빛을 가린다.",
        "built_space": "왼쪽 중경에 열린 문 한 개가 있고, 창유리 한 장과 안쪽 손잡이·수납부가 보인다. 왼쪽 끝에는 좌석 일부, 오른쪽에는 측면 거울 한 개, 아래에는 차량 내부의 테두리가 걸린다. 이현우의 등과 어깨가 왼쪽 전경을 크게 차지하고 신부와 라울은 문 밖 젖은 도로에 있다. 좁은 문 모서리를 외부에서 비스듬히 보는 구도라기보다 실내에서 밖을 보는 구도이며, 신부의 신발은 하단 밖으로 잘렸다.",
        "entities": "신부는 짧은 회색 머리의 고령 동아시아계 남성으로, 참고의 얼굴과 검은 사제복·흰 로만칼라에 대체로 부합한다. 라울은 갈색 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보이며, 남색 상의 위에 외투를 입었다. 이현우의 헝클어진 검은 머리와 어두운 회색 계열 옷은 이전 장면과 대체로 이어지지만 얼굴이 돌아가 있어 눈물과 부상은 확인하기 어렵다. 손전등 한 개는 신부 손에 있다. 이전 장면의 금발 아이는 없으나, 배경 차량 앞에 허용되지 않은 성인 인물이 적어도 두 명 보인다.",
        "hard_violations": [
         "화면 밖에 있어야 할 다른 도착자들이 배경에 적어도 두 명의 사람으로 노출된다.",
         "외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
        ],
        "physics": "신부의 손가락이 손전등 몸통을 감싸고 있어 물체의 지지가 분명하다. 신부와 라울은 수직으로 서 있고 다리가 하단으로 이어지지만 발 접촉은 잘려 보이지 않는다. 이를 공중 부양으로 볼 근거는 없다. 이현우는 차량 좌석 쪽에 몸을 둔 채 팔을 들어 올린 자연스러운 자세다. 젖은 도로의 빛 반사와 빗방울은 물리적으로 가능한 표현이다."
       },
       {
        "label": "B",
        "direction": "신부는 왼쪽 실내의 이현우를 보고 있고 라울도 같은 내부 방향을 주시한다. 신부의 오른손 손전등은 왼쪽 아래로 기울어 트럭 안쪽의 낮은 부분을 향한다. 이현우는 고개를 숙이고 손바닥을 문 쪽으로 들어 눈부심을 막는다. 금발 아이는 이현우에게 기대어 있으며 얼굴 방향은 대부분 가려져 있다.",
        "built_space": "중앙 왼쪽에 열린 문 한 개와 창유리·안쪽 손잡이·하부 스피커가 보인다. 전경 왼쪽에는 좌석과 그 위의 이현우 및 금발 아이가 크게 들어오고, 오른쪽과 아래에는 차량 개구부 테두리가 보인다. 신부는 중앙 오른쪽, 라울은 그 오른쪽 도로에 서며 두 사람의 신발까지 포함된다. 전신이라는 조건은 맞지만 실내가 화면의 상당 부분을 차지해, 외부 사선 카메라와 왼쪽 가장자리의 가느다란 실내 조각이라는 배치를 따르지 않는다.",
        "entities": "신부의 회색 짧은 머리, 고령 동아시아계 얼굴, 검은 사제복과 흰 로만칼라는 참고에 가깝다. 라울은 갈색 피부와 묶은 곱슬머리의 어린아이이며 남색 상의에 검은 외투를 걸쳤다. 이현우는 참고와 유사한 헝클어진 검은 머리와 회색 작업복을 유지하고, 얼굴에 젖은 흔적이 보인다. 그러나 이전 장면에서만 등장하고 이번에는 제외하라고 한 금발 아이가 이현우 품에 다시 등장한다. 배경에도 손전등을 든 추가 인물이 적어도 세 명 보인다.",
        "hard_violations": [
         "이번 장면에 등장시키지 말라고 명시한 이전 장면의 금발 아이를 이현우 품에 재등장시켰다.",
         "화면 밖에 머물러야 할 다른 도착자들을 배경의 여러 인물로 노출했다.",
         "외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
        ],
        "physics": "신부와 라울의 신발은 젖은 도로에 닿아 있고 체중을 지탱하는 자세가 자연스럽다. 신부는 오른손으로 손전등을 확실히 잡고 있다. 이현우는 좌석에 앉아 있으며 금발 아이는 그의 몸과 팔, 무릎에 지지된다. 금발 아이의 등장은 인물 제한 위반이지 부양이나 무지지 상태의 문제는 아니다. 빗속 광선과 노면 반사도 가능한 범위다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금발 아이를 제외하고 신부와 라울의 내부를 향한 시선을 구현했지만, 배경 인물 추가와 실내 시점, 신부의 발이 잘린 구도로 핵심 지시를 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "신부의 전신과 지면 접촉은 구현했지만, 명시적으로 제외된 금발 아이까지 재등장시키고 구조대원을 노출했으며 카메라도 외부 사선 시점이 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "신부와 라울은 화면 왼쪽 트럭 안의 이현우 쪽을 보고 있다. 신부가 오른손에 쥔 손전등은 왼쪽 아래 실내 방향으로 빛을 보내며, 이현우의 얼굴보다 낮은 쪽을 비춘다. 발광면은 카메라에도 보이지만 광선의 주축은 왼쪽으로 비껴간다. 이현우는 신부 쪽으로 고개를 돌리고 손을 들어 빛을 가린다.",
        "built_space": "왼쪽 중경에 열린 문 한 개가 있고, 창유리 한 장과 안쪽 손잡이·수납부가 보인다. 왼쪽 끝에는 좌석 일부, 오른쪽에는 측면 거울 한 개, 아래에는 차량 내부의 테두리가 걸린다. 이현우의 등과 어깨가 왼쪽 전경을 크게 차지하고 신부와 라울은 문 밖 젖은 도로에 있다. 좁은 문 모서리를 외부에서 비스듬히 보는 구도라기보다 실내에서 밖을 보는 구도이며, 신부의 신발은 하단 밖으로 잘렸다.",
        "entities": "신부는 짧은 회색 머리의 고령 동아시아계 남성으로, 참고의 얼굴과 검은 사제복·흰 로만칼라에 대체로 부합한다. 라울은 갈색 피부와 뒤로 묶은 곱슬머리의 어린 남자아이로 보이며, 남색 상의 위에 외투를 입었다. 이현우의 헝클어진 검은 머리와 어두운 회색 계열 옷은 이전 장면과 대체로 이어지지만 얼굴이 돌아가 있어 눈물과 부상은 확인하기 어렵다. 손전등 한 개는 신부 손에 있다. 이전 장면의 금발 아이는 없으나, 배경 차량 앞에 허용되지 않은 성인 인물이 적어도 두 명 보인다.",
        "hard_violations": [
         "화면 밖에 있어야 할 다른 도착자들이 배경에 적어도 두 명의 사람으로 노출된다.",
         "외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
        ],
        "physics": "신부의 손가락이 손전등 몸통을 감싸고 있어 물체의 지지가 분명하다. 신부와 라울은 수직으로 서 있고 다리가 하단으로 이어지지만 발 접촉은 잘려 보이지 않는다. 이를 공중 부양으로 볼 근거는 없다. 이현우는 차량 좌석 쪽에 몸을 둔 채 팔을 들어 올린 자연스러운 자세다. 젖은 도로의 빛 반사와 빗방울은 물리적으로 가능한 표현이다."
       },
       {
        "label": "A",
        "direction": "신부는 왼쪽 실내의 이현우를 보고 있고 라울도 같은 내부 방향을 주시한다. 신부의 오른손 손전등은 왼쪽 아래로 기울어 트럭 안쪽의 낮은 부분을 향한다. 이현우는 고개를 숙이고 손바닥을 문 쪽으로 들어 눈부심을 막는다. 금발 아이는 이현우에게 기대어 있으며 얼굴 방향은 대부분 가려져 있다.",
        "built_space": "중앙 왼쪽에 열린 문 한 개와 창유리·안쪽 손잡이·하부 스피커가 보인다. 전경 왼쪽에는 좌석과 그 위의 이현우 및 금발 아이가 크게 들어오고, 오른쪽과 아래에는 차량 개구부 테두리가 보인다. 신부는 중앙 오른쪽, 라울은 그 오른쪽 도로에 서며 두 사람의 신발까지 포함된다. 전신이라는 조건은 맞지만 실내가 화면의 상당 부분을 차지해, 외부 사선 카메라와 왼쪽 가장자리의 가느다란 실내 조각이라는 배치를 따르지 않는다.",
        "entities": "신부의 회색 짧은 머리, 고령 동아시아계 얼굴, 검은 사제복과 흰 로만칼라는 참고에 가깝다. 라울은 갈색 피부와 묶은 곱슬머리의 어린아이이며 남색 상의에 검은 외투를 걸쳤다. 이현우는 참고와 유사한 헝클어진 검은 머리와 회색 작업복을 유지하고, 얼굴에 젖은 흔적이 보인다. 그러나 이전 장면에서만 등장하고 이번에는 제외하라고 한 금발 아이가 이현우 품에 다시 등장한다. 배경에도 손전등을 든 추가 인물이 적어도 세 명 보인다.",
        "hard_violations": [
         "이번 장면에 등장시키지 말라고 명시한 이전 장면의 금발 아이를 이현우 품에 재등장시켰다.",
         "화면 밖에 머물러야 할 다른 도착자들을 배경의 여러 인물로 노출했다.",
         "외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
        ],
        "physics": "신부와 라울의 신발은 젖은 도로에 닿아 있고 체중을 지탱하는 자세가 자연스럽다. 신부는 오른손으로 손전등을 확실히 잡고 있다. 이현우는 좌석에 앉아 있으며 금발 아이는 그의 몸과 팔, 무릎에 지지된다. 금발 아이의 등장은 인물 제한 위반이지 부양이나 무지지 상태의 문제는 아니다. 빗속 광선과 노면 반사도 가능한 범위다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.095,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.845,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people or objects (프롬프트가 허용하지 않은 금발 아이 포함)",
     "[gpt-high] 이번 장면에 등장시키지 말라고 명시한 이전 장면의 금발 아이를 이현우 품에 재등장시켰다.",
     "[gpt-high] 화면 밖에 머물러야 할 다른 도착자들을 배경의 여러 인물로 노출했다.",
     "[gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
    ],
    "B": [
     "[gpt-high] 화면 밖에 있어야 할 다른 도착자들이 배경에 적어도 두 명의 사람으로 노출된다.",
     "[gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 845
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 대로 이전 샷의 다른 인물을 배제하고 명시된 인물만 정확히 배치했으며, 신부의 의상과 전체적인 구도 및 조명 연출이 프롬프트와 훌륭하게 일치합니다.  ★위반: [gpt-high] 화면 밖에 있어야 할 다른 도착자들이 배경에 적어도 두 명의 사람으로 노출된다. / [gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
   },
   {
    "label": "A",
    "score": 845,
    "verdict_ko": "프롬프트에서 명시적으로 제외할 것을 요구한 이전 샷의 금발 아이를 포함하는 심각한 위규를 범했으며, 신부의 복장도 레퍼런스와 다릅니다.  ★위반: [gemini-pro] invented people or objects (프롬프트가 허용하지 않은 금발 아이 포함) / [gpt-high] 이번 장면에 등장시키지 말라고 명시한 이전 장면의 금발 아이를 이현우 품에 재등장시켰다. / [gpt-high] 화면 밖에 머물러야 할 다른 도착자들을 배경의 여러 인물로 노출했다. / [gpt-high] 외부 조수석 문 옆으로 고정된 카메라 위치를 실내에서 바깥을 보는 시점으로 변경했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S72sh74_sel.png",
    "asset_id": "23632fdc-6e81-4b8c-9a29-b458968e8cc6",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1334467>",
    "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d85-64ce-7ffe-80fe-7642c30922fa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S72sh74"
  },
  "staged_characters_added": [
   "C03",
   "C06"
  ]
 },
 "S73sh6::signage": {
  "fp": "76050dd76dd3ebc6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::e0ae310ccd8caeab": {
  "subjects": [],
  "subject_text": "목포 보건소 내부\n열악한 장비가 갖춰진 지방 항구마을의 작은 의료 시설이다.",
  "identity": "canonical",
  "scope_id": "L125",
  "scope_role": "location_interior",
  "scope_sha": "41bda54ed1ed968b"
 },
 "S73sh6::bgfirst_bg": {
  "input_fingerprint": "4514461281e343ea",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 신부의 두 손이 이현우의 어깨를 따뜻하게 감싸 쥔 근접 구도.\n\nLOCATION (lock): Beside the recovering child's bed inside a harbor-town health clinic. The treatment room is lit for nighttime care.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly close to 이현우's nearer shoulder from his rear three-quarter side, slightly above shoulder height and looking gently downward. His bowed head and both shoulders occupy the lower center while 신부's two hands close around them from opposite sides, with only the priest's forearms and a small portion of his torso visible above. 이현우 looks down as he speaks about his promise, and the hands remain the focal point; emphasize proximity alone before the camera later withdraws.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination keeps skin and hand contact tactile, allowing the warmth to come from the gesture rather than an invented warm-colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 신부의 두 손이 이현우의 어깨를 따뜻하게 감싸 쥔 근접 구도.\n\nLOCATION (lock): Beside the recovering child's bed inside a harbor-town health clinic. The treatment room is lit for nighttime care.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly close to 이현우's nearer shoulder from his rear three-quarter side, slightly above shoulder height and looking gently downward. His bowed head and both shoulders occupy the lower center while 신부's two hands close around them from opposite sides, with only the priest's forearms and a small portion of his torso visible above. 이현우 looks down as he speaks about his promise, and the hands remain the focal point; emphasize proximity alone before the camera later withdraws.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination keeps skin and hand contact tactile, allowing the warmth to come from the gesture rather than an invented warm-colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S73sh6__bgfirst_bg.png",
  "asset_id": "d3d9a7aa-d260-45a4-bbe6-29e2a1aec38d",
  "input_asset_ids": [
   "cb7f47a2-8241-43a5-ab3a-ebc513c55373",
   "f773927f-983d-482b-92f2-821dee282724"
  ]
 },
 "S73sh6": {
  "input_fingerprint": "42ea3b01c3f27e88",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 두 손이 이현우의 어깨를 따뜻하게 감싸 쥔 근접 구도.\n\nLOCATION (lock): Beside the recovering child's bed inside a harbor-town health clinic. The treatment room is lit for nighttime care. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly close to 이현우's nearer shoulder from his rear three-quarter side, slightly above shoulder height and looking gently downward. His bowed head and both shoulders occupy the lower center while 신부's two hands close around them from opposite sides, with only the priest's forearms and a small portion of his torso visible above. 이현우 looks down as he speaks about his promise, and the hands remain the focal point; emphasize proximity alone before the camera later withdraws.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination keeps skin and hand contact tactile, allowing the warmth to come from the gesture rather than an invented warm-colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting is the nighttime Mokpo clinic, with a bed for the emergency-treated patient. Charlie retains his unrepaired battle damage and possession of B-200's chest component. 이현우: He stands at the bedside, still dirty and disheveled with his earlier wounds. 신부: He is at the bedside with his hands raised in a comforting gesture, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 두 손이 이현우의 어깨를 따뜻하게 감싸 쥔 근접 구도.\n\nLOCATION (lock): Beside the recovering child's bed inside a harbor-town health clinic. The treatment room is lit for nighttime care. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly close to 이현우's nearer shoulder from his rear three-quarter side, slightly above shoulder height and looking gently downward. His bowed head and both shoulders occupy the lower center while 신부's two hands close around them from opposite sides, with only the priest's forearms and a small portion of his torso visible above. 이현우 looks down as he speaks about his promise, and the hands remain the focal point; emphasize proximity alone before the camera later withdraws.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination keeps skin and hand contact tactile, allowing the warmth to come from the gesture rather than an invented warm-colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting is the nighttime Mokpo clinic, with a bed for the emergency-treated patient. Charlie retains his unrepaired battle damage and possession of B-200's chest component. 이현우: He stands at the bedside, still dirty and disheveled with his earlier wounds. 신부: He is at the bedside with his hands raised in a comforting gesture, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 신부의 두 손이 이현우의 어깨를 따뜻하게 감싸 쥔 근접 구도.\n\nLOCATION (lock): Beside the recovering child's bed inside a harbor-town health clinic. The treatment room is lit for nighttime care. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly close to 이현우's nearer shoulder from his rear three-quarter side, slightly above shoulder height and looking gently downward. His bowed head and both shoulders occupy the lower center while 신부's two hands close around them from opposite sides, with only the priest's forearms and a small portion of his torso visible above. 이현우 looks down as he speaks about his promise, and the hands remain the focal point; emphasize proximity alone before the camera later withdraws.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination keeps skin and hand contact tactile, allowing the warmth to come from the gesture rather than an invented warm-colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The setting is the nighttime Mokpo clinic, with a bed for the emergency-treated patient. Charlie retains his unrepaired battle damage and possession of B-200's chest component. 이현우: He stands at the bedside, still dirty and disheveled with his earlier wounds. 신부: He is at the bedside with his hands raised in a comforting gesture, wearing his clerical collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S73sh6__bgfirst_bg.png",
     "asset_id": "d3d9a7aa-d260-45a4-bbe6-29e2a1aec38d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S73sh6.png",
     "asset_id": "cb7f47a2-8241-43a5-ab3a-ebc513c55373",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L125B01.png",
     "asset_id": "f773927f-983d-482b-92f2-821dee282724",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:773901>",
     "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 이현우의 뒤에서 아래를 향하며, 신부는 이현우를 내려다봄.",
    "built_space": "참조 이미지와 유사한 병실 내부로, 좌측에 침대와 중앙에 창문이 배치됨.",
    "entities": "이현우(더러운 셔츠, 인이어 누락), 신부(사제복 착용, 프레이밍 지시와 달리 얼굴 전체 노출).",
    "hard_violations": [],
    "physics": "신부의 두 손이 이현우의 양어깨에 안정적으로 올려져 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 이현우의 어깨 뒤에서 아래를 향함.",
    "built_space": "침대와 창문, 의료 기기가 있는 야간 병실 내부 공간이 올바르게 구성됨.",
    "entities": "이현우(더러운 셔츠, 우측 귀에 인이어 착용), 신부(사제복 착용, 프레이밍 지시대로 얼굴 윗부분이 크롭됨).",
    "hard_violations": [],
    "physics": "신부의 두 손이 이현우의 어깨를 자연스럽게 감싸 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 근접 구도와 프레이밍(신부의 상체 일부만 노출)을 훌륭히 따랐으며 인이어 무전기 디테일도 정확하게 구현되었습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "신부의 얼굴이 온전히 나타나 프레이밍 지시를 위반했고, 이현우의 인이어 무전기가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 이현우의 뒤에서 아래를 향하며, 신부는 이현우를 내려다봄.",
        "built_space": "참조 이미지와 유사한 병실 내부로, 좌측에 침대와 중앙에 창문이 배치됨.",
        "entities": "이현우(더러운 셔츠, 인이어 누락), 신부(사제복 착용, 프레이밍 지시와 달리 얼굴 전체 노출).",
        "hard_violations": [],
        "physics": "신부의 두 손이 이현우의 양어깨에 안정적으로 올려져 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 이현우의 어깨 뒤에서 아래를 향함.",
        "built_space": "침대와 창문, 의료 기기가 있는 야간 병실 내부 공간이 올바르게 구성됨.",
        "entities": "이현우(더러운 셔츠, 우측 귀에 인이어 착용), 신부(사제복 착용, 프레이밍 지시대로 얼굴 윗부분이 크롭됨).",
        "hard_violations": [],
        "physics": "신부의 두 손이 이현우의 어깨를 자연스럽게 감싸 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 근접 구도와 프레이밍(신부의 상체 일부만 노출)을 훌륭히 따랐으며 인이어 무전기 디테일도 정확하게 구현되었습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "신부의 얼굴이 온전히 나타나 프레이밍 지시를 위반했고, 이현우의 인이어 무전기가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 이현우의 뒤에서 아래를 향하며, 신부는 이현우를 내려다봄.",
        "built_space": "참조 이미지와 유사한 병실 내부로, 좌측에 침대와 중앙에 창문이 배치됨.",
        "entities": "이현우(더러운 셔츠, 인이어 누락), 신부(사제복 착용, 프레이밍 지시와 달리 얼굴 전체 노출).",
        "hard_violations": [],
        "physics": "신부의 두 손이 이현우의 양어깨에 안정적으로 올려져 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 이현우의 어깨 뒤에서 아래를 향함.",
        "built_space": "침대와 창문, 의료 기기가 있는 야간 병실 내부 공간이 올바르게 구성됨.",
        "entities": "이현우(더러운 셔츠, 우측 귀에 인이어 착용), 신부(사제복 착용, 프레이밍 지시대로 얼굴 윗부분이 크롭됨).",
        "hard_violations": [],
        "physics": "신부의 두 손이 이현우의 어깨를 자연스럽게 감싸 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "어깨 뒤 사선에서 두 손의 감싸 쥠을 크게 잡아 지정된 근접 구도에 더 충실하지만, 신부의 하관과 몸통이 요구보다 많이 보입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "양어깨를 감싸는 동작과 야간 진료실은 맞지만, 신부의 얼굴·몸통과 배경까지 넓게 보여 손 중심의 밀착 클로즈업에서 더 멀어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 카메라에 등과 왼쪽 귀를 보이며 고개를 아래로 숙이고 있습니다. 눈은 가려져 정확한 응시점은 보이지 않습니다. 신부의 두 손은 서로 반대편에서 각각 이현우의 어깨를 감싸며, 접촉 대상이 분명합니다. 신부의 눈은 프레임 밖입니다.",
        "built_space": "왼쪽에 침대 한 개의 일부와 베개 하나, 모니터 한 대와 그 아래 장치 하나, 수액대 일부가 보입니다. 뒤에는 협탁 하나, 어두운 창 하나와 커튼 일부가 있습니다. 참조의 침대·모니터·창 주변 배치와 재질에 부합하며, 두 사람은 침대 옆에 있습니다. 중복 설비나 불가능한 반사는 보이지 않습니다. 뒤 사선의 가까운 시점이지만 신부의 몸통 노출은 지시보다 큽니다.",
        "entities": "보이는 인물은 이현우와 신부 두 명뿐입니다. 이현우의 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성으로 보이는 옆얼굴 일부, 피와 흙먼지가 묻은 어두운 셔츠, 귀의 소형 무전기가 요구와 맞습니다. 얼굴 대부분이 가려져 정확한 얼굴 일치는 판정하기 어렵습니다. 신부의 주름진 손과 하관은 고령 남성에 부합하고 검은 사제복과 흰 로만칼라가 보입니다. 아이나 찰리, 부품은 이 구도에 나타나지 않으며 추가 인물도 없습니다.",
        "hard_violations": [],
        "physics": "두 손 모두 소매에서 이어지는 손목을 갖고 있으며 실제 어깨 표면에 밀착합니다. 손가락은 어깨 곡면을 따라 굽어 있고 옷감에도 접촉이 읽힙니다. 숙인 머리는 목과 몸통에 자연스럽게 연결됩니다. 하체와 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만, 공중에 뜨거나 지지 없이 놓인 신체·물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 옆얼굴 일부를 보이며 아래로 고개를 숙입니다. 신부는 이현우의 머리와 어깨 쪽을 내려다봅니다. 양손은 각각 반대쪽 어깨에 닿아 있어 위로하는 동작의 대상과 방향이 맞습니다.",
        "built_space": "왼쪽에 침대 한 개, 베개 하나와 난간 일부, 벽등 하나, 의료 연결 패널, 모니터 한 대와 아래 장치 하나, 수액대 하나가 보입니다. 뒤에는 창 하나와 협탁 하나, 오른쪽에는 처치 카트 하나와 벽 부착 장치들이 보입니다. 참조 진료실의 주요 설비와 위치 관계를 유지하며 두 사람도 침대 옆에 있습니다. 다만 배경과 신부의 상반신을 넓게 포함해 지정된 밀착 구도보다 넓습니다.",
        "entities": "이현우와 신부만 보입니다. 이현우의 헝클어진 검은 머리, 젊고 마른 인상, 피와 먼지가 묻은 어두운 셔츠는 요구에 맞지만, 드러난 귀에서 인이어 무전기는 확인되지 않습니다. 숙인 얼굴로 인해 이현우의 정확한 얼굴 일치는 확인하기 어렵습니다. 신부는 참조와 유사한 고령 동아시아계 남성의 얼굴과 회색 머리 일부를 보이며, 검은 사제복과 흰 로만칼라를 착용했습니다. 신부의 얼굴이 크게 드러나는 것은 요구된 부분 노출 구도와 다릅니다.",
        "hard_violations": [],
        "physics": "신부의 팔은 몸통에서 손목과 손까지 자연스럽게 이어지고 양손은 이현우의 양어깨 위에 지지됩니다. 어깨를 누르며 감싸는 손가락과 숙인 목의 자세는 물리적으로 가능합니다. 두 사람의 발은 잘려 있어 직접적인 바닥 지지는 보이지 않지만 부유를 나타내는 정황은 없습니다. 침대와 카트 등 배경 물체에도 비정상적인 부유는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "어깨 뒤 사선에서 두 손의 감싸 쥠을 크게 잡아 지정된 근접 구도에 더 충실하지만, 신부의 하관과 몸통이 요구보다 많이 보입니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "양어깨를 감싸는 동작과 야간 진료실은 맞지만, 신부의 얼굴·몸통과 배경까지 넓게 보여 손 중심의 밀착 클로즈업에서 더 멀어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 카메라에 등과 왼쪽 귀를 보이며 고개를 아래로 숙이고 있습니다. 눈은 가려져 정확한 응시점은 보이지 않습니다. 신부의 두 손은 서로 반대편에서 각각 이현우의 어깨를 감싸며, 접촉 대상이 분명합니다. 신부의 눈은 프레임 밖입니다.",
        "built_space": "왼쪽에 침대 한 개의 일부와 베개 하나, 모니터 한 대와 그 아래 장치 하나, 수액대 일부가 보입니다. 뒤에는 협탁 하나, 어두운 창 하나와 커튼 일부가 있습니다. 참조의 침대·모니터·창 주변 배치와 재질에 부합하며, 두 사람은 침대 옆에 있습니다. 중복 설비나 불가능한 반사는 보이지 않습니다. 뒤 사선의 가까운 시점이지만 신부의 몸통 노출은 지시보다 큽니다.",
        "entities": "보이는 인물은 이현우와 신부 두 명뿐입니다. 이현우의 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성으로 보이는 옆얼굴 일부, 피와 흙먼지가 묻은 어두운 셔츠, 귀의 소형 무전기가 요구와 맞습니다. 얼굴 대부분이 가려져 정확한 얼굴 일치는 판정하기 어렵습니다. 신부의 주름진 손과 하관은 고령 남성에 부합하고 검은 사제복과 흰 로만칼라가 보입니다. 아이나 찰리, 부품은 이 구도에 나타나지 않으며 추가 인물도 없습니다.",
        "hard_violations": [],
        "physics": "두 손 모두 소매에서 이어지는 손목을 갖고 있으며 실제 어깨 표면에 밀착합니다. 손가락은 어깨 곡면을 따라 굽어 있고 옷감에도 접촉이 읽힙니다. 숙인 머리는 목과 몸통에 자연스럽게 연결됩니다. 하체와 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만, 공중에 뜨거나 지지 없이 놓인 신체·물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 옆얼굴 일부를 보이며 아래로 고개를 숙입니다. 신부는 이현우의 머리와 어깨 쪽을 내려다봅니다. 양손은 각각 반대쪽 어깨에 닿아 있어 위로하는 동작의 대상과 방향이 맞습니다.",
        "built_space": "왼쪽에 침대 한 개, 베개 하나와 난간 일부, 벽등 하나, 의료 연결 패널, 모니터 한 대와 아래 장치 하나, 수액대 하나가 보입니다. 뒤에는 창 하나와 협탁 하나, 오른쪽에는 처치 카트 하나와 벽 부착 장치들이 보입니다. 참조 진료실의 주요 설비와 위치 관계를 유지하며 두 사람도 침대 옆에 있습니다. 다만 배경과 신부의 상반신을 넓게 포함해 지정된 밀착 구도보다 넓습니다.",
        "entities": "이현우와 신부만 보입니다. 이현우의 헝클어진 검은 머리, 젊고 마른 인상, 피와 먼지가 묻은 어두운 셔츠는 요구에 맞지만, 드러난 귀에서 인이어 무전기는 확인되지 않습니다. 숙인 얼굴로 인해 이현우의 정확한 얼굴 일치는 확인하기 어렵습니다. 신부는 참조와 유사한 고령 동아시아계 남성의 얼굴과 회색 머리 일부를 보이며, 검은 사제복과 흰 로만칼라를 착용했습니다. 신부의 얼굴이 크게 드러나는 것은 요구된 부분 노출 구도와 다릅니다.",
        "hard_violations": [],
        "physics": "신부의 팔은 몸통에서 손목과 손까지 자연스럽게 이어지고 양손은 이현우의 양어깨 위에 지지됩니다. 어깨를 누르며 감싸는 손가락과 숙인 목의 자세는 물리적으로 가능합니다. 두 사람의 발은 잘려 있어 직접적인 바닥 지지는 보이지 않지만 부유를 나타내는 정황은 없습니다. 침대와 카트 등 배경 물체에도 비정상적인 부유는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.464,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.464,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1464
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 근접 구도와 프레이밍(신부의 상체 일부만 노출)을 훌륭히 따랐으며 인이어 무전기 디테일도 정확하게 구현되었습니다."
   },
   {
    "label": "A",
    "score": 1464,
    "verdict_ko": "신부의 얼굴이 온전히 나타나 프레이밍 지시를 위반했고, 이현우의 인이어 무전기가 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L125B01.png",
    "asset_id": "f773927f-983d-482b-92f2-821dee282724",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d8c-db47-77bb-94cc-8a1dc0189d3c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S73sh6__bgfirst_bg.png",
   "bg_asset_id": "d3d9a7aa-d260-45a4-bbe6-29e2a1aec38d",
   "bg_record_key": "S73sh6::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S73sh12::signage": {
  "fp": "f9a3885b091149cd",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S73sh12": {
  "input_fingerprint": "1e84e6c694fc3901",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 보건소 한가운데서 서로를 와락 끌어안은 역동적인 찰나.\n\nLOCATION (lock): In the open floor area of the harbor-town clinic's treatment room, near the patient's bed and entrance. Nighttime clinic lighting is on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Decelerate the lateral track beside 이현우's route at upper-chest height, looking slightly upward and obliquely across the reunion axis. Hold 이현우 on the left and 쿠마 on the right from the waist upward just after their arms close, their torsos still leaning into the contact and their faces turned past each other's shoulders rather than toward the camera. Let the closing character positions supply the motion accent, keeping the observers outside the crop until the subsequent retreat.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 보건소 중앙 공간 (The reunion takes place in the middle of the room); used as Peripheral room context distinguishes the embrace from the preceding shoulder detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued interior illumination and controlled contrast so the reunion feels emotionally warmer without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains occupied in the nighttime treatment room. Charlie retains his battle-damaged body and B-200's recovered chest component. 이현우: He remains dirty, disheveled and wounded, now standing with his arms in an embrace. 쿠마: He has entered the clinic and stands with his arms in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 보건소 한가운데서 서로를 와락 끌어안은 역동적인 찰나.\n\nLOCATION (lock): In the open floor area of the harbor-town clinic's treatment room, near the patient's bed and entrance. Nighttime clinic lighting is on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Decelerate the lateral track beside 이현우's route at upper-chest height, looking slightly upward and obliquely across the reunion axis. Hold 이현우 on the left and 쿠마 on the right from the waist upward just after their arms close, their torsos still leaning into the contact and their faces turned past each other's shoulders rather than toward the camera. Let the closing character positions supply the motion accent, keeping the observers outside the crop until the subsequent retreat.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 보건소 중앙 공간 (The reunion takes place in the middle of the room); used as Peripheral room context distinguishes the embrace from the preceding shoulder detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued interior illumination and controlled contrast so the reunion feels emotionally warmer without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains occupied in the nighttime treatment room. Charlie retains his battle-damaged body and B-200's recovered chest component. 이현우: He remains dirty, disheveled and wounded, now standing with his arms in an embrace. 쿠마: He has entered the clinic and stands with his arms in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우와 쿠마가 보건소 한가운데서 서로를 와락 끌어안은 역동적인 찰나.\n\nLOCATION (lock): In the open floor area of the harbor-town clinic's treatment room, near the patient's bed and entrance. Nighttime clinic lighting is on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Decelerate the lateral track beside 이현우's route at upper-chest height, looking slightly upward and obliquely across the reunion axis. Hold 이현우 on the left and 쿠마 on the right from the waist upward just after their arms close, their torsos still leaning into the contact and their faces turned past each other's shoulders rather than toward the camera. Let the closing character positions supply the motion accent, keeping the observers outside the crop until the subsequent retreat.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 보건소 중앙 공간 (The reunion takes place in the middle of the room); used as Peripheral room context distinguishes the embrace from the preceding shoulder detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established subdued interior illumination and controlled contrast so the reunion feels emotionally warmer without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains occupied in the nighttime treatment room. Charlie retains his battle-damaged body and B-200's recovered chest component. 이현우: He remains dirty, disheveled and wounded, now standing with his arms in an embrace. 쿠마: He has entered the clinic and stands with his arms in an embrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — wearing: 쿠마 패거리 우두머리다운 투박하고 낡은 검은색 가죽 재킷과 그런지한 데님 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
    "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
    "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 하지만 이현우의 등에 놓인 아래쪽 손은 가죽 소매가 아닌 캔버스 소매로 잘못 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조: 이현우의 등에 얹어진 두 손이 모두 엄지가 위로 향하고 손가락이 왼쪽을 가리키는 오른손 형태를 띠고 있음."
    ],
    "physics": "두 인물은 서서 서로에게 기대어 체중을 지지하고 있음. 그러나 이현우의 등에 얹어진 아래쪽 손은 위치와 꺾인 각도를 볼 때 해부학적으로 불가능한 구조임."
   },
   {
    "label": "B",
    "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
    "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
    "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 이현우의 등에 놓인 아래쪽 손은 이현우의 셔츠와 동일한 소매를 입고 있음.",
    "hard_violations": [],
    "physics": "두 인물은 서서 서로에게 체중을 싣고 지지하고 있음. 쿠마의 오른팔은 이현우의 어깨를 감싸고 있으며, 이현우의 왼팔은 물리적으로 가능은 하나 어색하게 자신의 등 아래쪽을 짚고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "배경, 인물, 조명이 프롬프트 및 레퍼런스와 훌륭하게 일치하나, 이현우의 왼팔이 상대를 안지 않고 자신의 등을 감싸고 있는 포즈 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 등에 얹어진 두 손이 모두 오른손 형태로 생성되었고 소매 재질도 잘못 렌더링되어 해부학적 하드 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
        "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
        "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 하지만 이현우의 등에 놓인 아래쪽 손은 가죽 소매가 아닌 캔버스 소매로 잘못 묘사됨.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조: 이현우의 등에 얹어진 두 손이 모두 엄지가 위로 향하고 손가락이 왼쪽을 가리키는 오른손 형태를 띠고 있음."
        ],
        "physics": "두 인물은 서서 서로에게 기대어 체중을 지지하고 있음. 그러나 이현우의 등에 얹어진 아래쪽 손은 위치와 꺾인 각도를 볼 때 해부학적으로 불가능한 구조임."
       },
       {
        "label": "B",
        "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
        "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
        "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 이현우의 등에 놓인 아래쪽 손은 이현우의 셔츠와 동일한 소매를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물은 서서 서로에게 체중을 싣고 지지하고 있음. 쿠마의 오른팔은 이현우의 어깨를 감싸고 있으며, 이현우의 왼팔은 물리적으로 가능은 하나 어색하게 자신의 등 아래쪽을 짚고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "배경, 인물, 조명이 프롬프트 및 레퍼런스와 훌륭하게 일치하나, 이현우의 왼팔이 상대를 안지 않고 자신의 등을 감싸고 있는 포즈 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이현우의 등에 얹어진 두 손이 모두 오른손 형태로 생성되었고 소매 재질도 잘못 렌더링되어 해부학적 하드 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
        "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
        "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 하지만 이현우의 등에 놓인 아래쪽 손은 가죽 소매가 아닌 캔버스 소매로 잘못 묘사됨.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조: 이현우의 등에 얹어진 두 손이 모두 엄지가 위로 향하고 손가락이 왼쪽을 가리키는 오른손 형태를 띠고 있음."
        ],
        "physics": "두 인물은 서서 서로에게 기대어 체중을 지지하고 있음. 그러나 이현우의 등에 얹어진 아래쪽 손은 위치와 꺾인 각도를 볼 때 해부학적으로 불가능한 구조임."
       },
       {
        "label": "B",
        "direction": "두 인물의 시선과 얼굴은 서로의 어깨 너머를 향하고 있으며, 카메라는 포옹하는 이현우와 쿠마를 담고 있음.",
        "built_space": "보건소 치료실. 화면 왼쪽에 환자 침대와 모니터, 중앙 배경에 수납장, 오른쪽에 출입문이 위치해 이전 샷의 공간과 완벽히 일치함.",
        "entities": "이현우(왼쪽)는 검은 머리, 피 묻은 셔츠, 인이어를 착용해 일치함. 쿠마(오른쪽)는 비니와 가죽 재킷을 착용함. 이현우의 등에 놓인 아래쪽 손은 이현우의 셔츠와 동일한 소매를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물은 서서 서로에게 체중을 싣고 지지하고 있음. 쿠마의 오른팔은 이현우의 어깨를 감싸고 있으며, 이현우의 왼팔은 물리적으로 가능은 하나 어색하게 자신의 등 아래쪽을 짚고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "왼쪽 이현우·오른쪽 쿠마의 밀착 포옹과 인이어 무전기까지 더 충실하지만, 허리 위 미디엄 숏보다 타이트하고 약한 앙각 및 병상 점유 상태는 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물 배치와 어깨에 묻힌 얼굴 방향은 맞지만, 포옹이 더 정적으로 보이고 이현우의 인이어가 확인되지 않으며 타이트한 구도와 빈 병상 문제도 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽에서 등을 카메라에 보이며 얼굴을 쿠마의 어깨 쪽으로 숙인다. 오른쪽 쿠마는 눈을 감고 이현우의 목과 어깨 쪽에 얼굴을 밀착한다. 두 사람 모두 렌즈를 보지 않고 팔로 상대의 등을 감싼다. 서로의 어깨 너머로 얼굴을 돌리라는 지시에는 대체로 부합한다.",
        "built_space": "왼쪽에 병상 한 개와 모니터 한 대, 뒤쪽에 낮은 수납장 한 개와 야간 창문 한 개, 오른쪽에 출입문 한 개와 의료용 카트 일부가 보인다. 두 사람은 병상과 출입문 사이 열린 공간에서 포옹한다. 회색 벽, 푸른 침구, 모니터와 창문의 관계는 이전 장면과 유사하다. 다만 노출된 베개와 침상은 비어 보여 병상이 계속 점유되어 있다는 조건과 맞지 않는다. 화면은 허리보다 위에서 끊기는 타이트한 상반신 구도이며, 지정된 약한 앙각은 뚜렷하지 않다.",
        "entities": "젊은 동아시아계 외형의 남성 두 명만 보이고 관찰자는 없다. 이현우의 헝클어진 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠, 귀의 검은 인이어가 이전 장면과 잘 이어진다. 얼굴 대부분은 가려져 정확한 얼굴 일치는 판별하기 어렵다. 쿠마는 참조의 짙은 비니와 낡은 검은 가죽 재킷을 착용했고 드러난 얼굴도 대체로 부합한다. 바지와 찰리의 흉부 부품은 구도 밖이므로 판정할 수 없다.",
        "hard_violations": [],
        "physics": "쿠마의 손 두 개가 이현우의 등 위쪽과 아래쪽에 실제로 닿고, 이현우의 팔은 쿠마의 어깨를 감싼다. 몸통이 서로 기울어 접촉하며 옷 주름도 압박 방향을 따른다. 발과 바닥 접촉은 잘렸지만 공중에 뜬 자세는 아니며, 보이는 범위에서 불가능한 관절이나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 이현우는 고개를 숙여 쿠마의 어깨 쪽을 향하고, 오른쪽 쿠마는 눈을 감은 채 이현우의 어깨에 얼굴을 댄다. 양쪽 팔은 상대 몸통을 감싸며 카메라를 향한 시선은 없다. 얼굴 방향은 지시와 대체로 맞지만 이미 안긴 채 머무는 순간처럼 보인다.",
        "built_space": "왼쪽에 병상 한 개와 모니터 한 대, 뒤에 수납장 한 개와 야간 창문 한 개, 오른쪽에 출입문 한 개, 의료용 카트와 바퀴 달린 의자 한 개가 보인다. 두 사람은 병상 옆 열린 공간에 서 있고 설비 중복이나 불가능한 반사는 없다. 이전 장면의 주요 재료와 배치는 유지되지만, 보이는 베개와 넓은 침상 면에는 환자가 없어 점유 상태 조건과 어긋난다. 역시 허리 위 전체를 담기보다 등과 어깨 중심으로 타이트하며 약한 앙각이 명확하지 않다.",
        "entities": "참조와 유사한 젊은 동아시아계 외형의 남성 두 명만 등장한다. 이현우의 검은 헝클어진 머리와 피·먼지가 묻은 셔츠는 이어지지만, 노출된 귀에서 지정된 인이어는 확인되지 않는다. 이현우의 얼굴은 대부분 가려져 있다. 쿠마의 비니, 검은 가죽 재킷과 보이는 얼굴은 참조에 대체로 맞는다. 하의와 찰리의 소지품은 화면 밖이다.",
        "hard_violations": [],
        "physics": "쿠마의 두 손이 이현우의 위쪽 등과 허리 부근을 붙잡고, 이현우의 팔은 쿠마의 어깨 뒤로 이어진다. 팔과 몸통의 접촉은 물리적으로 가능하며 옷도 눌리고 접혀 있다. 하체가 잘려 발의 지지는 직접 확인되지 않지만 부유를 나타내는 징후는 없다. 다만 몸통의 기울기와 압박이 비교적 안정적이어서 와락 끌어안은 직후의 운동감은 약하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 이현우·오른쪽 쿠마의 밀착 포옹과 인이어 무전기까지 더 충실하지만, 허리 위 미디엄 숏보다 타이트하고 약한 앙각 및 병상 점유 상태는 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물 배치와 어깨에 묻힌 얼굴 방향은 맞지만, 포옹이 더 정적으로 보이고 이현우의 인이어가 확인되지 않으며 타이트한 구도와 빈 병상 문제도 남는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 왼쪽에서 등을 카메라에 보이며 얼굴을 쿠마의 어깨 쪽으로 숙인다. 오른쪽 쿠마는 눈을 감고 이현우의 목과 어깨 쪽에 얼굴을 밀착한다. 두 사람 모두 렌즈를 보지 않고 팔로 상대의 등을 감싼다. 서로의 어깨 너머로 얼굴을 돌리라는 지시에는 대체로 부합한다.",
        "built_space": "왼쪽에 병상 한 개와 모니터 한 대, 뒤쪽에 낮은 수납장 한 개와 야간 창문 한 개, 오른쪽에 출입문 한 개와 의료용 카트 일부가 보인다. 두 사람은 병상과 출입문 사이 열린 공간에서 포옹한다. 회색 벽, 푸른 침구, 모니터와 창문의 관계는 이전 장면과 유사하다. 다만 노출된 베개와 침상은 비어 보여 병상이 계속 점유되어 있다는 조건과 맞지 않는다. 화면은 허리보다 위에서 끊기는 타이트한 상반신 구도이며, 지정된 약한 앙각은 뚜렷하지 않다.",
        "entities": "젊은 동아시아계 외형의 남성 두 명만 보이고 관찰자는 없다. 이현우의 헝클어진 검은 머리, 마른 체격, 피와 먼지가 묻은 어두운 셔츠, 귀의 검은 인이어가 이전 장면과 잘 이어진다. 얼굴 대부분은 가려져 정확한 얼굴 일치는 판별하기 어렵다. 쿠마는 참조의 짙은 비니와 낡은 검은 가죽 재킷을 착용했고 드러난 얼굴도 대체로 부합한다. 바지와 찰리의 흉부 부품은 구도 밖이므로 판정할 수 없다.",
        "hard_violations": [],
        "physics": "쿠마의 손 두 개가 이현우의 등 위쪽과 아래쪽에 실제로 닿고, 이현우의 팔은 쿠마의 어깨를 감싼다. 몸통이 서로 기울어 접촉하며 옷 주름도 압박 방향을 따른다. 발과 바닥 접촉은 잘렸지만 공중에 뜬 자세는 아니며, 보이는 범위에서 불가능한 관절이나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 이현우는 고개를 숙여 쿠마의 어깨 쪽을 향하고, 오른쪽 쿠마는 눈을 감은 채 이현우의 어깨에 얼굴을 댄다. 양쪽 팔은 상대 몸통을 감싸며 카메라를 향한 시선은 없다. 얼굴 방향은 지시와 대체로 맞지만 이미 안긴 채 머무는 순간처럼 보인다.",
        "built_space": "왼쪽에 병상 한 개와 모니터 한 대, 뒤에 수납장 한 개와 야간 창문 한 개, 오른쪽에 출입문 한 개, 의료용 카트와 바퀴 달린 의자 한 개가 보인다. 두 사람은 병상 옆 열린 공간에 서 있고 설비 중복이나 불가능한 반사는 없다. 이전 장면의 주요 재료와 배치는 유지되지만, 보이는 베개와 넓은 침상 면에는 환자가 없어 점유 상태 조건과 어긋난다. 역시 허리 위 전체를 담기보다 등과 어깨 중심으로 타이트하며 약한 앙각이 명확하지 않다.",
        "entities": "참조와 유사한 젊은 동아시아계 외형의 남성 두 명만 등장한다. 이현우의 검은 헝클어진 머리와 피·먼지가 묻은 셔츠는 이어지지만, 노출된 귀에서 지정된 인이어는 확인되지 않는다. 이현우의 얼굴은 대부분 가려져 있다. 쿠마의 비니, 검은 가죽 재킷과 보이는 얼굴은 참조에 대체로 맞는다. 하의와 찰리의 소지품은 화면 밖이다.",
        "hard_violations": [],
        "physics": "쿠마의 두 손이 이현우의 위쪽 등과 허리 부근을 붙잡고, 이현우의 팔은 쿠마의 어깨 뒤로 이어진다. 팔과 몸통의 접촉은 물리적으로 가능하며 옷도 눌리고 접혀 있다. 하체가 잘려 발의 지지는 직접 확인되지 않지만 부유를 나타내는 징후는 없다. 다만 몸통의 기울기와 압박이 비교적 안정적이어서 와락 끌어안은 직후의 운동감은 약하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.036,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조: 이현우의 등에 얹어진 두 손이 모두 엄지가 위로 향하고 손가락이 왼쪽을 가리키는 오른손 형태를 띠고 있음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1036
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "배경, 인물, 조명이 프롬프트 및 레퍼런스와 훌륭하게 일치하나, 이현우의 왼팔이 상대를 안지 않고 자신의 등을 감싸고 있는 포즈 오류가 있습니다."
   },
   {
    "label": "A",
    "score": 1036,
    "verdict_ko": "이현우의 등에 얹어진 두 손이 모두 오른손 형태로 생성되었고 소매 재질도 잘못 렌더링되어 해부학적 하드 위반이 발생했습니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체 구조: 이현우의 등에 얹어진 두 손이 모두 엄지가 위로 향하고 손가락이 왼쪽을 가리키는 오른손 형태를 띠고 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S73sh6_sel.png",
    "asset_id": "1d501007-8693-4a31-8f31-b4f10295c991",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:919321>",
    "asset_id": "9fea3d15-6e4a-4e12-961b-3d1014ab4442",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d94-5489-7ba5-a153-e61202d45ef1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh6"
  }
 },
 "S74sh5::signage": {
  "fp": "0c35250df8797441",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b053396501d147ab": {
  "subjects": [],
  "subject_text": "목포 항구마을 식당 내부\n항구 길거리에 열린 작은 간이식당이다.",
  "identity": "canonical",
  "scope_id": "L127",
  "scope_role": "location_interior",
  "scope_sha": "be64b395fbc87c6e"
 },
 "groupbg::harbor_street_diner": {
  "input_fingerprint": "ae0a9b9d3b5aed7f",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "harbor_street_diner",
    "tags": [
     "S74sh5",
     "S74sh8"
    ]
   },
   "context_sig": "883630c3ec4da2a0"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a streetside dining table in the harbor village at night, beside the fish-trading activity.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구마을 식당 내부: 항구 길거리에 열린 작은 간이식당이다. (특징: 접시에 담긴 기형적 돌연변이 생선회 형태)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 보건소를 나오는 현우와 쿠마. 이곳은 목포 항구마을 밤. 길거리 식당..\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a streetside dining table in the harbor village at night, beside the fish-trading activity.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구마을 식당 내부: 항구 길거리에 열린 작은 간이식당이다. (특징: 접시에 담긴 기형적 돌연변이 생선회 형태)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 보건소를 나오는 현우와 쿠마. 이곳은 목포 항구마을 밤. 길거리 식당..\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_street_diner_053abf.png",
  "asset_id": "c968c842-1515-4e34-93d4-9e589b820604",
  "input_asset_ids": [
   "1fe93e53-8a1a-4570-a744-63aa97ed81c0"
  ],
  "origin_tag": "S74sh5",
  "place_text": "At a streetside dining table in the harbor village at night, beside the fish-trading activity.",
  "origin_inputs": {
   "place_text": "At a streetside dining table in the harbor village at night, beside the fish-trading activity.",
   "time_of_day_en": "night",
   "conti_asset_id": "1fe93e53-8a1a-4570-a744-63aa97ed81c0"
  }
 },
 "S74sh5::bgfirst_bg": {
  "input_fingerprint": "0ff95abff8825d21",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식당 테이블 위로 낯선 남자의 거친 손이 흉측한 돌연변이 생선회가 담긴 접시를 내려놓아 테이블에 막 닿은 찰나.\n\nLOCATION (lock): At a streetside dining table in the harbor village at night, beside the fish-trading activity.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the downward crane beside the near table edge on 쿠마's side, looking obliquely down at the exact instant the plate first touches the tabletop. The plate of visibly abnormal mutant-fish sashimi occupies roughly one third of the lower-center frame while 낯선 남자's rough hand enters from the upper right, still supporting its edge; his face and the diners remain outside the crop. Keep a broad tabletop margin as scale context, emphasizing the lowered camera distance before the same route rises toward 이현우.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 돌연변이 생선회 접시 (Its base has just touched the table, with the seller's hand still on the edge) — The food-bearing upper side is visible at an oblique downward angle; used as Primary object detail, kept small enough to retain hand and table context; 식당 테이블 (Receiving the plate during the conversation) — The tabletop extends around the plate, with its near edge at the bottom of the image; used as Landing plane and natural scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime restaurant illumination renders the abnormal fish plainly without inventing a colored glow or a visible light fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식당 테이블 위로 낯선 남자의 거친 손이 흉측한 돌연변이 생선회가 담긴 접시를 내려놓아 테이블에 막 닿은 찰나.\n\nLOCATION (lock): At a streetside dining table in the harbor village at night, beside the fish-trading activity.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the downward crane beside the near table edge on 쿠마's side, looking obliquely down at the exact instant the plate first touches the tabletop. The plate of visibly abnormal mutant-fish sashimi occupies roughly one third of the lower-center frame while 낯선 남자's rough hand enters from the upper right, still supporting its edge; his face and the diners remain outside the crop. Keep a broad tabletop margin as scale context, emphasizing the lowered camera distance before the same route rises toward 이현우.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 돌연변이 생선회 접시 (Its base has just touched the table, with the seller's hand still on the edge) — The food-bearing upper side is visible at an oblique downward angle; used as Primary object detail, kept small enough to retain hand and table context; 식당 테이블 (Receiving the plate during the conversation) — The tabletop extends around the plate, with its near edge at the bottom of the image; used as Landing plane and natural scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime restaurant illumination renders the abnormal fish plainly without inventing a colored glow or a visible light fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S74sh5__bgfirst_bg.png",
  "asset_id": "4955b534-6922-4fbf-93e0-ac09645c9a44",
  "input_asset_ids": [
   "1fe93e53-8a1a-4570-a744-63aa97ed81c0",
   "c968c842-1515-4e34-93d4-9e589b820604"
  ]
 },
 "S74sh5": {
  "input_fingerprint": "c69890caa792c5b5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 테이블 위로 낯선 남자의 거친 손이 흉측한 돌연변이 생선회가 담긴 접시를 내려놓아 테이블에 막 닿은 찰나.\n\nLOCATION (lock): At a streetside dining table in the harbor village at night, beside the fish-trading activity. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the downward crane beside the near table edge on 쿠마's side, looking obliquely down at the exact instant the plate first touches the tabletop. The plate of visibly abnormal mutant-fish sashimi occupies roughly one third of the lower-center frame while 낯선 남자's rough hand enters from the upper right, still supporting its edge; his face and the diners remain outside the crop. Keep a broad tabletop margin as scale context, emphasizing the lowered camera distance before the same route rises toward 이현우.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 돌연변이 생선회 접시 (Its base has just touched the table, with the seller's hand still on the edge) — The food-bearing upper side is visible at an oblique downward angle; used as Primary object detail, kept small enough to retain hand and table context; 식당 테이블 (Receiving the plate during the conversation) — The tabletop extends around the plate, with its near edge at the bottom of the image; used as Landing plane and natural scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime restaurant illumination renders the abnormal fish plainly without inventing a colored glow or a visible light fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the nighttime harbor-side eatery, a plate of raw fish is being placed on the table; the fish is recognizably mutated by radioactive contamination.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낯선 남자 right now, so 낯선 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낯선 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 테이블 위로 낯선 남자의 거친 손이 흉측한 돌연변이 생선회가 담긴 접시를 내려놓아 테이블에 막 닿은 찰나.\n\nLOCATION (lock): At a streetside dining table in the harbor village at night, beside the fish-trading activity. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the downward crane beside the near table edge on 쿠마's side, looking obliquely down at the exact instant the plate first touches the tabletop. The plate of visibly abnormal mutant-fish sashimi occupies roughly one third of the lower-center frame while 낯선 남자's rough hand enters from the upper right, still supporting its edge; his face and the diners remain outside the crop. Keep a broad tabletop margin as scale context, emphasizing the lowered camera distance before the same route rises toward 이현우.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 돌연변이 생선회 접시 (Its base has just touched the table, with the seller's hand still on the edge) — The food-bearing upper side is visible at an oblique downward angle; used as Primary object detail, kept small enough to retain hand and table context; 식당 테이블 (Receiving the plate during the conversation) — The tabletop extends around the plate, with its near edge at the bottom of the image; used as Landing plane and natural scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime restaurant illumination renders the abnormal fish plainly without inventing a colored glow or a visible light fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the nighttime harbor-side eatery, a plate of raw fish is being placed on the table; the fish is recognizably mutated by radioactive contamination.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낯선 남자 right now, so 낯선 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낯선 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 식당 테이블 위로 낯선 남자의 거친 손이 흉측한 돌연변이 생선회가 담긴 접시를 내려놓아 테이블에 막 닿은 찰나.\n\nLOCATION (lock): At a streetside dining table in the harbor village at night, beside the fish-trading activity. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the downward crane beside the near table edge on 쿠마's side, looking obliquely down at the exact instant the plate first touches the tabletop. The plate of visibly abnormal mutant-fish sashimi occupies roughly one third of the lower-center frame while 낯선 남자's rough hand enters from the upper right, still supporting its edge; his face and the diners remain outside the crop. Keep a broad tabletop margin as scale context, emphasizing the lowered camera distance before the same route rises toward 이현우.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 돌연변이 생선회 접시 (Its base has just touched the table, with the seller's hand still on the edge) — The food-bearing upper side is visible at an oblique downward angle; used as Primary object detail, kept small enough to retain hand and table context; 식당 테이블 (Receiving the plate during the conversation) — The tabletop extends around the plate, with its near edge at the bottom of the image; used as Landing plane and natural scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime restaurant illumination renders the abnormal fish plainly without inventing a colored glow or a visible light fixture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): At the nighttime harbor-side eatery, a plate of raw fish is being placed on the table; the fish is recognizably mutated by radioactive contamination.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낯선 남자 right now, so 낯선 남자's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낯선 남자: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S74sh5__bgfirst_bg.png",
     "asset_id": "4955b534-6922-4fbf-93e0-ac09645c9a44",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S74sh5.png",
     "asset_id": "1fe93e53-8a1a-4570-a744-63aa97ed81c0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_street_diner_053abf.png",
     "asset_id": "c968c842-1515-4e34-93d4-9e589b820604",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 테이블을 가로질러 항구 쪽을 향하며 약간 아래를 봄. 거친 손이 오른쪽 위에서 접시를 향함.",
    "built_space": "항구의 야외 식당. 빨간 식탁보가 깔린 테이블과 반찬들이 있으며, 배경으로 항구와 어선이 넓게 보임(레퍼런스와 동일한 넓은 구도).",
    "entities": "눈이 여러 개 달린 기괴한 돌연변이 물고기 회와 접시, 낯선 남자의 거친 손, 빨간 식탁보, 간장 및 반찬 종지.",
    "hard_violations": [],
    "physics": "접시가 테이블 위에 놓여 있고, 남자의 손이 접시 가장자리를 단단히 잡고 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 테이블 위를 비스듬히 굽어봄. 손은 오른쪽 가장자리에서 접시를 잡고 있음.",
    "built_space": "빨간 비닐이 깔린 식당 테이블 위. 넓은 테이블 여백이 확보되어 있으며 배경은 아웃포커싱됨.",
    "entities": "작은 손이나 촉수 같은 부위가 섞인 기괴한 돌연변이 회, 낯선 남자의 거친 손, 빨간 식탁보, 반찬 접시들.",
    "hard_violations": [],
    "physics": "접시가 테이블 바닥에 닿아 있고, 손이 그 가장자리를 자연스럽게 지탱함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "요구된 클로즈업 샷 크기와 비스듬히 내려다보는 카메라 각도를 완벽히 구현했으며, 지시대로 레퍼런스의 구도를 복사하지 않고 테이블 여백을 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트가 요구한 '클로즈업' 샷 크기를 무시하고, 복사하지 말라고 명시된 로케이션 레퍼런스의 넓은 배경 구도를 그대로 차용하여 프레이밍 지시를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블을 가로질러 항구 쪽을 향하며 약간 아래를 봄. 거친 손이 오른쪽 위에서 접시를 향함.",
        "built_space": "항구의 야외 식당. 빨간 식탁보가 깔린 테이블과 반찬들이 있으며, 배경으로 항구와 어선이 넓게 보임(레퍼런스와 동일한 넓은 구도).",
        "entities": "눈이 여러 개 달린 기괴한 돌연변이 물고기 회와 접시, 낯선 남자의 거친 손, 빨간 식탁보, 간장 및 반찬 종지.",
        "hard_violations": [],
        "physics": "접시가 테이블 위에 놓여 있고, 남자의 손이 접시 가장자리를 단단히 잡고 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위를 비스듬히 굽어봄. 손은 오른쪽 가장자리에서 접시를 잡고 있음.",
        "built_space": "빨간 비닐이 깔린 식당 테이블 위. 넓은 테이블 여백이 확보되어 있으며 배경은 아웃포커싱됨.",
        "entities": "작은 손이나 촉수 같은 부위가 섞인 기괴한 돌연변이 회, 낯선 남자의 거친 손, 빨간 식탁보, 반찬 접시들.",
        "hard_violations": [],
        "physics": "접시가 테이블 바닥에 닿아 있고, 손이 그 가장자리를 자연스럽게 지탱함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "요구된 클로즈업 샷 크기와 비스듬히 내려다보는 카메라 각도를 완벽히 구현했으며, 지시대로 레퍼런스의 구도를 복사하지 않고 테이블 여백을 잘 살렸습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트가 요구한 '클로즈업' 샷 크기를 무시하고, 복사하지 말라고 명시된 로케이션 레퍼런스의 넓은 배경 구도를 그대로 차용하여 프레이밍 지시를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블을 가로질러 항구 쪽을 향하며 약간 아래를 봄. 거친 손이 오른쪽 위에서 접시를 향함.",
        "built_space": "항구의 야외 식당. 빨간 식탁보가 깔린 테이블과 반찬들이 있으며, 배경으로 항구와 어선이 넓게 보임(레퍼런스와 동일한 넓은 구도).",
        "entities": "눈이 여러 개 달린 기괴한 돌연변이 물고기 회와 접시, 낯선 남자의 거친 손, 빨간 식탁보, 간장 및 반찬 종지.",
        "hard_violations": [],
        "physics": "접시가 테이블 위에 놓여 있고, 남자의 손이 접시 가장자리를 단단히 잡고 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위를 비스듬히 굽어봄. 손은 오른쪽 가장자리에서 접시를 잡고 있음.",
        "built_space": "빨간 비닐이 깔린 식당 테이블 위. 넓은 테이블 여백이 확보되어 있으며 배경은 아웃포커싱됨.",
        "entities": "작은 손이나 촉수 같은 부위가 섞인 기괴한 돌연변이 회, 낯선 남자의 거친 손, 빨간 식탁보, 반찬 접시들.",
        "hard_violations": [],
        "physics": "접시가 테이블 바닥에 닿아 있고, 손이 그 가장자리를 자연스럽게 지탱함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "비스듬한 하향 클로즈업에서 거친 손이 접시 가장자리를 받친 채 상판에 내려놓는 순간을 구현하며, 넓은 상판 여백과 실물 같은 기형 회의 질감도 요구에 가깝다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "손의 진입 방향과 접시 접촉은 맞지만, 항구 전경과 조명까지 보여주는 넓고 낮은 구도가 지정된 하향 클로즈업보다 장소 참고사진의 구도에 치우친다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "손과 팔이 오른쪽 위에서 왼쪽 아래의 접시 오른쪽 테두리로 들어온다. 엄지는 테두리 위에, 나머지 손가락은 아래에 놓여 접시를 내려놓는 방향과 맞는다. 음식이 담긴 윗면이 카메라를 향해 비스듬히 보이며, 얼굴이나 시선은 프레임 밖이다.",
        "built_space": "붉은 비닐 상판 하나가 화면 대부분을 차지하고 가까운 가장자리가 왼쪽 아래에서 하단으로 이어진다. 왼쪽에는 겹친 작은 그릇 한 묶음, 간장 그릇 묶음, 반찬 그릇 하나, 금속 컵 하나와 용기들이 있고 오른쪽 아래에는 빈 그릇 하나가 잘려 보인다. 참고 장소의 상판 재질과 식기 배치에 부합하며, 건축물은 클로즈업 밖이라 확인할 수 없다. 접시 주변 상판 여백이 충분하다.",
        "entities": "중앙 아래에 도자기 접시 하나와 젖은 생선회가 보인다. 회에는 검붉은 병변, 울퉁불퉁한 조직과 비정상적인 돌기가 있어 흉측한 돌연변이 생선이라는 요구를 전달한다. 오른쪽에는 거칠고 주름진 성인 남성으로 읽히는 손 하나와 낡은 어두운 소매가 있다. 손만으로 정확한 나이나 민족은 판단할 수 없다. 낯선 남자의 얼굴과 이현우를 포함한 식사객은 보이지 않아 지정된 크롭에 맞는다.",
        "hard_violations": [],
        "physics": "접시는 상판 바로 위에 안착한 상태로 보이며, 오른쪽 가장자리는 엄지와 아래쪽 손가락이 계속 받치고 있다. 접시 아래의 접촉 그림자가 착지면을 뒷받침한다. 회와 고명은 접시 위에 놓여 있고, 손목과 팔은 오른쪽 화면 밖으로 자연스럽게 이어진다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "오른쪽 위에서 들어온 손이 접시 오른쪽 테두리를 잡고 있으며 손의 목표는 명확히 접시다. 음식 윗면도 보이지만 카메라의 하향 각도가 얕아 뒤쪽 항구를 크게 드러낸다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "전경에 붉은 비닐 식탁 하나, 왼쪽 뒤에 별도의 식탁 하나가 보인다. 전경에는 수저통 하나, 흰 용기 하나, 금속 컵 하나, 작은 그릇 묶음 하나, 간장 그릇 하나, 반찬 그릇 하나와 오른쪽 빈 그릇 하나가 있다. 배경에는 수조 하나, 적재 상자, 가게 기둥과 비닐막, 부두와 어선이 보여 참고 장소와 잘 연결된다. 다만 항구와 가게가 화면 상부를 넓게 차지하고 상단의 밝은 조명도 노출되어, 조명기구 없이 상판 중심으로 잡으라는 클로즈업 요구에서 벗어난다.",
        "entities": "도자기 접시에 회와 생선 머리, 큰 지느러미가 놓여 있다. 머리의 여러 눈과 회 조각마다 반복되는 눈 모양 조직이 돌연변이를 명확하게 표현하지만, 반복 문양은 A의 불규칙한 병변보다 인공적으로 읽힌다. 거친 성인 남성의 손과 얼룩무늬 어두운 소매가 보이고 얼굴이나 다른 사람은 없다. 인물 참고의 얼굴은 요구대로 크롭 밖이다.",
        "hard_violations": [],
        "physics": "접시의 앞쪽 바닥은 상판에 닿고 오른쪽 가장자리는 손이 붙잡고 있어 약간 기운 착지 순간으로 성립한다. 엄지가 테두리 위를 누르고 손가락이 바깥과 아래쪽을 받친다. 생선 머리와 회는 접시 위에 지지되어 있고 지느러미는 생선 조직에 연결되어 있다. 지지 없는 부유나 불가능한 손의 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "비스듬한 하향 클로즈업에서 거친 손이 접시 가장자리를 받친 채 상판에 내려놓는 순간을 구현하며, 넓은 상판 여백과 실물 같은 기형 회의 질감도 요구에 가깝다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "손의 진입 방향과 접시 접촉은 맞지만, 항구 전경과 조명까지 보여주는 넓고 낮은 구도가 지정된 하향 클로즈업보다 장소 참고사진의 구도에 치우친다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "손과 팔이 오른쪽 위에서 왼쪽 아래의 접시 오른쪽 테두리로 들어온다. 엄지는 테두리 위에, 나머지 손가락은 아래에 놓여 접시를 내려놓는 방향과 맞는다. 음식이 담긴 윗면이 카메라를 향해 비스듬히 보이며, 얼굴이나 시선은 프레임 밖이다.",
        "built_space": "붉은 비닐 상판 하나가 화면 대부분을 차지하고 가까운 가장자리가 왼쪽 아래에서 하단으로 이어진다. 왼쪽에는 겹친 작은 그릇 한 묶음, 간장 그릇 묶음, 반찬 그릇 하나, 금속 컵 하나와 용기들이 있고 오른쪽 아래에는 빈 그릇 하나가 잘려 보인다. 참고 장소의 상판 재질과 식기 배치에 부합하며, 건축물은 클로즈업 밖이라 확인할 수 없다. 접시 주변 상판 여백이 충분하다.",
        "entities": "중앙 아래에 도자기 접시 하나와 젖은 생선회가 보인다. 회에는 검붉은 병변, 울퉁불퉁한 조직과 비정상적인 돌기가 있어 흉측한 돌연변이 생선이라는 요구를 전달한다. 오른쪽에는 거칠고 주름진 성인 남성으로 읽히는 손 하나와 낡은 어두운 소매가 있다. 손만으로 정확한 나이나 민족은 판단할 수 없다. 낯선 남자의 얼굴과 이현우를 포함한 식사객은 보이지 않아 지정된 크롭에 맞는다.",
        "hard_violations": [],
        "physics": "접시는 상판 바로 위에 안착한 상태로 보이며, 오른쪽 가장자리는 엄지와 아래쪽 손가락이 계속 받치고 있다. 접시 아래의 접촉 그림자가 착지면을 뒷받침한다. 회와 고명은 접시 위에 놓여 있고, 손목과 팔은 오른쪽 화면 밖으로 자연스럽게 이어진다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "오른쪽 위에서 들어온 손이 접시 오른쪽 테두리를 잡고 있으며 손의 목표는 명확히 접시다. 음식 윗면도 보이지만 카메라의 하향 각도가 얕아 뒤쪽 항구를 크게 드러낸다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "전경에 붉은 비닐 식탁 하나, 왼쪽 뒤에 별도의 식탁 하나가 보인다. 전경에는 수저통 하나, 흰 용기 하나, 금속 컵 하나, 작은 그릇 묶음 하나, 간장 그릇 하나, 반찬 그릇 하나와 오른쪽 빈 그릇 하나가 있다. 배경에는 수조 하나, 적재 상자, 가게 기둥과 비닐막, 부두와 어선이 보여 참고 장소와 잘 연결된다. 다만 항구와 가게가 화면 상부를 넓게 차지하고 상단의 밝은 조명도 노출되어, 조명기구 없이 상판 중심으로 잡으라는 클로즈업 요구에서 벗어난다.",
        "entities": "도자기 접시에 회와 생선 머리, 큰 지느러미가 놓여 있다. 머리의 여러 눈과 회 조각마다 반복되는 눈 모양 조직이 돌연변이를 명확하게 표현하지만, 반복 문양은 A의 불규칙한 병변보다 인공적으로 읽힌다. 거친 성인 남성의 손과 얼룩무늬 어두운 소매가 보이고 얼굴이나 다른 사람은 없다. 인물 참고의 얼굴은 요구대로 크롭 밖이다.",
        "hard_violations": [],
        "physics": "접시의 앞쪽 바닥은 상판에 닿고 오른쪽 가장자리는 손이 붙잡고 있어 약간 기운 착지 순간으로 성립한다. 엄지가 테두리 위를 누르고 손가락이 바깥과 아래쪽을 받친다. 생선 머리와 회는 접시 위에 지지되어 있고 지느러미는 생선 조직에 연결되어 있다. 지지 없는 부유나 불가능한 손의 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.167,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.167,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1167
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 클로즈업 샷 크기와 비스듬히 내려다보는 카메라 각도를 완벽히 구현했으며, 지시대로 레퍼런스의 구도를 복사하지 않고 테이블 여백을 잘 살렸습니다."
   },
   {
    "label": "A",
    "score": 1167,
    "verdict_ko": "프롬프트가 요구한 '클로즈업' 샷 크기를 무시하고, 복사하지 말라고 명시된 로케이션 레퍼런스의 넓은 배경 구도를 그대로 차용하여 프레이밍 지시를 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_street_diner_053abf.png",
    "asset_id": "c968c842-1515-4e34-93d4-9e589b820604",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0d9f-208f-71d3-8aeb-b1f0965e8ab8",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S74sh5__bgfirst_bg.png",
   "bg_asset_id": "4955b534-6922-4fbf-93e0-ac09645c9a44",
   "bg_record_key": "S74sh5::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "harbor_street_diner",
   "groupbg_asset_id": "c968c842-1515-4e34-93d4-9e589b820604"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S74sh8::signage": {
  "fp": "de2cdc8b4ece049a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S74sh8": {
  "input_fingerprint": "f1a69971febaf2ca",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이곳을 떠날 계획임을 밝히는 쿠마를 향해 놀란 눈으로 쳐다보는 이현우의 얼굴.\n\nLOCATION (lock): At the same outdoor streetside dining table in the subdued harbor village at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the upward crane immediately behind and outside 쿠마's shoulder at 이현우's seated eye height, remaining offset from their conversation axis. 이현우's three-quarter face occupies the center-right while only a narrow edge of 쿠마's shoulder remains at left; 이현우 has lifted his widened eyes from the food to 쿠마's face just outside the upper-left crop. Hold the endpoint without a further push or lighting change, making the redirected gaze the final accent.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued nighttime restaurant illumination, preserving 이현우's widened eyes and restrained facial contrast without a reaction spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The plate of contaminated mutant raw fish remains on the table at the nighttime harbor-side eatery. 이현우: He remains at the table with his dirty appearance and earlier wounds unchanged. 쿠마: He remains at the harbor-side table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이곳을 떠날 계획임을 밝히는 쿠마를 향해 놀란 눈으로 쳐다보는 이현우의 얼굴.\n\nLOCATION (lock): At the same outdoor streetside dining table in the subdued harbor village at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the upward crane immediately behind and outside 쿠마's shoulder at 이현우's seated eye height, remaining offset from their conversation axis. 이현우's three-quarter face occupies the center-right while only a narrow edge of 쿠마's shoulder remains at left; 이현우 has lifted his widened eyes from the food to 쿠마's face just outside the upper-left crop. Hold the endpoint without a further push or lighting change, making the redirected gaze the final accent.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued nighttime restaurant illumination, preserving 이현우's widened eyes and restrained facial contrast without a reaction spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The plate of contaminated mutant raw fish remains on the table at the nighttime harbor-side eatery. 이현우: He remains at the table with his dirty appearance and earlier wounds unchanged. 쿠마: He remains at the harbor-side table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이곳을 떠날 계획임을 밝히는 쿠마를 향해 놀란 눈으로 쳐다보는 이현우의 얼굴.\n\nLOCATION (lock): At the same outdoor streetside dining table in the subdued harbor village at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the upward crane immediately behind and outside 쿠마's shoulder at 이현우's seated eye height, remaining offset from their conversation axis. 이현우's three-quarter face occupies the center-right while only a narrow edge of 쿠마's shoulder remains at left; 이현우 has lifted his widened eyes from the food to 쿠마's face just outside the upper-left crop. Hold the endpoint without a further push or lighting change, making the redirected gaze the final accent.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued nighttime restaurant illumination, preserving 이현우's widened eyes and restrained facial contrast without a reaction spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The plate of contaminated mutant raw fish remains on the table at the nighttime harbor-side eatery. 이현우: He remains at the table with his dirty appearance and earlier wounds unchanged. 쿠마: He remains at the harbor-side table.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함.; 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖, 쿠마의 얼굴이 있을 위치를 정확히 향하고 있습니다.",
    "built_space": "항구 배경의 야외 테이블로, 빨간색 비닐 식탁보 위에 돌연변이 생선회 접시, 스테인리스 컵, 간장병, 밑반찬 종지 등이 적절히 배치되어 있습니다.",
    "entities": "이현우는 핏자국이 있는 낡은 셔츠와 소형 인이어 무전기를 착용한 채 화면 우측을 차지하고 있으며, 좌측 가장자리에는 레퍼런스와 일치하는 네이비색 티셔츠를 입은 쿠마의 어깨가 보입니다.",
    "hard_violations": [],
    "physics": "이현우는 의자에 앉아 안정적인 자세를 취하고 있으며, 테이블 위 모든 사물은 표면에 정상적으로 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖의 쿠마를 향하고 있습니다.",
    "built_space": "야간 항구를 배경으로 빨간 식탁보가 깔린 테이블 위에 생선회와 식기류가 놓여 있습니다.",
    "entities": "이현우의 외양과 테이블 위 음식은 지시사항을 잘 따랐으나, 좌측에 보이는 쿠마의 어깨가 네이비 티셔츠가 아닌 이현우의 것과 유사한 깃이 달린 셔츠로 묘사되었습니다.",
    "hard_violations": [],
    "physics": "이현우의 앉은 자세와 테이블 위 사물들의 접촉면이 물리적으로 자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 지시한 카메라 앵글과 시선 방향을 훌륭하게 구현했으며, 레퍼런스 컷의 테이블 세팅과 쿠마의 의상(네이비 티셔츠)까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 인물 묘사는 우수하나, 화면 좌측에 걸친 쿠마의 어깨 의상이 레퍼런스와 달리 깃이 있는 셔츠로 잘못 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖, 쿠마의 얼굴이 있을 위치를 정확히 향하고 있습니다.",
        "built_space": "항구 배경의 야외 테이블로, 빨간색 비닐 식탁보 위에 돌연변이 생선회 접시, 스테인리스 컵, 간장병, 밑반찬 종지 등이 적절히 배치되어 있습니다.",
        "entities": "이현우는 핏자국이 있는 낡은 셔츠와 소형 인이어 무전기를 착용한 채 화면 우측을 차지하고 있으며, 좌측 가장자리에는 레퍼런스와 일치하는 네이비색 티셔츠를 입은 쿠마의 어깨가 보입니다.",
        "hard_violations": [],
        "physics": "이현우는 의자에 앉아 안정적인 자세를 취하고 있으며, 테이블 위 모든 사물은 표면에 정상적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖의 쿠마를 향하고 있습니다.",
        "built_space": "야간 항구를 배경으로 빨간 식탁보가 깔린 테이블 위에 생선회와 식기류가 놓여 있습니다.",
        "entities": "이현우의 외양과 테이블 위 음식은 지시사항을 잘 따랐으나, 좌측에 보이는 쿠마의 어깨가 네이비 티셔츠가 아닌 이현우의 것과 유사한 깃이 달린 셔츠로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "이현우의 앉은 자세와 테이블 위 사물들의 접촉면이 물리적으로 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 지시한 카메라 앵글과 시선 방향을 훌륭하게 구현했으며, 레퍼런스 컷의 테이블 세팅과 쿠마의 의상(네이비 티셔츠)까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 구도와 인물 묘사는 우수하나, 화면 좌측에 걸친 쿠마의 어깨 의상이 레퍼런스와 달리 깃이 있는 셔츠로 잘못 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖, 쿠마의 얼굴이 있을 위치를 정확히 향하고 있습니다.",
        "built_space": "항구 배경의 야외 테이블로, 빨간색 비닐 식탁보 위에 돌연변이 생선회 접시, 스테인리스 컵, 간장병, 밑반찬 종지 등이 적절히 배치되어 있습니다.",
        "entities": "이현우는 핏자국이 있는 낡은 셔츠와 소형 인이어 무전기를 착용한 채 화면 우측을 차지하고 있으며, 좌측 가장자리에는 레퍼런스와 일치하는 네이비색 티셔츠를 입은 쿠마의 어깨가 보입니다.",
        "hard_violations": [],
        "physics": "이현우는 의자에 앉아 안정적인 자세를 취하고 있으며, 테이블 위 모든 사물은 표면에 정상적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 좌측 상단 프레임 밖의 쿠마를 향하고 있습니다.",
        "built_space": "야간 항구를 배경으로 빨간 식탁보가 깔린 테이블 위에 생선회와 식기류가 놓여 있습니다.",
        "entities": "이현우의 외양과 테이블 위 음식은 지시사항을 잘 따랐으나, 좌측에 보이는 쿠마의 어깨가 네이비 티셔츠가 아닌 이현우의 것과 유사한 깃이 달린 셔츠로 묘사되었습니다.",
        "hard_violations": [],
        "physics": "이현우의 앉은 자세와 테이블 위 사물들의 접촉면이 물리적으로 자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "쿠마를 올려다보는 시선과 놀란 표정은 맞지만, 내려다보는 카메라와 넓은 상반신 구도, 크게 들어온 쿠마의 머리·어깨가 지정된 눈높이 얼굴 클로즈업에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "눈높이에 가까운 대화 시점, 쿠마를 향한 커진 눈과 남색 의상이 더 충실하지만, 얼굴 클로즈업 대신 식탁과 상반신을 넓게 담고 쿠마도 과도하게 노출했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴은 왼쪽으로 돌아간 삼사분면이며, 두 눈은 화면 왼쪽 위 전경의 쿠마 얼굴 쪽을 향합니다. 음식을 보거나 렌즈를 응시하지 않아 시선 전환의 도착점은 맞습니다. 쿠마는 이현우 쪽으로 머리를 돌리고 있지만 눈은 보이지 않습니다.",
        "built_space": "붉은 비닐 식탁 하나가 왼쪽에서 오른쪽 아래로 대각선으로 놓이고, 이현우는 그 건너편 오른쪽에 있습니다. 식탁에는 회 접시 하나, 흰 그릇 더미 하나, 간장 종지 더미 하나, 반찬 그릇 하나, 금속 통 하나와 일부 가려진 보온 용기·양념 용기가 보입니다. 뒤에는 별도 붉은 탁자 하나와 의자, 수직 기둥 및 항구 시설이 보입니다. 전경 식탁의 재질은 참고와 이어지지만, 참고에 없는 넓은 배경의 정확한 구조는 대조할 수 없습니다. 식탁 윗면과 이현우를 다소 내려다보는 구도이며, 왼쪽에는 좁은 어깨 끝이 아니라 쿠마의 머리와 큰 어깨가 들어옵니다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 동아시아계 청년 남성으로, 참고와 유사한 얼굴과 마른 체격입니다. 더러운 어두운 셔츠, 핏자국, 뺨의 상처와 귀의 검은 인이어 장치가 보입니다. 국적은 외형만으로 확인할 수 없습니다. 쿠마는 뒷머리·귀·목·어깨만 보여 얼굴 정체성이나 혼혈 여부는 확인하기 어렵고, 상의는 참고의 남색보다 회흑색으로 읽힙니다. 변이 생선회 접시가 식탁 아래쪽에 일부 보이며 다른 사람이나 삽입 문자는 없습니다.",
        "hard_violations": [],
        "physics": "이현우의 머리와 목은 앞으로 기울인 몸통에 자연스럽게 연결되어 있습니다. 좌석과 골반은 프레임 밖이라 실제 착석 접촉은 확인할 수 없지만 공중에 뜬 신체는 보이지 않습니다. 쿠마의 어깨도 프레임 밖 몸통으로 이어집니다. 접시와 용기는 식탁 위에 놓여 있고 그릇들은 서로 포개져 지지됩니다. 비정상적인 관절이나 지지 없이 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "이현우는 삼사분면 얼굴로 왼쪽 위의 쿠마를 바라봅니다. 홍채가 향한 방향과 고개 방향이 일치하고, 넓어진 눈과 살짝 벌린 입이 놀란 반응을 드러냅니다. 쿠마의 얼굴 윗부분은 잘려 있지만 이현우를 향한 머리 방향은 읽힙니다. 시선의 대상은 음식이나 카메라가 아니라 쿠마입니다.",
        "built_space": "붉은 비닐 식탁 하나가 화면 아래를 가로지르고 이현우가 건너편 중앙 오른쪽, 쿠마가 왼쪽 전경에 있습니다. 회 접시 하나, 흰 그릇 더미 하나, 간장 종지 더미 하나, 반찬 그릇 하나, 금속 통 하나, 보온 용기 하나와 양념 용기가 보입니다. 뒤에는 붉은 탁자·의자와 수직 기둥들, 항구 수면과 배가 작게 배치됩니다. 수면의 불빛 반사는 가능한 방향입니다. 참고의 식탁 재질과 식기 구성이 이어지며 카메라도 앉은 눈높이에 비교적 가깝습니다. 다만 가슴 아래까지 담고 쿠마의 머리와 어깨가 왼쪽을 크게 차지하여 지정된 얼굴 클로즈업과 좁은 어깨 가장자리 조건은 충족하지 못합니다.",
        "entities": "이현우는 참고와 유사한 젊은 동아시아계 남성 얼굴, 헝클어진 짧은 검은 머리, 마른 상체를 갖습니다. 어두운 낡은 셔츠의 먼지와 혈흔, 얼굴 상처, 검은 인이어 장치가 보입니다. 쿠마는 짧은 검은 머리와 남색 상의가 보여 참고의 가시적 특징에 부합하지만, 얼굴 대부분이 가려져 정확한 정체성과 나이는 검증하기 어렵습니다. 오염된 변이 생선회로 읽히는 접시가 식탁 아래쪽에 남아 있습니다. 추가 인물이나 자막은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "이현우는 식탁 너머에서 몸을 약간 앞으로 기울인 자세이며 목과 어깨의 연결이 자연스럽습니다. 의자와 골반은 가려져 접촉점은 확인할 수 없으나 부유하거나 불가능하게 꺾인 몸은 없습니다. 전경 쿠마의 머리와 어깨도 연속된 신체로 보입니다. 회 접시와 용기들은 식탁이 받치고, 포개진 그릇은 아래 그릇이 지지합니다. 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "쿠마를 올려다보는 시선과 놀란 표정은 맞지만, 내려다보는 카메라와 넓은 상반신 구도, 크게 들어온 쿠마의 머리·어깨가 지정된 눈높이 얼굴 클로즈업에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "눈높이에 가까운 대화 시점, 쿠마를 향한 커진 눈과 남색 의상이 더 충실하지만, 얼굴 클로즈업 대신 식탁과 상반신을 넓게 담고 쿠마도 과도하게 노출했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴은 왼쪽으로 돌아간 삼사분면이며, 두 눈은 화면 왼쪽 위 전경의 쿠마 얼굴 쪽을 향합니다. 음식을 보거나 렌즈를 응시하지 않아 시선 전환의 도착점은 맞습니다. 쿠마는 이현우 쪽으로 머리를 돌리고 있지만 눈은 보이지 않습니다.",
        "built_space": "붉은 비닐 식탁 하나가 왼쪽에서 오른쪽 아래로 대각선으로 놓이고, 이현우는 그 건너편 오른쪽에 있습니다. 식탁에는 회 접시 하나, 흰 그릇 더미 하나, 간장 종지 더미 하나, 반찬 그릇 하나, 금속 통 하나와 일부 가려진 보온 용기·양념 용기가 보입니다. 뒤에는 별도 붉은 탁자 하나와 의자, 수직 기둥 및 항구 시설이 보입니다. 전경 식탁의 재질은 참고와 이어지지만, 참고에 없는 넓은 배경의 정확한 구조는 대조할 수 없습니다. 식탁 윗면과 이현우를 다소 내려다보는 구도이며, 왼쪽에는 좁은 어깨 끝이 아니라 쿠마의 머리와 큰 어깨가 들어옵니다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 동아시아계 청년 남성으로, 참고와 유사한 얼굴과 마른 체격입니다. 더러운 어두운 셔츠, 핏자국, 뺨의 상처와 귀의 검은 인이어 장치가 보입니다. 국적은 외형만으로 확인할 수 없습니다. 쿠마는 뒷머리·귀·목·어깨만 보여 얼굴 정체성이나 혼혈 여부는 확인하기 어렵고, 상의는 참고의 남색보다 회흑색으로 읽힙니다. 변이 생선회 접시가 식탁 아래쪽에 일부 보이며 다른 사람이나 삽입 문자는 없습니다.",
        "hard_violations": [],
        "physics": "이현우의 머리와 목은 앞으로 기울인 몸통에 자연스럽게 연결되어 있습니다. 좌석과 골반은 프레임 밖이라 실제 착석 접촉은 확인할 수 없지만 공중에 뜬 신체는 보이지 않습니다. 쿠마의 어깨도 프레임 밖 몸통으로 이어집니다. 접시와 용기는 식탁 위에 놓여 있고 그릇들은 서로 포개져 지지됩니다. 비정상적인 관절이나 지지 없이 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "이현우는 삼사분면 얼굴로 왼쪽 위의 쿠마를 바라봅니다. 홍채가 향한 방향과 고개 방향이 일치하고, 넓어진 눈과 살짝 벌린 입이 놀란 반응을 드러냅니다. 쿠마의 얼굴 윗부분은 잘려 있지만 이현우를 향한 머리 방향은 읽힙니다. 시선의 대상은 음식이나 카메라가 아니라 쿠마입니다.",
        "built_space": "붉은 비닐 식탁 하나가 화면 아래를 가로지르고 이현우가 건너편 중앙 오른쪽, 쿠마가 왼쪽 전경에 있습니다. 회 접시 하나, 흰 그릇 더미 하나, 간장 종지 더미 하나, 반찬 그릇 하나, 금속 통 하나, 보온 용기 하나와 양념 용기가 보입니다. 뒤에는 붉은 탁자·의자와 수직 기둥들, 항구 수면과 배가 작게 배치됩니다. 수면의 불빛 반사는 가능한 방향입니다. 참고의 식탁 재질과 식기 구성이 이어지며 카메라도 앉은 눈높이에 비교적 가깝습니다. 다만 가슴 아래까지 담고 쿠마의 머리와 어깨가 왼쪽을 크게 차지하여 지정된 얼굴 클로즈업과 좁은 어깨 가장자리 조건은 충족하지 못합니다.",
        "entities": "이현우는 참고와 유사한 젊은 동아시아계 남성 얼굴, 헝클어진 짧은 검은 머리, 마른 상체를 갖습니다. 어두운 낡은 셔츠의 먼지와 혈흔, 얼굴 상처, 검은 인이어 장치가 보입니다. 쿠마는 짧은 검은 머리와 남색 상의가 보여 참고의 가시적 특징에 부합하지만, 얼굴 대부분이 가려져 정확한 정체성과 나이는 검증하기 어렵습니다. 오염된 변이 생선회로 읽히는 접시가 식탁 아래쪽에 남아 있습니다. 추가 인물이나 자막은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "이현우는 식탁 너머에서 몸을 약간 앞으로 기울인 자세이며 목과 어깨의 연결이 자연스럽습니다. 의자와 골반은 가려져 접촉점은 확인할 수 없으나 부유하거나 불가능하게 꺾인 몸은 없습니다. 전경 쿠마의 머리와 어깨도 연속된 신체로 보입니다. 회 접시와 용기들은 식탁이 받치고, 포개진 그릇은 아래 그릇이 지지합니다. 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.714
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 지시한 카메라 앵글과 시선 방향을 훌륭하게 구현했으며, 레퍼런스 컷의 테이블 세팅과 쿠마의 의상(네이비 티셔츠)까지 정확하게 반영했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "전반적인 구도와 인물 묘사는 우수하나, 화면 좌측에 걸친 쿠마의 어깨 의상이 레퍼런스와 달리 깃이 있는 셔츠로 잘못 묘사되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S74sh5_sel.png",
    "asset_id": "75a44ab6-ac8d-4ea0-bdd5-46adac808ccd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1128383>",
    "asset_id": "ac0e1831-ec53-4d61-bd22-48dca7072302",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0da8-a682-743d-9ae8-73f99b5f1c2d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S74sh5"
  },
  "staged_characters_added": [
   "C13"
  ]
 },
 "S75sh4::signage": {
  "fp": "6e8e5cdf9a4a7ef7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S75sh4": {
  "input_fingerprint": "69720ac6dfe4623c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 격납고 안으로 수많은 폐경비행기들이 먼지를 뒤집어쓴 채 방치된 웅장한 내부 전경.\n\nLOCATION (lock): Inside the abandoned airfield's large hangar, beyond its opened doors, among rows of dusty light aircraft. The nighttime interior is dim, with no specific fixture established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the open hangar threshold, hold the upper portion of the crane rise in a high, oblique wide view, looking directly across successive layers of abandoned light aircraft. Frame above the entering men's positions so no people appear, retaining only narrow portions of the doorway at the lateral edges; overlapping wings and fuselages establish the interior's scale without any single aircraft exceeding two-fifths of the image. Let the increased viewing distance carry the reveal, preserving the established inward axis rather than introducing a new lateral viewpoint.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Open hangar entrance (Open) — The entrance is viewed obliquely from its threshold, with its sides confined to the frame edges; used as Peripheral spatial reference anchoring the elevated interior view; Abandoned light aircraft (Numerous, aged, dust-covered and not operational) — Overlapping upper and side surfaces are visible across foreground, middle distance and background; used as Repeated aircraft forms create depth and communicate the scale of the abandoned collection.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure and controlled contrast preserve the dusty aircraft contours without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned airport has moss, broken roads and a deteriorated sign, with fires burning outside. The hangar door is open, revealing numerous old, nonoperational light aircraft.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 격납고 안으로 수많은 폐경비행기들이 먼지를 뒤집어쓴 채 방치된 웅장한 내부 전경.\n\nLOCATION (lock): Inside the abandoned airfield's large hangar, beyond its opened doors, among rows of dusty light aircraft. The nighttime interior is dim, with no specific fixture established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the open hangar threshold, hold the upper portion of the crane rise in a high, oblique wide view, looking directly across successive layers of abandoned light aircraft. Frame above the entering men's positions so no people appear, retaining only narrow portions of the doorway at the lateral edges; overlapping wings and fuselages establish the interior's scale without any single aircraft exceeding two-fifths of the image. Let the increased viewing distance carry the reveal, preserving the established inward axis rather than introducing a new lateral viewpoint.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Open hangar entrance (Open) — The entrance is viewed obliquely from its threshold, with its sides confined to the frame edges; used as Peripheral spatial reference anchoring the elevated interior view; Abandoned light aircraft (Numerous, aged, dust-covered and not operational) — Overlapping upper and side surfaces are visible across foreground, middle distance and background; used as Repeated aircraft forms create depth and communicate the scale of the abandoned collection.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure and controlled contrast preserve the dusty aircraft contours without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned airport has moss, broken roads and a deteriorated sign, with fires burning outside. The hangar door is open, revealing numerous old, nonoperational light aircraft.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 격납고 안으로 수많은 폐경비행기들이 먼지를 뒤집어쓴 채 방치된 웅장한 내부 전경.\n\nLOCATION (lock): Inside the abandoned airfield's large hangar, beyond its opened doors, among rows of dusty light aircraft. The nighttime interior is dim, with no specific fixture established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the open hangar threshold, hold the upper portion of the crane rise in a high, oblique wide view, looking directly across successive layers of abandoned light aircraft. Frame above the entering men's positions so no people appear, retaining only narrow portions of the doorway at the lateral edges; overlapping wings and fuselages establish the interior's scale without any single aircraft exceeding two-fifths of the image. Let the increased viewing distance carry the reveal, preserving the established inward axis rather than introducing a new lateral viewpoint.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Open hangar entrance (Open) — The entrance is viewed obliquely from its threshold, with its sides confined to the frame edges; used as Peripheral spatial reference anchoring the elevated interior view; Abandoned light aircraft (Numerous, aged, dust-covered and not operational) — Overlapping upper and side surfaces are visible across foreground, middle distance and background; used as Repeated aircraft forms create depth and communicate the scale of the abandoned collection.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure and controlled contrast preserve the dusty aircraft contours without specifying an unsupported interior light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The abandoned airport has moss, broken roads and a deteriorated sign, with fires burning outside. The hangar door is open, revealing numerous old, nonoperational light aircraft.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
    "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 바닥이 비정상적으로 매끄럽고 젖어 있음.",
    "entities": "폐경비행기 다수. 인물 없음. 우측 전경 비행기의 날개 형태 불량.",
    "hard_violations": [
     "[gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반",
     "[gemini-pro] 우측 전경 비행기의 우측 날개가 물리적으로 불가능하게 잘려 나간 형태",
     "[gpt-high] 참조와 프롬프트가 정하지 않은 숫자·등록기호 형태의 문자를 여러 기체 동체에 추가하여 새 문구 금지 조건을 위반했습니다."
    ],
    "physics": "비행기들이 바닥에 정지해 있으나, 우측 전경 비행기의 날개가 물리적으로 불가능한 기형적 형태임."
   },
   {
    "label": "B",
    "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
    "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 다수의 비행기가 배치될 충분한 공간.",
    "entities": "먼지를 뒤집어쓴 폐경비행기 다수. 인물 없음.",
    "hard_violations": [
     "[gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 (Location 사진의 지붕을 그대로 차용함)",
     "[gpt-high] 고정된 왼쪽 측벽 창 구획을 장소 참조의 네 개에서 다섯 개로 늘려 건축 개구부를 추가했습니다."
    ],
    "physics": "비행기들이 콘크리트 바닥 위에 바퀴로 안정적으로 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 후보 모두 구조물 레퍼런스의 곡선형 지붕을 반영하지 못했으나, B는 먼지 쌓인 건조한 바닥 질감을 잘 살렸고 비행기의 형태가 온전하여 더 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구조물 지침을 위반했을 뿐만 아니라, 버려진 실내임에도 바닥이 물에 젖은 것처럼 과도하게 반사되며 우측 비행기 날개가 기형적으로 잘려 있어 감점됨."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
        "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 다수의 비행기가 배치될 충분한 공간.",
        "entities": "먼지를 뒤집어쓴 폐경비행기 다수. 인물 없음.",
        "hard_violations": [
         "구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 (Location 사진의 지붕을 그대로 차용함)"
        ],
        "physics": "비행기들이 콘크리트 바닥 위에 바퀴로 안정적으로 지지되어 있음."
       },
       {
        "label": "A",
        "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
        "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 바닥이 비정상적으로 매끄럽고 젖어 있음.",
        "entities": "폐경비행기 다수. 인물 없음. 우측 전경 비행기의 날개 형태 불량.",
        "hard_violations": [
         "구조물 레퍼런스의 곡선형 지붕 형태 지침 위반",
         "우측 전경 비행기의 우측 날개가 물리적으로 불가능하게 잘려 나간 형태"
        ],
        "physics": "비행기들이 바닥에 정지해 있으나, 우측 전경 비행기의 날개가 물리적으로 불가능한 기형적 형태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 후보 모두 구조물 레퍼런스의 곡선형 지붕을 반영하지 못했으나, B는 먼지 쌓인 건조한 바닥 질감을 잘 살렸고 비행기의 형태가 온전하여 더 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구조물 지침을 위반했을 뿐만 아니라, 버려진 실내임에도 바닥이 물에 젖은 것처럼 과도하게 반사되며 우측 비행기 날개가 기형적으로 잘려 있어 감점됨."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
        "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 다수의 비행기가 배치될 충분한 공간.",
        "entities": "먼지를 뒤집어쓴 폐경비행기 다수. 인물 없음.",
        "hard_violations": [
         "구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 (Location 사진의 지붕을 그대로 차용함)"
        ],
        "physics": "비행기들이 콘크리트 바닥 위에 바퀴로 안정적으로 지지되어 있음."
       },
       {
        "label": "A",
        "direction": "카메라 시선은 격납고 입구에서 내부의 폐경비행기들을 향해 안쪽으로 깊게 들어감.",
        "built_space": "양옆에 열린 문이 있는 뾰족한 지붕의 격납고 내부 (구조물 레퍼런스의 곡선형 지붕 미반영). 바닥이 비정상적으로 매끄럽고 젖어 있음.",
        "entities": "폐경비행기 다수. 인물 없음. 우측 전경 비행기의 날개 형태 불량.",
        "hard_violations": [
         "구조물 레퍼런스의 곡선형 지붕 형태 지침 위반",
         "우측 전경 비행기의 우측 날개가 물리적으로 불가능하게 잘려 나간 형태"
        ],
        "physics": "비행기들이 바닥에 정지해 있으나, 우측 전경 비행기의 날개가 물리적으로 불가능한 기형적 형태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "야간의 먼지 쌓인 폐기체와 내부 방향은 맞지만, 고정 창열이 늘었고 전경 왼쪽 기체의 가로 점유가 2/5를 넘어 지정 구도에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "문틀을 가장자리에 두고 작은 기체들을 깊이 있게 배열한 구도는 더 충실하지만, 참조에 없는 기체 식별 문자를 추가해 최종 사용에는 부적합합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 열린 입구에서 중앙 통로를 따라 안쪽 뒷벽을 바라봅니다. 양쪽 기체의 기수와 정지한 프로펠러는 대체로 입구와 중앙 통로 쪽을 향하며, 사람의 시선이나 이동체는 없습니다. 새로운 측면 시점으로 바뀌지는 않았습니다.",
        "built_space": "입구 하나의 양쪽 문설주와 바닥 레일이 보이고, 골강판 벽과 철골 지붕 아래 좌우로 기체가 배치되어 있습니다. 왼쪽 벽에는 구분되는 창 구획이 다섯 개 보여 장소 참조의 네 구획보다 늘었습니다. 오른쪽 창열도 참조보다 촘촘합니다. 상단의 직선 경사 지붕은 구조 참조의 아치형 격납고와 다릅니다. 전경 왼쪽 기체는 보이는 날개 폭만으로 화면 너비의 약 44%를 차지합니다.",
        "entities": "사람이나 얼굴은 없습니다. 전경 두 대와 중·후경의 여러 대를 합쳐 적어도 여덟 대의 소형 프로펠러기가 보입니다. 흰색 외피, 청색·적색 띠, 탁한 캐노피와 먼지 낀 표면은 참조의 노후 경비행기와 부합합니다. 벽 주변에는 참조에도 있는 유형의 선반, 공구와 용기가 놓여 있습니다. 외부 간판과 도로는 이 내부 구도에서 보이지 않습니다.",
        "hard_violations": [
         "고정된 왼쪽 측벽 창 구획을 장소 참조의 네 개에서 다섯 개로 늘려 건축 개구부를 추가했습니다."
        ],
        "physics": "전경 기체들은 착륙바퀴로 콘크리트 바닥에 지지되고 바퀴 앞뒤에 고임목이 보입니다. 중경 기체도 보이는 착륙장치로 서 있으며 후경의 지지부는 기체끼리 가려집니다. 선반과 용기는 바닥에 놓여 있고, 떠 있는 물체나 움직이는 프로펠러는 없습니다. 바닥의 약한 광택도 가능한 반사입니다."
       },
       {
        "label": "B",
        "direction": "카메라는 문턱에서 중앙 통로를 따라 격납고 안쪽을 내려다봅니다. 좌우 열의 기체들은 대체로 입구를 향하면서 기수가 통로 쪽으로 조금 돌아가 있습니다. 사람의 시선이나 진행 동작은 없으며, 입구에서 안쪽으로 이어지는 촬영축이 유지됩니다.",
        "built_space": "입구 하나의 문설주가 양쪽 가장자리에 좁게 남고 바닥에는 문 레일이 있습니다. 양쪽 측벽에서 각각 네 구획의 창이 구분되어 장소 참조의 배치에 더 가깝습니다. 뒤쪽 선반과 골강판 벽이 보이며 지붕 윤곽은 대부분 프레임 밖이라 아치형 구조와의 일치 여부를 확정하기 어렵습니다. 전경 두 기체의 보이는 가로 폭은 각각 화면의 2/5 이내이고, 그 뒤로 기체 열이 이어집니다.",
        "entities": "사람과 얼굴은 없고, 약 열 대의 오래된 소형 프로펠러기가 전경부터 후경까지 보입니다. 먼지 낀 캐노피, 바랜 흰색 도장과 청색·적색 띠는 참조에 가깝습니다. 다만 왼쪽 중경 동체와 오른쪽 전경 동체에 참조에서 확인되지 않는 검은 숫자·등록기호 형태의 문자가 추가되어 있습니다. 외부 간판과 불은 프레임에 들어오지 않습니다.",
        "hard_violations": [
         "참조와 프롬프트가 정하지 않은 숫자·등록기호 형태의 문자를 여러 기체 동체에 추가하여 새 문구 금지 조건을 위반했습니다."
        ],
        "physics": "전경과 중경 기체들은 착륙바퀴로 바닥에 서 있으며 여러 바퀴에 고임목이 보입니다. 후경 기체의 하부는 겹침과 어둠으로 일부 가려지지만 공중에 떠 있다는 징후는 없습니다. 프로펠러는 멈춰 있고, 선반과 용기는 바닥에 지지됩니다. 젖은 바닥의 밝은 반사는 가능하지만 장소 참조보다 젖은 면적이 훨씬 큽니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "야간의 먼지 쌓인 폐기체와 내부 방향은 맞지만, 고정 창열이 늘었고 전경 왼쪽 기체의 가로 점유가 2/5를 넘어 지정 구도에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "문틀을 가장자리에 두고 작은 기체들을 깊이 있게 배열한 구도는 더 충실하지만, 참조에 없는 기체 식별 문자를 추가해 최종 사용에는 부적합합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 열린 입구에서 중앙 통로를 따라 안쪽 뒷벽을 바라봅니다. 양쪽 기체의 기수와 정지한 프로펠러는 대체로 입구와 중앙 통로 쪽을 향하며, 사람의 시선이나 이동체는 없습니다. 새로운 측면 시점으로 바뀌지는 않았습니다.",
        "built_space": "입구 하나의 양쪽 문설주와 바닥 레일이 보이고, 골강판 벽과 철골 지붕 아래 좌우로 기체가 배치되어 있습니다. 왼쪽 벽에는 구분되는 창 구획이 다섯 개 보여 장소 참조의 네 구획보다 늘었습니다. 오른쪽 창열도 참조보다 촘촘합니다. 상단의 직선 경사 지붕은 구조 참조의 아치형 격납고와 다릅니다. 전경 왼쪽 기체는 보이는 날개 폭만으로 화면 너비의 약 44%를 차지합니다.",
        "entities": "사람이나 얼굴은 없습니다. 전경 두 대와 중·후경의 여러 대를 합쳐 적어도 여덟 대의 소형 프로펠러기가 보입니다. 흰색 외피, 청색·적색 띠, 탁한 캐노피와 먼지 낀 표면은 참조의 노후 경비행기와 부합합니다. 벽 주변에는 참조에도 있는 유형의 선반, 공구와 용기가 놓여 있습니다. 외부 간판과 도로는 이 내부 구도에서 보이지 않습니다.",
        "hard_violations": [
         "고정된 왼쪽 측벽 창 구획을 장소 참조의 네 개에서 다섯 개로 늘려 건축 개구부를 추가했습니다."
        ],
        "physics": "전경 기체들은 착륙바퀴로 콘크리트 바닥에 지지되고 바퀴 앞뒤에 고임목이 보입니다. 중경 기체도 보이는 착륙장치로 서 있으며 후경의 지지부는 기체끼리 가려집니다. 선반과 용기는 바닥에 놓여 있고, 떠 있는 물체나 움직이는 프로펠러는 없습니다. 바닥의 약한 광택도 가능한 반사입니다."
       },
       {
        "label": "A",
        "direction": "카메라는 문턱에서 중앙 통로를 따라 격납고 안쪽을 내려다봅니다. 좌우 열의 기체들은 대체로 입구를 향하면서 기수가 통로 쪽으로 조금 돌아가 있습니다. 사람의 시선이나 진행 동작은 없으며, 입구에서 안쪽으로 이어지는 촬영축이 유지됩니다.",
        "built_space": "입구 하나의 문설주가 양쪽 가장자리에 좁게 남고 바닥에는 문 레일이 있습니다. 양쪽 측벽에서 각각 네 구획의 창이 구분되어 장소 참조의 배치에 더 가깝습니다. 뒤쪽 선반과 골강판 벽이 보이며 지붕 윤곽은 대부분 프레임 밖이라 아치형 구조와의 일치 여부를 확정하기 어렵습니다. 전경 두 기체의 보이는 가로 폭은 각각 화면의 2/5 이내이고, 그 뒤로 기체 열이 이어집니다.",
        "entities": "사람과 얼굴은 없고, 약 열 대의 오래된 소형 프로펠러기가 전경부터 후경까지 보입니다. 먼지 낀 캐노피, 바랜 흰색 도장과 청색·적색 띠는 참조에 가깝습니다. 다만 왼쪽 중경 동체와 오른쪽 전경 동체에 참조에서 확인되지 않는 검은 숫자·등록기호 형태의 문자가 추가되어 있습니다. 외부 간판과 불은 프레임에 들어오지 않습니다.",
        "hard_violations": [
         "참조와 프롬프트가 정하지 않은 숫자·등록기호 형태의 문자를 여러 기체 동체에 추가하여 새 문구 금지 조건을 위반했습니다."
        ],
        "physics": "전경과 중경 기체들은 착륙바퀴로 바닥에 서 있으며 여러 바퀴에 고임목이 보입니다. 후경 기체의 하부는 겹침과 어둠으로 일부 가려지지만 공중에 떠 있다는 징후는 없습니다. 프로펠러는 멈춰 있고, 선반과 용기는 바닥에 지지됩니다. 젖은 바닥의 밝은 반사는 가능하지만 장소 참조보다 젖은 면적이 훨씬 큽니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 (Location 사진의 지붕을 그대로 차용함)",
     "[gpt-high] 고정된 왼쪽 측벽 창 구획을 장소 참조의 네 개에서 다섯 개로 늘려 건축 개구부를 추가했습니다."
    ],
    "A": [
     "[gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반",
     "[gemini-pro] 우측 전경 비행기의 우측 날개가 물리적으로 불가능하게 잘려 나간 형태",
     "[gpt-high] 참조와 프롬프트가 정하지 않은 숫자·등록기호 형태의 문자를 여러 기체 동체에 추가하여 새 문구 금지 조건을 위반했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1500,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "두 후보 모두 구조물 레퍼런스의 곡선형 지붕을 반영하지 못했으나, B는 먼지 쌓인 건조한 바닥 질감을 잘 살렸고 비행기의 형태가 온전하여 더 우수함.  ★위반: [gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 (Location 사진의 지붕을 그대로 차용함) / [gpt-high] 고정된 왼쪽 측벽 창 구획을 장소 참조의 네 개에서 다섯 개로 늘려 건축 개구부를 추가했습니다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "구조물 지침을 위반했을 뿐만 아니라, 버려진 실내임에도 바닥이 물에 젖은 것처럼 과도하게 반사되며 우측 비행기 날개가 기형적으로 잘려 있어 감점됨.  ★위반: [gemini-pro] 구조물 레퍼런스의 곡선형 지붕 형태 지침 위반 / [gemini-pro] 우측 전경 비행기의 우측 날개가 물리적으로 불가능하게 잘려 나간 형태 / [gpt-high] 참조와 프롬프트가 정하지 않은 숫자·등록기호 형태의 문자를 여러 기체 동체에 추가하여 새 문구 금지 조건을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L128B02.png",
    "asset_id": "69c67933-e348-4e35-b1a8-bcf6fe38f3b4",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/background_chain/seed_bg_abandoned_airport_sel.png",
    "asset_id": "45ef8494-8d7b-45e5-a2fc-02ce9fedcbf4",
    "role": "structure_seed_look"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dae-5510-7f0e-b52e-768ea57a2efd",
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S75sh10::signage": {
  "fp": "46acbf4aa0f937ec",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S75sh10": {
  "input_fingerprint": "3522b20a3e1911ea",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 격납고 안쪽을 향해 한 발이 허공에 뜬 달리기 mid-action 자세로, 다급하게 입을 벌린 신부의 전신 구도.\n\nLOCATION (lock): At the open hangar entrance of the abandoned airfield, heading into the dim aircraft-storage hall at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established interior position, pan toward the entrance at approximately standing eye level, keeping the camera laterally outside 신부's inward running line. Catch his full figure at center-right in a three-quarter view, mouth open and one foot airborne, with generous clearance above his head and beneath his feet and open space toward the left for his advance. His attention remains on 이현우 outside the left edge, and the pan changes the direction of attention rather than adding a simultaneous push-in or lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hangar entrance (Open) — Seen from inside, behind and to the right of the incoming priest; used as Establishes where the interruption enters the conversation space; Abandoned light aircraft (Aged and stationary) — Partial aircraft sides remain at the lateral margins, clear of the running figure; used as Preserves the hangar context without obstructing the head-to-foot silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hangar's subdued nighttime exposure and controlled contrast, keeping the urgent expression and airborne stride readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open hangar, dust-covered light aircraft, and nighttime interior appearance from the reference. Exclude working engines and exhaust smoke; the aircraft have not yet been repaired.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with its old aircraft still nonoperational; fires continue outside amid the abandoned airport's moss and broken paving. 신부: He has arrived at the airport in his clerical collar, visibly urgent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 격납고 안쪽을 향해 한 발이 허공에 뜬 달리기 mid-action 자세로, 다급하게 입을 벌린 신부의 전신 구도.\n\nLOCATION (lock): At the open hangar entrance of the abandoned airfield, heading into the dim aircraft-storage hall at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established interior position, pan toward the entrance at approximately standing eye level, keeping the camera laterally outside 신부's inward running line. Catch his full figure at center-right in a three-quarter view, mouth open and one foot airborne, with generous clearance above his head and beneath his feet and open space toward the left for his advance. His attention remains on 이현우 outside the left edge, and the pan changes the direction of attention rather than adding a simultaneous push-in or lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hangar entrance (Open) — Seen from inside, behind and to the right of the incoming priest; used as Establishes where the interruption enters the conversation space; Abandoned light aircraft (Aged and stationary) — Partial aircraft sides remain at the lateral margins, clear of the running figure; used as Preserves the hangar context without obstructing the head-to-foot silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hangar's subdued nighttime exposure and controlled contrast, keeping the urgent expression and airborne stride readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open hangar, dust-covered light aircraft, and nighttime interior appearance from the reference. Exclude working engines and exhaust smoke; the aircraft have not yet been repaired.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with its old aircraft still nonoperational; fires continue outside amid the abandoned airport's moss and broken paving. 신부: He has arrived at the airport in his clerical collar, visibly urgent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 격납고 안쪽을 향해 한 발이 허공에 뜬 달리기 mid-action 자세로, 다급하게 입을 벌린 신부의 전신 구도.\n\nLOCATION (lock): At the open hangar entrance of the abandoned airfield, heading into the dim aircraft-storage hall at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established interior position, pan toward the entrance at approximately standing eye level, keeping the camera laterally outside 신부's inward running line. Catch his full figure at center-right in a three-quarter view, mouth open and one foot airborne, with generous clearance above his head and beneath his feet and open space toward the left for his advance. His attention remains on 이현우 outside the left edge, and the pan changes the direction of attention rather than adding a simultaneous push-in or lighting change.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Hangar entrance (Open) — Seen from inside, behind and to the right of the incoming priest; used as Establishes where the interruption enters the conversation space; Abandoned light aircraft (Aged and stationary) — Partial aircraft sides remain at the lateral margins, clear of the running figure; used as Preserves the hangar context without obstructing the head-to-foot silhouette.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding hangar's subdued nighttime exposure and controlled contrast, keeping the urgent expression and airborne stride readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the open hangar, dust-covered light aircraft, and nighttime interior appearance from the reference. Exclude working engines and exhaust smoke; the aircraft have not yet been repaired.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hangar remains open with its old aircraft still nonoperational; fires continue outside amid the abandoned airport's moss and broken paving. 신부: He has arrived at the airport in his clerical collar, visibly urgent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 신부 (한국인 남성, 60대의 얼굴, 짧은 머리) — wearing: 하얀 로만칼라가 돋보이는 낡고 단정한 검은색 사제복 상하의. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "신부는 화면 왼쪽 바깥을 향해 시선을 두고 격납고 안쪽으로 쇄도하고 있음.",
    "built_space": "격납고 내부에서 밖을 바라보는 뷰이며 양측에 경비행기가 배치됨. 단, 우측 비행기의 줄무늬 색상이 레퍼런스(붉은색)와 다름.",
    "entities": "얼굴, 헤어스타일, 사제복 및 로만칼라 등 신부의 외형이 지정된 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "달리는 동작이나 지시문과 달리 양발이 모두 바닥에서 떨어져 허공에 떠 있음."
   },
   {
    "label": "B",
    "direction": "신부는 화면 왼쪽 바깥의 보이지 않는 대상을 향해 시선을 고정하고 격납고 안쪽을 향해 달려오고 있음.",
    "built_space": "야간의 격납고 내부에서 열린 문 밖을 내다보는 구도. 양쪽 가장자리에 경비행기가 위치하며 우측 비행기는 레퍼런스와 동일한 붉은 줄무늬를 가짐.",
    "entities": "60대 한국인 남성의 얼굴, 짧은 머리, 낡고 단정한 검은색 사제복과 하얀 로만칼라 등 신부의 특성이 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "왼발이 바닥을 단단히 딛고 체중을 지탱하며, 오른발은 허공에 뜬 자연스러운 달리기 자세임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "'한 발이 허공에 뜬' 달리기 자세를 정확히 구현했으며, 프레이밍 조건과 우측 비행기의 붉은 줄무늬 등 레퍼런스의 배경 디테일을 충실히 반영함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 인물의 외형은 훌륭하나, 지시문과 달리 양발이 모두 공중에 떠 있고 우측 비행기의 색상이 레퍼런스와 일치하지 않음."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "신부는 화면 왼쪽 바깥의 보이지 않는 대상을 향해 시선을 고정하고 격납고 안쪽을 향해 달려오고 있음.",
        "built_space": "야간의 격납고 내부에서 열린 문 밖을 내다보는 구도. 양쪽 가장자리에 경비행기가 위치하며 우측 비행기는 레퍼런스와 동일한 붉은 줄무늬를 가짐.",
        "entities": "60대 한국인 남성의 얼굴, 짧은 머리, 낡고 단정한 검은색 사제복과 하얀 로만칼라 등 신부의 특성이 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "왼발이 바닥을 단단히 딛고 체중을 지탱하며, 오른발은 허공에 뜬 자연스러운 달리기 자세임."
       },
       {
        "label": "A",
        "direction": "신부는 화면 왼쪽 바깥을 향해 시선을 두고 격납고 안쪽으로 쇄도하고 있음.",
        "built_space": "격납고 내부에서 밖을 바라보는 뷰이며 양측에 경비행기가 배치됨. 단, 우측 비행기의 줄무늬 색상이 레퍼런스(붉은색)와 다름.",
        "entities": "얼굴, 헤어스타일, 사제복 및 로만칼라 등 신부의 외형이 지정된 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "달리는 동작이나 지시문과 달리 양발이 모두 바닥에서 떨어져 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "'한 발이 허공에 뜬' 달리기 자세를 정확히 구현했으며, 프레이밍 조건과 우측 비행기의 붉은 줄무늬 등 레퍼런스의 배경 디테일을 충실히 반영함."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "구도와 인물의 외형은 훌륭하나, 지시문과 달리 양발이 모두 공중에 떠 있고 우측 비행기의 색상이 레퍼런스와 일치하지 않음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 화면 왼쪽 바깥의 보이지 않는 대상을 향해 시선을 고정하고 격납고 안쪽을 향해 달려오고 있음.",
        "built_space": "야간의 격납고 내부에서 열린 문 밖을 내다보는 구도. 양쪽 가장자리에 경비행기가 위치하며 우측 비행기는 레퍼런스와 동일한 붉은 줄무늬를 가짐.",
        "entities": "60대 한국인 남성의 얼굴, 짧은 머리, 낡고 단정한 검은색 사제복과 하얀 로만칼라 등 신부의 특성이 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "왼발이 바닥을 단단히 딛고 체중을 지탱하며, 오른발은 허공에 뜬 자연스러운 달리기 자세임."
       },
       {
        "label": "A",
        "direction": "신부는 화면 왼쪽 바깥을 향해 시선을 두고 격납고 안쪽으로 쇄도하고 있음.",
        "built_space": "격납고 내부에서 밖을 바라보는 뷰이며 양측에 경비행기가 배치됨. 단, 우측 비행기의 줄무늬 색상이 레퍼런스(붉은색)와 다름.",
        "entities": "얼굴, 헤어스타일, 사제복 및 로만칼라 등 신부의 외형이 지정된 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "달리는 동작이나 지시문과 달리 양발이 모두 바닥에서 떨어져 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "머리 위와 발밑 여백이 넉넉한 중앙 오른쪽 전신 구도, 왼쪽을 향한 다급한 시선, 한 발을 든 달리기가 더 정확하지만 양옆 항공기의 색띠 배치는 이전 장면과 다릅니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물과 야간 격납고는 잘 맞지만 달리기가 더 정면으로 향하고 발밑 여백이 줄었으며, 한 발만 든 순간 대신 양발이 뜬 질주 순간을 보여줍니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "신부는 뒤쪽 오른편의 열린 입구에서 실내 전경으로 들어오며 약간 화면 왼쪽으로 향한다. 얼굴과 시선은 카메라보다 왼쪽을 향해 있어 화면 밖 이현우를 바라보라는 지시와 부합한다. 다만 몸의 진행 방향에는 카메라 쪽 성분도 상당하다. 양옆 항공기의 프로펠러는 정지해 있다.",
        "built_space": "중앙 오른쪽부터 오른쪽 가장자리까지 열린 출입구 하나가 있고, 신부는 문턱보다 실내 쪽에 있다. 왼쪽 벽에는 완전히 보이는 큰 창 구획 하나와 가장자리의 일부 창, 그 아래 정비 선반과 사다리가 보인다. 경비행기는 좌우 가장자리에 한 대씩 부분적으로 보이며 인물의 전신을 가리지 않는다. 골조와 골강판 벽, 낡은 바닥은 참고 장소와 유사하다. 다만 입구를 안에서 바라보는 역방향 구도라면 참고의 파란 띠 항공기는 화면 오른쪽에 와야 하는데, 여기서는 왼쪽에 보여 기존 배치의 연속성이 약하다. 젖은 바닥의 광원 반사는 가능한 위치다.",
        "entities": "인물은 한 명뿐이며 짧은 회색 머리의 60대 한국인 남성 신부로 읽힌다. 얼굴과 체격은 인물 참고에 가깝고, 흰 로만칼라와 검은 사제복 상하의, 검은 구두가 보인다. 입을 벌린 다급한 표정도 명확하다. 양옆의 낡고 오염된 프로펠러 경비행기, 야간 외부의 불길과 파손된 포장이 보이며 가동 중인 엔진이나 배기가스는 없다. 이현우나 다른 사람은 등장하지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽으로 내디딘 구두가 바닥에 닿아 체중을 받치고, 반대쪽 다리는 무릎을 굽혀 뒤로 들려 있다. 몸의 전진 기울기와 팔의 움직임이 달리기 중 한 발 지지 순간으로 성립한다. 항공기는 보이는 착륙바퀴와 바퀴 고임목으로 바닥에 지지되어 있고, 선반과 정비 물품도 바닥이나 선반에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "신부의 얼굴과 시선은 화면 왼쪽 바깥을 향해 이현우의 위치 지시와 맞는다. 그러나 가슴과 내민 다리는 비교적 카메라 정면으로 향해, 왼쪽의 빈 공간으로 비껴 들어오는 삼사분면 달리기보다 관객을 향한 질주로 더 강하게 읽힌다. 항공기 프로펠러는 정지 상태다.",
        "built_space": "신부 뒤쪽 오른편에 열린 출입구 하나가 있고, 왼쪽에는 큰 창 구획 하나와 가장자리의 부분 창, 정비 선반과 사다리가 있다. 좌우 가장자리에 경비행기 한 대씩이 부분적으로 보이며 인물과 겹치지 않는다. 왼쪽의 붉은 띠와 오른쪽의 파란 띠는 참고 장면을 반대 방향에서 보는 배치에 더 가깝다. 출입구 밖에는 항공기 한 대가 추가로 보이지만 그 위치는 참고 사진으로 확인되지 않는다. 카메라는 실내 눈높이에 가깝고, 바닥의 반사도 가능한 형태다. 전신은 들어오지만 발밑 여백은 A보다 좁다.",
        "entities": "짧은 회색 머리와 참고에 가까운 얼굴의 노년 한국인 남성 한 명이 흰 로만칼라, 검은 사제복 상하의와 구두를 착용했다. 입을 열고 긴박한 표정을 짓는다. 양옆 항공기는 먼지와 표면 마모가 있는 경비행기로 읽히며 작동 흔적은 없다. 외부에는 밤하늘과 불길, 낡은 포장이 보인다. 다른 사람이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "앞으로 내민 구두의 밑창과 바닥 사이에 틈이 있고 뒤쪽 구두도 들려 있어 양발이 공중에 있는 순간으로 보인다. 뒤로 접힌 다리와 앞으로 뻗은 착지 다리, 달리기 팔동작이 있어 지면을 박차고 다음 발을 내딛는 짧은 비행 단계로 물리적으로 성립한다. 근거 없이 떠 있는 몸은 아니지만 한 발을 든 순간이라는 지시에는 A보다 덜 정확하다. 양옆 항공기는 착륙바퀴로 바닥에 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "머리 위와 발밑 여백이 넉넉한 중앙 오른쪽 전신 구도, 왼쪽을 향한 다급한 시선, 한 발을 든 달리기가 더 정확하지만 양옆 항공기의 색띠 배치는 이전 장면과 다릅니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "인물과 야간 격납고는 잘 맞지만 달리기가 더 정면으로 향하고 발밑 여백이 줄었으며, 한 발만 든 순간 대신 양발이 뜬 질주 순간을 보여줍니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "신부는 뒤쪽 오른편의 열린 입구에서 실내 전경으로 들어오며 약간 화면 왼쪽으로 향한다. 얼굴과 시선은 카메라보다 왼쪽을 향해 있어 화면 밖 이현우를 바라보라는 지시와 부합한다. 다만 몸의 진행 방향에는 카메라 쪽 성분도 상당하다. 양옆 항공기의 프로펠러는 정지해 있다.",
        "built_space": "중앙 오른쪽부터 오른쪽 가장자리까지 열린 출입구 하나가 있고, 신부는 문턱보다 실내 쪽에 있다. 왼쪽 벽에는 완전히 보이는 큰 창 구획 하나와 가장자리의 일부 창, 그 아래 정비 선반과 사다리가 보인다. 경비행기는 좌우 가장자리에 한 대씩 부분적으로 보이며 인물의 전신을 가리지 않는다. 골조와 골강판 벽, 낡은 바닥은 참고 장소와 유사하다. 다만 입구를 안에서 바라보는 역방향 구도라면 참고의 파란 띠 항공기는 화면 오른쪽에 와야 하는데, 여기서는 왼쪽에 보여 기존 배치의 연속성이 약하다. 젖은 바닥의 광원 반사는 가능한 위치다.",
        "entities": "인물은 한 명뿐이며 짧은 회색 머리의 60대 한국인 남성 신부로 읽힌다. 얼굴과 체격은 인물 참고에 가깝고, 흰 로만칼라와 검은 사제복 상하의, 검은 구두가 보인다. 입을 벌린 다급한 표정도 명확하다. 양옆의 낡고 오염된 프로펠러 경비행기, 야간 외부의 불길과 파손된 포장이 보이며 가동 중인 엔진이나 배기가스는 없다. 이현우나 다른 사람은 등장하지 않는다.",
        "hard_violations": [],
        "physics": "화면 왼쪽으로 내디딘 구두가 바닥에 닿아 체중을 받치고, 반대쪽 다리는 무릎을 굽혀 뒤로 들려 있다. 몸의 전진 기울기와 팔의 움직임이 달리기 중 한 발 지지 순간으로 성립한다. 항공기는 보이는 착륙바퀴와 바퀴 고임목으로 바닥에 지지되어 있고, 선반과 정비 물품도 바닥이나 선반에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "신부의 얼굴과 시선은 화면 왼쪽 바깥을 향해 이현우의 위치 지시와 맞는다. 그러나 가슴과 내민 다리는 비교적 카메라 정면으로 향해, 왼쪽의 빈 공간으로 비껴 들어오는 삼사분면 달리기보다 관객을 향한 질주로 더 강하게 읽힌다. 항공기 프로펠러는 정지 상태다.",
        "built_space": "신부 뒤쪽 오른편에 열린 출입구 하나가 있고, 왼쪽에는 큰 창 구획 하나와 가장자리의 부분 창, 정비 선반과 사다리가 있다. 좌우 가장자리에 경비행기 한 대씩이 부분적으로 보이며 인물과 겹치지 않는다. 왼쪽의 붉은 띠와 오른쪽의 파란 띠는 참고 장면을 반대 방향에서 보는 배치에 더 가깝다. 출입구 밖에는 항공기 한 대가 추가로 보이지만 그 위치는 참고 사진으로 확인되지 않는다. 카메라는 실내 눈높이에 가깝고, 바닥의 반사도 가능한 형태다. 전신은 들어오지만 발밑 여백은 A보다 좁다.",
        "entities": "짧은 회색 머리와 참고에 가까운 얼굴의 노년 한국인 남성 한 명이 흰 로만칼라, 검은 사제복 상하의와 구두를 착용했다. 입을 열고 긴박한 표정을 짓는다. 양옆 항공기는 먼지와 표면 마모가 있는 경비행기로 읽히며 작동 흔적은 없다. 외부에는 밤하늘과 불길, 낡은 포장이 보인다. 다른 사람이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "앞으로 내민 구두의 밑창과 바닥 사이에 틈이 있고 뒤쪽 구두도 들려 있어 양발이 공중에 있는 순간으로 보인다. 뒤로 접힌 다리와 앞으로 뻗은 착지 다리, 달리기 팔동작이 있어 지면을 박차고 다음 발을 내딛는 짧은 비행 단계로 물리적으로 성립한다. 근거 없이 떠 있는 몸은 아니지만 한 발을 든 순간이라는 지시에는 A보다 덜 정확하다. 양옆 항공기는 착륙바퀴로 바닥에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.625,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.625,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1625
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "'한 발이 허공에 뜬' 달리기 자세를 정확히 구현했으며, 프레이밍 조건과 우측 비행기의 붉은 줄무늬 등 레퍼런스의 배경 디테일을 충실히 반영함."
   },
   {
    "label": "A",
    "score": 1625,
    "verdict_ko": "구도와 인물의 외형은 훌륭하나, 지시문과 달리 양발이 모두 공중에 떠 있고 우측 비행기의 색상이 레퍼런스와 일치하지 않음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S75sh4_sel.png",
    "asset_id": "5b49ddab-4018-45fd-98aa-f7de617e4d38",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 신부: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:773901>",
    "asset_id": "4792771b-ca0f-457e-b541-6803a18eba46",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0db5-d5db-76b2-a167-c5547ba573d2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh4"
  }
 },
 "S76sh1::signage": {
  "fp": "5d31688198071a0c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S76sh1": {
  "input_fingerprint": "b6fa9040a8ed625f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위에서 식은땀을 흘린 채 괴로운 표정으로 숨을 들이마셔 가슴이 부풀어 오른 앰버의 상체.\n\nLOCATION (lock): At the feverish child's bed inside the harbor-town health clinic, under nighttime treatment-room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless beside 앰버's shoulder, close and above her face, looking diagonally down across her upper torso rather than along the bed's centerline. Her strained face occupies the upper-left portion and her inhaling chest the lower center, with a narrow border of bed maintaining physical context; shallow focus keeps both the expression and breathing effort legible. Her eyes are closed against the discomfort, making this a direct observation of illness rather than a subjective or distorted view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed (Occupied by 앰버) — Only the supporting surface around her shoulder and upper torso is visible from the diagonal bedside angle; used as Provides a restrained contextual border around the close human subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior illumination with subdued brightness and gentle contrast reveals her cold sweat and feverish distress without stylized perceptual effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is recumbent in the health-center bed, sweating and suffering from a high fever. The bed supports her body, but the scene text does not specify the orientation of her torso and head or the placement of her arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains the treatment location at night. Charlie retains his unrepaired body and chest damage and possession of B-200's chest component. 앰버: She lies in bed with a high fever and cold sweat after emergency treatment of her side wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위에서 식은땀을 흘린 채 괴로운 표정으로 숨을 들이마셔 가슴이 부풀어 오른 앰버의 상체.\n\nLOCATION (lock): At the feverish child's bed inside the harbor-town health clinic, under nighttime treatment-room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless beside 앰버's shoulder, close and above her face, looking diagonally down across her upper torso rather than along the bed's centerline. Her strained face occupies the upper-left portion and her inhaling chest the lower center, with a narrow border of bed maintaining physical context; shallow focus keeps both the expression and breathing effort legible. Her eyes are closed against the discomfort, making this a direct observation of illness rather than a subjective or distorted view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed (Occupied by 앰버) — Only the supporting surface around her shoulder and upper torso is visible from the diagonal bedside angle; used as Provides a restrained contextual border around the close human subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior illumination with subdued brightness and gentle contrast reveals her cold sweat and feverish distress without stylized perceptual effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is recumbent in the health-center bed, sweating and suffering from a high fever. The bed supports her body, but the scene text does not specify the orientation of her torso and head or the placement of her arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains the treatment location at night. Charlie retains his unrepaired body and chest damage and possession of B-200's chest component. 앰버: She lies in bed with a high fever and cold sweat after emergency treatment of her side wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 침대 위에서 식은땀을 흘린 채 괴로운 표정으로 숨을 들이마셔 가슴이 부풀어 오른 앰버의 상체.\n\nLOCATION (lock): At the feverish child's bed inside the harbor-town health clinic, under nighttime treatment-room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin motionless beside 앰버's shoulder, close and above her face, looking diagonally down across her upper torso rather than along the bed's centerline. Her strained face occupies the upper-left portion and her inhaling chest the lower center, with a narrow border of bed maintaining physical context; shallow focus keeps both the expression and breathing effort legible. Her eyes are closed against the discomfort, making this a direct observation of illness rather than a subjective or distorted view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bed (Occupied by 앰버) — Only the supporting surface around her shoulder and upper torso is visible from the diagonal bedside angle; used as Provides a restrained contextual border around the close human subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral interior illumination with subdued brightness and gentle contrast reveals her cold sweat and feverish distress without stylized perceptual effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Amber is recumbent in the health-center bed, sweating and suffering from a high fever. The bed supports her body, but the scene text does not specify the orientation of her torso and head or the placement of her arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The clinic bed remains the treatment location at night. Charlie retains his unrepaired body and chest damage and possession of B-200's chest component. 앰버: She lies in bed with a high fever and cold sweat after emergency treatment of her side wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "B",
    "direction": "앰버는 눈을 꽉 감고 얼굴을 위쪽과 화면 오른쪽으로 향한다. 특정 대상을 바라보지 않아 눈을 감으라는 지시와 맞는다. 카메라는 침대 옆에서 상체를 비스듬히 보지만, 얼굴 바로 위에서 대각선 아래로 내려다보는 시점보다는 낮다. 얼굴은 왼쪽 위, 가슴은 아래 중앙에 놓인다.",
    "built_space": "침대 하나의 연푸른 침구와 머리 받침이 보이며, 뒤에는 모니터 한 대와 하부 장치, 창 하나, 수납대 하나가 있다. 이전 장면의 병실 재질과 야간 창밖 분위기는 대체로 이어진다. 중복된 설비나 불가능한 반사는 없다. 다만 침구가 좁은 테두리에 그치지 않고 넓게 보이며, 허리와 병실 배경까지 포함해 요구한 상체 클로즈업보다 범위가 크다.",
    "entities": "금발의 어린 여자아이 한 명만 보이고, 창백한 피부와 이마·볼의 땀, 찡그린 눈썹과 벌어진 입이 확인된다. 얼굴은 참조보다 다소 성숙하고 윤곽이 달라 정확한 동일인 재현은 약하다. 기름때 묻은 카키 작업복, 가죽 공구 벨트, 복합 필터형 방진 마스크가 있다. 마스크는 참조처럼 목과 가슴에 내려져 있으며 세부 형태는 다르다. 눈 크기는 감긴 상태라 확인할 수 없다.",
    "hard_violations": [],
    "physics": "머리와 등은 경사진 침대와 베개에 받쳐져 있고, 보이는 팔과 손도 몸 옆 침대에 놓여 있다. 마스크는 목 끈에 연결된 채 가슴에 얹혀 있으며 공구는 허리 벨트가 지지한다. 떠 있는 신체나 물체는 없다. 벌어진 입과 긴장한 얼굴, 도드라진 가슴 윤곽은 힘겹게 숨을 들이마시는 순간으로 읽히지만 흉곽 팽창 자체는 정지 화면에서 제한적으로만 확인된다."
   },
   {
    "label": "A",
    "direction": "앰버는 눈을 감고 얼굴을 위로 향하지만 눈썹과 입 주변이 비교적 이완되어 있다. 오른쪽 전경의 추가 인물은 앰버 쪽으로 고개를 숙인다. 카메라가 그 인물의 어깨 너머에서 앰버를 관찰하는 구도가 되어, 앰버의 어깨 옆과 얼굴 위에 직접 놓인 지정 시점과 다르다.",
    "built_space": "침대 하나, 큰 베개 하나, 오른쪽 난간 하나와 머리판 일부가 보인다. 배경에는 모니터 한 대, 창 하나, 수납대 하나가 있어 병실의 기본 맥락은 이어진다. 그러나 넓은 침구와 배경, 전경 인물까지 들어와 침대를 좁은 지지면 테두리로만 남기라는 구성을 벗어난다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
    "entities": "금발과 어린 얼굴, 카키색 작업복, 목에 내려진 정교한 방진 마스크는 앰버 참조와 비교적 가깝다. 이마에는 땀이 있으나 괴로운 표정보다는 잠든 표정이다. 가슴 아래는 푸른 담요로 덮였고, 허리 벨트는 프레임에서 확인되지 않아 누락으로 평가하지 않는다. 오른쪽에는 검은 머리와 어두운 옷의 다른 인물이 명백히 추가되어, 앰버만 보여야 한다는 조건을 위반한다.",
    "hard_violations": [
     "앰버만 등장해야 하는 장면에 다른 인물의 머리와 어깨·상체를 오른쪽 전경에 추가했다."
    ],
    "physics": "앰버의 머리는 베개에, 등과 팔은 침대에 지지된다. 마스크는 목 끈과 가슴에 받쳐져 있고 담요는 몸 위에 자연스럽게 놓여 있다. 추가 인물의 하체는 프레임 밖이지만 공중에 떠 있다는 증거는 없다. 물리적으로 불가능한 자세는 없으나, 닫힌 입과 이완된 표정 및 담요에 가려진 흉곽 때문에 요구된 힘겨운 들숨과 가슴 팽창은 읽히지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "식은땀과 고통스러운 들숨, 침대에 지지된 상체를 구현했지만, 허리와 병실 배경까지 보여 지정된 밀착 클로즈업보다 넓고 내려다보는 각도도 약하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 다른 인물을 전경에 추가해 어깨너머 구도로 바꾸었으며, 앰버도 괴롭게 들이마시는 순간보다 편안히 잠든 모습에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 눈을 꽉 감고 얼굴을 위쪽과 화면 오른쪽으로 향한다. 특정 대상을 바라보지 않아 눈을 감으라는 지시와 맞는다. 카메라는 침대 옆에서 상체를 비스듬히 보지만, 얼굴 바로 위에서 대각선 아래로 내려다보는 시점보다는 낮다. 얼굴은 왼쪽 위, 가슴은 아래 중앙에 놓인다.",
        "built_space": "침대 하나의 연푸른 침구와 머리 받침이 보이며, 뒤에는 모니터 한 대와 하부 장치, 창 하나, 수납대 하나가 있다. 이전 장면의 병실 재질과 야간 창밖 분위기는 대체로 이어진다. 중복된 설비나 불가능한 반사는 없다. 다만 침구가 좁은 테두리에 그치지 않고 넓게 보이며, 허리와 병실 배경까지 포함해 요구한 상체 클로즈업보다 범위가 크다.",
        "entities": "금발의 어린 여자아이 한 명만 보이고, 창백한 피부와 이마·볼의 땀, 찡그린 눈썹과 벌어진 입이 확인된다. 얼굴은 참조보다 다소 성숙하고 윤곽이 달라 정확한 동일인 재현은 약하다. 기름때 묻은 카키 작업복, 가죽 공구 벨트, 복합 필터형 방진 마스크가 있다. 마스크는 참조처럼 목과 가슴에 내려져 있으며 세부 형태는 다르다. 눈 크기는 감긴 상태라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "머리와 등은 경사진 침대와 베개에 받쳐져 있고, 보이는 팔과 손도 몸 옆 침대에 놓여 있다. 마스크는 목 끈에 연결된 채 가슴에 얹혀 있으며 공구는 허리 벨트가 지지한다. 떠 있는 신체나 물체는 없다. 벌어진 입과 긴장한 얼굴, 도드라진 가슴 윤곽은 힘겹게 숨을 들이마시는 순간으로 읽히지만 흉곽 팽창 자체는 정지 화면에서 제한적으로만 확인된다."
       },
       {
        "label": "B",
        "direction": "앰버는 눈을 감고 얼굴을 위로 향하지만 눈썹과 입 주변이 비교적 이완되어 있다. 오른쪽 전경의 추가 인물은 앰버 쪽으로 고개를 숙인다. 카메라가 그 인물의 어깨 너머에서 앰버를 관찰하는 구도가 되어, 앰버의 어깨 옆과 얼굴 위에 직접 놓인 지정 시점과 다르다.",
        "built_space": "침대 하나, 큰 베개 하나, 오른쪽 난간 하나와 머리판 일부가 보인다. 배경에는 모니터 한 대, 창 하나, 수납대 하나가 있어 병실의 기본 맥락은 이어진다. 그러나 넓은 침구와 배경, 전경 인물까지 들어와 침대를 좁은 지지면 테두리로만 남기라는 구성을 벗어난다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "금발과 어린 얼굴, 카키색 작업복, 목에 내려진 정교한 방진 마스크는 앰버 참조와 비교적 가깝다. 이마에는 땀이 있으나 괴로운 표정보다는 잠든 표정이다. 가슴 아래는 푸른 담요로 덮였고, 허리 벨트는 프레임에서 확인되지 않아 누락으로 평가하지 않는다. 오른쪽에는 검은 머리와 어두운 옷의 다른 인물이 명백히 추가되어, 앰버만 보여야 한다는 조건을 위반한다.",
        "hard_violations": [
         "앰버만 등장해야 하는 장면에 다른 인물의 머리와 어깨·상체를 오른쪽 전경에 추가했다."
        ],
        "physics": "앰버의 머리는 베개에, 등과 팔은 침대에 지지된다. 마스크는 목 끈과 가슴에 받쳐져 있고 담요는 몸 위에 자연스럽게 놓여 있다. 추가 인물의 하체는 프레임 밖이지만 공중에 떠 있다는 증거는 없다. 물리적으로 불가능한 자세는 없으나, 닫힌 입과 이완된 표정 및 담요에 가려진 흉곽 때문에 요구된 힘겨운 들숨과 가슴 팽창은 읽히지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "식은땀과 고통스러운 들숨, 침대에 지지된 상체를 구현했지만, 허리와 병실 배경까지 보여 지정된 밀착 클로즈업보다 넓고 내려다보는 각도도 약하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 다른 인물을 전경에 추가해 어깨너머 구도로 바꾸었으며, 앰버도 괴롭게 들이마시는 순간보다 편안히 잠든 모습에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 눈을 꽉 감고 얼굴을 위쪽과 화면 오른쪽으로 향한다. 특정 대상을 바라보지 않아 눈을 감으라는 지시와 맞는다. 카메라는 침대 옆에서 상체를 비스듬히 보지만, 얼굴 바로 위에서 대각선 아래로 내려다보는 시점보다는 낮다. 얼굴은 왼쪽 위, 가슴은 아래 중앙에 놓인다.",
        "built_space": "침대 하나의 연푸른 침구와 머리 받침이 보이며, 뒤에는 모니터 한 대와 하부 장치, 창 하나, 수납대 하나가 있다. 이전 장면의 병실 재질과 야간 창밖 분위기는 대체로 이어진다. 중복된 설비나 불가능한 반사는 없다. 다만 침구가 좁은 테두리에 그치지 않고 넓게 보이며, 허리와 병실 배경까지 포함해 요구한 상체 클로즈업보다 범위가 크다.",
        "entities": "금발의 어린 여자아이 한 명만 보이고, 창백한 피부와 이마·볼의 땀, 찡그린 눈썹과 벌어진 입이 확인된다. 얼굴은 참조보다 다소 성숙하고 윤곽이 달라 정확한 동일인 재현은 약하다. 기름때 묻은 카키 작업복, 가죽 공구 벨트, 복합 필터형 방진 마스크가 있다. 마스크는 참조처럼 목과 가슴에 내려져 있으며 세부 형태는 다르다. 눈 크기는 감긴 상태라 확인할 수 없다.",
        "hard_violations": [],
        "physics": "머리와 등은 경사진 침대와 베개에 받쳐져 있고, 보이는 팔과 손도 몸 옆 침대에 놓여 있다. 마스크는 목 끈에 연결된 채 가슴에 얹혀 있으며 공구는 허리 벨트가 지지한다. 떠 있는 신체나 물체는 없다. 벌어진 입과 긴장한 얼굴, 도드라진 가슴 윤곽은 힘겹게 숨을 들이마시는 순간으로 읽히지만 흉곽 팽창 자체는 정지 화면에서 제한적으로만 확인된다."
       },
       {
        "label": "A",
        "direction": "앰버는 눈을 감고 얼굴을 위로 향하지만 눈썹과 입 주변이 비교적 이완되어 있다. 오른쪽 전경의 추가 인물은 앰버 쪽으로 고개를 숙인다. 카메라가 그 인물의 어깨 너머에서 앰버를 관찰하는 구도가 되어, 앰버의 어깨 옆과 얼굴 위에 직접 놓인 지정 시점과 다르다.",
        "built_space": "침대 하나, 큰 베개 하나, 오른쪽 난간 하나와 머리판 일부가 보인다. 배경에는 모니터 한 대, 창 하나, 수납대 하나가 있어 병실의 기본 맥락은 이어진다. 그러나 넓은 침구와 배경, 전경 인물까지 들어와 침대를 좁은 지지면 테두리로만 남기라는 구성을 벗어난다. 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "금발과 어린 얼굴, 카키색 작업복, 목에 내려진 정교한 방진 마스크는 앰버 참조와 비교적 가깝다. 이마에는 땀이 있으나 괴로운 표정보다는 잠든 표정이다. 가슴 아래는 푸른 담요로 덮였고, 허리 벨트는 프레임에서 확인되지 않아 누락으로 평가하지 않는다. 오른쪽에는 검은 머리와 어두운 옷의 다른 인물이 명백히 추가되어, 앰버만 보여야 한다는 조건을 위반한다.",
        "hard_violations": [
         "앰버만 등장해야 하는 장면에 다른 인물의 머리와 어깨·상체를 오른쪽 전경에 추가했다."
        ],
        "physics": "앰버의 머리는 베개에, 등과 팔은 침대에 지지된다. 마스크는 목 끈과 가슴에 받쳐져 있고 담요는 몸 위에 자연스럽게 놓여 있다. 추가 인물의 하체는 프레임 밖이지만 공중에 떠 있다는 증거는 없다. 물리적으로 불가능한 자세는 없으나, 닫힌 입과 이완된 표정 및 담요에 가려진 흉곽 때문에 요구된 힘겨운 들숨과 가슴 팽창은 읽히지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 7,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "식은땀과 고통스러운 들숨, 침대에 지지된 상체를 구현했지만, 허리와 병실 배경까지 보여 지정된 밀착 클로즈업보다 넓고 내려다보는 각도도 약하다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "금지된 다른 인물을 전경에 추가해 어깨너머 구도로 바꾸었으며, 앰버도 괴롭게 들이마시는 순간보다 편안히 잠든 모습에 가깝다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S73sh6_sel.png",
    "asset_id": "1d501007-8693-4a31-8f31-b4f10295c991",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dbb-44f4-7688-9e6a-74ea79118d2d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh6"
  }
 },
 "S76sh7::signage": {
  "fp": "96537650b50a8823",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S76sh7": {
  "input_fingerprint": "a0a00b0b6cec11bc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 백의의 남자가 쿠마를 향해 세 손가락을 펼쳐 보인 채 멈춘 근접 찰나.\n\nLOCATION (lock): Beside the patient's bed in the harbor-town clinic's treatment room, lit for nighttime medical care. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly approach just behind and outside 쿠마's shoulder, with the lens slightly below the doctor's eye line and the sightline passing beside 쿠마 rather than through his silhouette. Keep 쿠마's shoulder and a sliver of his turned head at the left edge, while 백의의 남자's three-quarter face occupies the upper center and his three extended fingers remain unobstructed beside his chest at lower right. The doctor addresses 쿠마 just outside the crop, while 쿠마 inclines his head toward the doctor's raised hand with his gaze lowered; increased proximity alone supplies the emphasis on the deadline.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the neutral, subdued interior illumination, using controlled contrast to distinguish all three fingers while retaining the doctor's sober expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clinic bed, bedding, nearby interior surfaces, and nighttime lighting from the reference. Exclude aircraft, hangar equipment, and objects from the harbor restaurant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged chest and body remain unrepaired, and B-200's recovered component remains in his possession. 쿠마: He remains standing in the clinic.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 백의의 남자가 쿠마를 향해 세 손가락을 펼쳐 보인 채 멈춘 근접 찰나.\n\nLOCATION (lock): Beside the patient's bed in the harbor-town clinic's treatment room, lit for nighttime medical care. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly approach just behind and outside 쿠마's shoulder, with the lens slightly below the doctor's eye line and the sightline passing beside 쿠마 rather than through his silhouette. Keep 쿠마's shoulder and a sliver of his turned head at the left edge, while 백의의 남자's three-quarter face occupies the upper center and his three extended fingers remain unobstructed beside his chest at lower right. The doctor addresses 쿠마 just outside the crop, while 쿠마 inclines his head toward the doctor's raised hand with his gaze lowered; increased proximity alone supplies the emphasis on the deadline.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the neutral, subdued interior illumination, using controlled contrast to distinguish all three fingers while retaining the doctor's sober expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clinic bed, bedding, nearby interior surfaces, and nighttime lighting from the reference. Exclude aircraft, hangar equipment, and objects from the harbor restaurant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged chest and body remain unrepaired, and B-200's recovered component remains in his possession. 쿠마: He remains standing in the clinic.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 백의의 남자가 쿠마를 향해 세 손가락을 펼쳐 보인 채 멈춘 근접 찰나.\n\nLOCATION (lock): Beside the patient's bed in the harbor-town clinic's treatment room, lit for nighttime medical care. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly approach just behind and outside 쿠마's shoulder, with the lens slightly below the doctor's eye line and the sightline passing beside 쿠마 rather than through his silhouette. Keep 쿠마's shoulder and a sliver of his turned head at the left edge, while 백의의 남자's three-quarter face occupies the upper center and his three extended fingers remain unobstructed beside his chest at lower right. The doctor addresses 쿠마 just outside the crop, while 쿠마 inclines his head toward the doctor's raised hand with his gaze lowered; increased proximity alone supplies the emphasis on the deadline.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the neutral, subdued interior illumination, using controlled contrast to distinguish all three fingers while retaining the doctor's sober expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clinic bed, bedding, nearby interior surfaces, and nighttime lighting from the reference. Exclude aircraft, hangar equipment, and objects from the harbor restaurant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged chest and body remain unrepaired, and B-200's recovered component remains in his possession. 쿠마: He remains standing in the clinic.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 쿠마 (중국계 혼혈 남성, 20대의 얼굴, 짧은 짙은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 남자가 왼쪽 전경의 인물에게 시선을 향하고 세 손가락을 펴고 있음.",
    "built_space": "야간 병실 배경, 심박수 모니터와 창문 등 기준 이미지의 환경 요소가 올바르게 배치됨.",
    "entities": "프롬프트 지시와 다르게 왼쪽 전경에 쿠마 대신 앰버가 위치하고, 중앙의 의사에게 쿠마의 얼굴이 적용됨.",
    "hard_violations": [
     "[gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치",
     "[gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치",
     "[gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
    ],
    "physics": "인물의 자세는 안정적이나 손가락과 손바닥의 구조가 인체공학적으로 부자연스러움."
   },
   {
    "label": "B",
    "direction": "중앙의 남자가 왼쪽 전경 인물을 내려다보며 세 손가락을 명확히 펼치고 있음.",
    "built_space": "기준 이미지의 야간 병실 환경, 조명, 구조물들이 일치하게 구현됨.",
    "entities": "왼쪽 전경에 쿠마 대신 앰버가, 중앙 의사로 쿠마가 잘못 배치됨.",
    "hard_violations": [
     "[gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치",
     "[gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치",
     "[gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
    ],
    "physics": "손 제스처와 인물들의 자세가 중력에 맞게 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 후보 모두 프롬프트가 지정한 인물 배치를 완전히 어겼으나, B의 손가락 해부학적 묘사가 더 정확함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 인물 배치를 어긴 치명적 오류가 있으며, 편 손가락의 형태도 어색하게 뭉개짐."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 남자가 왼쪽 전경의 인물에게 시선을 향하고 세 손가락을 펴고 있음.",
        "built_space": "야간 병실 배경, 심박수 모니터와 창문 등 기준 이미지의 환경 요소가 올바르게 배치됨.",
        "entities": "프롬프트 지시와 다르게 왼쪽 전경에 쿠마 대신 앰버가 위치하고, 중앙의 의사에게 쿠마의 얼굴이 적용됨.",
        "hard_violations": [
         "왼쪽 전경에 쿠마 대신 앰버 배치",
         "중앙의 백의의 남자 자리에 쿠마 배치"
        ],
        "physics": "인물의 자세는 안정적이나 손가락과 손바닥의 구조가 인체공학적으로 부자연스러움."
       },
       {
        "label": "B",
        "direction": "중앙의 남자가 왼쪽 전경 인물을 내려다보며 세 손가락을 명확히 펼치고 있음.",
        "built_space": "기준 이미지의 야간 병실 환경, 조명, 구조물들이 일치하게 구현됨.",
        "entities": "왼쪽 전경에 쿠마 대신 앰버가, 중앙 의사로 쿠마가 잘못 배치됨.",
        "hard_violations": [
         "왼쪽 전경에 쿠마 대신 앰버 배치",
         "중앙의 백의의 남자 자리에 쿠마 배치"
        ],
        "physics": "손 제스처와 인물들의 자세가 중력에 맞게 자연스럽게 지탱됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 후보 모두 프롬프트가 지정한 인물 배치를 완전히 어겼으나, B의 손가락 해부학적 묘사가 더 정확함."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 인물 배치를 어긴 치명적 오류가 있으며, 편 손가락의 형태도 어색하게 뭉개짐."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 남자가 왼쪽 전경의 인물에게 시선을 향하고 세 손가락을 펴고 있음.",
        "built_space": "야간 병실 배경, 심박수 모니터와 창문 등 기준 이미지의 환경 요소가 올바르게 배치됨.",
        "entities": "프롬프트 지시와 다르게 왼쪽 전경에 쿠마 대신 앰버가 위치하고, 중앙의 의사에게 쿠마의 얼굴이 적용됨.",
        "hard_violations": [
         "왼쪽 전경에 쿠마 대신 앰버 배치",
         "중앙의 백의의 남자 자리에 쿠마 배치"
        ],
        "physics": "인물의 자세는 안정적이나 손가락과 손바닥의 구조가 인체공학적으로 부자연스러움."
       },
       {
        "label": "B",
        "direction": "중앙의 남자가 왼쪽 전경 인물을 내려다보며 세 손가락을 명확히 펼치고 있음.",
        "built_space": "기준 이미지의 야간 병실 환경, 조명, 구조물들이 일치하게 구현됨.",
        "entities": "왼쪽 전경에 쿠마 대신 앰버가, 중앙 의사로 쿠마가 잘못 배치됨.",
        "hard_violations": [
         "왼쪽 전경에 쿠마 대신 앰버 배치",
         "중앙의 백의의 남자 자리에 쿠마 배치"
        ],
        "physics": "손 제스처와 인물들의 자세가 중력에 맞게 자연스럽게 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "근접 구도와 세 손가락은 구현했지만, 서 있는 쿠마 대신 침대의 앰버를 대화 상대로 넣었고 손바닥도 상대보다 카메라를 향한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "쿠마 대신 앰버를 배치한 중대한 오류는 같지만, 상대에게 향한 손과 상대의 내려간 시선은 지정된 동작 관계에 더 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "백의 남자는 왼쪽의 금발 여자아이 얼굴을 바라본다. 세 손가락은 위로 펴져 있지만 손바닥은 여자아이보다 카메라 쪽을 향한다. 여자아이는 남자의 얼굴 쪽으로 고개를 돌렸고, 손을 향해 시선을 내리는 동작은 확인되지 않는다. 지시된 시선과 몸짓의 대상인 쿠마가 없다.",
        "built_space": "왼쪽 아래에 침대와 연한 청색 침구 일부, 왼쪽 뒤에 의료 모니터 한 대, 오른쪽 뒤에 야간 창 하나와 작업대 일부가 보인다. 회색 벽과 밤 조명은 참고 진료실에 부합하고 설비 중복이나 불가능한 반사는 없다. 다만 카메라는 서 있는 쿠마의 어깨 뒤가 아니라 침대에 기대 있는 여자아이 뒤에 있다.",
        "entities": "두 사람이 보인다. 중앙에는 짧은 검은 머리의 젊은 동아시아계 남자가 흰 가운을 입고 있고, 얼굴은 쿠마 참고와 닮았다. 왼쪽에는 금발 어린 여자아이가 참고의 앰버처럼 갈색 옷과 목의 금속 호흡 장비를 착용하고 있다. 요청된 쿠마의 짙은색 상의와 짧은 검은 머리의 전경 인물은 없다. 세 손가락은 분명하게 구별된다. 찰리와 회수 부품은 근접 화면 밖이므로 누락 자체는 감점 사항이 아니다.",
        "hard_violations": [
         "이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
        ],
        "physics": "여자아이의 등과 머리 뒤는 경사진 침구로 받쳐져 있다. 남자의 든 손은 손목과 가운 소매로 이어지고, 세 손가락을 펴고 나머지를 접은 자세는 가능하다. 목의 호흡 장비는 끈과 가슴에, 청진기는 남자의 목과 가운에 지지된다. 하체는 화면 밖이며 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "백의 남자는 왼쪽 여자아이의 얼굴을 내려다본다. 세 손가락을 편 손은 카메라에 손등을 보이고 손바닥을 여자아이 쪽으로 향해, 상대에게 숫자를 보이는 방향이 자연스럽다. 여자아이는 고개와 눈을 아래로 내려 손 쪽을 보는 동작에 가깝다. 그러나 실제 수신자는 쿠마가 아니라 앰버다.",
        "built_space": "왼쪽 아래에는 경사진 침대와 청색 침구, 중앙 왼쪽 뒤에는 의료 모니터 한 대와 그 아래 장치, 오른쪽 뒤에는 야간 창 하나와 작업대 일부가 보인다. 참고의 재질과 야간 진료실 분위기는 유지하며 설비의 명백한 중복은 없다. 백의 남자의 얼굴은 상단 중앙 부근, 손은 오른쪽 아래에 있지만 전경 여자아이의 머리와 몸통이 크게 들어와 쿠마의 어깨와 머리 일부만 남기는 지정 구도와 다르다.",
        "entities": "백의의 젊은 동아시아계 남자와 금발 여자아이 두 명이 보인다. 남자는 쿠마 참고를 닮은 얼굴에 흰 가운과 청진기를 착용했고, 여자아이는 앰버의 금발과 이전 장면의 갈색 옷 및 목의 호흡 장비를 지닌다. 따라서 백의 남자와 별도로 있어야 하는 쿠마의 정체성과 복장이 구현되지 않았다. 펴진 손가락은 세 개이며 서로 가려지지 않는다. 찰리와 부품은 이 근접 구도에서 보일 필요가 없다.",
        "hard_violations": [
         "이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
        ],
        "physics": "여자아이의 몸은 경사진 침대와 침구에 기대어 지지된다. 남자의 손은 손목과 소매를 통해 팔에 연결되고, 손가락 세 개를 편 동작도 해부학적으로 가능하다. 호흡 장비는 목의 끈과 상체에 받쳐져 있고 청진기는 목에 걸려 있다. 보이는 범위에서 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "근접 구도와 세 손가락은 구현했지만, 서 있는 쿠마 대신 침대의 앰버를 대화 상대로 넣었고 손바닥도 상대보다 카메라를 향한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "쿠마 대신 앰버를 배치한 중대한 오류는 같지만, 상대에게 향한 손과 상대의 내려간 시선은 지정된 동작 관계에 더 가깝다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "백의 남자는 왼쪽의 금발 여자아이 얼굴을 바라본다. 세 손가락은 위로 펴져 있지만 손바닥은 여자아이보다 카메라 쪽을 향한다. 여자아이는 남자의 얼굴 쪽으로 고개를 돌렸고, 손을 향해 시선을 내리는 동작은 확인되지 않는다. 지시된 시선과 몸짓의 대상인 쿠마가 없다.",
        "built_space": "왼쪽 아래에 침대와 연한 청색 침구 일부, 왼쪽 뒤에 의료 모니터 한 대, 오른쪽 뒤에 야간 창 하나와 작업대 일부가 보인다. 회색 벽과 밤 조명은 참고 진료실에 부합하고 설비 중복이나 불가능한 반사는 없다. 다만 카메라는 서 있는 쿠마의 어깨 뒤가 아니라 침대에 기대 있는 여자아이 뒤에 있다.",
        "entities": "두 사람이 보인다. 중앙에는 짧은 검은 머리의 젊은 동아시아계 남자가 흰 가운을 입고 있고, 얼굴은 쿠마 참고와 닮았다. 왼쪽에는 금발 어린 여자아이가 참고의 앰버처럼 갈색 옷과 목의 금속 호흡 장비를 착용하고 있다. 요청된 쿠마의 짙은색 상의와 짧은 검은 머리의 전경 인물은 없다. 세 손가락은 분명하게 구별된다. 찰리와 회수 부품은 근접 화면 밖이므로 누락 자체는 감점 사항이 아니다.",
        "hard_violations": [
         "이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
        ],
        "physics": "여자아이의 등과 머리 뒤는 경사진 침구로 받쳐져 있다. 남자의 든 손은 손목과 가운 소매로 이어지고, 세 손가락을 펴고 나머지를 접은 자세는 가능하다. 목의 호흡 장비는 끈과 가슴에, 청진기는 남자의 목과 가운에 지지된다. 하체는 화면 밖이며 공중에 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "백의 남자는 왼쪽 여자아이의 얼굴을 내려다본다. 세 손가락을 편 손은 카메라에 손등을 보이고 손바닥을 여자아이 쪽으로 향해, 상대에게 숫자를 보이는 방향이 자연스럽다. 여자아이는 고개와 눈을 아래로 내려 손 쪽을 보는 동작에 가깝다. 그러나 실제 수신자는 쿠마가 아니라 앰버다.",
        "built_space": "왼쪽 아래에는 경사진 침대와 청색 침구, 중앙 왼쪽 뒤에는 의료 모니터 한 대와 그 아래 장치, 오른쪽 뒤에는 야간 창 하나와 작업대 일부가 보인다. 참고의 재질과 야간 진료실 분위기는 유지하며 설비의 명백한 중복은 없다. 백의 남자의 얼굴은 상단 중앙 부근, 손은 오른쪽 아래에 있지만 전경 여자아이의 머리와 몸통이 크게 들어와 쿠마의 어깨와 머리 일부만 남기는 지정 구도와 다르다.",
        "entities": "백의의 젊은 동아시아계 남자와 금발 여자아이 두 명이 보인다. 남자는 쿠마 참고를 닮은 얼굴에 흰 가운과 청진기를 착용했고, 여자아이는 앰버의 금발과 이전 장면의 갈색 옷 및 목의 호흡 장비를 지닌다. 따라서 백의 남자와 별도로 있어야 하는 쿠마의 정체성과 복장이 구현되지 않았다. 펴진 손가락은 세 개이며 서로 가려지지 않는다. 찰리와 부품은 이 근접 구도에서 보일 필요가 없다.",
        "hard_violations": [
         "이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
        ],
        "physics": "여자아이의 몸은 경사진 침대와 침구에 기대어 지지된다. 남자의 손은 손목과 소매를 통해 팔에 연결되고, 손가락 세 개를 편 동작도 해부학적으로 가능하다. 호흡 장비는 목의 끈과 상체에 받쳐져 있고 청진기는 목에 걸려 있다. 보이는 범위에서 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치",
     "[gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치",
     "[gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
    ],
    "B": [
     "[gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치",
     "[gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치",
     "[gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1417,
   "A": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "두 후보 모두 프롬프트가 지정한 인물 배치를 완전히 어겼으나, B의 손가락 해부학적 묘사가 더 정확함.  ★위반: [gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치 / [gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치 / [gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "지정된 인물 배치를 어긴 치명적 오류가 있으며, 편 손가락의 형태도 어색하게 뭉개짐.  ★위반: [gemini-pro] 왼쪽 전경에 쿠마 대신 앰버 배치 / [gemini-pro] 중앙의 백의의 남자 자리에 쿠마 배치 / [gpt-high] 이 샷에 보이도록 지정되지 않은 앰버를 전경 대화 상대로 추가하고, 그 자리에 있어야 할 서 있는 쿠마를 대체했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S76sh1_sel.png",
    "asset_id": "de6de452-e752-46f6-8bf9-e9f82f1dfce2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1305657>",
    "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 쿠마: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1128383>",
    "asset_id": "ac0e1831-ec53-4d61-bd22-48dca7072302",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dcf-2303-7dff-8b74-8f79cc862120",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S76sh1"
  },
  "staged_characters_added": [
   "C13"
  ]
 },
 "S76sh11::signage": {
  "fp": "e692923dd39405c7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S76sh11": {
  "input_fingerprint": "8ad0747efc3665ff",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신만만한 표정의 이모티콘을 띄운 채 한쪽 주먹을 불끈 쥔 찰리의 정면 상체.\n\nLOCATION (lock): In the standing space beside the clinic bed, among the gathered visitors under nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral move on the group's outer side near 이현우 and 쿠마, with the lens below 찰리's shoulder level and tilted slightly upward in a medium three-quarter view. Place 찰리's upper body just right of center, keeping the confident facial emoticon readable and the newly clenched fist beside the torso rather than overlapping the display. His body and attention face 이현우 and 쿠마 beyond the left edge, not the lens; the transfer of attention to his offer carries the beat without an additional dramatic lighting change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained neutral illumination, preserving precise robot contours and readable facial graphics without treating the display as an invented room light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged body and exposed chest interior remain unrepaired despite his readiness to help, and he retains B-200's chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신만만한 표정의 이모티콘을 띄운 채 한쪽 주먹을 불끈 쥔 찰리의 정면 상체.\n\nLOCATION (lock): In the standing space beside the clinic bed, among the gathered visitors under nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral move on the group's outer side near 이현우 and 쿠마, with the lens below 찰리's shoulder level and tilted slightly upward in a medium three-quarter view. Place 찰리's upper body just right of center, keeping the confident facial emoticon readable and the newly clenched fist beside the torso rather than overlapping the display. His body and attention face 이현우 and 쿠마 beyond the left edge, not the lens; the transfer of attention to his offer carries the beat without an additional dramatic lighting change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained neutral illumination, preserving precise robot contours and readable facial graphics without treating the display as an invented room light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged body and exposed chest interior remain unrepaired despite his readiness to help, and he retains B-200's chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자신만만한 표정의 이모티콘을 띄운 채 한쪽 주먹을 불끈 쥔 찰리의 정면 상체.\n\nLOCATION (lock): In the standing space beside the clinic bed, among the gathered visitors under nighttime room lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral move on the group's outer side near 이현우 and 쿠마, with the lens below 찰리's shoulder level and tilted slightly upward in a medium three-quarter view. Place 찰리's upper body just right of center, keeping the confident facial emoticon readable and the newly clenched fist beside the torso rather than overlapping the display. His body and attention face 이현우 and 쿠마 beyond the left edge, not the lens; the transfer of attention to his offer carries the beat without an additional dramatic lighting change.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the health center's restrained neutral illumination, preserving precise robot contours and readable facial graphics without treating the display as an invented room light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nighttime clinic bed remains in use. Charlie's damaged body and exposed chest interior remain unrepaired despite his readiness to help, and he retains B-200's chest component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선과 몸체 방향이 정확히 프레임 왼쪽 바깥(이현우와 쿠마가 위치한 곳)을 향하고 있으며, 쥐어진 오른손 주먹은 몸통 옆을 겨냥합니다.",
    "built_space": "야간 병실 구조가 레퍼런스와 일치합니다. 화면 좌측 후경에 환자용 모니터와 병상 일부가 배치되었고, 약간 아래에서 위를 올려다보는 앵글이 안정적으로 적용되었습니다.",
    "entities": "찰리의 장갑판 질감과 푸른색 원자로 심볼이 정확히 묘사되었습니다. 특히 마스크의 입 선이 위로 휘어진 곡선 형태로 표현되어 '자신만만한 표정'의 이모티콘을 명확하게 보여줍니다.",
    "hard_violations": [
     "[gemini-pro] 숏 텍스트에 명시되지 않은 인물(화면 왼쪽 전경의 검은 실루엣)이 프레임 내에 침범하여 'never add a person the shot text does not show' 제약을 위반함.",
     "[gpt-high] 화면 왼쪽 전경에 찰리가 아닌 사람의 몸 일부를 추가해, 찰리만 보이도록 한 인물 제한을 위반했다."
    ],
    "physics": "찰리는 지면에 두 발로 단단히 서서 체중을 지탱하고 있으며, 들어 올린 팔과 주먹의 자세 또한 물리적으로 자연스럽게 유지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "찰리의 몸과 시선이 화면 좌측을 향하며, 오른손 주먹을 가슴 옆으로 들어 올린 상태입니다.",
    "built_space": "레퍼런스와 동일한 조명 및 병실 배경이 구현되었습니다. 모니터와 창문, 병상 일부가 정확한 위치에 렌더링되었습니다.",
    "entities": "찰리의 로봇 외형 디테일은 레퍼런스에 부합하나, 마스크 입 부분이 표정 변화 없는 일자 형태로 묘사되어 '자신만만한 이모티콘'이라는 핵심 지시를 누락했습니다.",
    "hard_violations": [
     "[gemini-pro] 화면 왼쪽과 오른쪽 양쪽 전경에 숏 텍스트에 없는 인물의 신체 일부(어깨 실루엣)가 등장하여 엄격한 인물 제한 규칙을 위반함.",
     "[gpt-high] 화면 왼쪽과 오른쪽 전경에 사람의 몸 일부를 추가했다. 이 숏에서 허용된 가시 인물은 찰리뿐이다."
    ],
    "physics": "찰리의 발이 바닥을 안정적으로 디디고 서 있으며, 팔을 들어 주먹을 쥔 동작에 필요한 관절 꺾임과 중력 지지가 올바르게 표현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "마스크 입 부분의 선을 곡선으로 처리해 '자신만만한 이모티콘' 표정을 훌륭히 구현했으나, 프롬프트에 명시되지 않은 타 인물의 실루엣이 노출되는 치명적인 규칙 위반이 존재합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 표정 변화 없이 레퍼런스의 일자 입 선을 그대로 유지했으며, 화면 양측 전경에 지시문이 금지한 타 인물의 실루엣을 배치한 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 몸체 방향이 정확히 프레임 왼쪽 바깥(이현우와 쿠마가 위치한 곳)을 향하고 있으며, 쥐어진 오른손 주먹은 몸통 옆을 겨냥합니다.",
        "built_space": "야간 병실 구조가 레퍼런스와 일치합니다. 화면 좌측 후경에 환자용 모니터와 병상 일부가 배치되었고, 약간 아래에서 위를 올려다보는 앵글이 안정적으로 적용되었습니다.",
        "entities": "찰리의 장갑판 질감과 푸른색 원자로 심볼이 정확히 묘사되었습니다. 특히 마스크의 입 선이 위로 휘어진 곡선 형태로 표현되어 '자신만만한 표정'의 이모티콘을 명확하게 보여줍니다.",
        "hard_violations": [
         "숏 텍스트에 명시되지 않은 인물(화면 왼쪽 전경의 검은 실루엣)이 프레임 내에 침범하여 'never add a person the shot text does not show' 제약을 위반함."
        ],
        "physics": "찰리는 지면에 두 발로 단단히 서서 체중을 지탱하고 있으며, 들어 올린 팔과 주먹의 자세 또한 물리적으로 자연스럽게 유지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리의 몸과 시선이 화면 좌측을 향하며, 오른손 주먹을 가슴 옆으로 들어 올린 상태입니다.",
        "built_space": "레퍼런스와 동일한 조명 및 병실 배경이 구현되었습니다. 모니터와 창문, 병상 일부가 정확한 위치에 렌더링되었습니다.",
        "entities": "찰리의 로봇 외형 디테일은 레퍼런스에 부합하나, 마스크 입 부분이 표정 변화 없는 일자 형태로 묘사되어 '자신만만한 이모티콘'이라는 핵심 지시를 누락했습니다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 양쪽 전경에 숏 텍스트에 없는 인물의 신체 일부(어깨 실루엣)가 등장하여 엄격한 인물 제한 규칙을 위반함."
        ],
        "physics": "찰리의 발이 바닥을 안정적으로 디디고 서 있으며, 팔을 들어 주먹을 쥔 동작에 필요한 관절 꺾임과 중력 지지가 올바르게 표현되었습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "마스크 입 부분의 선을 곡선으로 처리해 '자신만만한 이모티콘' 표정을 훌륭히 구현했으나, 프롬프트에 명시되지 않은 타 인물의 실루엣이 노출되는 치명적인 규칙 위반이 존재합니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 표정 변화 없이 레퍼런스의 일자 입 선을 그대로 유지했으며, 화면 양측 전경에 지시문이 금지한 타 인물의 실루엣을 배치한 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선과 몸체 방향이 정확히 프레임 왼쪽 바깥(이현우와 쿠마가 위치한 곳)을 향하고 있으며, 쥐어진 오른손 주먹은 몸통 옆을 겨냥합니다.",
        "built_space": "야간 병실 구조가 레퍼런스와 일치합니다. 화면 좌측 후경에 환자용 모니터와 병상 일부가 배치되었고, 약간 아래에서 위를 올려다보는 앵글이 안정적으로 적용되었습니다.",
        "entities": "찰리의 장갑판 질감과 푸른색 원자로 심볼이 정확히 묘사되었습니다. 특히 마스크의 입 선이 위로 휘어진 곡선 형태로 표현되어 '자신만만한 표정'의 이모티콘을 명확하게 보여줍니다.",
        "hard_violations": [
         "숏 텍스트에 명시되지 않은 인물(화면 왼쪽 전경의 검은 실루엣)이 프레임 내에 침범하여 'never add a person the shot text does not show' 제약을 위반함."
        ],
        "physics": "찰리는 지면에 두 발로 단단히 서서 체중을 지탱하고 있으며, 들어 올린 팔과 주먹의 자세 또한 물리적으로 자연스럽게 유지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리의 몸과 시선이 화면 좌측을 향하며, 오른손 주먹을 가슴 옆으로 들어 올린 상태입니다.",
        "built_space": "레퍼런스와 동일한 조명 및 병실 배경이 구현되었습니다. 모니터와 창문, 병상 일부가 정확한 위치에 렌더링되었습니다.",
        "entities": "찰리의 로봇 외형 디테일은 레퍼런스에 부합하나, 마스크 입 부분이 표정 변화 없는 일자 형태로 묘사되어 '자신만만한 이모티콘'이라는 핵심 지시를 누락했습니다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 양쪽 전경에 숏 텍스트에 없는 인물의 신체 일부(어깨 실루엣)가 등장하여 엄격한 인물 제한 규칙을 위반함."
        ],
        "physics": "찰리의 발이 바닥을 안정적으로 디디고 서 있으며, 팔을 들어 주먹을 쥔 동작에 필요한 관절 꺾임과 중력 지지가 올바르게 표현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "몸통 옆 주먹과 왼쪽을 향한 자세는 맞지만, 양쪽 전경의 추가 인물이 금지 조건을 어기고 손상된 가슴 내부도 보존하지 않았다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "낮은 카메라의 상체 삼사분면 구도와 자신만만한 미소는 더 충실하지만, 왼쪽 전경의 추가 인물과 복구된 듯 닫힌 가슴 때문에 적합한 최종 컷은 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴과 몸통은 화면 왼쪽 전경의 사람 쪽으로 조금 돌아가 있어 렌즈 정면 응시는 아니다. 이현우와 쿠마인지는 확인할 수 없다. 쥔 주먹은 화면 왼쪽의 몸통 옆에 올라와 얼굴 표시를 가리지 않는다.",
        "built_space": "왼쪽에 침대 일부 한 개, 그 뒤에 모니터와 하부 장치로 이루어진 의료 장비 한 세트, 오른쪽 뒤에 야경이 보이는 창이 있다. 찰리는 침대 옆 공간에 서 있다. 회색 벽과 침구, 중립적인 야간 조명은 장소 참조와 대체로 이어지며, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "샌드 베이지 장갑, 흰 마스크, 주황색 점 형태의 두 눈, 선 형태의 입, 푸른 원형 가슴 장치가 찰리 참조와 맞는다. 표정은 자신만만하다기보다 평온한 미소에 가깝다. 표면 마모는 있으나 가슴 장갑은 닫혀 있어 노출된 손상 내부가 유지되지 않았다. B-200의 가슴 부품 보유는 식별할 수 없다. 양쪽 전경에는 찰리 외의 검은 옷을 입은 사람 일부가 보이며 신원은 확인되지 않는다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 전경에 사람의 몸 일부를 추가했다. 이 숏에서 허용된 가시 인물은 찰리뿐이다."
        ],
        "physics": "들어 올린 주먹은 손목과 굽힌 팔꿈치, 어깨 관절로 몸통에 연결되어 있어 기계 팔의 동작으로 성립한다. 하체는 화면 아래로 이어지고 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만, 공중에 떠 있다는 징후는 없다. 가슴 장치는 몸체에 장착되어 있다."
       },
       {
        "label": "B",
        "direction": "찰리의 얼굴과 몸통은 화면 왼쪽 상대 쪽으로 돌아가 있고, 얼굴은 살짝 기울어져 있다. 렌즈보다는 왼쪽 전경의 사람을 향하는 것으로 읽히지만 상대의 신원은 확인되지 않는다. 쥔 주먹은 몸통 왼쪽 옆에 있으며 얼굴 표시와 겹치지 않는다.",
        "built_space": "왼쪽 아래에 침대 일부 한 개와 그 뒤 의료 모니터 한 세트, 뒤쪽에 야경이 보이는 창, 오른쪽에 수액 주머니가 달린 지지대 한 개가 보인다. 찰리는 침대 옆에 서 있다. 낮은 시점 때문에 천장 일부가 보이며, 어깨 아래에서 올려다보라는 지시에 A보다 가깝다. 수액 설비는 참조에서 확인되지 않지만 중복 설비나 불가능한 배치로 단정할 근거는 없다.",
        "entities": "찰리의 베이지색 각진 장갑, 육중한 팔, 흰 마스크, 두 주황색 눈과 웃는 입선, 푸른 가슴 장치는 참조 정체성을 유지한다. 기울어진 얼굴과 올라간 입꼬리가 자신만만한 감정에 A보다 가깝다. 다만 가슴은 장갑으로 닫혀 있어 손상 내부 노출 조건과 다르며 B-200의 부품 보유도 식별되지 않는다. 왼쪽 전경에 신원 불명의 검은 옷을 입은 사람 일부가 추가되어 있다.",
        "hard_violations": [
         "화면 왼쪽 전경에 찰리가 아닌 사람의 몸 일부를 추가해, 찰리만 보이도록 한 인물 제한을 위반했다."
        ],
        "physics": "주먹은 손목과 팔꿈치 관절에 연결되고 굽힌 팔은 어깨가 지지하므로 들어 올려 힘주어 쥐는 동작이 가능하다. 몸통과 골반은 이어져 있으며 발은 구도 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않고, 수액 주머니는 지지대에 매달려 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "몸통 옆 주먹과 왼쪽을 향한 자세는 맞지만, 양쪽 전경의 추가 인물이 금지 조건을 어기고 손상된 가슴 내부도 보존하지 않았다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "낮은 카메라의 상체 삼사분면 구도와 자신만만한 미소는 더 충실하지만, 왼쪽 전경의 추가 인물과 복구된 듯 닫힌 가슴 때문에 적합한 최종 컷은 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴과 몸통은 화면 왼쪽 전경의 사람 쪽으로 조금 돌아가 있어 렌즈 정면 응시는 아니다. 이현우와 쿠마인지는 확인할 수 없다. 쥔 주먹은 화면 왼쪽의 몸통 옆에 올라와 얼굴 표시를 가리지 않는다.",
        "built_space": "왼쪽에 침대 일부 한 개, 그 뒤에 모니터와 하부 장치로 이루어진 의료 장비 한 세트, 오른쪽 뒤에 야경이 보이는 창이 있다. 찰리는 침대 옆 공간에 서 있다. 회색 벽과 침구, 중립적인 야간 조명은 장소 참조와 대체로 이어지며, 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "샌드 베이지 장갑, 흰 마스크, 주황색 점 형태의 두 눈, 선 형태의 입, 푸른 원형 가슴 장치가 찰리 참조와 맞는다. 표정은 자신만만하다기보다 평온한 미소에 가깝다. 표면 마모는 있으나 가슴 장갑은 닫혀 있어 노출된 손상 내부가 유지되지 않았다. B-200의 가슴 부품 보유는 식별할 수 없다. 양쪽 전경에는 찰리 외의 검은 옷을 입은 사람 일부가 보이며 신원은 확인되지 않는다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 전경에 사람의 몸 일부를 추가했다. 이 숏에서 허용된 가시 인물은 찰리뿐이다."
        ],
        "physics": "들어 올린 주먹은 손목과 굽힌 팔꿈치, 어깨 관절로 몸통에 연결되어 있어 기계 팔의 동작으로 성립한다. 하체는 화면 아래로 이어지고 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만, 공중에 떠 있다는 징후는 없다. 가슴 장치는 몸체에 장착되어 있다."
       },
       {
        "label": "A",
        "direction": "찰리의 얼굴과 몸통은 화면 왼쪽 상대 쪽으로 돌아가 있고, 얼굴은 살짝 기울어져 있다. 렌즈보다는 왼쪽 전경의 사람을 향하는 것으로 읽히지만 상대의 신원은 확인되지 않는다. 쥔 주먹은 몸통 왼쪽 옆에 있으며 얼굴 표시와 겹치지 않는다.",
        "built_space": "왼쪽 아래에 침대 일부 한 개와 그 뒤 의료 모니터 한 세트, 뒤쪽에 야경이 보이는 창, 오른쪽에 수액 주머니가 달린 지지대 한 개가 보인다. 찰리는 침대 옆에 서 있다. 낮은 시점 때문에 천장 일부가 보이며, 어깨 아래에서 올려다보라는 지시에 A보다 가깝다. 수액 설비는 참조에서 확인되지 않지만 중복 설비나 불가능한 배치로 단정할 근거는 없다.",
        "entities": "찰리의 베이지색 각진 장갑, 육중한 팔, 흰 마스크, 두 주황색 눈과 웃는 입선, 푸른 가슴 장치는 참조 정체성을 유지한다. 기울어진 얼굴과 올라간 입꼬리가 자신만만한 감정에 A보다 가깝다. 다만 가슴은 장갑으로 닫혀 있어 손상 내부 노출 조건과 다르며 B-200의 부품 보유도 식별되지 않는다. 왼쪽 전경에 신원 불명의 검은 옷을 입은 사람 일부가 추가되어 있다.",
        "hard_violations": [
         "화면 왼쪽 전경에 찰리가 아닌 사람의 몸 일부를 추가해, 찰리만 보이도록 한 인물 제한을 위반했다."
        ],
        "physics": "주먹은 손목과 팔꿈치 관절에 연결되고 굽힌 팔은 어깨가 지지하므로 들어 올려 힘주어 쥐는 동작이 가능하다. 몸통과 골반은 이어져 있으며 발은 구도 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않고, 수액 주머니는 지지대에 매달려 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 숏 텍스트에 명시되지 않은 인물(화면 왼쪽 전경의 검은 실루엣)이 프레임 내에 침범하여 'never add a person the shot text does not show' 제약을 위반함.",
     "[gpt-high] 화면 왼쪽 전경에 찰리가 아닌 사람의 몸 일부를 추가해, 찰리만 보이도록 한 인물 제한을 위반했다."
    ],
    "B": [
     "[gemini-pro] 화면 왼쪽과 오른쪽 양쪽 전경에 숏 텍스트에 없는 인물의 신체 일부(어깨 실루엣)가 등장하여 엄격한 인물 제한 규칙을 위반함.",
     "[gpt-high] 화면 왼쪽과 오른쪽 전경에 사람의 몸 일부를 추가했다. 이 숏에서 허용된 가시 인물은 찰리뿐이다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "마스크 입 부분의 선을 곡선으로 처리해 '자신만만한 이모티콘' 표정을 훌륭히 구현했으나, 프롬프트에 명시되지 않은 타 인물의 실루엣이 노출되는 치명적인 규칙 위반이 존재합니다.  ★위반: [gemini-pro] 숏 텍스트에 명시되지 않은 인물(화면 왼쪽 전경의 검은 실루엣)이 프레임 내에 침범하여 'never add a person the shot text does not show' 제약을 위반함. / [gpt-high] 화면 왼쪽 전경에 찰리가 아닌 사람의 몸 일부를 추가해, 찰리만 보이도록 한 인물 제한을 위반했다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "요구된 표정 변화 없이 레퍼런스의 일자 입 선을 그대로 유지했으며, 화면 양측 전경에 지시문이 금지한 타 인물의 실루엣을 배치한 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 화면 왼쪽과 오른쪽 양쪽 전경에 숏 텍스트에 없는 인물의 신체 일부(어깨 실루엣)가 등장하여 엄격한 인물 제한 규칙을 위반함. / [gpt-high] 화면 왼쪽과 오른쪽 전경에 사람의 몸 일부를 추가했다. 이 숏에서 허용된 가시 인물은 찰리뿐이다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S76sh1_sel.png",
    "asset_id": "de6de452-e752-46f6-8bf9-e9f82f1dfce2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dd4-0c21-7c25-9203-995b9dd773f7",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S76sh1"
  }
 },
 "S77sh22::signage": {
  "fp": "34bdb3da9ebb5346",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S77sh22": {
  "input_fingerprint": "dcf13e2636635010",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 작동하는 비행기들 사이에서 찰리가 양손을 허리에 얹고 슈퍼히어로처럼 당당한 포즈를 취한 전신.\n\nLOCATION (lock): Among the newly running light aircraft inside the abandoned airfield's hangar. Overhead bulbs have been surging in brightness, with one bursting during the power-up. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the pullback from a low position, offset from 찰리's front and tilted gently upward, showing his complete hands-on-hips pose before anyone reaches him. Place him slightly left of center with headroom and clear ground beneath his feet, arranging operating aircraft in separate depth layers on either side and leaving the right side open for the approaching group, still outside the image. He angles his head toward those companions off-screen right, turning the success into a shared invitation rather than a pose for the lens.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Operating light aircraft (Engines newly started, emitting black exhaust) — Aircraft sides and oblique front portions flank 찰리 at different depths; used as Visible evidence of the repair, with no individual aircraft occupying more than two-fifths of the frame; Hangar interior (Occupied by the repaired aircraft) — The open interior extends behind the full-body figure; used as Preserves the collective scale of the achievement and space for the imminent embrace.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime illumination and controlled contrast to retain precise robot surfaces and the black exhaust from the newly operating aircraft.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several old aircraft now have running engines and are emitting dark exhaust inside the nighttime hangar; at least one light bulb has burst during the power surge. Charlie remains physically battle-damaged despite successfully supplying power, and he retains B-200's component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 작동하는 비행기들 사이에서 찰리가 양손을 허리에 얹고 슈퍼히어로처럼 당당한 포즈를 취한 전신.\n\nLOCATION (lock): Among the newly running light aircraft inside the abandoned airfield's hangar. Overhead bulbs have been surging in brightness, with one bursting during the power-up. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the pullback from a low position, offset from 찰리's front and tilted gently upward, showing his complete hands-on-hips pose before anyone reaches him. Place him slightly left of center with headroom and clear ground beneath his feet, arranging operating aircraft in separate depth layers on either side and leaving the right side open for the approaching group, still outside the image. He angles his head toward those companions off-screen right, turning the success into a shared invitation rather than a pose for the lens.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Operating light aircraft (Engines newly started, emitting black exhaust) — Aircraft sides and oblique front portions flank 찰리 at different depths; used as Visible evidence of the repair, with no individual aircraft occupying more than two-fifths of the frame; Hangar interior (Occupied by the repaired aircraft) — The open interior extends behind the full-body figure; used as Preserves the collective scale of the achievement and space for the imminent embrace.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime illumination and controlled contrast to retain precise robot surfaces and the black exhaust from the newly operating aircraft.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several old aircraft now have running engines and are emitting dark exhaust inside the nighttime hangar; at least one light bulb has burst during the power surge. Charlie remains physically battle-damaged despite successfully supplying power, and he retains B-200's component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 작동하는 비행기들 사이에서 찰리가 양손을 허리에 얹고 슈퍼히어로처럼 당당한 포즈를 취한 전신.\n\nLOCATION (lock): Among the newly running light aircraft inside the abandoned airfield's hangar. Overhead bulbs have been surging in brightness, with one bursting during the power-up. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the entrance to the pullback from a low position, offset from 찰리's front and tilted gently upward, showing his complete hands-on-hips pose before anyone reaches him. Place him slightly left of center with headroom and clear ground beneath his feet, arranging operating aircraft in separate depth layers on either side and leaving the right side open for the approaching group, still outside the image. He angles his head toward those companions off-screen right, turning the success into a shared invitation rather than a pose for the lens.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Operating light aircraft (Engines newly started, emitting black exhaust) — Aircraft sides and oblique front portions flank 찰리 at different depths; used as Visible evidence of the repair, with no individual aircraft occupying more than two-fifths of the frame; Hangar interior (Occupied by the repaired aircraft) — The open interior extends behind the full-body figure; used as Preserves the collective scale of the achievement and space for the imminent embrace.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime illumination and controlled contrast to retain precise robot surfaces and the black exhaust from the newly operating aircraft.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several old aircraft now have running engines and are emitting dark exhaust inside the nighttime hangar; at least one light bulb has burst during the power surge. Charlie remains physically battle-damaged despite successfully supplying power, and he retains B-200's component.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 고개는 화면 우측(프레임 밖)을 향하고 있으며, 카메라는 낮은 위치에서 살짝 올려다보는 구도를 취하고 있습니다.",
    "built_space": "레퍼런스와 일치하는 격납고 내부. 좌우와 뒤쪽에 여러 대의 경비행기가 배치되어 있으며, 우측 상단 천장에서 전구가 터지며 스파크가 튀고 있습니다.",
    "entities": "찰리(거대한 고릴라 비율, 모래색 장갑, 흰색 마스크, 푸른 원자로)의 외형이 레퍼런스와 매우 일치하며, 작동 중인 비행기에서 검은 배기가스가 발생하고 있습니다.",
    "hard_violations": [],
    "physics": "찰리는 젖은 격납고 바닥에 두 발을 딛고 안정적으로 서 있으며, 비행기 역시 바닥에 지지되어 있습니다."
   },
   {
    "label": "B",
    "direction": "찰리는 화면 우측으로 고개를 돌리고 있으며, 카메라는 낮은 위치에서 피사체를 올려다보고 있습니다.",
    "built_space": "레퍼런스와 동일한 구조의 격납고 안이며, 비행기들이 여러 깊이 층에 주차되어 있고 우측 천장에 스파크가 튀는 전구가 있습니다.",
    "entities": "찰리의 전체적인 형태와 마스크는 일치하나, 가슴 원자로의 내부 소용돌이 디테일이 다소 간략화되었습니다. 작동 중인 비행기와 배기가스가 보입니다.",
    "hard_violations": [],
    "physics": "찰리와 비행기들 모두 격납고 바닥에 물리적으로 올바르게 지지되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캐릭터의 디자인(특히 가슴 원자로의 디테일과 마스크)이 레퍼런스와 더 정확히 일치하며, 지시된 구도와 앵글을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 동작을 잘 따랐으나, 캐릭터의 가슴 원자로와 마스크 디테일이 A에 비해 다소 덜 정확합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 고개는 화면 우측(프레임 밖)을 향하고 있으며, 카메라는 낮은 위치에서 살짝 올려다보는 구도를 취하고 있습니다.",
        "built_space": "레퍼런스와 일치하는 격납고 내부. 좌우와 뒤쪽에 여러 대의 경비행기가 배치되어 있으며, 우측 상단 천장에서 전구가 터지며 스파크가 튀고 있습니다.",
        "entities": "찰리(거대한 고릴라 비율, 모래색 장갑, 흰색 마스크, 푸른 원자로)의 외형이 레퍼런스와 매우 일치하며, 작동 중인 비행기에서 검은 배기가스가 발생하고 있습니다.",
        "hard_violations": [],
        "physics": "찰리는 젖은 격납고 바닥에 두 발을 딛고 안정적으로 서 있으며, 비행기 역시 바닥에 지지되어 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 우측으로 고개를 돌리고 있으며, 카메라는 낮은 위치에서 피사체를 올려다보고 있습니다.",
        "built_space": "레퍼런스와 동일한 구조의 격납고 안이며, 비행기들이 여러 깊이 층에 주차되어 있고 우측 천장에 스파크가 튀는 전구가 있습니다.",
        "entities": "찰리의 전체적인 형태와 마스크는 일치하나, 가슴 원자로의 내부 소용돌이 디테일이 다소 간략화되었습니다. 작동 중인 비행기와 배기가스가 보입니다.",
        "hard_violations": [],
        "physics": "찰리와 비행기들 모두 격납고 바닥에 물리적으로 올바르게 지지되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캐릭터의 디자인(특히 가슴 원자로의 디테일과 마스크)이 레퍼런스와 더 정확히 일치하며, 지시된 구도와 앵글을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 구도와 동작을 잘 따랐으나, 캐릭터의 가슴 원자로와 마스크 디테일이 A에 비해 다소 덜 정확합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 고개는 화면 우측(프레임 밖)을 향하고 있으며, 카메라는 낮은 위치에서 살짝 올려다보는 구도를 취하고 있습니다.",
        "built_space": "레퍼런스와 일치하는 격납고 내부. 좌우와 뒤쪽에 여러 대의 경비행기가 배치되어 있으며, 우측 상단 천장에서 전구가 터지며 스파크가 튀고 있습니다.",
        "entities": "찰리(거대한 고릴라 비율, 모래색 장갑, 흰색 마스크, 푸른 원자로)의 외형이 레퍼런스와 매우 일치하며, 작동 중인 비행기에서 검은 배기가스가 발생하고 있습니다.",
        "hard_violations": [],
        "physics": "찰리는 젖은 격납고 바닥에 두 발을 딛고 안정적으로 서 있으며, 비행기 역시 바닥에 지지되어 있습니다."
       },
       {
        "label": "B",
        "direction": "찰리는 화면 우측으로 고개를 돌리고 있으며, 카메라는 낮은 위치에서 피사체를 올려다보고 있습니다.",
        "built_space": "레퍼런스와 동일한 구조의 격납고 안이며, 비행기들이 여러 깊이 층에 주차되어 있고 우측 천장에 스파크가 튀는 전구가 있습니다.",
        "entities": "찰리의 전체적인 형태와 마스크는 일치하나, 가슴 원자로의 내부 소용돌이 디테일이 다소 간략화되었습니다. 작동 중인 비행기와 배기가스가 보입니다.",
        "hard_violations": [],
        "physics": "찰리와 비행기들 모두 격납고 바닥에 물리적으로 올바르게 지지되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 사선 시점, 왼쪽의 전신 포즈, 화면 밖 오른쪽을 향한 얼굴과 열린 접근 공간이 지시에 더 충실하나, 참조보다 다리가 길고 B-200 부품 보유는 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "양손을 허리에 댄 전신과 가동 중인 항공기는 충실하지만, A보다 정면에 가까운 시점과 조금 더 큰 인물 비중으로 지정된 사선 와이드 구도가 약해지며 B-200 부품도 식별되지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 몸통은 카메라에 비스듬하고 얼굴은 화면 오른쪽을 향해 돌아가 있어, 렌즈가 아닌 화면 밖 동료들이 올 방향을 바라본다. 양팔의 팔꿈치는 바깥으로 벌어지고 두 손은 허리를 향한다. 좌우 전경 항공기의 기수는 각각 중앙 통로 쪽으로 비스듬히 향하며, 뒤쪽 항공기도 앞쪽 사선으로 놓여 있다.",
        "built_space": "중앙보다 왼쪽에 찰리 전신이 있고 머리 위와 발밑에 여백이 남는다. 오른쪽 바닥 통로는 비어 있다. 좌우 전경에 항공기 한 대씩, 후방에 적어도 두 대가 보여 서로 다른 깊이를 이룬다. 천장에는 점등된 등 네 개와 불꽃이 튀는 등 한 개가 식별된다. 골강판 벽, 박공지붕 철골, 양측 높은 창열, 벽가 선반과 작업대, 얼룩진 콘크리트 바닥은 장소 참조와 같은 계열이다. 카메라는 바닥 가까이에서 완만히 올려다보며, 젖은 바닥의 조명 반사도 가능한 위치다.",
        "entities": "등장 인물은 기계 찰리 하나뿐이다. 샌드 베이지의 마모된 각진 장갑, 흰 마스크, 주황색 점 눈 두 개와 선 입, 푸른 원형 가슴 동력원이 참조와 부합한다. 인간 피부나 치아는 없다. 다만 참조의 짧은 다리와 거대한 고릴라형 팔에 비해 다리가 길고 팔의 비중이 작다. 낡은 경비행기, 회전하는 프로펠러, 검은 배기가스와 천장 전구의 파열 불꽃이 보인다. 장갑 손상은 있지만 B-200의 부품으로 식별할 수 있는 물체는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 벌린 양발의 발바닥으로 바닥을 딛고 있으며 무게중심이 두 발 사이에 있다. 양손은 허리 장갑에 닿고 팔꿈치와 손목의 연결도 성립한다. 항공기는 착륙바퀴로 지지되며 전경 바퀴에는 고임목이 보인다. 프로펠러의 흐림은 엔진 축을 중심으로 한 회전으로 읽힌다. 배기가스는 엔진 주변에서 퍼지고 파열 불꽃은 천장 등 위치에서 떨어진다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 몸통은 거의 정면이지만 얼굴은 화면 오른쪽으로 돌아가 있어 화면 밖 접근 방향을 본다. 양손은 허리에 놓이고 팔꿈치는 좌우로 벌어져 있다. 좌우 전경 항공기는 중앙 통로를 향한 사선 기수를 보이며 뒤쪽 항공기들도 통로 양편에서 앞쪽을 향한다.",
        "built_space": "찰리는 중앙보다 왼쪽에 전신으로 서 있고 발 아래 바닥과 머리 위 여백이 남는다. 오른쪽 접근 공간도 비어 있다. 좌우 전경 두 대와 중후경의 적어도 세 대가 식별되어 항공기 깊이층은 유지된다. 점등된 천장 등 다섯 개와 별도의 파열 불꽃 한 곳이 보인다. 양측 창열, 골강판 벽, 박공지붕 철골과 작업용 집기, 젖고 낡은 바닥은 장소 참조에 대체로 부합한다. 낮은 시점이지만 A보다 찰리의 정면 축에 가깝고 찰리의 화면 점유율도 조금 크다. 바닥 반사는 조명과 젖은 표면의 관계상 자연스럽다.",
        "entities": "찰리 한 개체만 등장하며 추가 사람은 없다. 베이지 장갑의 긁힘과 패임, 흰 마스크, 주황색 두 눈과 선 입, 푸른 가슴 원자로는 참조와 맞는다. 몸통은 육중하지만 참조만큼 팔이 길거나 다리가 짧지는 않다. 낡은 경비행기들의 프로펠러 회전과 짙은 배기가스, 천장의 파열 불꽃이 보인다. B-200 부품은 별도 물체로 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발은 바닥에 접촉하고 넓은 지지면 안에 몸통이 놓인다. 두 손은 허리 양측 장갑을 짚고 있어 지정된 정지 포즈가 가능하다. 항공기들은 바퀴로 바닥에 지지되고 전경에는 고임목도 있다. 프로펠러는 기수 축에 연결된 회전 흐림으로 표현되며 배기는 엔진 부근에서 상승한다. 천장 불꽃은 등 부근에서 발생한다. 지지 없는 부유나 불가능한 신체 연결은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 사선 시점, 왼쪽의 전신 포즈, 화면 밖 오른쪽을 향한 얼굴과 열린 접근 공간이 지시에 더 충실하나, 참조보다 다리가 길고 B-200 부품 보유는 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "양손을 허리에 댄 전신과 가동 중인 항공기는 충실하지만, A보다 정면에 가까운 시점과 조금 더 큰 인물 비중으로 지정된 사선 와이드 구도가 약해지며 B-200 부품도 식별되지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 몸통은 카메라에 비스듬하고 얼굴은 화면 오른쪽을 향해 돌아가 있어, 렌즈가 아닌 화면 밖 동료들이 올 방향을 바라본다. 양팔의 팔꿈치는 바깥으로 벌어지고 두 손은 허리를 향한다. 좌우 전경 항공기의 기수는 각각 중앙 통로 쪽으로 비스듬히 향하며, 뒤쪽 항공기도 앞쪽 사선으로 놓여 있다.",
        "built_space": "중앙보다 왼쪽에 찰리 전신이 있고 머리 위와 발밑에 여백이 남는다. 오른쪽 바닥 통로는 비어 있다. 좌우 전경에 항공기 한 대씩, 후방에 적어도 두 대가 보여 서로 다른 깊이를 이룬다. 천장에는 점등된 등 네 개와 불꽃이 튀는 등 한 개가 식별된다. 골강판 벽, 박공지붕 철골, 양측 높은 창열, 벽가 선반과 작업대, 얼룩진 콘크리트 바닥은 장소 참조와 같은 계열이다. 카메라는 바닥 가까이에서 완만히 올려다보며, 젖은 바닥의 조명 반사도 가능한 위치다.",
        "entities": "등장 인물은 기계 찰리 하나뿐이다. 샌드 베이지의 마모된 각진 장갑, 흰 마스크, 주황색 점 눈 두 개와 선 입, 푸른 원형 가슴 동력원이 참조와 부합한다. 인간 피부나 치아는 없다. 다만 참조의 짧은 다리와 거대한 고릴라형 팔에 비해 다리가 길고 팔의 비중이 작다. 낡은 경비행기, 회전하는 프로펠러, 검은 배기가스와 천장 전구의 파열 불꽃이 보인다. 장갑 손상은 있지만 B-200의 부품으로 식별할 수 있는 물체는 보이지 않는다.",
        "hard_violations": [],
        "physics": "찰리는 벌린 양발의 발바닥으로 바닥을 딛고 있으며 무게중심이 두 발 사이에 있다. 양손은 허리 장갑에 닿고 팔꿈치와 손목의 연결도 성립한다. 항공기는 착륙바퀴로 지지되며 전경 바퀴에는 고임목이 보인다. 프로펠러의 흐림은 엔진 축을 중심으로 한 회전으로 읽힌다. 배기가스는 엔진 주변에서 퍼지고 파열 불꽃은 천장 등 위치에서 떨어진다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 몸통은 거의 정면이지만 얼굴은 화면 오른쪽으로 돌아가 있어 화면 밖 접근 방향을 본다. 양손은 허리에 놓이고 팔꿈치는 좌우로 벌어져 있다. 좌우 전경 항공기는 중앙 통로를 향한 사선 기수를 보이며 뒤쪽 항공기들도 통로 양편에서 앞쪽을 향한다.",
        "built_space": "찰리는 중앙보다 왼쪽에 전신으로 서 있고 발 아래 바닥과 머리 위 여백이 남는다. 오른쪽 접근 공간도 비어 있다. 좌우 전경 두 대와 중후경의 적어도 세 대가 식별되어 항공기 깊이층은 유지된다. 점등된 천장 등 다섯 개와 별도의 파열 불꽃 한 곳이 보인다. 양측 창열, 골강판 벽, 박공지붕 철골과 작업용 집기, 젖고 낡은 바닥은 장소 참조에 대체로 부합한다. 낮은 시점이지만 A보다 찰리의 정면 축에 가깝고 찰리의 화면 점유율도 조금 크다. 바닥 반사는 조명과 젖은 표면의 관계상 자연스럽다.",
        "entities": "찰리 한 개체만 등장하며 추가 사람은 없다. 베이지 장갑의 긁힘과 패임, 흰 마스크, 주황색 두 눈과 선 입, 푸른 가슴 원자로는 참조와 맞는다. 몸통은 육중하지만 참조만큼 팔이 길거나 다리가 짧지는 않다. 낡은 경비행기들의 프로펠러 회전과 짙은 배기가스, 천장의 파열 불꽃이 보인다. B-200 부품은 별도 물체로 확인되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 두 발은 바닥에 접촉하고 넓은 지지면 안에 몸통이 놓인다. 두 손은 허리 양측 장갑을 짚고 있어 지정된 정지 포즈가 가능하다. 항공기들은 바퀴로 바닥에 지지되고 전경에는 고임목도 있다. 프로펠러는 기수 축에 연결된 회전 흐림으로 표현되며 배기는 엔진 부근에서 상승한다. 천장 불꽃은 등 부근에서 발생한다. 지지 없는 부유나 불가능한 신체 연결은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1875,
   "B": 1875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "캐릭터의 디자인(특히 가슴 원자로의 디테일과 마스크)이 레퍼런스와 더 정확히 일치하며, 지시된 구도와 앵글을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "지시된 구도와 동작을 잘 따랐으나, 캐릭터의 가슴 원자로와 마스크 디테일이 A에 비해 다소 덜 정확합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S75sh4_sel.png",
    "asset_id": "5b49ddab-4018-45fd-98aa-f7de617e4d38",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dda-38ae-7dbb-abce-23b38da63967",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh4"
  }
 },
 "S77sh39::signage": {
  "fp": "e294e283bfdfb1a7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4e412ec1827de752": {
  "subjects": [],
  "subject_text": "목포공항 격납고 내부\n거대한 창고 건물 안으로 먼지 덮인 비행체들이 보관되어 있다.",
  "identity": "canonical",
  "scope_id": "L129",
  "scope_role": "location_interior",
  "scope_sha": "a6301c07c943732d"
 },
 "S77sh39::bgfirst_bg": {
  "input_fingerprint": "a36c14d0a4882119",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지상에 내려 찰리 곁에 나란히 선 이현우의 전신 구도.\n\nLOCATION (lock): On the airfield apron beside the helicopter's boarding area in daylight, where the robot waits on the ground.\n\nTIME OF DAY (lock): night to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the ground-level tracking move as 이현우 completes his approach, maintaining the slightly low, side-and-rear three-quarter view rather than crossing in front of the pair. Show 이현우 head to foot at center-left with his weight settling onto his arriving foot and 찰리 at center-right beginning to pivot toward him, leaving clear ground beneath both and a visible gap to the helicopter in the background. 이현우's gaze is lowered toward the ground where he stops, while 찰리 looks down toward his companion's arriving step; their changing proximity carries the reunion without a push-in.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (On the ground before departure) — An oblique exterior portion is visible behind the pair, with the cabin interior outside the framing; used as Establishes the departure they have stepped away from while remaining smaller than two-fifths of the image; Airport ground (The pair have reached the same ground-level position); used as Visible space beneath their feet and between them and the helicopter clarifies the completed approach.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with controlled contrast supports the quiet reunion while preserving natural human detail and clean robot articulation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지상에 내려 찰리 곁에 나란히 선 이현우의 전신 구도.\n\nLOCATION (lock): On the airfield apron beside the helicopter's boarding area in daylight, where the robot waits on the ground.\n\nTIME OF DAY (lock): night to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the ground-level tracking move as 이현우 completes his approach, maintaining the slightly low, side-and-rear three-quarter view rather than crossing in front of the pair. Show 이현우 head to foot at center-left with his weight settling onto his arriving foot and 찰리 at center-right beginning to pivot toward him, leaving clear ground beneath both and a visible gap to the helicopter in the background. 이현우's gaze is lowered toward the ground where he stops, while 찰리 looks down toward his companion's arriving step; their changing proximity carries the reunion without a push-in.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (On the ground before departure) — An oblique exterior portion is visible behind the pair, with the cabin interior outside the framing; used as Establishes the departure they have stepped away from while remaining smaller than two-fifths of the image; Airport ground (The pair have reached the same ground-level position); used as Visible space beneath their feet and between them and the helicopter clarifies the completed approach.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with controlled contrast supports the quiet reunion while preserving natural human detail and clean robot articulation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh39__bgfirst_bg.png",
  "asset_id": "857f728b-9bf1-4567-a6d6-ad8daa25db50",
  "input_asset_ids": [
   "609125c2-5374-42e8-b221-6537e2b40d31",
   "a5599059-47f4-4453-91ca-cf9f1f6002e6"
  ]
 },
 "S77sh39": {
  "input_fingerprint": "8a35a7634f1a0ede",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 지상에 내려 찰리 곁에 나란히 선 이현우의 전신 구도.\n\nLOCATION (lock): On the airfield apron beside the helicopter's boarding area in daylight, where the robot waits on the ground. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the ground-level tracking move as 이현우 completes his approach, maintaining the slightly low, side-and-rear three-quarter view rather than crossing in front of the pair. Show 이현우 head to foot at center-left with his weight settling onto his arriving foot and 찰리 at center-right beginning to pivot toward him, leaving clear ground beneath both and a visible gap to the helicopter in the background. 이현우's gaze is lowered toward the ground where he stops, while 찰리 looks down toward his companion's arriving step; their changing proximity carries the reunion without a push-in.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (On the ground before departure) — An oblique exterior portion is visible behind the pair, with the cabin interior outside the framing; used as Establishes the departure they have stepped away from while remaining smaller than two-fifths of the image; Airport ground (The pair have reached the same ground-level position); used as Visible space beneath their feet and between them and the helicopter clarifies the completed approach.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with controlled contrast supports the quiet reunion while preserving natural human detail and clean robot articulation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight has replaced the nighttime hangar sequence, and the departure helicopter remains on the ground. Charlie stands outside it with his existing body and chest damage and retains B-200's component. 이현우: He has stepped down onto the ground, still bearing his earlier wounds and dirty appearance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 지상에 내려 찰리 곁에 나란히 선 이현우의 전신 구도.\n\nLOCATION (lock): On the airfield apron beside the helicopter's boarding area in daylight, where the robot waits on the ground. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the ground-level tracking move as 이현우 completes his approach, maintaining the slightly low, side-and-rear three-quarter view rather than crossing in front of the pair. Show 이현우 head to foot at center-left with his weight settling onto his arriving foot and 찰리 at center-right beginning to pivot toward him, leaving clear ground beneath both and a visible gap to the helicopter in the background. 이현우's gaze is lowered toward the ground where he stops, while 찰리 looks down toward his companion's arriving step; their changing proximity carries the reunion without a push-in.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (On the ground before departure) — An oblique exterior portion is visible behind the pair, with the cabin interior outside the framing; used as Establishes the departure they have stepped away from while remaining smaller than two-fifths of the image; Airport ground (The pair have reached the same ground-level position); used as Visible space beneath their feet and between them and the helicopter clarifies the completed approach.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with controlled contrast supports the quiet reunion while preserving natural human detail and clean robot articulation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight has replaced the nighttime hangar sequence, and the departure helicopter remains on the ground. Charlie stands outside it with his existing body and chest damage and retains B-200's component. 이현우: He has stepped down onto the ground, still bearing his earlier wounds and dirty appearance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 지상에 내려 찰리 곁에 나란히 선 이현우의 전신 구도.\n\nLOCATION (lock): On the airfield apron beside the helicopter's boarding area in daylight, where the robot waits on the ground. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the ground-level tracking move as 이현우 completes his approach, maintaining the slightly low, side-and-rear three-quarter view rather than crossing in front of the pair. Show 이현우 head to foot at center-left with his weight settling onto his arriving foot and 찰리 at center-right beginning to pivot toward him, leaving clear ground beneath both and a visible gap to the helicopter in the background. 이현우's gaze is lowered toward the ground where he stops, while 찰리 looks down toward his companion's arriving step; their changing proximity carries the reunion without a push-in.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Helicopter (On the ground before departure) — An oblique exterior portion is visible behind the pair, with the cabin interior outside the framing; used as Establishes the departure they have stepped away from while remaining smaller than two-fifths of the image; Airport ground (The pair have reached the same ground-level position); used as Visible space beneath their feet and between them and the helicopter clarifies the completed approach.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daylight with controlled contrast supports the quiet reunion while preserving natural human detail and clean robot articulation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight has replaced the nighttime hangar sequence, and the departure helicopter remains on the ground. Charlie stands outside it with his existing body and chest damage and retains B-200's component. 이현우: He has stepped down onto the ground, still bearing his earlier wounds and dirty appearance.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh39__bgfirst_bg.png",
     "asset_id": "857f728b-9bf1-4567-a6d6-ad8daa25db50",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S77sh39.png",
     "asset_id": "609125c2-5374-42e8-b221-6537e2b40d31",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L129B01.png",
     "asset_id": "a5599059-47f4-4453-91ca-cf9f1f6002e6",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 발밑 바닥을 향해 시선을 내리고 있으며, 찰리는 이현우가 내딛는 발을 향해 고개를 숙여 응시합니다.",
    "built_space": "비행장 앞치마 위에 위치하며, 왼쪽 배경에 격납고, 오른쪽 뒤편에 헬기가 배치되어 지정된 공간 비율을 준수합니다.",
    "entities": "이현우는 오염된 의상과 인이어 무전기를 착용했고, 찰리는 지시된 고릴라 형태의 샌드 베이지 장갑판과 가슴의 푸른 코어를 갖췄습니다.",
    "hard_violations": [],
    "physics": "이현우는 뒷발로 체중을 지탱하며 앞발 뒷꿈치가 바닥에 닿으려는 자연스러운 걸음걸이 물리량을 보여주며, 찰리 역시 안정적으로 지면에 서 있습니다."
   },
   {
    "label": "B",
    "direction": "이현우와 찰리 모두 바닥을 내려다보고 있으나 찰리의 시선이 이현우의 발보다는 두 사람 사이의 빈 공간을 향합니다.",
    "built_space": "격납고와 헬기가 배치된 야외 활주로 공간으로, 참조 이미지의 건축적 특징과 위치를 잘 반영하고 있습니다.",
    "entities": "이현우는 오염된 의상을 입었으나 인이어가 보이지 않으며, 찰리와 헬기는 참조된 외형을 정확히 따르고 있습니다.",
    "hard_violations": [],
    "physics": "이현우의 두 발이 완전히 평평하게 지면에 닿아 있어 움직임이 완료된 정지 상태를 나타내며 떠 있는 요소는 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우가 내딛는 발에 체중이 실리는 순간과 두 캐릭터의 시선 교차를 프롬프트의 지시대로 정확하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우가 이미 두 발을 땅에 완전히 딛고 서 있어 '도착하는 발에 체중을 싣는' 동작 지시를 누락했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 발밑 바닥을 향해 시선을 내리고 있으며, 찰리는 이현우가 내딛는 발을 향해 고개를 숙여 응시합니다.",
        "built_space": "비행장 앞치마 위에 위치하며, 왼쪽 배경에 격납고, 오른쪽 뒤편에 헬기가 배치되어 지정된 공간 비율을 준수합니다.",
        "entities": "이현우는 오염된 의상과 인이어 무전기를 착용했고, 찰리는 지시된 고릴라 형태의 샌드 베이지 장갑판과 가슴의 푸른 코어를 갖췄습니다.",
        "hard_violations": [],
        "physics": "이현우는 뒷발로 체중을 지탱하며 앞발 뒷꿈치가 바닥에 닿으려는 자연스러운 걸음걸이 물리량을 보여주며, 찰리 역시 안정적으로 지면에 서 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리 모두 바닥을 내려다보고 있으나 찰리의 시선이 이현우의 발보다는 두 사람 사이의 빈 공간을 향합니다.",
        "built_space": "격납고와 헬기가 배치된 야외 활주로 공간으로, 참조 이미지의 건축적 특징과 위치를 잘 반영하고 있습니다.",
        "entities": "이현우는 오염된 의상을 입었으나 인이어가 보이지 않으며, 찰리와 헬기는 참조된 외형을 정확히 따르고 있습니다.",
        "hard_violations": [],
        "physics": "이현우의 두 발이 완전히 평평하게 지면에 닿아 있어 움직임이 완료된 정지 상태를 나타내며 떠 있는 요소는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우가 내딛는 발에 체중이 실리는 순간과 두 캐릭터의 시선 교차를 프롬프트의 지시대로 정확하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "이현우가 이미 두 발을 땅에 완전히 딛고 서 있어 '도착하는 발에 체중을 싣는' 동작 지시를 누락했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 발밑 바닥을 향해 시선을 내리고 있으며, 찰리는 이현우가 내딛는 발을 향해 고개를 숙여 응시합니다.",
        "built_space": "비행장 앞치마 위에 위치하며, 왼쪽 배경에 격납고, 오른쪽 뒤편에 헬기가 배치되어 지정된 공간 비율을 준수합니다.",
        "entities": "이현우는 오염된 의상과 인이어 무전기를 착용했고, 찰리는 지시된 고릴라 형태의 샌드 베이지 장갑판과 가슴의 푸른 코어를 갖췄습니다.",
        "hard_violations": [],
        "physics": "이현우는 뒷발로 체중을 지탱하며 앞발 뒷꿈치가 바닥에 닿으려는 자연스러운 걸음걸이 물리량을 보여주며, 찰리 역시 안정적으로 지면에 서 있습니다."
       },
       {
        "label": "B",
        "direction": "이현우와 찰리 모두 바닥을 내려다보고 있으나 찰리의 시선이 이현우의 발보다는 두 사람 사이의 빈 공간을 향합니다.",
        "built_space": "격납고와 헬기가 배치된 야외 활주로 공간으로, 참조 이미지의 건축적 특징과 위치를 잘 반영하고 있습니다.",
        "entities": "이현우는 오염된 의상을 입었으나 인이어가 보이지 않으며, 찰리와 헬기는 참조된 외형을 정확히 따르고 있습니다.",
        "hard_violations": [],
        "physics": "이현우의 두 발이 완전히 평평하게 지면에 닿아 있어 움직임이 완료된 정지 상태를 나타내며 떠 있는 요소는 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "찰리의 낮고 육중한 체형은 더 정확하지만, 정면 쪽 시점과 이미 멈춘 자세가 지정된 후측면 구도 및 도착하는 발의 체중 이동을 놓친다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "도착하는 발동작과 헬리콥터까지의 공간 분리가 더 충실하지만, 후측면 시점이 아니며 찰리가 지나치게 크고 다리가 길다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 고개와 눈을 자기 발 앞 지면으로 내리고, 찰리도 얼굴을 왼쪽 아래의 이현우 발 부근으로 향한다. 시선의 목표는 대체로 맞지만 두 몸의 앞면이 보여, 카메라가 요구된 후측면에 있지는 않다. 찰리가 몸을 돌리기 시작하는 순간도 뚜렷하지 않다.",
        "built_space": "왼쪽에 열린 주 격납고 한 동, 중앙 뒤편에 낮은 부속 건물과 울타리, 오른쪽에 착륙한 헬리콥터 한 대가 보인다. 균열 난 콘크리트와 노란 지상선은 장소 참조와 부합한다. 이현우는 중앙 왼쪽, 찰리는 중앙 오른쪽에 전신으로 서고 발 아래 지면과 뒤쪽 헬리콥터까지의 간격도 보인다. 다만 헬리콥터의 열린 출입구 안쪽이 보여 객실 내부를 제외하라는 지시와 어긋난다.",
        "entities": "인물은 이현우와 찰리뿐이다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 어두운 셔츠·바지, 흙먼지와 얼굴 상처가 참조에 부합한다. 인이어 무전기는 명확히 확인되지 않는다. 찰리는 베이지 장갑, 긴 팔과 짧은 다리, 흰 마스크와 두 발광 눈·입선, 푸른 원자로를 갖춰 참조 체형에 가깝다. 표면 마모는 있지만 가슴의 기존 파손은 뚜렷하지 않고, 양손에 B-200 부품은 보이지 않는다. 낮의 비행장이지만 빛은 요구된 절제된 주광보다 다소 강하다.",
        "hard_violations": [],
        "physics": "이현우의 양 신발과 찰리의 양발이 지면에 닿아 체중을 지탱한다. 팔과 손은 관절에 자연스럽게 연결되어 있고 공중에 떠 있는 물체는 없다. 헬리콥터도 착륙 바퀴로 지지된다. 다만 이현우는 두 발을 내려놓고 멈춘 상태에 가까워, 도착하는 발로 무게가 실리는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "이현우는 앞으로 내딛는 발 주변의 지면을 내려다보고, 찰리의 얼굴도 왼쪽 아래 이현우의 도착 지점으로 향한다. 이현우는 오른쪽으로 접근하며 멈추는 모습이다. 그러나 셔츠 앞면과 찰리의 가슴이 드러나는 전측면 시점이라, 지정된 후측면 카메라 축은 충족하지 못한다.",
        "built_space": "왼쪽에 열린 주 격납고 한 동, 그 뒤로 낮은 부속 건물들, 울타리와 조명 기둥이 있고 오른쪽 배경에는 헬리콥터 한 대가 있다. 콘크리트 계류장과 노란 선, 주변 산지가 장소 참조와 대응한다. 두 인물은 중앙 좌우에 머리부터 발끝까지 들어오며, 발 아래 여백과 헬리콥터까지의 빈 지면이 명확하다. 헬리콥터는 A보다 작은 배경 요소로 남지만, 여기서도 열린 객실 출입구의 어두운 내부가 보인다.",
        "entities": "이현우와 찰리 외에 추가 인물은 없다. 이현우의 젊은 동아시아계 남성 외형, 검은 머리, 마른 체격, 오염된 어두운 셔츠와 바지, 귀의 소형 장치는 설정에 부합한다. 얼굴 상처와 핏자국은 A보다 덜 명확하다. 찰리의 베이지 장갑과 흰 마스크, 발광 눈, 입선과 푸른 가슴 원자로는 맞지만, 키가 이현우에 근접하고 다리가 길어 참조의 작고 육중한 고릴라형 비율에서 벗어난다. 가슴 파손은 분명하지 않으며 두 손은 비어 있어 B-200 부품 유지가 표현되지 않았다. 주광은 다소 선명하고 강하다.",
        "hard_violations": [],
        "physics": "이현우는 뒤쪽 신발로 몸을 지지하면서 앞쪽 신발의 뒤꿈치를 지면에 대고 발끝을 조금 들어, 마지막 걸음을 내려놓는 동작을 보인다. 완전히 무게가 옮겨진 순간은 아니지만 접근을 마치는 행동으로 가능하다. 찰리는 벌린 두 발로 서 있으며 숙인 상체도 지지 범위 안에 있다. 헬리콥터는 바퀴로 지면에 놓여 있고 지지 없이 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "찰리의 낮고 육중한 체형은 더 정확하지만, 정면 쪽 시점과 이미 멈춘 자세가 지정된 후측면 구도 및 도착하는 발의 체중 이동을 놓친다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "도착하는 발동작과 헬리콥터까지의 공간 분리가 더 충실하지만, 후측면 시점이 아니며 찰리가 지나치게 크고 다리가 길다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 고개와 눈을 자기 발 앞 지면으로 내리고, 찰리도 얼굴을 왼쪽 아래의 이현우 발 부근으로 향한다. 시선의 목표는 대체로 맞지만 두 몸의 앞면이 보여, 카메라가 요구된 후측면에 있지는 않다. 찰리가 몸을 돌리기 시작하는 순간도 뚜렷하지 않다.",
        "built_space": "왼쪽에 열린 주 격납고 한 동, 중앙 뒤편에 낮은 부속 건물과 울타리, 오른쪽에 착륙한 헬리콥터 한 대가 보인다. 균열 난 콘크리트와 노란 지상선은 장소 참조와 부합한다. 이현우는 중앙 왼쪽, 찰리는 중앙 오른쪽에 전신으로 서고 발 아래 지면과 뒤쪽 헬리콥터까지의 간격도 보인다. 다만 헬리콥터의 열린 출입구 안쪽이 보여 객실 내부를 제외하라는 지시와 어긋난다.",
        "entities": "인물은 이현우와 찰리뿐이다. 이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로, 마른 체격과 어두운 셔츠·바지, 흙먼지와 얼굴 상처가 참조에 부합한다. 인이어 무전기는 명확히 확인되지 않는다. 찰리는 베이지 장갑, 긴 팔과 짧은 다리, 흰 마스크와 두 발광 눈·입선, 푸른 원자로를 갖춰 참조 체형에 가깝다. 표면 마모는 있지만 가슴의 기존 파손은 뚜렷하지 않고, 양손에 B-200 부품은 보이지 않는다. 낮의 비행장이지만 빛은 요구된 절제된 주광보다 다소 강하다.",
        "hard_violations": [],
        "physics": "이현우의 양 신발과 찰리의 양발이 지면에 닿아 체중을 지탱한다. 팔과 손은 관절에 자연스럽게 연결되어 있고 공중에 떠 있는 물체는 없다. 헬리콥터도 착륙 바퀴로 지지된다. 다만 이현우는 두 발을 내려놓고 멈춘 상태에 가까워, 도착하는 발로 무게가 실리는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "이현우는 앞으로 내딛는 발 주변의 지면을 내려다보고, 찰리의 얼굴도 왼쪽 아래 이현우의 도착 지점으로 향한다. 이현우는 오른쪽으로 접근하며 멈추는 모습이다. 그러나 셔츠 앞면과 찰리의 가슴이 드러나는 전측면 시점이라, 지정된 후측면 카메라 축은 충족하지 못한다.",
        "built_space": "왼쪽에 열린 주 격납고 한 동, 그 뒤로 낮은 부속 건물들, 울타리와 조명 기둥이 있고 오른쪽 배경에는 헬리콥터 한 대가 있다. 콘크리트 계류장과 노란 선, 주변 산지가 장소 참조와 대응한다. 두 인물은 중앙 좌우에 머리부터 발끝까지 들어오며, 발 아래 여백과 헬리콥터까지의 빈 지면이 명확하다. 헬리콥터는 A보다 작은 배경 요소로 남지만, 여기서도 열린 객실 출입구의 어두운 내부가 보인다.",
        "entities": "이현우와 찰리 외에 추가 인물은 없다. 이현우의 젊은 동아시아계 남성 외형, 검은 머리, 마른 체격, 오염된 어두운 셔츠와 바지, 귀의 소형 장치는 설정에 부합한다. 얼굴 상처와 핏자국은 A보다 덜 명확하다. 찰리의 베이지 장갑과 흰 마스크, 발광 눈, 입선과 푸른 가슴 원자로는 맞지만, 키가 이현우에 근접하고 다리가 길어 참조의 작고 육중한 고릴라형 비율에서 벗어난다. 가슴 파손은 분명하지 않으며 두 손은 비어 있어 B-200 부품 유지가 표현되지 않았다. 주광은 다소 선명하고 강하다.",
        "hard_violations": [],
        "physics": "이현우는 뒤쪽 신발로 몸을 지지하면서 앞쪽 신발의 뒤꿈치를 지면에 대고 발끝을 조금 들어, 마지막 걸음을 내려놓는 동작을 보인다. 완전히 무게가 옮겨진 순간은 아니지만 접근을 마치는 행동으로 가능하다. 찰리는 벌린 두 발로 서 있으며 숙인 상체도 지지 범위 안에 있다. 헬리콥터는 바퀴로 지면에 놓여 있고 지지 없이 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.405
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.405
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1405
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이현우가 내딛는 발에 체중이 실리는 순간과 두 캐릭터의 시선 교차를 프롬프트의 지시대로 정확하게 연출했습니다."
   },
   {
    "label": "B",
    "score": 1405,
    "verdict_ko": "이현우가 이미 두 발을 땅에 완전히 딛고 서 있어 '도착하는 발에 체중을 싣는' 동작 지시를 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L129B01.png",
    "asset_id": "a5599059-47f4-4453-91ca-cf9f1f6002e6",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0de0-77b3-7471-acec-0dcd58760080",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh39__bgfirst_bg.png",
   "bg_asset_id": "857f728b-9bf1-4567-a6d6-ad8daa25db50",
   "bg_record_key": "S77sh39::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S77sh56::signage": {
  "fp": "e43744a05097ff37",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S77sh56::bgfirst_bg": {
  "input_fingerprint": "9b9fe70e28f75e28",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활주로 끝을 향해 나란히 발을 내디딘 mid-stride 자세를 한 이현우와 찰리의 경쾌한 뒷모습 전경.\n\nLOCATION (lock): On the abandoned airfield's open runway, walking toward its far end after the helicopter's departure.\n\nTIME OF DAY (lock): night to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static, near-eye-level wide view from behind and slightly outside the companions' shared walking line, looking directly along the runway toward its far end. Place 이현우 left of center and 찰리 right of center in the lower middle of the image, both fully visible with backpacks and clear space below their feet; catch different phases of their strides rather than synchronized steps. Their attention follows the distant route ahead as they walk toward the upper center, letting their recession change scale naturally without a zoom or a new camera angle.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Far end of the runway, destination of both receding companions in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Runway (The companions are walking toward its far end) — Its length recedes from the lower foreground toward the upper-center distance; used as Provides the shared direction of travel and generous space ahead; Backpacks (Worn by both companions) — Their rear-facing outer sides are visible against the walkers' backs; used as Keeps the departure visually grounded in the continuing journey.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daylight and restrained contrast, allowing the lightness of their companionship to emerge through movement rather than a warmer lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활주로 끝을 향해 나란히 발을 내디딘 mid-stride 자세를 한 이현우와 찰리의 경쾌한 뒷모습 전경.\n\nLOCATION (lock): On the abandoned airfield's open runway, walking toward its far end after the helicopter's departure.\n\nTIME OF DAY (lock): night to day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static, near-eye-level wide view from behind and slightly outside the companions' shared walking line, looking directly along the runway toward its far end. Place 이현우 left of center and 찰리 right of center in the lower middle of the image, both fully visible with backpacks and clear space below their feet; catch different phases of their strides rather than synchronized steps. Their attention follows the distant route ahead as they walk toward the upper center, letting their recession change scale naturally without a zoom or a new camera angle.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Far end of the runway, destination of both receding companions in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Runway (The companions are walking toward its far end) — Its length recedes from the lower foreground toward the upper-center distance; used as Provides the shared direction of travel and generous space ahead; Backpacks (Worn by both companions) — Their rear-facing outer sides are visible against the walkers' backs; used as Keeps the departure visually grounded in the continuing journey.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daylight and restrained contrast, allowing the lightness of their companionship to emerge through movement rather than a warmer lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh56__bgfirst_bg.png",
  "asset_id": "c48ad37e-50d3-4d5c-91fc-d0a5ce6ad684",
  "input_asset_ids": [
   "24c6117e-2683-4c62-916b-8fd226c9f787",
   "624ede91-b541-482e-9417-34ae46ee8a0b"
  ]
 },
 "S77sh56": {
  "input_fingerprint": "00dd8799272d8a7d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 활주로 끝을 향해 나란히 발을 내디딘 mid-stride 자세를 한 이현우와 찰리의 경쾌한 뒷모습 전경.\n\nLOCATION (lock): On the abandoned airfield's open runway, walking toward its far end after the helicopter's departure. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static, near-eye-level wide view from behind and slightly outside the companions' shared walking line, looking directly along the runway toward its far end. Place 이현우 left of center and 찰리 right of center in the lower middle of the image, both fully visible with backpacks and clear space below their feet; catch different phases of their strides rather than synchronized steps. Their attention follows the distant route ahead as they walk toward the upper center, letting their recession change scale naturally without a zoom or a new camera angle.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Far end of the runway, destination of both receding companions in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Runway (The companions are walking toward its far end) — Its length recedes from the lower foreground toward the upper-center distance; used as Provides the shared direction of travel and generous space ahead; Backpacks (Worn by both companions) — Their rear-facing outer sides are visible against the walkers' backs; used as Keeps the departure visually grounded in the continuing journey.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daylight and restrained contrast, allowing the lightness of their companionship to emerge through movement rather than a warmer lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has departed and disappeared into the distance over the daylight airport. Charlie now carries a backpack, retains his battle damage and B-200's component, and no longer retains the violet given away before departure. 이현우: He is walking away from the airport with a backpack, still visibly wounded and disheveled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 활주로 끝을 향해 나란히 발을 내디딘 mid-stride 자세를 한 이현우와 찰리의 경쾌한 뒷모습 전경.\n\nLOCATION (lock): On the abandoned airfield's open runway, walking toward its far end after the helicopter's departure. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static, near-eye-level wide view from behind and slightly outside the companions' shared walking line, looking directly along the runway toward its far end. Place 이현우 left of center and 찰리 right of center in the lower middle of the image, both fully visible with backpacks and clear space below their feet; catch different phases of their strides rather than synchronized steps. Their attention follows the distant route ahead as they walk toward the upper center, letting their recession change scale naturally without a zoom or a new camera angle.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Far end of the runway, destination of both receding companions in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Runway (The companions are walking toward its far end) — Its length recedes from the lower foreground toward the upper-center distance; used as Provides the shared direction of travel and generous space ahead; Backpacks (Worn by both companions) — Their rear-facing outer sides are visible against the walkers' backs; used as Keeps the departure visually grounded in the continuing journey.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daylight and restrained contrast, allowing the lightness of their companionship to emerge through movement rather than a warmer lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has departed and disappeared into the distance over the daylight airport. Charlie now carries a backpack, retains his battle damage and B-200's component, and no longer retains the violet given away before departure. 이현우: He is walking away from the airport with a backpack, still visibly wounded and disheveled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night to day.\n\nSHOT TEXT (authoritative, Korean): 활주로 끝을 향해 나란히 발을 내디딘 mid-stride 자세를 한 이현우와 찰리의 경쾌한 뒷모습 전경.\n\nLOCATION (lock): On the abandoned airfield's open runway, walking toward its far end after the helicopter's departure. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static, near-eye-level wide view from behind and slightly outside the companions' shared walking line, looking directly along the runway toward its far end. Place 이현우 left of center and 찰리 right of center in the lower middle of the image, both fully visible with backpacks and clear space below their feet; catch different phases of their strides rather than synchronized steps. Their attention follows the distant route ahead as they walk toward the upper center, letting their recession change scale naturally without a zoom or a new camera angle.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Far end of the runway, destination of both receding companions in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Runway (The companions are walking toward its far end) — Its length recedes from the lower foreground toward the upper-center distance; used as Provides the shared direction of travel and generous space ahead; Backpacks (Worn by both companions) — Their rear-facing outer sides are visible against the walkers' backs; used as Keeps the departure visually grounded in the continuing journey.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the subdued daylight and restrained contrast, allowing the lightness of their companionship to emerge through movement rather than a warmer lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The helicopter has departed and disappeared into the distance over the daylight airport. Charlie now carries a backpack, retains his battle damage and B-200's component, and no longer retains the violet given away before departure. 이현우: He is walking away from the airport with a backpack, still visibly wounded and disheveled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh56__bgfirst_bg.png",
     "asset_id": "c48ad37e-50d3-4d5c-91fc-d0a5ce6ad684",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S77sh56.png",
     "asset_id": "24c6117e-2683-4c62-916b-8fd226c9f787",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L129B02.png",
     "asset_id": "624ede91-b541-482e-9417-34ae46ee8a0b",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
    "built_space": "참조된 관제탑, 격납고, 활주로가 정확한 위치에 묘사됨. 구도와 인물 배치 적절함.",
    "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 멘 뒷모습. 그러나 찰리가 이현우보다 훨씬 크고 긴 다리를 가져 참조 설정에 어긋남.",
    "hard_violations": [],
    "physics": "두 인물의 발이 지면에 닿아 있고 자연스러운 보행 동작을 지탱함."
   },
   {
    "label": "B",
    "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
    "built_space": "참조 이미지의 공간과 구조물들이 정확한 원근감으로 배경에 구현됨.",
    "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 메고 있음. 찰리가 짧은 다리와 굵은 팔을 지닌 고릴라 비율로 이현우보다 작게 묘사되어 설정과 일치함.",
    "hard_violations": [],
    "physics": "지면에 발을 안정적으로 딛고 걷는 자세가 물리적으로 타당함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍과 배경은 좋으나, 찰리가 설정(고릴라 체형, 작은 키)과 달리 이현우보다 크고 인간형에 가깝게 묘사되어 캐릭터 재현도가 매우 낮음."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 카메라 앵글과 구도를 잘 따랐으며, 찰리의 독특한 고릴라 비율과 상대적으로 작은 키를 이현우와 대비시켜 정확히 구현함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
        "built_space": "참조된 관제탑, 격납고, 활주로가 정확한 위치에 묘사됨. 구도와 인물 배치 적절함.",
        "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 멘 뒷모습. 그러나 찰리가 이현우보다 훨씬 크고 긴 다리를 가져 참조 설정에 어긋남.",
        "hard_violations": [],
        "physics": "두 인물의 발이 지면에 닿아 있고 자연스러운 보행 동작을 지탱함."
       },
       {
        "label": "B",
        "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
        "built_space": "참조 이미지의 공간과 구조물들이 정확한 원근감으로 배경에 구현됨.",
        "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 메고 있음. 찰리가 짧은 다리와 굵은 팔을 지닌 고릴라 비율로 이현우보다 작게 묘사되어 설정과 일치함.",
        "hard_violations": [],
        "physics": "지면에 발을 안정적으로 딛고 걷는 자세가 물리적으로 타당함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍과 배경은 좋으나, 찰리가 설정(고릴라 체형, 작은 키)과 달리 이현우보다 크고 인간형에 가깝게 묘사되어 캐릭터 재현도가 매우 낮음."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 카메라 앵글과 구도를 잘 따랐으며, 찰리의 독특한 고릴라 비율과 상대적으로 작은 키를 이현우와 대비시켜 정확히 구현함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
        "built_space": "참조된 관제탑, 격납고, 활주로가 정확한 위치에 묘사됨. 구도와 인물 배치 적절함.",
        "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 멘 뒷모습. 그러나 찰리가 이현우보다 훨씬 크고 긴 다리를 가져 참조 설정에 어긋남.",
        "hard_violations": [],
        "physics": "두 인물의 발이 지면에 닿아 있고 자연스러운 보행 동작을 지탱함."
       },
       {
        "label": "B",
        "direction": "두 인물이 카메라를 등지고 활주로 먼 끝을 향해 나란히 걷고 있음.",
        "built_space": "참조 이미지의 공간과 구조물들이 정확한 원근감으로 배경에 구현됨.",
        "entities": "왼쪽 이현우, 오른쪽 찰리가 배낭을 메고 있음. 찰리가 짧은 다리와 굵은 팔을 지닌 고릴라 비율로 이현우보다 작게 묘사되어 설정과 일치함.",
        "hard_violations": [],
        "physics": "지면에 발을 안정적으로 딛고 걷는 자세가 물리적으로 타당함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "하단 중앙의 전신 뒷모습과 작은 찰리의 상대 크기가 지시된 와이드 구도에 더 충실하지만, 따뜻한 역광과 비슷한 보행 위상은 요구에서 벗어난다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "주광과 폐활주로는 잘 맞지만, 두 피사체가 더 크게 배치되고 찰리가 현우와 거의 같은 키이며 보폭까지 동기화되어 지정된 구도와 동작에 덜 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 현우와 오른쪽 찰리 모두 등을 보이고 머리와 몸을 활주로 상단 중앙의 소실점 쪽으로 향한다. 발의 진행 방향도 같은 목적지를 따른다. 눈은 뒷모습이라 확인할 수 없으며, 별도로 겨누는 물체는 없다.",
        "built_space": "갈라진 포장 활주로 하나와 낡은 중앙선이 전경에서 먼 중앙으로 이어진다. 왼쪽에 관제탑 하나, 큰 격납고 하나와 열린 출입구 하나, 낮은 부속 건물들 및 약 다섯 개의 조명주가 보인다. 양옆 초지와 뒤쪽 산줄기도 장소 참조와 부합한다. 두 인물은 활주로 중앙선 양옆의 하단 중앙에 전신으로 놓이며 발 아래 여백이 있다. 눈높이에 가까운 후방 와이드 구도이고, 불가능한 반사나 중복된 주요 시설은 보이지 않는다.",
        "entities": "인물은 현우와 기계 찰리 둘뿐이다. 현우는 짧고 흐트러진 검은 머리, 마른 체격, 어두운 셔츠와 흙 묻은 바지를 갖췄다. 후면만 보여 한국계 미국인이라는 정체성, 정확한 연령과 얼굴 일치는 판별할 수 없고 인이어도 식별되지 않는다. 찰리는 현우보다 뚜렷하게 작고 베이지 장갑판과 긴 팔을 갖췄지만 참조보다 몸통과 팔의 육중함이 약하다. 얼굴과 가슴 원자로는 후면 구도상 가려지는 것이 맞다. 각자 배낭 하나를 메고 바깥 면이 카메라에 보인다. 장갑판 마모는 보이나 현우의 출혈 흔적과 B-200 부품은 식별되지 않는다. 헬리콥터와 보라색 꽃은 없다. 빛은 요구된 절제된 주광보다 따뜻하고 역광 효과가 강하다.",
        "hard_violations": [],
        "physics": "두 인물 모두 오른발 쪽으로 지면을 지지하면서 왼발을 뒤에 남긴 보행 순간으로 읽힌다. 현우의 들린 뒤꿈치와 찰리의 기울어진 발은 보행으로 가능한 자세이며 몸 전체가 뜨지는 않는다. 다만 둘 다 같은 쪽 다리가 뒤에 있어 서로 다른 보행 위상이라는 요구가 충분히 드러나지 않는다. 배낭은 어깨끈으로 지지되고 등에 밀착한다. 팔과 손도 각 몸체에 자연스럽게 연결되어 있다."
       },
       {
        "label": "B",
        "direction": "현우는 중앙선 왼쪽, 찰리는 오른쪽에서 머리와 몸을 먼 활주로 중앙으로 향한다. 두 사람의 이동 목표는 상단 중앙의 활주로 끝으로 명확하다. 얼굴이나 눈을 카메라 쪽으로 돌리지 않았으며 배낭의 뒤쪽 면이 보인다.",
        "built_space": "중앙선이 있는 활주로 하나, 왼쪽 관제탑 하나와 격납고 하나, 격납고의 검은 출입구 하나가 보인다. 낮은 부속 건물들과 약 다섯 개의 조명주, 주변 초지와 산줄기가 장소 참조의 구성을 유지한다. 격납고는 참조보다 작고 먼 비중으로 표현되지만 주요 시설의 중복은 없다. 후방 눈높이 와이드이며 전신과 발 아래 여백은 확보되지만, 두 피사체가 A보다 화면 높이를 크게 차지하여 하단 중앙에 물러나 있는 구도는 약해진다.",
        "entities": "현우 한 명과 찰리 한 대만 보인다. 현우의 검은 헝클어진 머리, 마른 체격, 더러운 어두운 셔츠와 찢어진 바지는 설정에 부합한다. 후면이라 얼굴의 연령·민족적 외관과 인이어는 확인할 수 없고, 명확한 피나 상처도 식별하기 어렵다. 찰리는 마모된 샌드 베이지 장갑판의 기계이지만 현우와 거의 같은 키이며 다리가 길어, 작고 육중한 고릴라형 비율과 다르다. 가려진 얼굴과 가슴은 평가할 수 없다. 배낭은 각각 하나씩 보이며 B-200 부품은 식별되지 않는다. 헬리콥터와 보라색 꽃은 없다. 낮의 비교적 중립적인 색감은 A보다 조명 요구에 가깝다.",
        "hard_violations": [],
        "physics": "현우와 찰리 모두 오른발을 앞쪽 지면에 두고 왼발 뒤꿈치를 들어 밑창을 카메라에 보이는 거의 동일한 걸음 단계다. 지지하는 발과 지면 접촉이 있어 보행 자체는 가능하지만, 서로 다른 위상을 포착하라는 지시와는 맞지 않는다. 두 배낭은 어깨끈과 등에 의해 지지된다. 공중에 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "하단 중앙의 전신 뒷모습과 작은 찰리의 상대 크기가 지시된 와이드 구도에 더 충실하지만, 따뜻한 역광과 비슷한 보행 위상은 요구에서 벗어난다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "주광과 폐활주로는 잘 맞지만, 두 피사체가 더 크게 배치되고 찰리가 현우와 거의 같은 키이며 보폭까지 동기화되어 지정된 구도와 동작에 덜 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 현우와 오른쪽 찰리 모두 등을 보이고 머리와 몸을 활주로 상단 중앙의 소실점 쪽으로 향한다. 발의 진행 방향도 같은 목적지를 따른다. 눈은 뒷모습이라 확인할 수 없으며, 별도로 겨누는 물체는 없다.",
        "built_space": "갈라진 포장 활주로 하나와 낡은 중앙선이 전경에서 먼 중앙으로 이어진다. 왼쪽에 관제탑 하나, 큰 격납고 하나와 열린 출입구 하나, 낮은 부속 건물들 및 약 다섯 개의 조명주가 보인다. 양옆 초지와 뒤쪽 산줄기도 장소 참조와 부합한다. 두 인물은 활주로 중앙선 양옆의 하단 중앙에 전신으로 놓이며 발 아래 여백이 있다. 눈높이에 가까운 후방 와이드 구도이고, 불가능한 반사나 중복된 주요 시설은 보이지 않는다.",
        "entities": "인물은 현우와 기계 찰리 둘뿐이다. 현우는 짧고 흐트러진 검은 머리, 마른 체격, 어두운 셔츠와 흙 묻은 바지를 갖췄다. 후면만 보여 한국계 미국인이라는 정체성, 정확한 연령과 얼굴 일치는 판별할 수 없고 인이어도 식별되지 않는다. 찰리는 현우보다 뚜렷하게 작고 베이지 장갑판과 긴 팔을 갖췄지만 참조보다 몸통과 팔의 육중함이 약하다. 얼굴과 가슴 원자로는 후면 구도상 가려지는 것이 맞다. 각자 배낭 하나를 메고 바깥 면이 카메라에 보인다. 장갑판 마모는 보이나 현우의 출혈 흔적과 B-200 부품은 식별되지 않는다. 헬리콥터와 보라색 꽃은 없다. 빛은 요구된 절제된 주광보다 따뜻하고 역광 효과가 강하다.",
        "hard_violations": [],
        "physics": "두 인물 모두 오른발 쪽으로 지면을 지지하면서 왼발을 뒤에 남긴 보행 순간으로 읽힌다. 현우의 들린 뒤꿈치와 찰리의 기울어진 발은 보행으로 가능한 자세이며 몸 전체가 뜨지는 않는다. 다만 둘 다 같은 쪽 다리가 뒤에 있어 서로 다른 보행 위상이라는 요구가 충분히 드러나지 않는다. 배낭은 어깨끈으로 지지되고 등에 밀착한다. 팔과 손도 각 몸체에 자연스럽게 연결되어 있다."
       },
       {
        "label": "A",
        "direction": "현우는 중앙선 왼쪽, 찰리는 오른쪽에서 머리와 몸을 먼 활주로 중앙으로 향한다. 두 사람의 이동 목표는 상단 중앙의 활주로 끝으로 명확하다. 얼굴이나 눈을 카메라 쪽으로 돌리지 않았으며 배낭의 뒤쪽 면이 보인다.",
        "built_space": "중앙선이 있는 활주로 하나, 왼쪽 관제탑 하나와 격납고 하나, 격납고의 검은 출입구 하나가 보인다. 낮은 부속 건물들과 약 다섯 개의 조명주, 주변 초지와 산줄기가 장소 참조의 구성을 유지한다. 격납고는 참조보다 작고 먼 비중으로 표현되지만 주요 시설의 중복은 없다. 후방 눈높이 와이드이며 전신과 발 아래 여백은 확보되지만, 두 피사체가 A보다 화면 높이를 크게 차지하여 하단 중앙에 물러나 있는 구도는 약해진다.",
        "entities": "현우 한 명과 찰리 한 대만 보인다. 현우의 검은 헝클어진 머리, 마른 체격, 더러운 어두운 셔츠와 찢어진 바지는 설정에 부합한다. 후면이라 얼굴의 연령·민족적 외관과 인이어는 확인할 수 없고, 명확한 피나 상처도 식별하기 어렵다. 찰리는 마모된 샌드 베이지 장갑판의 기계이지만 현우와 거의 같은 키이며 다리가 길어, 작고 육중한 고릴라형 비율과 다르다. 가려진 얼굴과 가슴은 평가할 수 없다. 배낭은 각각 하나씩 보이며 B-200 부품은 식별되지 않는다. 헬리콥터와 보라색 꽃은 없다. 낮의 비교적 중립적인 색감은 A보다 조명 요구에 가깝다.",
        "hard_violations": [],
        "physics": "현우와 찰리 모두 오른발을 앞쪽 지면에 두고 왼발 뒤꿈치를 들어 밑창을 카메라에 보이는 거의 동일한 걸음 단계다. 지지하는 발과 지면 접촉이 있어 보행 자체는 가능하지만, 서로 다른 위상을 포착하라는 지시와는 맞지 않는다. 두 배낭은 어깨끈과 등에 의해 지지된다. 공중에 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.321,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.321,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1321,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "프레이밍과 배경은 좋으나, 찰리가 설정(고릴라 체형, 작은 키)과 달리 이현우보다 크고 인간형에 가깝게 묘사되어 캐릭터 재현도가 매우 낮음."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 카메라 앵글과 구도를 잘 따랐으며, 찰리의 독특한 고릴라 비율과 상대적으로 작은 키를 이현우와 대비시켜 정확히 구현함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L129B02.png",
    "asset_id": "624ede91-b541-482e-9417-34ae46ee8a0b",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0de8-38a8-7a1d-8aeb-3b16857d86cb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S77sh56__bgfirst_bg.png",
   "bg_asset_id": "c48ad37e-50d3-4d5c-91fc-d0a5ce6ad684",
   "bg_record_key": "S77sh56::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S78sh2::signage": {
  "fp": "82874e717c30fcea",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::edcadc8e8f584e4f": {
  "subjects": [
   {
    "subject_native": "한국 연안 소형 어선",
    "search_terms_native": [
     "한국 소형 어선",
     "연안 FRP 어선",
     "포구 정박 어선"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 추가 검색어를 덧붙이지 마십시오.",
    "reason_ko": "서구 데이터 중심의 생성 모델은 한국 포구 특유의 소형 FRP 어선 선형과 조타실 설비 대신 서양식 모터보트나 목선을 그리기 쉬움."
   }
  ],
  "subject_text": "목포 항구 부두·생선 좌판·정박장, 출항 수역\n어둠이 깔린 항만 부두로 작은 어선들이 밧줄로 묶여 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L183",
  "scope_role": null,
  "scope_sha": null
 },
 "era_fail::6aea8298ebc1ce11": {
  "stage": "research",
  "subject": "한국 연안 소형 어선",
  "terms": [
   "한국 소형 어선",
   "연안 FRP 어선",
   "포구 정박 어선"
  ],
  "status": "no_usable",
  "queries": [
   [
    "한국 소형 어선",
    "포구 정박 어선"
   ]
  ],
  "candidate_urls": [
   "https://japanese.visitkorea.or.kr/public/images/2023/10/17/9198d946f1f64498b453b0382b1db704.jpg",
   "https://mblogthumb-phinf.pstatic.net/MjAyNDA3MDRfMjYy/MDAxNzIwMDY4OTA4NzIy.cjAP7ykzOya68pBb6Gjzd2KA-QlfEZp48VQNRpAwM4Ag.USxqlshFQnGOHVfS1381-1OjVB_aDtKTcolM4tLpQdAg.JPEG/0_%282%29.jpg?type=w400",
   "https://upload.wikimedia.org/wikipedia/commons/0/07/Vissersboot_in_Korea_III.jpg",
   "https://media.bunjang.co.kr/product/280899830_%7Bcnt%7D_1722407328_w%7Bres%7D.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 1
 },
 "groupbg::harbor_waiting_quay": {
  "input_fingerprint": "bafb57511444ef11",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "harbor_waiting_quay",
    "tags": [
     "S78sh2"
    ]
   },
   "context_sig": "c14eefab8dc8c4c7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두·생선 좌판·정박장, 출항 수역: 어둠이 깔린 항만 부두로 작은 어선들이 밧줄로 묶여 있다. (특징: 커다란 모포를 머리끝까지 뒤집어쓴 찰리의 실루엣; 생선 좌판의 생선과 밧줄을 묶고 있는 선원; 물살을 가르며 항구를 떠나는 크리스의 소형 모터 어선)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 커다란 모포를 뒤집어쓴 찰리와 현우가 부둣가에 걸터앉아있다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두·생선 좌판·정박장, 출항 수역: 어둠이 깔린 항만 부두로 작은 어선들이 밧줄로 묶여 있다. (특징: 커다란 모포를 머리끝까지 뒤집어쓴 찰리의 실루엣; 생선 좌판의 생선과 밧줄을 묶고 있는 선원; 물살을 가르며 항구를 떠나는 크리스의 소형 모터 어선)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 커다란 모포를 뒤집어쓴 찰리와 현우가 부둣가에 걸터앉아있다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_waiting_quay_bf7967.png",
  "asset_id": "d3dca578-711c-463b-ace2-899a2477cb77",
  "input_asset_ids": [
   "eeff0814-a4ae-456c-8b10-ac8bb5118b81"
  ],
  "origin_tag": "S78sh2",
  "place_text": "On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.",
  "origin_inputs": {
   "place_text": "On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.",
   "time_of_day_en": "dusk",
   "conti_asset_id": "eeff0814-a4ae-456c-8b10-ac8bb5118b81"
  }
 },
 "S78sh2::bgfirst_bg": {
  "input_fingerprint": "8397d829f61f6d46",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 커다란 회색 모포를 뒤집어쓴 찰리와 배낭을 멘 이현우가 부둣가에 나란히 걸터앉은 구도.\n\nLOCATION (lock): On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane descent on the landward side of the pier, holding a rear three-quarter wide view with a gentle downward tilt toward the seated pair. Place blanket-covered 찰리 at lower left and backpack-wearing 이현우 beside him nearer center, both perched on the pier edge with their seated bodies and hanging lower legs legible; leave the upper half open to the harbor and moored boats. 찰리 leans slightly forward while 이현우 turns his head farther across the harbor, their attention searching the distant boats rather than returning to the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pier edge (Occupied by the two seated companions) — Runs diagonally beneath the figures, separating the landward camera position from the water beyond; used as Anchors their shared seated position and establishes the harbor-facing direction; Moored boats (Stationary in the harbor) — Distributed as separate oblique hull views across the background; used as Supplies the visible area of their search without allowing one boat to overwhelm the pair; Harbor water (Undulating); used as Separates the pier from the boats and supplies depth behind the seated figures; Large gray blanket (Draped over 찰리) — Its back and side folds follow the seated robot's silhouette; used as Makes his concealment readable without obscuring the relationship between the companions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light and restrained contrast preserve the gray blanket and harbor depth without adding unsupported dock lights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 커다란 회색 모포를 뒤집어쓴 찰리와 배낭을 멘 이현우가 부둣가에 나란히 걸터앉은 구도.\n\nLOCATION (lock): On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane descent on the landward side of the pier, holding a rear three-quarter wide view with a gentle downward tilt toward the seated pair. Place blanket-covered 찰리 at lower left and backpack-wearing 이현우 beside him nearer center, both perched on the pier edge with their seated bodies and hanging lower legs legible; leave the upper half open to the harbor and moored boats. 찰리 leans slightly forward while 이현우 turns his head farther across the harbor, their attention searching the distant boats rather than returning to the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pier edge (Occupied by the two seated companions) — Runs diagonally beneath the figures, separating the landward camera position from the water beyond; used as Anchors their shared seated position and establishes the harbor-facing direction; Moored boats (Stationary in the harbor) — Distributed as separate oblique hull views across the background; used as Supplies the visible area of their search without allowing one boat to overwhelm the pair; Harbor water (Undulating); used as Separates the pier from the boats and supplies depth behind the seated figures; Large gray blanket (Draped over 찰리) — Its back and side folds follow the seated robot's silhouette; used as Makes his concealment readable without obscuring the relationship between the companions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light and restrained contrast preserve the gray blanket and harbor depth without adding unsupported dock lights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh2__bgfirst_bg.png",
  "asset_id": "d8c6b125-3b6c-44cd-8e72-bbf05a34e7c1",
  "input_asset_ids": [
   "eeff0814-a4ae-456c-8b10-ac8bb5118b81",
   "d3dca578-711c-463b-ace2-899a2477cb77"
  ]
 },
 "S78sh2": {
  "input_fingerprint": "77db8a085300fa9f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 회색 모포를 뒤집어쓴 찰리와 배낭을 멘 이현우가 부둣가에 나란히 걸터앉은 구도.\n\nLOCATION (lock): On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane descent on the landward side of the pier, holding a rear three-quarter wide view with a gentle downward tilt toward the seated pair. Place blanket-covered 찰리 at lower left and backpack-wearing 이현우 beside him nearer center, both perched on the pier edge with their seated bodies and hanging lower legs legible; leave the upper half open to the harbor and moored boats. 찰리 leans slightly forward while 이현우 turns his head farther across the harbor, their attention searching the distant boats rather than returning to the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pier edge (Occupied by the two seated companions) — Runs diagonally beneath the figures, separating the landward camera position from the water beyond; used as Anchors their shared seated position and establishes the harbor-facing direction; Moored boats (Stationary in the harbor) — Distributed as separate oblique hull views across the background; used as Supplies the visible area of their search without allowing one boat to overwhelm the pair; Harbor water (Undulating); used as Separates the pier from the boats and supplies depth behind the seated figures; Large gray blanket (Draped over 찰리) — Its back and side folds follow the seated robot's silhouette; used as Makes his concealment readable without obscuring the relationship between the companions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light and restrained contrast preserve the gray blanket and harbor depth without adding unsupported dock lights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats lie moored at Mokpo harbor in the evening, with the sea moving beside the quay. Charlie sits wrapped in a large blanket, retaining his backpack, damaged body and B-200's recovered chest component beneath the covering. 이현우: He sits on the quay with his backpack and retains his earlier wounds and disheveled appearance.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 회색 모포를 뒤집어쓴 찰리와 배낭을 멘 이현우가 부둣가에 나란히 걸터앉은 구도.\n\nLOCATION (lock): On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane descent on the landward side of the pier, holding a rear three-quarter wide view with a gentle downward tilt toward the seated pair. Place blanket-covered 찰리 at lower left and backpack-wearing 이현우 beside him nearer center, both perched on the pier edge with their seated bodies and hanging lower legs legible; leave the upper half open to the harbor and moored boats. 찰리 leans slightly forward while 이현우 turns his head farther across the harbor, their attention searching the distant boats rather than returning to the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pier edge (Occupied by the two seated companions) — Runs diagonally beneath the figures, separating the landward camera position from the water beyond; used as Anchors their shared seated position and establishes the harbor-facing direction; Moored boats (Stationary in the harbor) — Distributed as separate oblique hull views across the background; used as Supplies the visible area of their search without allowing one boat to overwhelm the pair; Harbor water (Undulating); used as Separates the pier from the boats and supplies depth behind the seated figures; Large gray blanket (Draped over 찰리) — Its back and side folds follow the seated robot's silhouette; used as Makes his concealment readable without obscuring the relationship between the companions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light and restrained contrast preserve the gray blanket and harbor depth without adding unsupported dock lights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats lie moored at Mokpo harbor in the evening, with the sea moving beside the quay. Charlie sits wrapped in a large blanket, retaining his backpack, damaged body and B-200's recovered chest component beneath the covering. 이현우: He sits on the quay with his backpack and retains his earlier wounds and disheveled appearance.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 커다란 회색 모포를 뒤집어쓴 찰리와 배낭을 멘 이현우가 부둣가에 나란히 걸터앉은 구도.\n\nLOCATION (lock): On the harbor quay's edge at dusk, beside the moored boats and a nearby fish seller. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the crane descent on the landward side of the pier, holding a rear three-quarter wide view with a gentle downward tilt toward the seated pair. Place blanket-covered 찰리 at lower left and backpack-wearing 이현우 beside him nearer center, both perched on the pier edge with their seated bodies and hanging lower legs legible; leave the upper half open to the harbor and moored boats. 찰리 leans slightly forward while 이현우 turns his head farther across the harbor, their attention searching the distant boats rather than returning to the camera.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Pier edge (Occupied by the two seated companions) — Runs diagonally beneath the figures, separating the landward camera position from the water beyond; used as Anchors their shared seated position and establishes the harbor-facing direction; Moored boats (Stationary in the harbor) — Distributed as separate oblique hull views across the background; used as Supplies the visible area of their search without allowing one boat to overwhelm the pair; Harbor water (Undulating); used as Separates the pier from the boats and supplies depth behind the seated figures; Large gray blanket (Draped over 찰리) — Its back and side folds follow the seated robot's silhouette; used as Makes his concealment readable without obscuring the relationship between the companions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim evening ambient light and restrained contrast preserve the gray blanket and harbor depth without adding unsupported dock lights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Boats lie moored at Mokpo harbor in the evening, with the sea moving beside the quay. Charlie sits wrapped in a large blanket, retaining his backpack, damaged body and B-200's recovered chest component beneath the covering. 이현우: He sits on the quay with his backpack and retains his earlier wounds and disheveled appearance.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh2__bgfirst_bg.png",
     "asset_id": "d8c6b125-3b6c-44cd-8e72-bbf05a34e7c1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S78sh2.png",
     "asset_id": "eeff0814-a4ae-456c-8b10-ac8bb5118b81",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_waiting_quay_bf7967.png",
     "asset_id": "d3dca578-711c-463b-ace2-899a2477cb77",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 시선과 얼굴 방향이 항구에 정박된 배들을 향해 있음.",
    "built_space": "참조 사진의 부둣가 대각선 모서리, 왼쪽 천막, 항구와 배들이 정확한 위치에 재현되었으며 추가 인물은 없음.",
    "entities": "모포를 두른 찰리(로봇 머리 외형 일부 노출)와 배낭을 멘 이현우가 지시된 인상착의와 일치함.",
    "hard_violations": [],
    "physics": "부둣가 모서리에 안정적으로 걸터앉아 있으나, 이현우의 오른발이 턱 아래로 내려가지 않고 위로 올라와 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 항구 너머의 배들을 바라보고 있음.",
    "built_space": "부둣가 모서리와 배경은 잘 구현되었으나, 좌측 가판대 쪽에 사람이 서 있음.",
    "entities": "찰리와 이현우가 명시된 착장(모포, 배낭)으로 앉아 있으나 프롬프트에 없는 인물이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 프롬프트의 인물 규정을 위반하여 샷 텍스트에 없는 인물(왼쪽 가판대의 사람)을 임의로 생성함.",
     "[gpt-high] 등장인물을 찰리와 이현우로 한정하고 다른 사람을 추가하지 말라는 지시와 달리, 왼쪽 판매대에 상인 한 명을 추가했다."
    ],
    "physics": "두 인물이 부둣가 턱에 지지하여 다리를 자연스럽게 아래로 늘어뜨리고 앉아 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 조명, 배경을 훌륭하게 재현했으나, 이현우의 다리가 아래로 늘어뜨려지지 않고 위로 굽혀져 있어 자세 지시를 완벽히 따르지는 못했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "자세와 구도는 지시사항에 부합하나, 샷 텍스트에 존재하지 않는 추가 인물(상인)을 임의로 등장시켜 규정을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 시선과 얼굴 방향이 항구에 정박된 배들을 향해 있음.",
        "built_space": "참조 사진의 부둣가 대각선 모서리, 왼쪽 천막, 항구와 배들이 정확한 위치에 재현되었으며 추가 인물은 없음.",
        "entities": "모포를 두른 찰리(로봇 머리 외형 일부 노출)와 배낭을 멘 이현우가 지시된 인상착의와 일치함.",
        "hard_violations": [],
        "physics": "부둣가 모서리에 안정적으로 걸터앉아 있으나, 이현우의 오른발이 턱 아래로 내려가지 않고 위로 올라와 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 항구 너머의 배들을 바라보고 있음.",
        "built_space": "부둣가 모서리와 배경은 잘 구현되었으나, 좌측 가판대 쪽에 사람이 서 있음.",
        "entities": "찰리와 이현우가 명시된 착장(모포, 배낭)으로 앉아 있으나 프롬프트에 없는 인물이 추가됨.",
        "hard_violations": [
         "프롬프트의 인물 규정을 위반하여 샷 텍스트에 없는 인물(왼쪽 가판대의 사람)을 임의로 생성함."
        ],
        "physics": "두 인물이 부둣가 턱에 지지하여 다리를 자연스럽게 아래로 늘어뜨리고 앉아 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 앵글과 조명, 배경을 훌륭하게 재현했으나, 이현우의 다리가 아래로 늘어뜨려지지 않고 위로 굽혀져 있어 자세 지시를 완벽히 따르지는 못했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "자세와 구도는 지시사항에 부합하나, 샷 텍스트에 존재하지 않는 추가 인물(상인)을 임의로 등장시켜 규정을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 시선과 얼굴 방향이 항구에 정박된 배들을 향해 있음.",
        "built_space": "참조 사진의 부둣가 대각선 모서리, 왼쪽 천막, 항구와 배들이 정확한 위치에 재현되었으며 추가 인물은 없음.",
        "entities": "모포를 두른 찰리(로봇 머리 외형 일부 노출)와 배낭을 멘 이현우가 지시된 인상착의와 일치함.",
        "hard_violations": [],
        "physics": "부둣가 모서리에 안정적으로 걸터앉아 있으나, 이현우의 오른발이 턱 아래로 내려가지 않고 위로 올라와 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 항구 너머의 배들을 바라보고 있음.",
        "built_space": "부둣가 모서리와 배경은 잘 구현되었으나, 좌측 가판대 쪽에 사람이 서 있음.",
        "entities": "찰리와 이현우가 명시된 착장(모포, 배낭)으로 앉아 있으나 프롬프트에 없는 인물이 추가됨.",
        "hard_violations": [
         "프롬프트의 인물 규정을 위반하여 샷 텍스트에 없는 인물(왼쪽 가판대의 사람)을 임의로 생성함."
        ],
        "physics": "두 인물이 부둣가 턱에 지지하여 다리를 자연스럽게 아래로 늘어뜨리고 앉아 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "후방 사선 구도와 항구를 보는 방향은 맞지만, 허용되지 않은 생선 상인을 추가했으며 두 인물의 늘어뜨린 다리도 충분히 읽히지 않는다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "추가 인물 없이 왼쪽 찰리와 중앙의 배낭 멘 이현우를 정확히 배치하고 이현우의 매달린 다리까지 보여주지만, 찰리의 다리와 이현우의 손 동작은 불명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 상체를 앞으로 숙이고 머리를 수면 쪽으로 향한다. 이현우는 고개를 오른쪽으로 돌려 항구 오른쪽의 배들이 있는 방향을 보며, 카메라를 돌아보지 않는다. 눈 자체는 보이지 않아 특정 배를 응시하는지는 확인할 수 없다. 왼쪽에 추가된 상인은 작업대 쪽으로 고개를 숙이고 있다.",
        "built_space": "육지 쪽에서 두 인물의 뒤와 옆을 보는 와이드 구도다. 대각선 콘크리트 가장자리에 찰리가 왼쪽, 이현우가 그 옆 중앙에 앉아 있다. 전경 좌우에 계선주 두 개, 왼쪽에 천막 판매대 하나, 원경 왼쪽에 붉은 등대 하나가 있으며 참조의 배치와 대체로 맞는다. 배들은 좌우 가까운 곳과 중앙 원경에 나뉘어 있다. 다만 전경 부두가 다리를 가려 이현우의 하퇴 일부만 보이고 찰리의 다리는 보이지 않는다.",
        "entities": "회색 모포를 두른 찰리와 배낭을 멘 이현우 외에 왼쪽 판매대 뒤 성인 상인 한 명이 보인다. 찰리는 베이지색 각진 기계 머리와 둥근 측면 부품이 드러나 참조의 기계 정체성에 부합한다. 얼굴, 팔, 가슴 부품과 모포 안의 소지품은 가려져 확인할 수 없다. 이현우는 검은 헝클어진 머리, 마른 체격, 오염된 어두운 옷의 젊은 남성으로 보이나 얼굴과 인이어 무전기는 판별하기 어렵다. 손이 특정 물건을 잡거나 사용하는 모습은 명확하지 않다.",
        "hard_violations": [
         "등장인물을 찰리와 이현우로 한정하고 다른 사람을 추가하지 말라는 지시와 달리, 왼쪽 판매대에 상인 한 명을 추가했다."
        ],
        "physics": "두 인물의 엉덩이는 콘크리트 부두에 지지되어 있고, 모포는 찰리의 몸을 따라 내려와 부두 위에 놓인다. 이현우의 배낭은 어깨끈으로 등에 고정되어 있다. 보이는 다리 부분은 앉은 자세에서 자연스럽게 아래로 내려간다. 배들은 수면에 떠 있고 가까운 배에는 계류줄이 보인다. 지지 없이 공중에 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "찰리는 머리와 상체를 조금 앞으로 기울여 항구 쪽을 향한다. 이현우는 머리를 오른쪽 먼 항구로 돌려 오른쪽 정박선 무리 방향을 본다. 둘 다 카메라를 향하지 않으며, 항구를 함께 살피는 방향 관계가 성립한다.",
        "built_space": "육지 쪽의 완만한 내려다보기와 후방 사선 와이드 구도이며, 찰리는 하단 왼쪽, 이현우는 중앙 가까이에 나란히 앉아 있다. 두 사람 아래로 부두 가장자리가 대각선으로 이어지고 이현우 아래에는 수직 콘크리트 벽면도 드러난다. 계선주 두 개가 전경 좌우에, 천막 판매대 하나가 왼쪽에, 붉은 등대 하나가 먼 왼쪽에 있다. 상반부에는 항구와 여러 방향의 정박선이 펼쳐져 참조 장소를 잘 유지한다. 이현우의 하퇴와 신발은 보이지만 찰리의 다리는 모포에 가려져 있다.",
        "entities": "분명하게 보이는 인물은 찰리와 이현우 두 명뿐이다. 찰리의 베이지색 기계 머리와 넓은 모포 실루엣은 참조와 부합하며, 가려진 얼굴과 가슴 및 소지품은 확인할 수 없다. 이현우는 짧고 흐트러진 검은 머리, 마른 체격, 낡고 어두운 셔츠와 바지, 등에 멘 배낭을 갖췄다. 후면이라 정확한 얼굴 일치, 상처와 인이어는 확인하기 어렵다. 손 일부가 무릎 부근에 보이지만 특정 대상에 대한 파지나 조작은 식별되지 않는다.",
        "hard_violations": [],
        "physics": "두 인물은 부두 가장자리에 엉덩이를 두고 체중을 싣고 있다. 이현우의 보이는 다리는 무릎에서 꺾여 벽면 바깥으로 자연스럽게 내려가며 신발까지 이어진다. 찰리의 모포는 몸과 부두 표면에 지지되어 두꺼운 주름을 만든다. 배낭은 어깨끈으로 고정되고 선박은 수면의 부력과 계류줄로 지지된다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "후방 사선 구도와 항구를 보는 방향은 맞지만, 허용되지 않은 생선 상인을 추가했으며 두 인물의 늘어뜨린 다리도 충분히 읽히지 않는다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "추가 인물 없이 왼쪽 찰리와 중앙의 배낭 멘 이현우를 정확히 배치하고 이현우의 매달린 다리까지 보여주지만, 찰리의 다리와 이현우의 손 동작은 불명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 상체를 앞으로 숙이고 머리를 수면 쪽으로 향한다. 이현우는 고개를 오른쪽으로 돌려 항구 오른쪽의 배들이 있는 방향을 보며, 카메라를 돌아보지 않는다. 눈 자체는 보이지 않아 특정 배를 응시하는지는 확인할 수 없다. 왼쪽에 추가된 상인은 작업대 쪽으로 고개를 숙이고 있다.",
        "built_space": "육지 쪽에서 두 인물의 뒤와 옆을 보는 와이드 구도다. 대각선 콘크리트 가장자리에 찰리가 왼쪽, 이현우가 그 옆 중앙에 앉아 있다. 전경 좌우에 계선주 두 개, 왼쪽에 천막 판매대 하나, 원경 왼쪽에 붉은 등대 하나가 있으며 참조의 배치와 대체로 맞는다. 배들은 좌우 가까운 곳과 중앙 원경에 나뉘어 있다. 다만 전경 부두가 다리를 가려 이현우의 하퇴 일부만 보이고 찰리의 다리는 보이지 않는다.",
        "entities": "회색 모포를 두른 찰리와 배낭을 멘 이현우 외에 왼쪽 판매대 뒤 성인 상인 한 명이 보인다. 찰리는 베이지색 각진 기계 머리와 둥근 측면 부품이 드러나 참조의 기계 정체성에 부합한다. 얼굴, 팔, 가슴 부품과 모포 안의 소지품은 가려져 확인할 수 없다. 이현우는 검은 헝클어진 머리, 마른 체격, 오염된 어두운 옷의 젊은 남성으로 보이나 얼굴과 인이어 무전기는 판별하기 어렵다. 손이 특정 물건을 잡거나 사용하는 모습은 명확하지 않다.",
        "hard_violations": [
         "등장인물을 찰리와 이현우로 한정하고 다른 사람을 추가하지 말라는 지시와 달리, 왼쪽 판매대에 상인 한 명을 추가했다."
        ],
        "physics": "두 인물의 엉덩이는 콘크리트 부두에 지지되어 있고, 모포는 찰리의 몸을 따라 내려와 부두 위에 놓인다. 이현우의 배낭은 어깨끈으로 등에 고정되어 있다. 보이는 다리 부분은 앉은 자세에서 자연스럽게 아래로 내려간다. 배들은 수면에 떠 있고 가까운 배에는 계류줄이 보인다. 지지 없이 공중에 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "찰리는 머리와 상체를 조금 앞으로 기울여 항구 쪽을 향한다. 이현우는 머리를 오른쪽 먼 항구로 돌려 오른쪽 정박선 무리 방향을 본다. 둘 다 카메라를 향하지 않으며, 항구를 함께 살피는 방향 관계가 성립한다.",
        "built_space": "육지 쪽의 완만한 내려다보기와 후방 사선 와이드 구도이며, 찰리는 하단 왼쪽, 이현우는 중앙 가까이에 나란히 앉아 있다. 두 사람 아래로 부두 가장자리가 대각선으로 이어지고 이현우 아래에는 수직 콘크리트 벽면도 드러난다. 계선주 두 개가 전경 좌우에, 천막 판매대 하나가 왼쪽에, 붉은 등대 하나가 먼 왼쪽에 있다. 상반부에는 항구와 여러 방향의 정박선이 펼쳐져 참조 장소를 잘 유지한다. 이현우의 하퇴와 신발은 보이지만 찰리의 다리는 모포에 가려져 있다.",
        "entities": "분명하게 보이는 인물은 찰리와 이현우 두 명뿐이다. 찰리의 베이지색 기계 머리와 넓은 모포 실루엣은 참조와 부합하며, 가려진 얼굴과 가슴 및 소지품은 확인할 수 없다. 이현우는 짧고 흐트러진 검은 머리, 마른 체격, 낡고 어두운 셔츠와 바지, 등에 멘 배낭을 갖췄다. 후면이라 정확한 얼굴 일치, 상처와 인이어는 확인하기 어렵다. 손 일부가 무릎 부근에 보이지만 특정 대상에 대한 파지나 조작은 식별되지 않는다.",
        "hard_violations": [],
        "physics": "두 인물은 부두 가장자리에 엉덩이를 두고 체중을 싣고 있다. 이현우의 보이는 다리는 무릎에서 꺾여 벽면 바깥으로 자연스럽게 내려가며 신발까지 이어진다. 찰리의 모포는 몸과 부두 표면에 지지되어 두꺼운 주름을 만든다. 배낭은 어깨끈으로 고정되고 선박은 수면의 부력과 계류줄로 지지된다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트의 인물 규정을 위반하여 샷 텍스트에 없는 인물(왼쪽 가판대의 사람)을 임의로 생성함.",
     "[gpt-high] 등장인물을 찰리와 이현우로 한정하고 다른 사람을 추가하지 말라는 지시와 달리, 왼쪽 판매대에 상인 한 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 앵글과 조명, 배경을 훌륭하게 재현했으나, 이현우의 다리가 아래로 늘어뜨려지지 않고 위로 굽혀져 있어 자세 지시를 완벽히 따르지는 못했습니다."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "자세와 구도는 지시사항에 부합하나, 샷 텍스트에 존재하지 않는 추가 인물(상인)을 임의로 등장시켜 규정을 크게 위반했습니다.  ★위반: [gemini-pro] 프롬프트의 인물 규정을 위반하여 샷 텍스트에 없는 인물(왼쪽 가판대의 사람)을 임의로 생성함. / [gpt-high] 등장인물을 찰리와 이현우로 한정하고 다른 사람을 추가하지 말라는 지시와 달리, 왼쪽 판매대에 상인 한 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_harbor_waiting_quay_bf7967.png",
    "asset_id": "d3dca578-711c-463b-ace2-899a2477cb77",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0df0-e1d0-74b3-91cd-e1a3401439ab",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh2__bgfirst_bg.png",
   "bg_asset_id": "d8c6b125-3b6c-44cd-8e72-bbf05a34e7c1",
   "bg_record_key": "S78sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "harbor_waiting_quay",
   "groupbg_asset_id": "d3dca578-711c-463b-ace2-899a2477cb77"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S78sh15::signage": {
  "fp": "a9f4f03434528d04",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::fishing_boat_berth": {
  "input_fingerprint": "bb8086a12c7a869e",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "fishing_boat_berth",
    "tags": [
     "S78sh15"
    ]
   },
   "context_sig": "08814f13e7433db9"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the boarding point beside an old fishing boat moored in the harbor.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두·생선 좌판·정박장, 출항 수역: 어둠이 깔린 항만 부두로 작은 어선들이 밧줄로 묶여 있다. (특징: 커다란 모포를 머리끝까지 뒤집어쓴 찰리의 실루엣; 생선 좌판의 생선과 밧줄을 묶고 있는 선원; 물살을 가르며 항구를 떠나는 크리스의 소형 모터 어선)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선원이 어디론가로 현우를 인도한다.\n- 낡은 배를 정비하는 40대 초반의 크리스와 몇몇 선원들 보인다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the boarding point beside an old fishing boat moored in the harbor.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n목포 항구 부두·생선 좌판·정박장, 출항 수역: 어둠이 깔린 항만 부두로 작은 어선들이 밧줄로 묶여 있다. (특징: 커다란 모포를 머리끝까지 뒤집어쓴 찰리의 실루엣; 생선 좌판의 생선과 밧줄을 묶고 있는 선원; 물살을 가르며 항구를 떠나는 크리스의 소형 모터 어선)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선원이 어디론가로 현우를 인도한다.\n- 낡은 배를 정비하는 40대 초반의 크리스와 몇몇 선원들 보인다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_fishing_boat_berth_e53793.png",
  "asset_id": "745a1b05-7af3-4956-b2c1-4ad5920aefe9",
  "input_asset_ids": [
   "2e79fe5f-bf77-4dde-be5b-1a5074cd2bad"
  ],
  "origin_tag": "S78sh15",
  "place_text": "At the boarding point beside an old fishing boat moored in the harbor.",
  "origin_inputs": {
   "place_text": "At the boarding point beside an old fishing boat moored in the harbor.",
   "time_of_day_en": "dusk",
   "conti_asset_id": "2e79fe5f-bf77-4dde-be5b-1a5074cd2bad"
  }
 },
 "S78sh15::bgfirst_bg": {
  "input_fingerprint": "31082bdb23c4ebd0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 크리스가 자신의 배 안쪽을 향해 턱을 위로 든 채 시선을 맞춰 타라는 신호를 보내고 멈춘 찰나.\n\nLOCATION (lock): At the boarding point beside an old fishing boat moored in the harbor.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly beside the boat just behind 이현우's shoulder, with the lens below 크리스's eye line and a slight upward tilt across, not along, his frontal axis. Keep a narrow sliver of 이현우's shoulder and turned head at the left edge, placing 크리스's upper body at center-right and a limited portion of the boat interior farther right. Capture 크리스 with his chin lifted toward that interior while his eyes remain lowered toward 이현우, who looks up at him; the completed chin gesture, rather than a new distance or lighting change, carries the permission to board.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Boat interior indicated by 크리스's chin gesture in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 크리스's boat (Aged and still moored before boarding) — A partial view into the boat sits to the right of 크리스, in the direction of his chin gesture; used as Makes the invitation's destination visible while keeping the boat subordinate to the human exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim evening ambient exposure with controlled facial contrast so the subtle boarding signal reads without an added light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 크리스가 자신의 배 안쪽을 향해 턱을 위로 든 채 시선을 맞춰 타라는 신호를 보내고 멈춘 찰나.\n\nLOCATION (lock): At the boarding point beside an old fishing boat moored in the harbor.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly beside the boat just behind 이현우's shoulder, with the lens below 크리스's eye line and a slight upward tilt across, not along, his frontal axis. Keep a narrow sliver of 이현우's shoulder and turned head at the left edge, placing 크리스's upper body at center-right and a limited portion of the boat interior farther right. Capture 크리스 with his chin lifted toward that interior while his eyes remain lowered toward 이현우, who looks up at him; the completed chin gesture, rather than a new distance or lighting change, carries the permission to board.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Boat interior indicated by 크리스's chin gesture in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 크리스's boat (Aged and still moored before boarding) — A partial view into the boat sits to the right of 크리스, in the direction of his chin gesture; used as Makes the invitation's destination visible while keeping the boat subordinate to the human exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim evening ambient exposure with controlled facial contrast so the subtle boarding signal reads without an added light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh15__bgfirst_bg.png",
  "asset_id": "6d833f7f-fb0d-4dbe-97c6-eed2908c7762",
  "input_asset_ids": [
   "2e79fe5f-bf77-4dde-be5b-1a5074cd2bad",
   "745a1b05-7af3-4956-b2c1-4ad5920aefe9"
  ]
 },
 "S78sh15": {
  "input_fingerprint": "b1e07fbcc1045267",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 자신의 배 안쪽을 향해 턱을 위로 든 채 시선을 맞춰 타라는 신호를 보내고 멈춘 찰나.\n\nLOCATION (lock): At the boarding point beside an old fishing boat moored in the harbor. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly beside the boat just behind 이현우's shoulder, with the lens below 크리스's eye line and a slight upward tilt across, not along, his frontal axis. Keep a narrow sliver of 이현우's shoulder and turned head at the left edge, placing 크리스's upper body at center-right and a limited portion of the boat interior farther right. Capture 크리스 with his chin lifted toward that interior while his eyes remain lowered toward 이현우, who looks up at him; the completed chin gesture, rather than a new distance or lighting change, carries the permission to board.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Boat interior indicated by 크리스's chin gesture in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 크리스's boat (Aged and still moored before boarding) — A partial view into the boat sits to the right of 크리스, in the direction of his chin gesture; used as Makes the invitation's destination visible while keeping the boat subordinate to the human exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim evening ambient exposure with controlled facial contrast so the subtle boarding signal reads without an added light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old boat remains moored at the evening harbor. Charlie is still covered by the large blanket, now drawn over his face, and retains his backpack, unrepaired damage and B-200's component. 이현우: He stands beside the old boat with his backpack, still visibly dirty and wounded. 크리스: He remains beside the old boat he was servicing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (한국인 남성, 40대 초반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 자신의 배 안쪽을 향해 턱을 위로 든 채 시선을 맞춰 타라는 신호를 보내고 멈춘 찰나.\n\nLOCATION (lock): At the boarding point beside an old fishing boat moored in the harbor. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly beside the boat just behind 이현우's shoulder, with the lens below 크리스's eye line and a slight upward tilt across, not along, his frontal axis. Keep a narrow sliver of 이현우's shoulder and turned head at the left edge, placing 크리스's upper body at center-right and a limited portion of the boat interior farther right. Capture 크리스 with his chin lifted toward that interior while his eyes remain lowered toward 이현우, who looks up at him; the completed chin gesture, rather than a new distance or lighting change, carries the permission to board.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Boat interior indicated by 크리스's chin gesture in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 크리스's boat (Aged and still moored before boarding) — A partial view into the boat sits to the right of 크리스, in the direction of his chin gesture; used as Makes the invitation's destination visible while keeping the boat subordinate to the human exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim evening ambient exposure with controlled facial contrast so the subtle boarding signal reads without an added light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old boat remains moored at the evening harbor. Charlie is still covered by the large blanket, now drawn over his face, and retains his backpack, unrepaired damage and B-200's component. 이현우: He stands beside the old boat with his backpack, still visibly dirty and wounded. 크리스: He remains beside the old boat he was servicing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (한국인 남성, 40대 초반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 크리스가 자신의 배 안쪽을 향해 턱을 위로 든 채 시선을 맞춰 타라는 신호를 보내고 멈춘 찰나.\n\nLOCATION (lock): At the boarding point beside an old fishing boat moored in the harbor. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly beside the boat just behind 이현우's shoulder, with the lens below 크리스's eye line and a slight upward tilt across, not along, his frontal axis. Keep a narrow sliver of 이현우's shoulder and turned head at the left edge, placing 크리스's upper body at center-right and a limited portion of the boat interior farther right. Capture 크리스 with his chin lifted toward that interior while his eyes remain lowered toward 이현우, who looks up at him; the completed chin gesture, rather than a new distance or lighting change, carries the permission to board.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Boat interior indicated by 크리스's chin gesture in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 크리스's boat (Aged and still moored before boarding) — A partial view into the boat sits to the right of 크리스, in the direction of his chin gesture; used as Makes the invitation's destination visible while keeping the boat subordinate to the human exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim evening ambient exposure with controlled facial contrast so the subtle boarding signal reads without an added light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old boat remains moored at the evening harbor. Charlie is still covered by the large blanket, now drawn over his face, and retains his backpack, unrepaired damage and B-200's component. 이현우: He stands beside the old boat with his backpack, still visibly dirty and wounded. 크리스: He remains beside the old boat he was servicing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 크리스 (한국인 남성, 40대 초반의 얼굴, 짧은 검은 머리); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh15__bgfirst_bg.png",
     "asset_id": "6d833f7f-fb0d-4dbe-97c6-eed2908c7762",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S78sh15.png",
     "asset_id": "2e79fe5f-bf77-4dde-be5b-1a5074cd2bad",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 크리스: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1188250>",
     "asset_id": "32d2abe2-c7f3-44b7-b735-acbe733675b1",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_fishing_boat_berth_e53793.png",
     "asset_id": "745a1b05-7af3-4956-b2c1-4ad5920aefe9",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 크리스: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1188250>",
     "asset_id": "32d2abe2-c7f3-44b7-b735-acbe733675b1",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "크리스는 배 쪽(오른쪽)을 향해 손을 뻗어 가리키고 있으며, 시선 역시 배 쪽이나 먼 곳을 향하고 있어 이현우를 내려다보지 않습니다. 이현우는 크리스를 바라봅니다.",
    "built_space": "항구 부두에 정박된 어선과 배경의 등대 등 주어진 장소 레퍼런스의 구조와 일치하며, 카메라 위치와 인물 배치가 적절합니다.",
    "entities": "크리스와 흙먼지가 묻은 배낭을 멘 이현우 모두 주어진 캐릭터 레퍼런스와 일치합니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 크리스의 팔을 뻗은 동작에 무리가 없습니다."
   },
   {
    "label": "B",
    "direction": "크리스는 턱을 배 안쪽(오른쪽)으로 치켜들고 있으면서 시선은 정확히 아래로 내리깔아 이현우를 향하고 있습니다. 이현우는 크리스를 올려다봅니다.",
    "built_space": "부두와 정박된 어선, 어선 내부의 모습 및 배경의 구조물들이 레퍼런스와 일치하며 요구된 프레임 구성을 잘 따릅니다.",
    "entities": "크리스의 얼굴과 체형, 이현우의 뒷모습과 배낭 등 지시된 인물들의 특징이 정확하게 반영되었습니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 부두 위에 자연스럽게 서 있으며 자세에 물리적 오류가 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "크리스가 턱 대신 손으로 배를 가리키고 있으며, 시선이 이현우를 향해 있지 않아 핵심적인 행동 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "턱을 배 쪽으로 들어 올리면서 시선은 이현우를 향해 내리까는 크리스의 미묘한 표정과 시선 교환을 프롬프트대로 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "크리스는 배 쪽(오른쪽)을 향해 손을 뻗어 가리키고 있으며, 시선 역시 배 쪽이나 먼 곳을 향하고 있어 이현우를 내려다보지 않습니다. 이현우는 크리스를 바라봅니다.",
        "built_space": "항구 부두에 정박된 어선과 배경의 등대 등 주어진 장소 레퍼런스의 구조와 일치하며, 카메라 위치와 인물 배치가 적절합니다.",
        "entities": "크리스와 흙먼지가 묻은 배낭을 멘 이현우 모두 주어진 캐릭터 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 크리스의 팔을 뻗은 동작에 무리가 없습니다."
       },
       {
        "label": "B",
        "direction": "크리스는 턱을 배 안쪽(오른쪽)으로 치켜들고 있으면서 시선은 정확히 아래로 내리깔아 이현우를 향하고 있습니다. 이현우는 크리스를 올려다봅니다.",
        "built_space": "부두와 정박된 어선, 어선 내부의 모습 및 배경의 구조물들이 레퍼런스와 일치하며 요구된 프레임 구성을 잘 따릅니다.",
        "entities": "크리스의 얼굴과 체형, 이현우의 뒷모습과 배낭 등 지시된 인물들의 특징이 정확하게 반영되었습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 부두 위에 자연스럽게 서 있으며 자세에 물리적 오류가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "크리스가 턱 대신 손으로 배를 가리키고 있으며, 시선이 이현우를 향해 있지 않아 핵심적인 행동 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "턱을 배 쪽으로 들어 올리면서 시선은 이현우를 향해 내리까는 크리스의 미묘한 표정과 시선 교환을 프롬프트대로 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "크리스는 배 쪽(오른쪽)을 향해 손을 뻗어 가리키고 있으며, 시선 역시 배 쪽이나 먼 곳을 향하고 있어 이현우를 내려다보지 않습니다. 이현우는 크리스를 바라봅니다.",
        "built_space": "항구 부두에 정박된 어선과 배경의 등대 등 주어진 장소 레퍼런스의 구조와 일치하며, 카메라 위치와 인물 배치가 적절합니다.",
        "entities": "크리스와 흙먼지가 묻은 배낭을 멘 이현우 모두 주어진 캐릭터 레퍼런스와 일치합니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 크리스의 팔을 뻗은 동작에 무리가 없습니다."
       },
       {
        "label": "B",
        "direction": "크리스는 턱을 배 안쪽(오른쪽)으로 치켜들고 있으면서 시선은 정확히 아래로 내리깔아 이현우를 향하고 있습니다. 이현우는 크리스를 올려다봅니다.",
        "built_space": "부두와 정박된 어선, 어선 내부의 모습 및 배경의 구조물들이 레퍼런스와 일치하며 요구된 프레임 구성을 잘 따릅니다.",
        "entities": "크리스의 얼굴과 체형, 이현우의 뒷모습과 배낭 등 지시된 인물들의 특징이 정확하게 반영되었습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 부두 위에 자연스럽게 서 있으며 자세에 물리적 오류가 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낮은 시점의 미디엄 숏과 턱을 든 채 이현우를 바라보는 정지 순간은 더 충실하지만, 턱이 오른쪽 배 안을 가리키는 방향성과 왼쪽 인물의 좁은 크롭은 미흡하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "턱짓 대신 크게 뻗은 손으로 승선을 권하고, 허벅지까지 넓어진 구도와 과도한 전경 인물 면적으로 지정된 순간과 프레이밍에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "크리스는 턱을 올리고 눈을 화면 왼쪽의 이현우 쪽으로 내려 바라본다. 이현우도 고개를 돌려 크리스를 올려다본다. 다만 크리스의 얼굴과 턱 앞쪽은 여전히 이현우 쪽에 가까워, 화면 오른쪽 선실 내부를 턱으로 지목한다는 방향은 명확하지 않다. 손으로 대신 가리키는 동작은 없다.",
        "built_space": "오른쪽에는 낡은 흰색 선실 하나, 열린 출입구 하나, 일부 보이는 조타륜 하나, 파란 난간과 상자들이 있다. 뒤에는 부두, 정박 어선들과 등대 하나가 보이며 장소 참조의 재료와 배치를 대체로 유지한다. 크리스는 배 난간 옆 중앙 오른쪽에 있고 이현우는 왼쪽 전경에 있다. 눈높이보다 낮은 카메라와 상반신 중심 구도는 적합하지만, 이현우의 머리와 어깨가 화면 왼쪽 약 3분의 1을 차지해 요구한 가느다란 가장자리가 아니다. 불가능한 반사나 명백히 중복된 주요 설비는 보이지 않는다.",
        "entities": "보이는 사람은 크리스와 이현우 두 명뿐이다. 크리스는 짧은 검은 머리의 중년 동아시아계 남성으로 참조 얼굴과 대체로 닮았지만, 참조의 정장 대신 오염된 후드 작업복과 멜빵 작업복을 입었다. 이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 옆얼굴에 오염이나 상처가 보이며, 참조의 남색 티셔츠 대신 어두운 겉옷을 입었다. 얼굴이 일부만 보여 정확한 동일인 여부에는 한계가 있다. 배낭으로 이어지는 듯한 어깨끈은 일부 보인다. 낡은 어선과 황혼 항구는 부합하며, 찰리나 추가 인물 및 삽입 문자는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 부두에서 상체를 세우고 서 있는 자세로 읽힌다. 발은 프레임 밖이지만 부양이나 불가능한 체중 지지의 징후는 없다. 크리스의 턱 들기와 목 회전은 가능한 동작이다. 배의 밧줄은 난간 부근에 걸리고 이어져 있으며 선실 설비와 상자는 배 구조물 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "크리스의 눈과 얼굴은 왼쪽 전경의 이현우를 향하고, 이현우는 크리스를 올려다본다. 턱은 조금 올라가 있지만 오른쪽 배 내부를 명확히 지목하지 않는다. 대신 오른쪽으로 길게 뻗은 팔과 펼친 손바닥이 선실 출입구를 안내한다. 목적지는 손짓으로 분명해졌으나 요구한 완료된 턱짓이 동작의 중심이 아니다.",
        "built_space": "오른쪽에 흰색 선실 하나와 출입구 하나, 조타륜 하나가 있고, 그 앞에는 밧줄이 감긴 계류 기둥 하나와 파란 난간이 있다. 선실의 전구와 파란 상자, 붉은 통, 뒤쪽 등대와 부두 지붕 등은 장소 참조와 대체로 일치한다. 두 사람 모두 부두 쪽에 위치한다. 그러나 크리스가 허벅지까지 보이고 이현우의 등과 배낭이 왼쪽을 크게 점유해, 지정한 상반신 중심 미디엄 숏과 좁은 어깨 너머 구도보다 넓다. 명백한 설비 중복이나 불가능한 반사는 없다.",
        "entities": "크리스와 이현우 외의 인물은 보이지 않는다. 크리스는 짧은 검은 머리의 중년 동아시아계 남성으로 참조 얼굴에 대체로 부합하나, 정장 대신 어두운 작업 재킷과 티셔츠를 입었다. 이현우는 헝클어진 검은 머리와 젊은 옆얼굴을 보이며 배낭을 메고 있다. 겉옷은 심하게 오염되어 있지만 얼굴의 부상은 이 구도에서 분명하지 않다. 참조의 티셔츠와 달리 후드 겉옷을 입었다. 낡은 정박 어선과 황혼 배경은 부합하며 추가 문자나 찰리는 나타나지 않는다.",
        "hard_violations": [],
        "physics": "크리스는 부두에 서서 한쪽 팔을 어깨와 팔꿈치로 지탱해 선실 쪽으로 뻗고 있다. 펼친 손의 자세는 물리적으로 가능하며, 문제는 불가능한 동작이 아니라 지시와 다른 동작이라는 점이다. 이현우의 배낭은 어깨끈으로 지지된다. 두 사람의 발은 화면 밖이지만 부양을 시사하지 않는다. 계류 밧줄은 기둥에 감겨 있고 상자와 통은 배 내부 바닥이나 선반에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낮은 시점의 미디엄 숏과 턱을 든 채 이현우를 바라보는 정지 순간은 더 충실하지만, 턱이 오른쪽 배 안을 가리키는 방향성과 왼쪽 인물의 좁은 크롭은 미흡하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "턱짓 대신 크게 뻗은 손으로 승선을 권하고, 허벅지까지 넓어진 구도와 과도한 전경 인물 면적으로 지정된 순간과 프레이밍에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "크리스는 턱을 올리고 눈을 화면 왼쪽의 이현우 쪽으로 내려 바라본다. 이현우도 고개를 돌려 크리스를 올려다본다. 다만 크리스의 얼굴과 턱 앞쪽은 여전히 이현우 쪽에 가까워, 화면 오른쪽 선실 내부를 턱으로 지목한다는 방향은 명확하지 않다. 손으로 대신 가리키는 동작은 없다.",
        "built_space": "오른쪽에는 낡은 흰색 선실 하나, 열린 출입구 하나, 일부 보이는 조타륜 하나, 파란 난간과 상자들이 있다. 뒤에는 부두, 정박 어선들과 등대 하나가 보이며 장소 참조의 재료와 배치를 대체로 유지한다. 크리스는 배 난간 옆 중앙 오른쪽에 있고 이현우는 왼쪽 전경에 있다. 눈높이보다 낮은 카메라와 상반신 중심 구도는 적합하지만, 이현우의 머리와 어깨가 화면 왼쪽 약 3분의 1을 차지해 요구한 가느다란 가장자리가 아니다. 불가능한 반사나 명백히 중복된 주요 설비는 보이지 않는다.",
        "entities": "보이는 사람은 크리스와 이현우 두 명뿐이다. 크리스는 짧은 검은 머리의 중년 동아시아계 남성으로 참조 얼굴과 대체로 닮았지만, 참조의 정장 대신 오염된 후드 작업복과 멜빵 작업복을 입었다. 이현우는 헝클어진 검은 머리의 젊은 동아시아계 남성으로 옆얼굴에 오염이나 상처가 보이며, 참조의 남색 티셔츠 대신 어두운 겉옷을 입었다. 얼굴이 일부만 보여 정확한 동일인 여부에는 한계가 있다. 배낭으로 이어지는 듯한 어깨끈은 일부 보인다. 낡은 어선과 황혼 항구는 부합하며, 찰리나 추가 인물 및 삽입 문자는 없다.",
        "hard_violations": [],
        "physics": "두 사람은 부두에서 상체를 세우고 서 있는 자세로 읽힌다. 발은 프레임 밖이지만 부양이나 불가능한 체중 지지의 징후는 없다. 크리스의 턱 들기와 목 회전은 가능한 동작이다. 배의 밧줄은 난간 부근에 걸리고 이어져 있으며 선실 설비와 상자는 배 구조물 위에 놓여 있다. 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "크리스의 눈과 얼굴은 왼쪽 전경의 이현우를 향하고, 이현우는 크리스를 올려다본다. 턱은 조금 올라가 있지만 오른쪽 배 내부를 명확히 지목하지 않는다. 대신 오른쪽으로 길게 뻗은 팔과 펼친 손바닥이 선실 출입구를 안내한다. 목적지는 손짓으로 분명해졌으나 요구한 완료된 턱짓이 동작의 중심이 아니다.",
        "built_space": "오른쪽에 흰색 선실 하나와 출입구 하나, 조타륜 하나가 있고, 그 앞에는 밧줄이 감긴 계류 기둥 하나와 파란 난간이 있다. 선실의 전구와 파란 상자, 붉은 통, 뒤쪽 등대와 부두 지붕 등은 장소 참조와 대체로 일치한다. 두 사람 모두 부두 쪽에 위치한다. 그러나 크리스가 허벅지까지 보이고 이현우의 등과 배낭이 왼쪽을 크게 점유해, 지정한 상반신 중심 미디엄 숏과 좁은 어깨 너머 구도보다 넓다. 명백한 설비 중복이나 불가능한 반사는 없다.",
        "entities": "크리스와 이현우 외의 인물은 보이지 않는다. 크리스는 짧은 검은 머리의 중년 동아시아계 남성으로 참조 얼굴에 대체로 부합하나, 정장 대신 어두운 작업 재킷과 티셔츠를 입었다. 이현우는 헝클어진 검은 머리와 젊은 옆얼굴을 보이며 배낭을 메고 있다. 겉옷은 심하게 오염되어 있지만 얼굴의 부상은 이 구도에서 분명하지 않다. 참조의 티셔츠와 달리 후드 겉옷을 입었다. 낡은 정박 어선과 황혼 배경은 부합하며 추가 문자나 찰리는 나타나지 않는다.",
        "hard_violations": [],
        "physics": "크리스는 부두에 서서 한쪽 팔을 어깨와 팔꿈치로 지탱해 선실 쪽으로 뻗고 있다. 펼친 손의 자세는 물리적으로 가능하며, 문제는 불가능한 동작이 아니라 지시와 다른 동작이라는 점이다. 이현우의 배낭은 어깨끈으로 지지된다. 두 사람의 발은 화면 밖이지만 부양을 시사하지 않는다. 계류 밧줄은 기둥에 감겨 있고 상자와 통은 배 내부 바닥이나 선반에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.167,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.167,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1167,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1167,
    "verdict_ko": "크리스가 턱 대신 손으로 배를 가리키고 있으며, 시선이 이현우를 향해 있지 않아 핵심적인 행동 지시를 위반했습니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "턱을 배 쪽으로 들어 올리면서 시선은 이현우를 향해 내리까는 크리스의 미묘한 표정과 시선 교환을 프롬프트대로 정확하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_fishing_boat_berth_e53793.png",
    "asset_id": "745a1b05-7af3-4956-b2c1-4ad5920aefe9",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 크리스: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1188250>",
    "asset_id": "32d2abe2-c7f3-44b7-b735-acbe733675b1",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0dfb-d2aa-7077-80dd-cc01bfd0c9ea",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S78sh15__bgfirst_bg.png",
   "bg_asset_id": "6d833f7f-fb0d-4dbe-97c6-eed2908c7762",
   "bg_record_key": "S78sh15::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "fishing_boat_berth",
   "groupbg_asset_id": "745a1b05-7af3-4956-b2c1-4ad5920aefe9"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S78sh19::signage": {
  "fp": "9fffa64f69b24533",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S78sh19": {
  "input_fingerprint": "c5ca0b500755a18d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우와 찰리를 태운 낡은 어선이 어두운 밤바다를 향해 뱃머리를 둔 채 궤적의 물살을 길게 남기고 있는 전경.\n\nLOCATION (lock): In the harbor's departure waters at night, with the small fishing boat heading toward the open sea. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the crane withdrawal, observe directly from high behind the boat's pier-side quarter, looking diagonally down its outward course. Place the old fishing boat in the upper middle, occupying less than a third of the frame, with its long wake extending toward the lower foreground; 이현우 and 찰리 remain small aboard, wrapped in their blankets and oriented toward the distant sea rather than the camera. Let the increasing camera-to-subject distance carry the departure, holding the established travel direction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing fishing boat in the upper-center of the frame, background, moves toward open sea beyond the upper frame; Trailing wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old fishing boat (Leaving the harbor with both passengers aboard) — Stern and pier-side quarter visible, bow pointing away toward open water; used as Small shared vessel establishing the passengers' departure and scale; Wake (Extending behind the moving boat) — Runs diagonally from the stern toward the lower foreground; used as Connects the distant boat to the camera's withdrawal position; Sea (Dark, with moving water around the departing vessel); used as Open negative space around the boat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the night sea subdued and the departing figures legible without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small, worn boat is leaving the harbor at night. Charlie remains covered by a large blanket and retains his backpack; his previously damaged chest remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우와 찰리를 태운 낡은 어선이 어두운 밤바다를 향해 뱃머리를 둔 채 궤적의 물살을 길게 남기고 있는 전경.\n\nLOCATION (lock): In the harbor's departure waters at night, with the small fishing boat heading toward the open sea. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the crane withdrawal, observe directly from high behind the boat's pier-side quarter, looking diagonally down its outward course. Place the old fishing boat in the upper middle, occupying less than a third of the frame, with its long wake extending toward the lower foreground; 이현우 and 찰리 remain small aboard, wrapped in their blankets and oriented toward the distant sea rather than the camera. Let the increasing camera-to-subject distance carry the departure, holding the established travel direction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing fishing boat in the upper-center of the frame, background, moves toward open sea beyond the upper frame; Trailing wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old fishing boat (Leaving the harbor with both passengers aboard) — Stern and pier-side quarter visible, bow pointing away toward open water; used as Small shared vessel establishing the passengers' departure and scale; Wake (Extending behind the moving boat) — Runs diagonally from the stern toward the lower foreground; used as Connects the distant boat to the camera's withdrawal position; Sea (Dark, with moving water around the departing vessel); used as Open negative space around the boat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the night sea subdued and the departing figures legible without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small, worn boat is leaving the harbor at night. Charlie remains covered by a large blanket and retains his backpack; his previously damaged chest remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 이현우와 찰리를 태운 낡은 어선이 어두운 밤바다를 향해 뱃머리를 둔 채 궤적의 물살을 길게 남기고 있는 전경.\n\nLOCATION (lock): In the harbor's departure waters at night, with the small fishing boat heading toward the open sea. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the crane withdrawal, observe directly from high behind the boat's pier-side quarter, looking diagonally down its outward course. Place the old fishing boat in the upper middle, occupying less than a third of the frame, with its long wake extending toward the lower foreground; 이현우 and 찰리 remain small aboard, wrapped in their blankets and oriented toward the distant sea rather than the camera. Let the increasing camera-to-subject distance carry the departure, holding the established travel direction.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Departing fishing boat in the upper-center of the frame, background, moves toward open sea beyond the upper frame; Trailing wake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Old fishing boat (Leaving the harbor with both passengers aboard) — Stern and pier-side quarter visible, bow pointing away toward open water; used as Small shared vessel establishing the passengers' departure and scale; Wake (Extending behind the moving boat) — Runs diagonally from the stern toward the lower foreground; used as Connects the distant boat to the camera's withdrawal position; Sea (Dark, with moving water around the departing vessel); used as Open negative space around the boat.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the night sea subdued and the departing figures legible without introducing an unsupported visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small, worn boat is leaving the harbor at night. Charlie remains covered by a large blanket and retains his backpack; his previously damaged chest remains unrepaired.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 멀어지는 어선을 향해 비스듬히 아래를 내려다보고 있으며, 배는 먼 바다를 향함.",
    "built_space": "프레임 상단 중앙에 어선이 위치하고, 좌측 원경에 긴 방파제가 보임.",
    "entities": "낡은 어선과 길게 이어진 물살이 명확히 보이나, 인물은 크기가 작아 식별되지 않음.",
    "hard_violations": [
     "[gpt-high] 지정되지 않은 해안의 다수 점광원과 조명 시설을 추가하여 장소의 추가 발명 및 근거 없는 가시 광원 금지 조건을 위반했다."
    ],
    "physics": "배가 정상적으로 물 위에 떠 있으며, 전진에 따른 항적(물살)이 자연스럽게 형성됨."
   },
   {
    "label": "B",
    "direction": "카메라는 바다로 나아가는 어선의 뒤쪽에서 비스듬히 아래를 향하고 있음.",
    "built_space": "상단 중앙에 어선이 있고, 원경 양쪽에 항구를 나타내는 등대와 방파제가 배치됨.",
    "entities": "어선과 물살이 요구사항에 맞게 묘사되었으며, 인물은 스케일 상 눈에 띄지 않음.",
    "hard_violations": [
     "[gpt-high] 장소를 추가로 발명하지 말라는 제한에도 왼쪽 가장자리에 별도 선박을 넣었다.",
     "[gpt-high] 명시되지 않은 두 등대의 불빛과 선실의 가시 조명을 추가하여 근거 없는 가시 광원을 금지한 조건을 위반했다."
    ],
    "physics": "배의 이동에 의해 물살이 일어나는 물리적 상태가 자연스럽게 지탱 및 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도와 물살의 방향을 정확히 구현했으며, 방파제와 조명 묘사가 더 시네마틱하고 사실적입니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 구도와 어선의 위치를 잘 따랐으나, 항구의 디테일과 조명이 다소 단조롭습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 멀어지는 어선을 향해 비스듬히 아래를 내려다보고 있으며, 배는 먼 바다를 향함.",
        "built_space": "프레임 상단 중앙에 어선이 위치하고, 좌측 원경에 긴 방파제가 보임.",
        "entities": "낡은 어선과 길게 이어진 물살이 명확히 보이나, 인물은 크기가 작아 식별되지 않음.",
        "hard_violations": [],
        "physics": "배가 정상적으로 물 위에 떠 있으며, 전진에 따른 항적(물살)이 자연스럽게 형성됨."
       },
       {
        "label": "B",
        "direction": "카메라는 바다로 나아가는 어선의 뒤쪽에서 비스듬히 아래를 향하고 있음.",
        "built_space": "상단 중앙에 어선이 있고, 원경 양쪽에 항구를 나타내는 등대와 방파제가 배치됨.",
        "entities": "어선과 물살이 요구사항에 맞게 묘사되었으며, 인물은 스케일 상 눈에 띄지 않음.",
        "hard_violations": [],
        "physics": "배의 이동에 의해 물살이 일어나는 물리적 상태가 자연스럽게 지탱 및 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 구도와 물살의 방향을 정확히 구현했으며, 방파제와 조명 묘사가 더 시네마틱하고 사실적입니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "요구된 구도와 어선의 위치를 잘 따랐으나, 항구의 디테일과 조명이 다소 단조롭습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 멀어지는 어선을 향해 비스듬히 아래를 내려다보고 있으며, 배는 먼 바다를 향함.",
        "built_space": "프레임 상단 중앙에 어선이 위치하고, 좌측 원경에 긴 방파제가 보임.",
        "entities": "낡은 어선과 길게 이어진 물살이 명확히 보이나, 인물은 크기가 작아 식별되지 않음.",
        "hard_violations": [],
        "physics": "배가 정상적으로 물 위에 떠 있으며, 전진에 따른 항적(물살)이 자연스럽게 형성됨."
       },
       {
        "label": "B",
        "direction": "카메라는 바다로 나아가는 어선의 뒤쪽에서 비스듬히 아래를 향하고 있음.",
        "built_space": "상단 중앙에 어선이 있고, 원경 양쪽에 항구를 나타내는 등대와 방파제가 배치됨.",
        "entities": "어선과 물살이 요구사항에 맞게 묘사되었으며, 인물은 스케일 상 눈에 띄지 않음.",
        "hard_violations": [],
        "physics": "배의 이동에 의해 물살이 일어나는 물리적 상태가 자연스럽게 지탱 및 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "높은 후방 시점과 긴 항적은 맞지만, 별도 선박과 양쪽 등대·점등된 선실을 추가해 장소 및 가시 광원 제한을 위반한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "작은 어선의 상단 중앙 배치와 외해를 향한 진행을 충실히 구현하고 인물도 명확히 노출하지 않지만, 지정되지 않은 항만 시설과 다수의 불빛을 추가했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "선미는 카메라 쪽이고 선수는 화면 위쪽에서 약간 오른쪽의 열린 바다를 향한다. 항적은 선미에서 하단 전경까지 길게 이어져 출항 방향과 일치한다. 선상 형체의 눈이나 시선은 판독되지 않는다.",
        "built_space": "어선 한 척의 선미와 오른쪽 측면, 작은 선실 하나와 마스트가 보인다. 배는 상단 중앙에서 화면의 3분의 1보다 작고 카메라는 높은 후방에 있어 지정 구도에 부합한다. 배경에는 좌우 방파제와 각각의 등대, 오른쪽 산지와 해안 불빛이 있으며 왼쪽 가장자리에는 별도 소형 선박도 보인다. 수면의 등빛 반사는 광원 아래로 이어져 광학적으로 가능하다.",
        "entities": "낡은 어선, 어두운 바다, 긴 흰 항적은 분명하다. 선미 쪽 녹색 및 회색 덩어리는 담요나 적재물처럼 보이지만 두 인물이라고 확정할 얼굴이나 신체는 식별되지 않는다. 이현우와 찰리의 정체, 배낭, 손상된 가슴은 이 크기에서 확인할 수 없으며 제공된 참조 이미지도 없다. 수평선에는 약한 황혼빛이 남아 있다.",
        "hard_violations": [
         "장소를 추가로 발명하지 말라는 제한에도 왼쪽 가장자리에 별도 선박을 넣었다.",
         "명시되지 않은 두 등대의 불빛과 선실의 가시 조명을 추가하여 근거 없는 가시 광원을 금지한 조건을 위반했다."
        ],
        "physics": "선체는 수면에 잠겨 부력으로 지지되고, 선상 설비와 덮인 덩어리들은 갑판에 놓여 있다. 항적이 선미에 연결되고 뒤로 넓어지는 모습은 동력 출항으로 설명된다. 지지 없이 떠 있는 물체나 불가능한 신체 자세는 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "선수는 화면 위쪽에서 약간 오른쪽의 외해를 향하고 선미는 카메라를 향한다. 선미에서 하단 중앙으로 이어지는 긴 항적이 이동 방향을 뒷받침한다. 식별 가능한 얼굴이나 시선은 없다.",
        "built_space": "어선에는 선실 하나, 마스트와 안테나, 선미 난간과 갑판 적재물이 보이며 선미와 오른쪽 측면이 함께 드러난다. 높은 후방 사선 시점, 상단 중앙의 작은 배, 하단 전경의 항적이라는 배치는 요구에 맞는다. 왼쪽 배경에는 방파제 하나와 끝의 표지탑 하나, 여러 해안 불빛이 있고 양쪽 수평선에 산지가 있다. 왼쪽 불빛의 수면 반사는 가능한 위치다.",
        "entities": "작고 낡은 어선 한 척, 어두운 바다와 긴 항적이 보인다. 갑판에는 원통형 장비와 황갈색 묶음이 있지만 사람의 얼굴이나 명확한 인체는 보이지 않아 인물 비노출 조건에 더 잘 맞는다. 두 승객의 정체와 담요·배낭·가슴 상태는 확인할 수 없다. 하늘은 청회색의 늦은 황혼으로 읽힌다.",
        "hard_violations": [
         "지정되지 않은 해안의 다수 점광원과 조명 시설을 추가하여 장소의 추가 발명 및 근거 없는 가시 광원 금지 조건을 위반했다."
        ],
        "physics": "선체는 수면의 부력으로 지지되고 선실·마스트·갑판 장비는 배에 고정되거나 갑판에 놓여 있다. 선미 바로 뒤에서 시작하는 포말과 길게 퍼지는 항적은 배의 전진으로 자연스럽게 설명된다. 지지 없는 물체나 공중에 뜬 인체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "높은 후방 시점과 긴 항적은 맞지만, 별도 선박과 양쪽 등대·점등된 선실을 추가해 장소 및 가시 광원 제한을 위반한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "작은 어선의 상단 중앙 배치와 외해를 향한 진행을 충실히 구현하고 인물도 명확히 노출하지 않지만, 지정되지 않은 항만 시설과 다수의 불빛을 추가했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "선미는 카메라 쪽이고 선수는 화면 위쪽에서 약간 오른쪽의 열린 바다를 향한다. 항적은 선미에서 하단 전경까지 길게 이어져 출항 방향과 일치한다. 선상 형체의 눈이나 시선은 판독되지 않는다.",
        "built_space": "어선 한 척의 선미와 오른쪽 측면, 작은 선실 하나와 마스트가 보인다. 배는 상단 중앙에서 화면의 3분의 1보다 작고 카메라는 높은 후방에 있어 지정 구도에 부합한다. 배경에는 좌우 방파제와 각각의 등대, 오른쪽 산지와 해안 불빛이 있으며 왼쪽 가장자리에는 별도 소형 선박도 보인다. 수면의 등빛 반사는 광원 아래로 이어져 광학적으로 가능하다.",
        "entities": "낡은 어선, 어두운 바다, 긴 흰 항적은 분명하다. 선미 쪽 녹색 및 회색 덩어리는 담요나 적재물처럼 보이지만 두 인물이라고 확정할 얼굴이나 신체는 식별되지 않는다. 이현우와 찰리의 정체, 배낭, 손상된 가슴은 이 크기에서 확인할 수 없으며 제공된 참조 이미지도 없다. 수평선에는 약한 황혼빛이 남아 있다.",
        "hard_violations": [
         "장소를 추가로 발명하지 말라는 제한에도 왼쪽 가장자리에 별도 선박을 넣었다.",
         "명시되지 않은 두 등대의 불빛과 선실의 가시 조명을 추가하여 근거 없는 가시 광원을 금지한 조건을 위반했다."
        ],
        "physics": "선체는 수면에 잠겨 부력으로 지지되고, 선상 설비와 덮인 덩어리들은 갑판에 놓여 있다. 항적이 선미에 연결되고 뒤로 넓어지는 모습은 동력 출항으로 설명된다. 지지 없이 떠 있는 물체나 불가능한 신체 자세는 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "선수는 화면 위쪽에서 약간 오른쪽의 외해를 향하고 선미는 카메라를 향한다. 선미에서 하단 중앙으로 이어지는 긴 항적이 이동 방향을 뒷받침한다. 식별 가능한 얼굴이나 시선은 없다.",
        "built_space": "어선에는 선실 하나, 마스트와 안테나, 선미 난간과 갑판 적재물이 보이며 선미와 오른쪽 측면이 함께 드러난다. 높은 후방 사선 시점, 상단 중앙의 작은 배, 하단 전경의 항적이라는 배치는 요구에 맞는다. 왼쪽 배경에는 방파제 하나와 끝의 표지탑 하나, 여러 해안 불빛이 있고 양쪽 수평선에 산지가 있다. 왼쪽 불빛의 수면 반사는 가능한 위치다.",
        "entities": "작고 낡은 어선 한 척, 어두운 바다와 긴 항적이 보인다. 갑판에는 원통형 장비와 황갈색 묶음이 있지만 사람의 얼굴이나 명확한 인체는 보이지 않아 인물 비노출 조건에 더 잘 맞는다. 두 승객의 정체와 담요·배낭·가슴 상태는 확인할 수 없다. 하늘은 청회색의 늦은 황혼으로 읽힌다.",
        "hard_violations": [
         "지정되지 않은 해안의 다수 점광원과 조명 시설을 추가하여 장소의 추가 발명 및 근거 없는 가시 광원 금지 조건을 위반했다."
        ],
        "physics": "선체는 수면의 부력으로 지지되고 선실·마스트·갑판 장비는 배에 고정되거나 갑판에 놓여 있다. 선미 바로 뒤에서 시작하는 포말과 길게 퍼지는 항적은 배의 전진으로 자연스럽게 설명된다. 지지 없는 물체나 공중에 뜬 인체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.607,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gpt-high] 장소를 추가로 발명하지 말라는 제한에도 왼쪽 가장자리에 별도 선박을 넣었다.",
     "[gpt-high] 명시되지 않은 두 등대의 불빛과 선실의 가시 조명을 추가하여 근거 없는 가시 광원을 금지한 조건을 위반했다."
    ],
    "A": [
     "[gpt-high] 지정되지 않은 해안의 다수 점광원과 조명 시설을 추가하여 장소의 추가 발명 및 근거 없는 가시 광원 금지 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1500,
   "A": 1607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "지정된 구도와 물살의 방향을 정확히 구현했으며, 방파제와 조명 묘사가 더 시네마틱하고 사실적입니다.  ★위반: [gpt-high] 장소를 추가로 발명하지 말라는 제한에도 왼쪽 가장자리에 별도 선박을 넣었다. / [gpt-high] 명시되지 않은 두 등대의 불빛과 선실의 가시 조명을 추가하여 근거 없는 가시 광원을 금지한 조건을 위반했다."
   },
   {
    "label": "A",
    "score": 1607,
    "verdict_ko": "요구된 구도와 어선의 위치를 잘 따랐으나, 항구의 디테일과 조명이 다소 단조롭습니다.  ★위반: [gpt-high] 지정되지 않은 해안의 다수 점광원과 조명 시설을 추가하여 장소의 추가 발명 및 근거 없는 가시 광원 금지 조건을 위반했다."
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e04-d845-7b4a-96ca-5c53faff1738",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S79sh8::signage": {
  "fp": "6f85b28c7b2c6e33",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::887c905f2549a219": {
  "subjects": [],
  "subject_text": "크리스의 어선 선원실\n소파와 선반이 있는 매우 비좁은 2인용 휴식 공간이다.",
  "identity": "canonical",
  "scope_id": "L134",
  "scope_role": "location_interior",
  "scope_sha": "bd38a81e5d76024a"
 },
 "groupbg::crew_cabin": {
  "input_fingerprint": "9df8a30857f8934b",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "crew_cabin",
    "tags": [
     "S79sh10",
     "S79sh8"
    ]
   },
   "context_sig": "5e2afcdcc09b435c"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 선원실: 소파와 선반이 있는 매우 비좁은 2인용 휴식 공간이다. (특징: 작은 인조가죽 소파에 기대앉은 현우와 찰리; 배가 기울며 바닥으로 쏟아지는 벽면 집기들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 드르르르륵. 문을 열면, 2명이 잘 법한 작은 선원실.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 선원실: 소파와 선반이 있는 매우 비좁은 2인용 휴식 공간이다. (특징: 작은 인조가죽 소파에 기대앉은 현우와 찰리; 배가 기울며 바닥으로 쏟아지는 벽면 집기들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 드르르르륵. 문을 열면, 2명이 잘 법한 작은 선원실.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_crew_cabin_950558.png",
  "asset_id": "9461daf3-d476-4a75-85de-3ff852115719",
  "input_asset_ids": [
   "e4180a8e-d044-4b4c-9b87-e44eaa9ad977"
  ],
  "origin_tag": "S79sh8",
  "place_text": "On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.",
  "origin_inputs": {
   "place_text": "On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.",
   "time_of_day_en": "night",
   "conti_asset_id": "e4180a8e-d044-4b4c-9b87-e44eaa9ad977"
  }
 },
 "S79sh8::bgfirst_bg": {
  "input_fingerprint": "b065f258234eb8e6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 눈을 감고 잠든 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit time cut, hold a static, direct view just above 이현우's reclining face, looking gently downward from the established three-quarter side. His peaceful face occupies the central-right portion, with a narrow strip of sofa retaining the setting; his eyes are closed in sleep, with no active visual target. Emphasize the reduced camera distance rather than adding a new lighting or staging change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting the sleeping passenger) — Only the supporting section beside his head and shoulder is visible; used as A narrow contextual edge beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the cabin, with gentle facial contrast and no dreamlike distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우가 눈을 감고 잠든 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit time cut, hold a static, direct view just above 이현우's reclining face, looking gently downward from the established three-quarter side. His peaceful face occupies the central-right portion, with a narrow strip of sofa retaining the setting; his eyes are closed in sleep, with no active visual target. Emphasize the reduced camera distance rather than adding a new lighting or staging change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting the sleeping passenger) — Only the supporting section beside his head and shoulder is visible; used as A narrow contextual edge beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the cabin, with gentle facial contrast and no dreamlike distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S79sh8__bgfirst_bg.png",
  "asset_id": "62711b63-a0f4-4400-9796-f7e2f076dcbb",
  "input_asset_ids": [
   "e4180a8e-d044-4b4c-9b87-e44eaa9ad977",
   "9461daf3-d476-4a75-85de-3ff852115719"
  ]
 },
 "S79sh8": {
  "input_fingerprint": "d2a5f490567c56a7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 눈을 감고 잠든 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit time cut, hold a static, direct view just above 이현우's reclining face, looking gently downward from the established three-quarter side. His peaceful face occupies the central-right portion, with a narrow strip of sofa retaining the setting; his eyes are closed in sleep, with no active visual target. Emphasize the reduced camera distance rather than adding a new lighting or staging change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting the sleeping passenger) — Only the supporting section beside his head and shoulder is visible; used as A narrow contextual edge beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the cabin, with gentle facial contrast and no dreamlike distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep, reclining against the sofa in the crew cabin with his legs extended and his eyes closed. The sofa supports his leaning torso, while the precise turn of his head and the positions of his arms are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin has a sofa, and its door remains closed after the departure. Charlie rests leaned back, with the blanket and backpack brought aboard still in his possession. 이현우: He is asleep, reclining against the sofa with his legs stretched out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 눈을 감고 잠든 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit time cut, hold a static, direct view just above 이현우's reclining face, looking gently downward from the established three-quarter side. His peaceful face occupies the central-right portion, with a narrow strip of sofa retaining the setting; his eyes are closed in sleep, with no active visual target. Emphasize the reduced camera distance rather than adding a new lighting or staging change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting the sleeping passenger) — Only the supporting section beside his head and shoulder is visible; used as A narrow contextual edge beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the cabin, with gentle facial contrast and no dreamlike distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep, reclining against the sofa in the crew cabin with his legs extended and his eyes closed. The sofa supports his leaning torso, while the precise turn of his head and the positions of his arms are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin has a sofa, and its door remains closed after the departure. Charlie rests leaned back, with the blanket and backpack brought aboard still in his possession. 이현우: He is asleep, reclining against the sofa with his legs stretched out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 눈을 감고 잠든 평온한 얼굴 클로즈업.\n\nLOCATION (lock): On the sofa inside the fishing boat's small two-person crew cabin. Modest cabin lighting is sufficient for the nighttime rest area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the explicit time cut, hold a static, direct view just above 이현우's reclining face, looking gently downward from the established three-quarter side. His peaceful face occupies the central-right portion, with a narrow strip of sofa retaining the setting; his eyes are closed in sleep, with no active visual target. Emphasize the reduced camera distance rather than adding a new lighting or staging change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sofa (Supporting the sleeping passenger) — Only the supporting section beside his head and shoulder is visible; used as A narrow contextual edge beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained ambient illumination appropriate to the cabin, with gentle facial contrast and no dreamlike distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is asleep, reclining against the sofa in the crew cabin with his legs extended and his eyes closed. The sofa supports his leaning torso, while the precise turn of his head and the positions of his arms are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small two-person cabin has a sofa, and its door remains closed after the departure. Charlie rests leaned back, with the blanket and backpack brought aboard still in his possession. 이현우: He is asleep, reclining against the sofa with his legs stretched out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S79sh8__bgfirst_bg.png",
     "asset_id": "62711b63-a0f4-4400-9796-f7e2f076dcbb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S79sh8.png",
     "asset_id": "e4180a8e-d044-4b4c-9b87-e44eaa9ad977",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_crew_cabin_950558.png",
     "asset_id": "9461daf3-d476-4a75-85de-3ff852115719",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:971162>",
     "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
    "built_space": "선실 소파에 앉아 있으나, 요구된 좁은 배경 대신 등 뒤로 선실의 문과 선반 등 깊은 공간이 모두 노출됨.",
    "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 있는 낡은 어두운 셔츠와 오른쪽 귀의 인이어 무전기 확인됨.",
    "hard_violations": [],
    "physics": "머리와 등, 어깨가 소파 등받이에 중력에 맞게 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
    "built_space": "선실 소파. 프롬프트 지시대로 인물의 얼굴 주변으로 소파의 일부와 벽면만이 좁게 노출된 타이트한 공간 구성을 보여줌.",
    "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 묻은 어두운 셔츠와 왼쪽 귀의 인이어 무전기 확인됨.",
    "hard_violations": [],
    "physics": "머리와 상체가 소파 표면에 자연스럽게 기대어 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 샷 스케일과 배경을 소파의 좁은 가장자리로 제한하라는 프레이밍 지시를 정확히 구현하여 가장 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물의 외형과 포즈는 맞으나, 클로즈업 지시를 무시하고 선실 문과 선반 등 배경을 넓게 노출하여 프레이밍에서 크게 감점됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
        "built_space": "선실 소파에 앉아 있으나, 요구된 좁은 배경 대신 등 뒤로 선실의 문과 선반 등 깊은 공간이 모두 노출됨.",
        "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 있는 낡은 어두운 셔츠와 오른쪽 귀의 인이어 무전기 확인됨.",
        "hard_violations": [],
        "physics": "머리와 등, 어깨가 소파 등받이에 중력에 맞게 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
        "built_space": "선실 소파. 프롬프트 지시대로 인물의 얼굴 주변으로 소파의 일부와 벽면만이 좁게 노출된 타이트한 공간 구성을 보여줌.",
        "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 묻은 어두운 셔츠와 왼쪽 귀의 인이어 무전기 확인됨.",
        "hard_violations": [],
        "physics": "머리와 상체가 소파 표면에 자연스럽게 기대어 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 샷 스케일과 배경을 소파의 좁은 가장자리로 제한하라는 프레이밍 지시를 정확히 구현하여 가장 우수함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물의 외형과 포즈는 맞으나, 클로즈업 지시를 무시하고 선실 문과 선반 등 배경을 넓게 노출하여 프레이밍에서 크게 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
        "built_space": "선실 소파에 앉아 있으나, 요구된 좁은 배경 대신 등 뒤로 선실의 문과 선반 등 깊은 공간이 모두 노출됨.",
        "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 있는 낡은 어두운 셔츠와 오른쪽 귀의 인이어 무전기 확인됨.",
        "hard_violations": [],
        "physics": "머리와 등, 어깨가 소파 등받이에 중력에 맞게 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 있어 특정 방향을 향한 시선은 없음.",
        "built_space": "선실 소파. 프롬프트 지시대로 인물의 얼굴 주변으로 소파의 일부와 벽면만이 좁게 노출된 타이트한 공간 구성을 보여줌.",
        "entities": "이현우의 얼굴 및 체격 일치. 핏자국이 묻은 어두운 셔츠와 왼쪽 귀의 인이어 무전기 확인됨.",
        "hard_violations": [],
        "physics": "머리와 상체가 소파 표면에 자연스럽게 기대어 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "잠든 얼굴을 중앙 오른쪽에 두고 위에서 부드럽게 내려다보는 근접 시점이 더 정확하지만, 소파와 상체는 지정된 좁은 가장자리보다 넓게 보입니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "수면 표정과 선실 재질은 맞지만, 시점이 얼굴 높이에 가깝고 문·선반·창까지 펼쳐 보여 얼굴 중심의 제한된 배경 구도를 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 완전히 감겨 있고 능동적인 시선 대상은 없습니다. 얼굴은 화면 왼쪽으로 조금 돌아간 채 위를 향하며, 카메라는 비스듬한 측면 위에서 내려다봅니다. 요구한 평온한 수면과 삼사분 측면 시점에 부합합니다.",
        "built_space": "갈색 가죽 소파 한 개의 등받이가 머리와 어깨 뒤에 보이고, 위쪽에는 밝은 벽과 목재 띠 일부가 있습니다. 왼쪽에는 선실 통로와 문틀 일부만 흐릿하게 보입니다. 가죽의 색과 마모, 벽의 재료는 장소 참조와 일치합니다. 중복 설비나 부자연스러운 반사는 없지만 소파가 좁은 맥락용 가장자리보다 넓습니다.",
        "entities": "인물은 한 명뿐이며, 참조와 유사한 앳된 동아시아계 남성 얼굴과 짧고 헝클어진 검은 머리입니다. 한국계 미국인이라는 국적·배경 자체는 외형으로 확인할 수 없습니다. 먼지와 짙은 핏자국이 묻은 어두운 셔츠, 보이는 귀의 검은 소형 인이어 장치가 요구와 맞습니다. 하체와 휴대품은 프레임 밖이므로 평가하지 않습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 기울어진 상체가 소파 등받이에 기대어 지지됩니다. 목과 어깨는 이 지지면을 따라 자연스럽게 놓여 있고, 공중에 들린 팔다리나 지지 없는 물체는 보이지 않습니다. 인이어 장치는 귀에 끼워져 있습니다."
       },
       {
        "label": "B",
        "direction": "두 눈을 감고 입 주변의 힘을 푼 수면 표정이며 시선 대상은 없습니다. 얼굴은 화면 왼쪽으로 약간 돌아가 위를 향합니다. 카메라는 A보다 얼굴 높이에 가까워, 얼굴 바로 위에서 부드럽게 내려다보라는 지시가 덜 드러납니다.",
        "built_space": "갈색 소파 한 개 뒤로 왼쪽 끝의 닫힌 문 한 개, 그 오른쪽의 좁은 측면 문 한 개, 여러 단으로 나뉜 수납 선반 한 개, 오른쪽 상단 창 한 개, 천장등 한 개가 보입니다. 배치와 재료는 장소 참조에 대체로 맞고 창밖은 어둡습니다. 다만 선실 전경을 넓게 포함하여 머리와 어깨 옆의 소파만 제한적으로 보이라는 구도에서 벗어납니다. 불가능한 반사나 설비 중복은 보이지 않습니다.",
        "entities": "한 명의 앳된 동아시아계 남성이 보이며 검은 헝클어진 머리와 얼굴은 참조 인물에 대체로 부합합니다. 외형만으로 국적은 확인할 수 없습니다. 어두운 낡은 셔츠에는 먼지와 붉은 얼룩이 있고, 귀에는 검은 소형 인이어 장치가 있습니다. 추가 인물이나 그래픽 문구는 없으며, 하체와 휴대품은 구도 밖입니다.",
        "hard_violations": [],
        "physics": "뒤통수는 소파 등받이 상단에 닿아 있고 어깨와 등도 소파에 기대어 있습니다. 기댄 몸의 무게를 가구가 받는 자세로 읽히며, 지지 없이 떠 있는 신체 부위는 보이지 않습니다. 귀의 장치도 정상적으로 고정되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "잠든 얼굴을 중앙 오른쪽에 두고 위에서 부드럽게 내려다보는 근접 시점이 더 정확하지만, 소파와 상체는 지정된 좁은 가장자리보다 넓게 보입니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "수면 표정과 선실 재질은 맞지만, 시점이 얼굴 높이에 가깝고 문·선반·창까지 펼쳐 보여 얼굴 중심의 제한된 배경 구도를 벗어납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 완전히 감겨 있고 능동적인 시선 대상은 없습니다. 얼굴은 화면 왼쪽으로 조금 돌아간 채 위를 향하며, 카메라는 비스듬한 측면 위에서 내려다봅니다. 요구한 평온한 수면과 삼사분 측면 시점에 부합합니다.",
        "built_space": "갈색 가죽 소파 한 개의 등받이가 머리와 어깨 뒤에 보이고, 위쪽에는 밝은 벽과 목재 띠 일부가 있습니다. 왼쪽에는 선실 통로와 문틀 일부만 흐릿하게 보입니다. 가죽의 색과 마모, 벽의 재료는 장소 참조와 일치합니다. 중복 설비나 부자연스러운 반사는 없지만 소파가 좁은 맥락용 가장자리보다 넓습니다.",
        "entities": "인물은 한 명뿐이며, 참조와 유사한 앳된 동아시아계 남성 얼굴과 짧고 헝클어진 검은 머리입니다. 한국계 미국인이라는 국적·배경 자체는 외형으로 확인할 수 없습니다. 먼지와 짙은 핏자국이 묻은 어두운 셔츠, 보이는 귀의 검은 소형 인이어 장치가 요구와 맞습니다. 하체와 휴대품은 프레임 밖이므로 평가하지 않습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 기울어진 상체가 소파 등받이에 기대어 지지됩니다. 목과 어깨는 이 지지면을 따라 자연스럽게 놓여 있고, 공중에 들린 팔다리나 지지 없는 물체는 보이지 않습니다. 인이어 장치는 귀에 끼워져 있습니다."
       },
       {
        "label": "A",
        "direction": "두 눈을 감고 입 주변의 힘을 푼 수면 표정이며 시선 대상은 없습니다. 얼굴은 화면 왼쪽으로 약간 돌아가 위를 향합니다. 카메라는 A보다 얼굴 높이에 가까워, 얼굴 바로 위에서 부드럽게 내려다보라는 지시가 덜 드러납니다.",
        "built_space": "갈색 소파 한 개 뒤로 왼쪽 끝의 닫힌 문 한 개, 그 오른쪽의 좁은 측면 문 한 개, 여러 단으로 나뉜 수납 선반 한 개, 오른쪽 상단 창 한 개, 천장등 한 개가 보입니다. 배치와 재료는 장소 참조에 대체로 맞고 창밖은 어둡습니다. 다만 선실 전경을 넓게 포함하여 머리와 어깨 옆의 소파만 제한적으로 보이라는 구도에서 벗어납니다. 불가능한 반사나 설비 중복은 보이지 않습니다.",
        "entities": "한 명의 앳된 동아시아계 남성이 보이며 검은 헝클어진 머리와 얼굴은 참조 인물에 대체로 부합합니다. 외형만으로 국적은 확인할 수 없습니다. 어두운 낡은 셔츠에는 먼지와 붉은 얼룩이 있고, 귀에는 검은 소형 인이어 장치가 있습니다. 추가 인물이나 그래픽 문구는 없으며, 하체와 휴대품은 구도 밖입니다.",
        "hard_violations": [],
        "physics": "뒤통수는 소파 등받이 상단에 닿아 있고 어깨와 등도 소파에 기대어 있습니다. 기댄 몸의 무게를 가구가 받는 자세로 읽히며, 지지 없이 떠 있는 신체 부위는 보이지 않습니다. 귀의 장치도 정상적으로 고정되어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.321,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.321,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "클로즈업 샷 스케일과 배경을 소파의 좁은 가장자리로 제한하라는 프레이밍 지시를 정확히 구현하여 가장 우수함."
   },
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "인물의 외형과 포즈는 맞으나, 클로즈업 지시를 무시하고 선실 문과 선반 등 배경을 넓게 노출하여 프레이밍에서 크게 감점됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_crew_cabin_950558.png",
    "asset_id": "9461daf3-d476-4a75-85de-3ff852115719",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e09-c970-7a6f-9f00-208186654a00",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S79sh8__bgfirst_bg.png",
   "bg_asset_id": "62711b63-a0f4-4400-9796-f7e2f076dcbb",
   "bg_record_key": "S79sh8::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "crew_cabin",
   "groupbg_asset_id": "9461daf3-d476-4a75-85de-3ff852115719"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S79sh10::signage": {
  "fp": "ea2336fcae57b26a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S79sh10": {
  "input_fingerprint": "0368aaf3cafff6d0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 크게 기울어진 바닥에서 발이 미끄러지며 벽 쪽으로 몸이 쏠린 이현우와 찰리의 mid-action.\n\nLOCATION (lock): Inside the fishing boat's small crew resting cabin, between the sofa and the nearby wall as the vessel lists. The interior retains its modest nighttime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established pullback and lowering into a knee-height, oblique handheld view across both feet and torsos, retaining the cabin's conversational side rather than reversing the axis. Capture 이현우 left of center with his feet sliding out from under him and 찰리 to his right with knees buckling at a different phase, both thrown toward the wall at frame right and looking toward the off-screen point of impact. Make their displacement across the tilted floor the dominant change, leaving the sofa partially visible behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Wall toward which both passengers slide in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Sharply tilted with the listing ship) — Cuts obliquely beneath the slipping feet; used as Makes the loss of balance physically readable; Cabin wall (In the path of the passengers' involuntary movement) — Visible at the right edge, with the impact area continuing outside the frame; used as Defines the destination of the sideways fall; Sofa (Partially visible behind the displaced passengers) — Seen obliquely from the same side as the preceding cabin coverage; used as Preserves spatial continuity with the resting scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cabin's restrained ambient exposure through the violent motion, without introducing flashes or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin and its sofa are now sharply tilted with the vessel. Charlie is being thrown toward the wall; no removal of his blanket or backpack has been established. 이현우: He is awake and being thrown toward the cabin wall, no longer resting securely on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 크게 기울어진 바닥에서 발이 미끄러지며 벽 쪽으로 몸이 쏠린 이현우와 찰리의 mid-action.\n\nLOCATION (lock): Inside the fishing boat's small crew resting cabin, between the sofa and the nearby wall as the vessel lists. The interior retains its modest nighttime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established pullback and lowering into a knee-height, oblique handheld view across both feet and torsos, retaining the cabin's conversational side rather than reversing the axis. Capture 이현우 left of center with his feet sliding out from under him and 찰리 to his right with knees buckling at a different phase, both thrown toward the wall at frame right and looking toward the off-screen point of impact. Make their displacement across the tilted floor the dominant change, leaving the sofa partially visible behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Wall toward which both passengers slide in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Sharply tilted with the listing ship) — Cuts obliquely beneath the slipping feet; used as Makes the loss of balance physically readable; Cabin wall (In the path of the passengers' involuntary movement) — Visible at the right edge, with the impact area continuing outside the frame; used as Defines the destination of the sideways fall; Sofa (Partially visible behind the displaced passengers) — Seen obliquely from the same side as the preceding cabin coverage; used as Preserves spatial continuity with the resting scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cabin's restrained ambient exposure through the violent motion, without introducing flashes or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin and its sofa are now sharply tilted with the vessel. Charlie is being thrown toward the wall; no removal of his blanket or backpack has been established. 이현우: He is awake and being thrown toward the cabin wall, no longer resting securely on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 크게 기울어진 바닥에서 발이 미끄러지며 벽 쪽으로 몸이 쏠린 이현우와 찰리의 mid-action.\n\nLOCATION (lock): Inside the fishing boat's small crew resting cabin, between the sofa and the nearby wall as the vessel lists. The interior retains its modest nighttime illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the established pullback and lowering into a knee-height, oblique handheld view across both feet and torsos, retaining the cabin's conversational side rather than reversing the axis. Capture 이현우 left of center with his feet sliding out from under him and 찰리 to his right with knees buckling at a different phase, both thrown toward the wall at frame right and looking toward the off-screen point of impact. Make their displacement across the tilted floor the dominant change, leaving the sofa partially visible behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Wall toward which both passengers slide in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Cabin floor (Sharply tilted with the listing ship) — Cuts obliquely beneath the slipping feet; used as Makes the loss of balance physically readable; Cabin wall (In the path of the passengers' involuntary movement) — Visible at the right edge, with the impact area continuing outside the frame; used as Defines the destination of the sideways fall; Sofa (Partially visible behind the displaced passengers) — Seen obliquely from the same side as the preceding cabin coverage; used as Preserves spatial continuity with the resting scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the cabin's restrained ambient exposure through the violent motion, without introducing flashes or a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The small cabin and its sofa are now sharply tilted with the vessel. Charlie is being thrown toward the wall; no removal of his blanket or backpack has been established. 이현우: He is awake and being thrown toward the cabin wall, no longer resting securely on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우와 찰리 모두 몸이 화면 오른쪽 벽을 향해 쏠려 있으며, 시선 역시 프롬프트에 명시된 대로 화면 오른쪽 프레임 밖의 충돌 지점을 정확히 향하고 있음.",
    "built_space": "카메라가 무릎 높이에서 비스듬히 바닥을 비추고 있으며, 바닥이 크게 기울어져 있음. 화면 오른쪽에 충돌 대상인 벽이 위치하고, 두 사람 뒤편으로 이전 샷과 동일한 재질의 가죽 소파가 부분적으로 정확히 묘사됨.",
    "entities": "이현우는 핏자국이 있는 낡은 셔츠와 바지를 입고 오른쪽 귀에 소형 인이어 무전기를 착용한 10대 후반 남성으로 잘 묘사되었음. 찰리 역시 샌드 베이지색 장갑판, 흰색 마스크, 가슴의 푸른 원자로를 지닌 고릴라 형태의 기계 외형과 정확히 일치함.",
    "hard_violations": [],
    "physics": "크게 기울어진 바닥으로 인해 이현우는 발이 미끄러져 몸이 낮아지고 있고, 찰리 역시 무릎이 꺾이며 왼손으로 벽을 짚고 바닥을 지탱하는 등 중력과 기울기에 따른 체중 이동 및 지지 상태가 물리적으로 매우 타당함."
   },
   {
    "label": "B",
    "direction": "찰리의 자세와 시선은 어느 정도 벽 쪽을 향하지만, 이현우의 시선은 프롬프트의 요구와 달리 충돌 지점(오른쪽)이 아닌 화면 왼쪽 아래를 향하고 있으며, 왼팔도 카메라 쪽으로 뻗어 있음.",
    "built_space": "기울어진 바닥, 뒤편의 소파, 오른쪽의 벽 등 요구된 공간적 요소와 구조물은 이전 샷의 기준과 일치하게 잘 배치되어 있음.",
    "entities": "이현우의 인물 특징과 낡은 옷차림이 잘 묘사됨. 찰리의 거대한 고릴라 비율 기계 외형, 마스크, 원자로 등도 레퍼런스와 일치하게 구현됨.",
    "hard_violations": [],
    "physics": "이현우가 통제력을 잃고 바닥에 미끄러지기보다는 오른손으로 찰리의 등이나 어깨를 짚고 의지하는 자세를 취하고 있어, 두 사람이 각각 미끄러지며 벽으로 튕겨 나간다는 액션의 물리적 묘사가 다소 어색함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 캐릭터가 화면 오른쪽 벽을 향해 쏠리며 충돌 지점을 바라보는 동적인 슬라이딩 액션과 물리적 제어가 완벽히 구현되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "이현우의 시선이 충돌 지점이 아닌 화면 왼쪽 아래를 향하고 있으며, 찰리에게 몸을 지탱하고 있어 요구된 역동적인 미끄러짐 액션이 잘 살아나지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 찰리 모두 몸이 화면 오른쪽 벽을 향해 쏠려 있으며, 시선 역시 프롬프트에 명시된 대로 화면 오른쪽 프레임 밖의 충돌 지점을 정확히 향하고 있음.",
        "built_space": "카메라가 무릎 높이에서 비스듬히 바닥을 비추고 있으며, 바닥이 크게 기울어져 있음. 화면 오른쪽에 충돌 대상인 벽이 위치하고, 두 사람 뒤편으로 이전 샷과 동일한 재질의 가죽 소파가 부분적으로 정확히 묘사됨.",
        "entities": "이현우는 핏자국이 있는 낡은 셔츠와 바지를 입고 오른쪽 귀에 소형 인이어 무전기를 착용한 10대 후반 남성으로 잘 묘사되었음. 찰리 역시 샌드 베이지색 장갑판, 흰색 마스크, 가슴의 푸른 원자로를 지닌 고릴라 형태의 기계 외형과 정확히 일치함.",
        "hard_violations": [],
        "physics": "크게 기울어진 바닥으로 인해 이현우는 발이 미끄러져 몸이 낮아지고 있고, 찰리 역시 무릎이 꺾이며 왼손으로 벽을 짚고 바닥을 지탱하는 등 중력과 기울기에 따른 체중 이동 및 지지 상태가 물리적으로 매우 타당함."
       },
       {
        "label": "B",
        "direction": "찰리의 자세와 시선은 어느 정도 벽 쪽을 향하지만, 이현우의 시선은 프롬프트의 요구와 달리 충돌 지점(오른쪽)이 아닌 화면 왼쪽 아래를 향하고 있으며, 왼팔도 카메라 쪽으로 뻗어 있음.",
        "built_space": "기울어진 바닥, 뒤편의 소파, 오른쪽의 벽 등 요구된 공간적 요소와 구조물은 이전 샷의 기준과 일치하게 잘 배치되어 있음.",
        "entities": "이현우의 인물 특징과 낡은 옷차림이 잘 묘사됨. 찰리의 거대한 고릴라 비율 기계 외형, 마스크, 원자로 등도 레퍼런스와 일치하게 구현됨.",
        "hard_violations": [],
        "physics": "이현우가 통제력을 잃고 바닥에 미끄러지기보다는 오른손으로 찰리의 등이나 어깨를 짚고 의지하는 자세를 취하고 있어, 두 사람이 각각 미끄러지며 벽으로 튕겨 나간다는 액션의 물리적 묘사가 다소 어색함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 캐릭터가 화면 오른쪽 벽을 향해 쏠리며 충돌 지점을 바라보는 동적인 슬라이딩 액션과 물리적 제어가 완벽히 구현되었습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "이현우의 시선이 충돌 지점이 아닌 화면 왼쪽 아래를 향하고 있으며, 찰리에게 몸을 지탱하고 있어 요구된 역동적인 미끄러짐 액션이 잘 살아나지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우와 찰리 모두 몸이 화면 오른쪽 벽을 향해 쏠려 있으며, 시선 역시 프롬프트에 명시된 대로 화면 오른쪽 프레임 밖의 충돌 지점을 정확히 향하고 있음.",
        "built_space": "카메라가 무릎 높이에서 비스듬히 바닥을 비추고 있으며, 바닥이 크게 기울어져 있음. 화면 오른쪽에 충돌 대상인 벽이 위치하고, 두 사람 뒤편으로 이전 샷과 동일한 재질의 가죽 소파가 부분적으로 정확히 묘사됨.",
        "entities": "이현우는 핏자국이 있는 낡은 셔츠와 바지를 입고 오른쪽 귀에 소형 인이어 무전기를 착용한 10대 후반 남성으로 잘 묘사되었음. 찰리 역시 샌드 베이지색 장갑판, 흰색 마스크, 가슴의 푸른 원자로를 지닌 고릴라 형태의 기계 외형과 정확히 일치함.",
        "hard_violations": [],
        "physics": "크게 기울어진 바닥으로 인해 이현우는 발이 미끄러져 몸이 낮아지고 있고, 찰리 역시 무릎이 꺾이며 왼손으로 벽을 짚고 바닥을 지탱하는 등 중력과 기울기에 따른 체중 이동 및 지지 상태가 물리적으로 매우 타당함."
       },
       {
        "label": "B",
        "direction": "찰리의 자세와 시선은 어느 정도 벽 쪽을 향하지만, 이현우의 시선은 프롬프트의 요구와 달리 충돌 지점(오른쪽)이 아닌 화면 왼쪽 아래를 향하고 있으며, 왼팔도 카메라 쪽으로 뻗어 있음.",
        "built_space": "기울어진 바닥, 뒤편의 소파, 오른쪽의 벽 등 요구된 공간적 요소와 구조물은 이전 샷의 기준과 일치하게 잘 배치되어 있음.",
        "entities": "이현우의 인물 특징과 낡은 옷차림이 잘 묘사됨. 찰리의 거대한 고릴라 비율 기계 외형, 마스크, 원자로 등도 레퍼런스와 일치하게 구현됨.",
        "hard_violations": [],
        "physics": "이현우가 통제력을 잃고 바닥에 미끄러지기보다는 오른손으로 찰리의 등이나 어깨를 짚고 의지하는 자세를 취하고 있어, 두 사람이 각각 미끄러지며 벽으로 튕겨 나간다는 액션의 물리적 묘사가 다소 어색함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "낮은 사선 와이드와 미끄러지는 발은 구현했지만, 이현우와 찰리 모두 오른쪽 충돌 지점을 바라보지 않으며 찰리의 담요·배낭 연속성도 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "이현우의 시선과 뻗은 팔, 오른쪽으로 쏠리는 몸과 서로 다른 무릎 굽힘이 지시된 순간에 더 충실하지만, 찰리의 시선과 담요·배낭 유지에는 미달한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 발은 왼쪽 전경으로 빠지고 상체는 오른쪽 찰리 쪽으로 기울지만, 눈은 오른쪽 화면 밖 충돌 지점이 아니라 앞쪽 아래를 향한다. 앞으로 내민 손도 전경 바닥 쪽이다. 찰리는 오른쪽 벽에 한 손을 대고 몸을 기울였으나 흰 얼굴은 카메라와 앞쪽 바닥을 향한다. 두 몸의 이동 목적지는 오른쪽 벽으로 읽히지만 두 시선의 목표는 지시와 다르다.",
        "built_space": "뒤쪽에 낡은 갈색 소파 한 개, 오른쪽에 충돌 대상 벽, 왼쪽에 원형 창이 있는 문 한 개, 후면에 사각 창 한 개와 벽등 한 개, 오른쪽 후면에 책꽂이 한 개가 보인다. 두 인물은 소파 앞과 오른쪽 벽 사이에 있고 바닥은 오른쪽으로 내려가는 사선을 이룬다. 낮은 카메라에서 발과 몸통을 함께 담으며 소파도 일부 남긴다. 회백색 벽판·목재 테두리·닳은 소파와 따뜻한 어둠은 이전 컷과 대체로 맞지만, 이전 컷에 없던 영역의 설비까지 동일한지는 확인할 수 없다.",
        "entities": "이현우 한 명과 찰리 한 대만 보인다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 마른 체격, 오염되고 피가 묻은 어두운 셔츠와 바지, 작은 검은 인이어를 갖춰 참조와 대체로 일치한다. 찰리는 긴 육중한 팔과 짧은 다리, 샌드 베이지 장갑판, 흰 마스크, 주황색 점눈 두 개와 선형 입, 파란 원형 가슴 장치를 유지한다. 찰리의 노출된 몸 주변에 담요는 보이지 않으며 배낭도 확인되지 않는다. 추가 인물이나 화면 위 자막은 없다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽으로 뻗은 신발은 바닥에 닿아 있고 무릎과 골반이 낮아져, 발이 빠지며 주저앉는 순간으로 해석할 수 있다. 몸 전체가 근거 없이 공중에 떠 있지는 않다. 찰리는 굽힌 다리의 발과 오른쪽 벽을 짚은 손으로 지지되며 다른 팔을 앞으로 내밀어 균형을 잡는다. 두 인물의 관절 굽힘과 접촉은 배가 기울며 균형을 잃는 동작으로 가능하다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 화면 밖 벽의 충돌 예상 지점을 바라보고 같은 방향으로 팔을 길게 뻗는다. 발은 왼쪽 전경으로 빠지고 상체는 오른쪽으로 쏠려 목표 방향이 명확하다. 찰리 역시 오른쪽 벽에 손을 대고 그쪽으로 몸이 기울지만, 얼굴은 벽의 충돌 지점보다는 카메라 쪽과 아래를 향한다. 따라서 두 인물이 모두 충돌 지점을 본다는 조건은 완전히 충족하지 못한다.",
        "built_space": "왼쪽 후면에 갈색 소파 한 개, 오른쪽 중경과 가장자리에 충돌 대상 벽, 뒤쪽에 사각 창 한 개와 벽등 한 개, 열린 수납부 한 열과 문틀이 보인다. 두 인물은 소파 앞에서 벽으로 밀려나는 위치에 있으며 소파가 몸 뒤로 이어진다. 무릎 높이의 낮은 사선 와이드에서 두 인물의 발과 몸통, 비스듬한 바닥을 함께 보여 준다. 회백색 벽, 목재 마감과 닳은 갈색 소파는 이전 컷의 재질에 부합하며 조명도 절제된 야간 분위기다. 추가로 드러난 후면 설비의 정확한 연속성은 좁은 이전 컷만으로 검증할 수 없다.",
        "entities": "이현우 한 명과 찰리 한 대가 있다. 이현우의 젊은 동아시아계 남성 외형, 짧고 헝클어진 검은 머리, 마른 체격, 얼룩진 어두운 셔츠와 바지 및 인이어는 참조와 대체로 맞는다. 찰리의 작고 육중한 고릴라형 비율, 긴 팔, 베이지 장갑판, 흰 마스크와 점눈 두 개·입 선, 푸른 가슴 원자로도 유지된다. 담요는 보이지 않고 배낭 역시 명확히 확인되지 않아 소지 상태의 연속성이 불충분하다. 불필요한 인물이나 그래픽 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 양쪽 신발 뒤꿈치와 가장자리가 바닥에 걸쳐 있고, 다리가 앞으로 빠지면서 골반이 내려가고 양팔이 벌어진 상태다. 이는 미끄러지며 넘어지는 짧은 순간으로 가능하며 지지 없는 부유로 볼 근거는 없다. 찰리는 한쪽 발을 바닥에 디디고 반대쪽 다리를 더 접은 채 벽을 손으로 짚는다. 벽 접촉과 발의 지지가 분명하고, 이현우와 다른 단계에서 무릎이 꺾이는 동작도 읽힌다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낮은 사선 와이드와 미끄러지는 발은 구현했지만, 이현우와 찰리 모두 오른쪽 충돌 지점을 바라보지 않으며 찰리의 담요·배낭 연속성도 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "이현우의 시선과 뻗은 팔, 오른쪽으로 쏠리는 몸과 서로 다른 무릎 굽힘이 지시된 순간에 더 충실하지만, 찰리의 시선과 담요·배낭 유지에는 미달한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 발은 왼쪽 전경으로 빠지고 상체는 오른쪽 찰리 쪽으로 기울지만, 눈은 오른쪽 화면 밖 충돌 지점이 아니라 앞쪽 아래를 향한다. 앞으로 내민 손도 전경 바닥 쪽이다. 찰리는 오른쪽 벽에 한 손을 대고 몸을 기울였으나 흰 얼굴은 카메라와 앞쪽 바닥을 향한다. 두 몸의 이동 목적지는 오른쪽 벽으로 읽히지만 두 시선의 목표는 지시와 다르다.",
        "built_space": "뒤쪽에 낡은 갈색 소파 한 개, 오른쪽에 충돌 대상 벽, 왼쪽에 원형 창이 있는 문 한 개, 후면에 사각 창 한 개와 벽등 한 개, 오른쪽 후면에 책꽂이 한 개가 보인다. 두 인물은 소파 앞과 오른쪽 벽 사이에 있고 바닥은 오른쪽으로 내려가는 사선을 이룬다. 낮은 카메라에서 발과 몸통을 함께 담으며 소파도 일부 남긴다. 회백색 벽판·목재 테두리·닳은 소파와 따뜻한 어둠은 이전 컷과 대체로 맞지만, 이전 컷에 없던 영역의 설비까지 동일한지는 확인할 수 없다.",
        "entities": "이현우 한 명과 찰리 한 대만 보인다. 이현우는 젊은 동아시아계 남성 외형, 헝클어진 검은 머리, 마른 체격, 오염되고 피가 묻은 어두운 셔츠와 바지, 작은 검은 인이어를 갖춰 참조와 대체로 일치한다. 찰리는 긴 육중한 팔과 짧은 다리, 샌드 베이지 장갑판, 흰 마스크, 주황색 점눈 두 개와 선형 입, 파란 원형 가슴 장치를 유지한다. 찰리의 노출된 몸 주변에 담요는 보이지 않으며 배낭도 확인되지 않는다. 추가 인물이나 화면 위 자막은 없다.",
        "hard_violations": [],
        "physics": "이현우의 앞쪽으로 뻗은 신발은 바닥에 닿아 있고 무릎과 골반이 낮아져, 발이 빠지며 주저앉는 순간으로 해석할 수 있다. 몸 전체가 근거 없이 공중에 떠 있지는 않다. 찰리는 굽힌 다리의 발과 오른쪽 벽을 짚은 손으로 지지되며 다른 팔을 앞으로 내밀어 균형을 잡는다. 두 인물의 관절 굽힘과 접촉은 배가 기울며 균형을 잃는 동작으로 가능하다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 화면 밖 벽의 충돌 예상 지점을 바라보고 같은 방향으로 팔을 길게 뻗는다. 발은 왼쪽 전경으로 빠지고 상체는 오른쪽으로 쏠려 목표 방향이 명확하다. 찰리 역시 오른쪽 벽에 손을 대고 그쪽으로 몸이 기울지만, 얼굴은 벽의 충돌 지점보다는 카메라 쪽과 아래를 향한다. 따라서 두 인물이 모두 충돌 지점을 본다는 조건은 완전히 충족하지 못한다.",
        "built_space": "왼쪽 후면에 갈색 소파 한 개, 오른쪽 중경과 가장자리에 충돌 대상 벽, 뒤쪽에 사각 창 한 개와 벽등 한 개, 열린 수납부 한 열과 문틀이 보인다. 두 인물은 소파 앞에서 벽으로 밀려나는 위치에 있으며 소파가 몸 뒤로 이어진다. 무릎 높이의 낮은 사선 와이드에서 두 인물의 발과 몸통, 비스듬한 바닥을 함께 보여 준다. 회백색 벽, 목재 마감과 닳은 갈색 소파는 이전 컷의 재질에 부합하며 조명도 절제된 야간 분위기다. 추가로 드러난 후면 설비의 정확한 연속성은 좁은 이전 컷만으로 검증할 수 없다.",
        "entities": "이현우 한 명과 찰리 한 대가 있다. 이현우의 젊은 동아시아계 남성 외형, 짧고 헝클어진 검은 머리, 마른 체격, 얼룩진 어두운 셔츠와 바지 및 인이어는 참조와 대체로 맞는다. 찰리의 작고 육중한 고릴라형 비율, 긴 팔, 베이지 장갑판, 흰 마스크와 점눈 두 개·입 선, 푸른 가슴 원자로도 유지된다. 담요는 보이지 않고 배낭 역시 명확히 확인되지 않아 소지 상태의 연속성이 불충분하다. 불필요한 인물이나 그래픽 오버레이는 없다.",
        "hard_violations": [],
        "physics": "이현우의 양쪽 신발 뒤꿈치와 가장자리가 바닥에 걸쳐 있고, 다리가 앞으로 빠지면서 골반이 내려가고 양팔이 벌어진 상태다. 이는 미끄러지며 넘어지는 짧은 순간으로 가능하며 지지 없는 부유로 볼 근거는 없다. 찰리는 한쪽 발을 바닥에 디디고 반대쪽 다리를 더 접은 채 벽을 손으로 짚는다. 벽 접촉과 발의 지지가 분명하고, 이현우와 다른 단계에서 무릎이 꺾이는 동작도 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.417
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.417
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "두 캐릭터가 화면 오른쪽 벽을 향해 쏠리며 충돌 지점을 바라보는 동적인 슬라이딩 액션과 물리적 제어가 완벽히 구현되었습니다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "이현우의 시선이 충돌 지점이 아닌 화면 왼쪽 아래를 향하고 있으며, 찰리에게 몸을 지탱하고 있어 요구된 역동적인 미끄러짐 액션이 잘 살아나지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S79sh8_sel.png",
    "asset_id": "412d99dd-9b6f-4da1-acd8-67bb235f7809",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e12-497d-7bf4-863a-e03964e6f87f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S79sh8"
  }
 },
 "S80sh3::signage": {
  "fp": "136bdb52e32512ca",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::5120180a4794f2c6": {
  "subjects": [],
  "subject_text": "크리스의 어선 갑판, 폭풍우 속 갑판\n좁은 뱃전 위로 거센 폭우와 파도가 들이치는 혼란스러운 외곽 공간이다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L180",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::storm_boat_deck": {
  "input_fingerprint": "85f0e9ca3a75e081",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "storm_boat_deck",
    "tags": [
     "S80sh11",
     "S80sh3"
    ]
   },
   "context_sig": "f4a1b851de717a84"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 갑판, 폭풍우 속 갑판: 좁은 뱃전 위로 거센 폭우와 파도가 들이치는 혼란스러운 외곽 공간이다. (특징: 장대비와 칠흑 같은 바다 위 거대한 파도의 시각적 묘사; 심하게 기울어지는 갑판 바닥과 금속 기둥; 한 손으로 기둥을 잡고 바다로 떨어지는 현우의 손을 잡은 찰리의 기계 팔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선박 위로 올라온 현우와 찰리. 좀 전의 평온했던 날씨와는 비교할 수 없는 엄청난 양의 폭우와 태풍!\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 갑판, 폭풍우 속 갑판: 좁은 뱃전 위로 거센 폭우와 파도가 들이치는 혼란스러운 외곽 공간이다. (특징: 장대비와 칠흑 같은 바다 위 거대한 파도의 시각적 묘사; 심하게 기울어지는 갑판 바닥과 금속 기둥; 한 손으로 기둥을 잡고 바다로 떨어지는 현우의 손을 잡은 찰리의 기계 팔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 선박 위로 올라온 현우와 찰리. 좀 전의 평온했던 날씨와는 비교할 수 없는 엄청난 양의 폭우와 태풍!\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_boat_deck_4a4535.png",
  "asset_id": "91c2b1a3-aa12-4361-854e-debefcdaada0",
  "input_asset_ids": [
   "898760a1-297c-4514-93fb-e3f449dc9611"
  ],
  "origin_tag": "S80sh3",
  "place_text": "On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.",
  "origin_inputs": {
   "place_text": "On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.",
   "time_of_day_en": "night, torrential rain and typhoon",
   "conti_asset_id": "898760a1-297c-4514-93fb-e3f449dc9611"
  }
 },
 "S80sh3::bgfirst_bg": {
  "input_fingerprint": "1acccc306da0a5fc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 파도가 갑판 위 선원들을 덮치는 찰나.\n\nLOCATION (lock): On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established elevated position above one side of the deck, look diagonally down across the crew toward the incoming wave, observing the event directly rather than through a character's eyes. Place the crew across the central deck, with 이현우 and 찰리 smaller at the near-left edge, and catch the wave reaching them from the upper-right background; all attention is directed toward its off-screen approaching crest, not the lens. Emphasize the bodies' displacement, giving the sailors different interrupted running phases, shoulder angles, and uneven spacing instead of a synchronized defensive pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ship's deck (Being swept by the first large wave) — Seen diagonally from above along the established camera-side edge; used as Shared ground plane for the differently phased bodies; Incoming wave (Breaking across the crew) — Advances from the upper-right background onto the central deck; used as Creates a readable source and direction for the impact; Torrential rain (Severely restricting visibility); used as Partially obscures depth without erasing the crew's distinct silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve nighttime exposure under the torrential rain, allowing storm-obscured visibility without adding lightning or an unsupported colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 파도가 갑판 위 선원들을 덮치는 찰나.\n\nLOCATION (lock): On the fishing boat's exposed working deck in violent nighttime rain and heavy seas.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established elevated position above one side of the deck, look diagonally down across the crew toward the incoming wave, observing the event directly rather than through a character's eyes. Place the crew across the central deck, with 이현우 and 찰리 smaller at the near-left edge, and catch the wave reaching them from the upper-right background; all attention is directed toward its off-screen approaching crest, not the lens. Emphasize the bodies' displacement, giving the sailors different interrupted running phases, shoulder angles, and uneven spacing instead of a synchronized defensive pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ship's deck (Being swept by the first large wave) — Seen diagonally from above along the established camera-side edge; used as Shared ground plane for the differently phased bodies; Incoming wave (Breaking across the crew) — Advances from the upper-right background onto the central deck; used as Creates a readable source and direction for the impact; Torrential rain (Severely restricting visibility); used as Partially obscures depth without erasing the crew's distinct silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve nighttime exposure under the torrential rain, allowing storm-obscured visibility without adding lightning or an unsupported colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh3__bgfirst_bg.png",
  "asset_id": "25e4395f-4bc9-4b8a-aea9-e8d125bd2c0d",
  "input_asset_ids": [
   "898760a1-297c-4514-93fb-e3f449dc9611",
   "91c2b1a3-aa12-4361-854e-debefcdaada0"
  ]
 },
 "S80sh3": {
  "input_fingerprint": "3b072de808d23352",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 거대한 파도가 갑판 위 선원들을 덮치는 찰나.\n\nLOCATION (lock): On the fishing boat's exposed working deck in violent nighttime rain and heavy seas. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established elevated position above one side of the deck, look diagonally down across the crew toward the incoming wave, observing the event directly rather than through a character's eyes. Place the crew across the central deck, with 이현우 and 찰리 smaller at the near-left edge, and catch the wave reaching them from the upper-right background; all attention is directed toward its off-screen approaching crest, not the lens. Emphasize the bodies' displacement, giving the sailors different interrupted running phases, shoulder angles, and uneven spacing instead of a synchronized defensive pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ship's deck (Being swept by the first large wave) — Seen diagonally from above along the established camera-side edge; used as Shared ground plane for the differently phased bodies; Incoming wave (Breaking across the crew) — Advances from the upper-right background onto the central deck; used as Creates a readable source and direction for the impact; Torrential rain (Severely restricting visibility); used as Partially obscures depth without erasing the crew's distinct silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve nighttime exposure under the torrential rain, allowing storm-obscured visibility without adding lightning or an unsupported colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat is engulfed in torrential rain, typhoon winds and huge waves, with visibility severely reduced. Charlie is now on deck, retaining his existing mechanical damage. 이현우: He is on the exposed deck, soaked by rain and seawater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 거대한 파도가 갑판 위 선원들을 덮치는 찰나.\n\nLOCATION (lock): On the fishing boat's exposed working deck in violent nighttime rain and heavy seas. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established elevated position above one side of the deck, look diagonally down across the crew toward the incoming wave, observing the event directly rather than through a character's eyes. Place the crew across the central deck, with 이현우 and 찰리 smaller at the near-left edge, and catch the wave reaching them from the upper-right background; all attention is directed toward its off-screen approaching crest, not the lens. Emphasize the bodies' displacement, giving the sailors different interrupted running phases, shoulder angles, and uneven spacing instead of a synchronized defensive pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ship's deck (Being swept by the first large wave) — Seen diagonally from above along the established camera-side edge; used as Shared ground plane for the differently phased bodies; Incoming wave (Breaking across the crew) — Advances from the upper-right background onto the central deck; used as Creates a readable source and direction for the impact; Torrential rain (Severely restricting visibility); used as Partially obscures depth without erasing the crew's distinct silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve nighttime exposure under the torrential rain, allowing storm-obscured visibility without adding lightning or an unsupported colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat is engulfed in torrential rain, typhoon winds and huge waves, with visibility severely reduced. Charlie is now on deck, retaining his existing mechanical damage. 이현우: He is on the exposed deck, soaked by rain and seawater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 거대한 파도가 갑판 위 선원들을 덮치는 찰나.\n\nLOCATION (lock): On the fishing boat's exposed working deck in violent nighttime rain and heavy seas. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the established elevated position above one side of the deck, look diagonally down across the crew toward the incoming wave, observing the event directly rather than through a character's eyes. Place the crew across the central deck, with 이현우 and 찰리 smaller at the near-left edge, and catch the wave reaching them from the upper-right background; all attention is directed toward its off-screen approaching crest, not the lens. Emphasize the bodies' displacement, giving the sailors different interrupted running phases, shoulder angles, and uneven spacing instead of a synchronized defensive pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ship's deck (Being swept by the first large wave) — Seen diagonally from above along the established camera-side edge; used as Shared ground plane for the differently phased bodies; Incoming wave (Breaking across the crew) — Advances from the upper-right background onto the central deck; used as Creates a readable source and direction for the impact; Torrential rain (Severely restricting visibility); used as Partially obscures depth without erasing the crew's distinct silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve nighttime exposure under the torrential rain, allowing storm-obscured visibility without adding lightning or an unsupported colored source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The boat is engulfed in torrential rain, typhoon winds and huge waves, with visibility severely reduced. Charlie is now on deck, retaining his existing mechanical damage. 이현우: He is on the exposed deck, soaked by rain and seawater.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼); 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh3__bgfirst_bg.png",
     "asset_id": "25e4395f-4bc9-4b8a-aea9-e8d125bd2c0d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S80sh3.png",
     "asset_id": "898760a1-297c-4514-93fb-e3f449dc9611",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1305657>",
     "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_boat_deck_4a4535.png",
     "asset_id": "91c2b1a3-aa12-4361-854e-debefcdaada0",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1220012>",
     "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1305657>",
     "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "파도는 우상단 바다에서 왼쪽 아래 중앙 갑판으로 넘어옵니다. 중앙과 오른쪽 선원들은 대체로 오른쪽 파도를 향하지만, 뒤쪽 선원의 고개는 아래로 숙여져 있습니다. 이현우는 오른쪽 아래 찰리 부근을 보고, 찰리의 얼굴은 파도가 아닌 화면 앞쪽을 향합니다. 오른쪽 선원의 뻗은 손은 난간을 향합니다. 모두가 화면 밖 접근하는 파도 마루에 주의를 둔 상태는 아닙니다.",
    "built_space": "왼쪽 선실 벽과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 켜진 조명 3개가 보입니다. 금속 상자는 중앙 왼쪽과 오른쪽 전경에 각각 1개씩 있고 뒤쪽에도 낮은 상자가 일부 보입니다. 오른쪽 현측 난간과 전경 계선주 1개가 분명하며 다른 난간 설비는 물보라에 가려집니다. 참고 장소의 낡은 금속과 밧줄 배치를 잘 유지합니다. 인물들은 난간 안쪽 작업 갑판에 있고 카메라는 왼쪽 위에서 대각선으로 내려다봅니다.",
    "entities": "이현우로 보이는 젊은 동아시아계 남성 1명, 다른 선원 4명, 찰리 1대가 보입니다. 이현우의 젖은 검은 머리와 젊은 얼굴은 참고와 대체로 맞지만 국적은 외관으로 확인할 수 없습니다. 남색 티셔츠 대신 검은 방수 외투와 주황색 작업복이 보입니다. 다른 선원들은 방수복과 모자로 얼굴이 가려져 나이와 민족을 판별하기 어렵습니다. 찰리는 짧고 육중한 고릴라형 기계이며 베이지 장갑, 긴 팔, 흰 마스크와 주황색 눈이 맞고 표면 마모도 있습니다. 비와 거대한 파도, 젖은 금속 갑판이 보이며 자막이나 번개는 없습니다.",
    "hard_violations": [],
    "physics": "이현우는 갑판에 발을 딛고 왼손으로 선실 쪽 구조물을 잡으며 다른 손을 찰리 위에 댑니다. 찰리는 짧은 다리와 갑판에 짚은 긴 팔로 몸을 지탱합니다. 중앙 선원들은 무릎을 굽히거나 한쪽 다리를 들어 이동 중이고, 오른쪽 선원은 난간을 짚어 버팁니다. 물보라에 일부 발이 가리지만 근거 없이 공중에 떠 있는 몸은 보이지 않습니다. 파도는 현측을 넘어 쏟아지고 상자와 밧줄은 갑판에 놓여 있습니다."
   },
   {
    "label": "A",
    "direction": "파도는 우상단 배경에서 오른쪽 현측을 넘어 중앙 갑판으로 밀려옵니다. 뒤쪽 두 선원, 중앙 주황색 선원, 오른쪽 두 선원의 머리와 어깨가 모두 파도 쪽으로 향합니다. 왼쪽 가까운 이현우와 찰리도 몸과 머리를 오른쪽 뒤 파도 방향으로 돌려 렌즈를 보지 않습니다. 달리는 몸의 진행 방향과 파도를 향한 주의가 읽힙니다. 다만 파도 마루의 상당 부분은 화면 안에도 보입니다.",
    "built_space": "왼쪽 선실과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 조명 3개가 참고와 같은 상대적 위치에 있습니다. 중앙 왼쪽과 오른쪽 전경의 금속 상자 2개 및 뒤쪽 작은 상자 1개가 보입니다. 오른쪽 난간 안쪽에는 앞쪽 계선주 3개가 구분되며 밧줄 뭉치들도 갑판 가장자리에 있습니다. 높은 왼쪽 시점에서 공유 갑판 면을 대각선으로 내려다보고, 이현우와 찰리는 가까운 왼쪽 가장자리, 나머지 선원들은 중앙에서 오른쪽에 불균등하게 배치됩니다. 찰리의 하체 일부는 화면 아래로 잘립니다.",
    "entities": "이현우로 보이는 젊은 검은 머리 남성 1명, 다른 선원 5명, 찰리 1대가 보입니다. 이현우는 옆뒤 모습이라 정확한 얼굴 일치는 확인하기 어렵지만 머리와 체격은 참고와 양립합니다. 겉옷은 참고의 티셔츠가 아니라 검은 방수복입니다. 선원들은 후드와 작업복을 착용하여 세부 나이와 민족은 확인하기 어렵습니다. 찰리는 베이지색 각진 장갑과 넓은 몸통, 긴 팔을 가진 기계로 보이며 마모가 있습니다. 얼굴이 돌아가 있어 흰 마스크와 눈의 세부는 평가할 수 없습니다. 야간 폭우와 큰 파도, 실제 선박 재질이 보이고 별도 문구나 번개는 없습니다.",
    "hard_violations": [],
    "physics": "이현우는 왼손으로 선실 쪽 구조물을 붙잡고 몸을 낮춥니다. 찰리도 왼쪽 구조물에 손을 대며 낮은 자세를 취하고, 아래쪽 다리의 끝은 프레임 밖이라 접지를 직접 확인할 수 없습니다. 중앙 주황색 선원은 한쪽 부츠를 갑판에 디디고 반대쪽 다리를 뒤로 접은 달리기 단계입니다. 뒤쪽 선원들도 다리의 굽힘과 보폭이 서로 다르고, 오른쪽 선원들은 난간에 손이나 팔을 대고 기울어진 몸을 받칩니다. 지지나 동작 근거 없이 떠 있는 인물은 없으며, 물은 난간을 넘어 갑판으로 쏟아집니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "높은 사선 와이드 구도와 파도의 갑판 강타는 충실하지만, 찰리가 화면 앞쪽을 보고 이현우도 파도 마루보다 아래쪽을 보아 시선 통일 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "우상단 파도로 향하는 인물들의 주의, 중앙 갑판의 불균등한 간격과 서로 다른 달리기 단계가 명확해 지정된 순간과 배치를 더 충실하게 구현합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "파도는 우상단 바다에서 왼쪽 아래 중앙 갑판으로 넘어옵니다. 중앙과 오른쪽 선원들은 대체로 오른쪽 파도를 향하지만, 뒤쪽 선원의 고개는 아래로 숙여져 있습니다. 이현우는 오른쪽 아래 찰리 부근을 보고, 찰리의 얼굴은 파도가 아닌 화면 앞쪽을 향합니다. 오른쪽 선원의 뻗은 손은 난간을 향합니다. 모두가 화면 밖 접근하는 파도 마루에 주의를 둔 상태는 아닙니다.",
        "built_space": "왼쪽 선실 벽과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 켜진 조명 3개가 보입니다. 금속 상자는 중앙 왼쪽과 오른쪽 전경에 각각 1개씩 있고 뒤쪽에도 낮은 상자가 일부 보입니다. 오른쪽 현측 난간과 전경 계선주 1개가 분명하며 다른 난간 설비는 물보라에 가려집니다. 참고 장소의 낡은 금속과 밧줄 배치를 잘 유지합니다. 인물들은 난간 안쪽 작업 갑판에 있고 카메라는 왼쪽 위에서 대각선으로 내려다봅니다.",
        "entities": "이현우로 보이는 젊은 동아시아계 남성 1명, 다른 선원 4명, 찰리 1대가 보입니다. 이현우의 젖은 검은 머리와 젊은 얼굴은 참고와 대체로 맞지만 국적은 외관으로 확인할 수 없습니다. 남색 티셔츠 대신 검은 방수 외투와 주황색 작업복이 보입니다. 다른 선원들은 방수복과 모자로 얼굴이 가려져 나이와 민족을 판별하기 어렵습니다. 찰리는 짧고 육중한 고릴라형 기계이며 베이지 장갑, 긴 팔, 흰 마스크와 주황색 눈이 맞고 표면 마모도 있습니다. 비와 거대한 파도, 젖은 금속 갑판이 보이며 자막이나 번개는 없습니다.",
        "hard_violations": [],
        "physics": "이현우는 갑판에 발을 딛고 왼손으로 선실 쪽 구조물을 잡으며 다른 손을 찰리 위에 댑니다. 찰리는 짧은 다리와 갑판에 짚은 긴 팔로 몸을 지탱합니다. 중앙 선원들은 무릎을 굽히거나 한쪽 다리를 들어 이동 중이고, 오른쪽 선원은 난간을 짚어 버팁니다. 물보라에 일부 발이 가리지만 근거 없이 공중에 떠 있는 몸은 보이지 않습니다. 파도는 현측을 넘어 쏟아지고 상자와 밧줄은 갑판에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "파도는 우상단 배경에서 오른쪽 현측을 넘어 중앙 갑판으로 밀려옵니다. 뒤쪽 두 선원, 중앙 주황색 선원, 오른쪽 두 선원의 머리와 어깨가 모두 파도 쪽으로 향합니다. 왼쪽 가까운 이현우와 찰리도 몸과 머리를 오른쪽 뒤 파도 방향으로 돌려 렌즈를 보지 않습니다. 달리는 몸의 진행 방향과 파도를 향한 주의가 읽힙니다. 다만 파도 마루의 상당 부분은 화면 안에도 보입니다.",
        "built_space": "왼쪽 선실과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 조명 3개가 참고와 같은 상대적 위치에 있습니다. 중앙 왼쪽과 오른쪽 전경의 금속 상자 2개 및 뒤쪽 작은 상자 1개가 보입니다. 오른쪽 난간 안쪽에는 앞쪽 계선주 3개가 구분되며 밧줄 뭉치들도 갑판 가장자리에 있습니다. 높은 왼쪽 시점에서 공유 갑판 면을 대각선으로 내려다보고, 이현우와 찰리는 가까운 왼쪽 가장자리, 나머지 선원들은 중앙에서 오른쪽에 불균등하게 배치됩니다. 찰리의 하체 일부는 화면 아래로 잘립니다.",
        "entities": "이현우로 보이는 젊은 검은 머리 남성 1명, 다른 선원 5명, 찰리 1대가 보입니다. 이현우는 옆뒤 모습이라 정확한 얼굴 일치는 확인하기 어렵지만 머리와 체격은 참고와 양립합니다. 겉옷은 참고의 티셔츠가 아니라 검은 방수복입니다. 선원들은 후드와 작업복을 착용하여 세부 나이와 민족은 확인하기 어렵습니다. 찰리는 베이지색 각진 장갑과 넓은 몸통, 긴 팔을 가진 기계로 보이며 마모가 있습니다. 얼굴이 돌아가 있어 흰 마스크와 눈의 세부는 평가할 수 없습니다. 야간 폭우와 큰 파도, 실제 선박 재질이 보이고 별도 문구나 번개는 없습니다.",
        "hard_violations": [],
        "physics": "이현우는 왼손으로 선실 쪽 구조물을 붙잡고 몸을 낮춥니다. 찰리도 왼쪽 구조물에 손을 대며 낮은 자세를 취하고, 아래쪽 다리의 끝은 프레임 밖이라 접지를 직접 확인할 수 없습니다. 중앙 주황색 선원은 한쪽 부츠를 갑판에 디디고 반대쪽 다리를 뒤로 접은 달리기 단계입니다. 뒤쪽 선원들도 다리의 굽힘과 보폭이 서로 다르고, 오른쪽 선원들은 난간에 손이나 팔을 대고 기울어진 몸을 받칩니다. 지지나 동작 근거 없이 떠 있는 인물은 없으며, 물은 난간을 넘어 갑판으로 쏟아집니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "높은 사선 와이드 구도와 파도의 갑판 강타는 충실하지만, 찰리가 화면 앞쪽을 보고 이현우도 파도 마루보다 아래쪽을 보아 시선 통일 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "우상단 파도로 향하는 인물들의 주의, 중앙 갑판의 불균등한 간격과 서로 다른 달리기 단계가 명확해 지정된 순간과 배치를 더 충실하게 구현합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "파도는 우상단 바다에서 왼쪽 아래 중앙 갑판으로 넘어옵니다. 중앙과 오른쪽 선원들은 대체로 오른쪽 파도를 향하지만, 뒤쪽 선원의 고개는 아래로 숙여져 있습니다. 이현우는 오른쪽 아래 찰리 부근을 보고, 찰리의 얼굴은 파도가 아닌 화면 앞쪽을 향합니다. 오른쪽 선원의 뻗은 손은 난간을 향합니다. 모두가 화면 밖 접근하는 파도 마루에 주의를 둔 상태는 아닙니다.",
        "built_space": "왼쪽 선실 벽과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 켜진 조명 3개가 보입니다. 금속 상자는 중앙 왼쪽과 오른쪽 전경에 각각 1개씩 있고 뒤쪽에도 낮은 상자가 일부 보입니다. 오른쪽 현측 난간과 전경 계선주 1개가 분명하며 다른 난간 설비는 물보라에 가려집니다. 참고 장소의 낡은 금속과 밧줄 배치를 잘 유지합니다. 인물들은 난간 안쪽 작업 갑판에 있고 카메라는 왼쪽 위에서 대각선으로 내려다봅니다.",
        "entities": "이현우로 보이는 젊은 동아시아계 남성 1명, 다른 선원 4명, 찰리 1대가 보입니다. 이현우의 젖은 검은 머리와 젊은 얼굴은 참고와 대체로 맞지만 국적은 외관으로 확인할 수 없습니다. 남색 티셔츠 대신 검은 방수 외투와 주황색 작업복이 보입니다. 다른 선원들은 방수복과 모자로 얼굴이 가려져 나이와 민족을 판별하기 어렵습니다. 찰리는 짧고 육중한 고릴라형 기계이며 베이지 장갑, 긴 팔, 흰 마스크와 주황색 눈이 맞고 표면 마모도 있습니다. 비와 거대한 파도, 젖은 금속 갑판이 보이며 자막이나 번개는 없습니다.",
        "hard_violations": [],
        "physics": "이현우는 갑판에 발을 딛고 왼손으로 선실 쪽 구조물을 잡으며 다른 손을 찰리 위에 댑니다. 찰리는 짧은 다리와 갑판에 짚은 긴 팔로 몸을 지탱합니다. 중앙 선원들은 무릎을 굽히거나 한쪽 다리를 들어 이동 중이고, 오른쪽 선원은 난간을 짚어 버팁니다. 물보라에 일부 발이 가리지만 근거 없이 공중에 떠 있는 몸은 보이지 않습니다. 파도는 현측을 넘어 쏟아지고 상자와 밧줄은 갑판에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "파도는 우상단 배경에서 오른쪽 현측을 넘어 중앙 갑판으로 밀려옵니다. 뒤쪽 두 선원, 중앙 주황색 선원, 오른쪽 두 선원의 머리와 어깨가 모두 파도 쪽으로 향합니다. 왼쪽 가까운 이현우와 찰리도 몸과 머리를 오른쪽 뒤 파도 방향으로 돌려 렌즈를 보지 않습니다. 달리는 몸의 진행 방향과 파도를 향한 주의가 읽힙니다. 다만 파도 마루의 상당 부분은 화면 안에도 보입니다.",
        "built_space": "왼쪽 선실과 배관, 권양기 1개, 중심 기둥 1개, 구명환 1개, 조명 3개가 참고와 같은 상대적 위치에 있습니다. 중앙 왼쪽과 오른쪽 전경의 금속 상자 2개 및 뒤쪽 작은 상자 1개가 보입니다. 오른쪽 난간 안쪽에는 앞쪽 계선주 3개가 구분되며 밧줄 뭉치들도 갑판 가장자리에 있습니다. 높은 왼쪽 시점에서 공유 갑판 면을 대각선으로 내려다보고, 이현우와 찰리는 가까운 왼쪽 가장자리, 나머지 선원들은 중앙에서 오른쪽에 불균등하게 배치됩니다. 찰리의 하체 일부는 화면 아래로 잘립니다.",
        "entities": "이현우로 보이는 젊은 검은 머리 남성 1명, 다른 선원 5명, 찰리 1대가 보입니다. 이현우는 옆뒤 모습이라 정확한 얼굴 일치는 확인하기 어렵지만 머리와 체격은 참고와 양립합니다. 겉옷은 참고의 티셔츠가 아니라 검은 방수복입니다. 선원들은 후드와 작업복을 착용하여 세부 나이와 민족은 확인하기 어렵습니다. 찰리는 베이지색 각진 장갑과 넓은 몸통, 긴 팔을 가진 기계로 보이며 마모가 있습니다. 얼굴이 돌아가 있어 흰 마스크와 눈의 세부는 평가할 수 없습니다. 야간 폭우와 큰 파도, 실제 선박 재질이 보이고 별도 문구나 번개는 없습니다.",
        "hard_violations": [],
        "physics": "이현우는 왼손으로 선실 쪽 구조물을 붙잡고 몸을 낮춥니다. 찰리도 왼쪽 구조물에 손을 대며 낮은 자세를 취하고, 아래쪽 다리의 끝은 프레임 밖이라 접지를 직접 확인할 수 없습니다. 중앙 주황색 선원은 한쪽 부츠를 갑판에 디디고 반대쪽 다리를 뒤로 접은 달리기 단계입니다. 뒤쪽 선원들도 다리의 굽힘과 보폭이 서로 다르고, 오른쪽 선원들은 난간에 손이나 팔을 대고 기울어진 몸을 받칩니다. 지지나 동작 근거 없이 떠 있는 인물은 없으며, 물은 난간을 넘어 갑판으로 쏟아집니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 7,
   "A": 9
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "높은 사선 와이드 구도와 파도의 갑판 강타는 충실하지만, 찰리가 화면 앞쪽을 보고 이현우도 파도 마루보다 아래쪽을 보아 시선 통일 지시에서 벗어납니다."
   },
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "우상단 파도로 향하는 인물들의 주의, 중앙 갑판의 불균등한 간격과 서로 다른 달리기 단계가 명확해 지정된 순간과 배치를 더 충실하게 구현합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_boat_deck_4a4535.png",
    "asset_id": "91c2b1a3-aa12-4361-854e-debefcdaada0",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1305657>",
    "asset_id": "c9ccdab1-56c8-4bc2-bf10-f9bd84346492",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e17-bfdb-70ab-86f5-627aac57ca1f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh3__bgfirst_bg.png",
   "bg_asset_id": "25e4395f-4bc9-4b8a-aea9-e8d125bd2c0d",
   "bg_record_key": "S80sh3::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "storm_boat_deck",
   "groupbg_asset_id": "91c2b1a3-aa12-4361-854e-debefcdaada0"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S80sh11::signage": {
  "fp": "5a52c5f8cfa05b08",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S80sh11": {
  "input_fingerprint": "30984f7cbd37478e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 파도의 충격으로 찰리의 기계 손에서 이현우의 손이 미끄러지듯 빠져나가는 근접 찰나.\n\nLOCATION (lock): At the edge of the storm-battered fishing boat's deck, beside the upright support being gripped against the waves. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from immediately inside the deck edge, slightly above the hands and looking down at the inherited three-quarter angle, beginning the downward tilt as the grip fails. 찰리's mechanical forearm enters from the upper left and 이현우's arm descends toward the lower right, with the slipping fingertips near center and enough deck edge and wave-driven water retained to explain the separation. Keep faces outside the crop, treat the breaking contact as the sole point of attention, and emphasize proximity rather than a new screen axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Deck edge (Beside the failing handhold) — A narrow oblique boundary behind the arms; used as Establishes the division between the ship and the fall; Wave-driven water (Crossing the arms during the impact); used as Provides immediate physical context for the grip release.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the storm's subdued nighttime exposure while separating human skin, precise mechanical contours, and crossing water without exaggerated specular brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the boat's wet deck surfaces, fixed deck structures, and storm-dark lighting from the reference. Exclude the exact crest and suspended spray of the earlier wave; do not repeat that transient water shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and enormous waves continue battering the vessel. Charlie remains braced against a deck pillar with one hand as his other hand loses its grip. 이현우: He is drenched and losing his handhold, being swept away by the wave.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 파도의 충격으로 찰리의 기계 손에서 이현우의 손이 미끄러지듯 빠져나가는 근접 찰나.\n\nLOCATION (lock): At the edge of the storm-battered fishing boat's deck, beside the upright support being gripped against the waves. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from immediately inside the deck edge, slightly above the hands and looking down at the inherited three-quarter angle, beginning the downward tilt as the grip fails. 찰리's mechanical forearm enters from the upper left and 이현우's arm descends toward the lower right, with the slipping fingertips near center and enough deck edge and wave-driven water retained to explain the separation. Keep faces outside the crop, treat the breaking contact as the sole point of attention, and emphasize proximity rather than a new screen axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Deck edge (Beside the failing handhold) — A narrow oblique boundary behind the arms; used as Establishes the division between the ship and the fall; Wave-driven water (Crossing the arms during the impact); used as Provides immediate physical context for the grip release.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the storm's subdued nighttime exposure while separating human skin, precise mechanical contours, and crossing water without exaggerated specular brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the boat's wet deck surfaces, fixed deck structures, and storm-dark lighting from the reference. Exclude the exact crest and suspended spray of the earlier wave; do not repeat that transient water shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and enormous waves continue battering the vessel. Charlie remains braced against a deck pillar with one hand as his other hand loses its grip. 이현우: He is drenched and losing his handhold, being swept away by the wave.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 파도의 충격으로 찰리의 기계 손에서 이현우의 손이 미끄러지듯 빠져나가는 근접 찰나.\n\nLOCATION (lock): At the edge of the storm-battered fishing boat's deck, beside the upright support being gripped against the waves. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from immediately inside the deck edge, slightly above the hands and looking down at the inherited three-quarter angle, beginning the downward tilt as the grip fails. 찰리's mechanical forearm enters from the upper left and 이현우's arm descends toward the lower right, with the slipping fingertips near center and enough deck edge and wave-driven water retained to explain the separation. Keep faces outside the crop, treat the breaking contact as the sole point of attention, and emphasize proximity rather than a new screen axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Deck edge (Beside the failing handhold) — A narrow oblique boundary behind the arms; used as Establishes the division between the ship and the fall; Wave-driven water (Crossing the arms during the impact); used as Provides immediate physical context for the grip release.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the storm's subdued nighttime exposure while separating human skin, precise mechanical contours, and crossing water without exaggerated specular brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the boat's wet deck surfaces, fixed deck structures, and storm-dark lighting from the reference. Exclude the exact crest and suspended spray of the earlier wave; do not repeat that transient water shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Torrential rain and enormous waves continue battering the vessel. Charlie remains braced against a deck pillar with one hand as his other hand loses its grip. 이현우: He is drenched and losing his handhold, being swept away by the wave.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 마르고 단단한 체격에 핏자국과 흙먼지가 묻은 낡고 어두운 색상의 캐주얼 셔츠와 바지, 귀에는 소형 인이어 무전기를 착용함. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "기계 팔은 좌측 상단에서 뻗어 나오고, 인간의 팔은 우측 하단에서 위로 뻗어 서로의 손끝이 방금 분리된 방향을 향하고 있음.",
    "built_space": "갑판 가장자리, 밧줄, 금속 클리트 등 참조 이미지의 선박 구조물들이 카메라 각도에 맞게 정확히 배치되어 있음.",
    "entities": "샌드 베이지 색상의 기계 팔(찰리)과 어두운 색 셔츠를 입고 물에 젖은 인간의 팔(이현우)이 명세와 일치함.",
    "hard_violations": [],
    "physics": "인간의 팔은 파도의 힘에 의해 뒤로 당겨지는 텐션을 보이며 화면 밖 신체에 지탱되어 있고, 기계 팔 역시 안정적으로 지탱되어 있음."
   },
   {
    "label": "B",
    "direction": "기계 팔과 인간의 팔이 서로를 향해 뻗어 있으나 맞잡던 궤적이 다소 어긋나 보임.",
    "built_space": "갑판과 난간 구조물이 보이나 디테일의 원근과 배치가 약간 평면적임.",
    "entities": "찰리의 기계 팔과 이현우의 팔이 등장하지만, 이현우의 손가락 개수가 정상 범위를 벗어남.",
    "hard_violations": [
     "[gemini-pro] 이현우의 손가락이 6개로 그려짐 (물리적으로 불가능한 해부학)"
    ],
    "physics": "두 팔 모두 화면 밖의 본체에 의해 지탱되어 허공에 떠 있으며, 물보라가 물리적 충격을 표현함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 구도와 클로즈업 프레이밍을 정확히 따르며, 해부학적 오류 없이 두 손이 미끄러지는 찰나의 순간을 사실적으로 잘 포착했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인간의 손가락이 6개로 그려지는 치명적인 해부학적 오류(Hard Violation)가 있어 사용할 수 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "기계 팔은 좌측 상단에서 뻗어 나오고, 인간의 팔은 우측 하단에서 위로 뻗어 서로의 손끝이 방금 분리된 방향을 향하고 있음.",
        "built_space": "갑판 가장자리, 밧줄, 금속 클리트 등 참조 이미지의 선박 구조물들이 카메라 각도에 맞게 정확히 배치되어 있음.",
        "entities": "샌드 베이지 색상의 기계 팔(찰리)과 어두운 색 셔츠를 입고 물에 젖은 인간의 팔(이현우)이 명세와 일치함.",
        "hard_violations": [],
        "physics": "인간의 팔은 파도의 힘에 의해 뒤로 당겨지는 텐션을 보이며 화면 밖 신체에 지탱되어 있고, 기계 팔 역시 안정적으로 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "기계 팔과 인간의 팔이 서로를 향해 뻗어 있으나 맞잡던 궤적이 다소 어긋나 보임.",
        "built_space": "갑판과 난간 구조물이 보이나 디테일의 원근과 배치가 약간 평면적임.",
        "entities": "찰리의 기계 팔과 이현우의 팔이 등장하지만, 이현우의 손가락 개수가 정상 범위를 벗어남.",
        "hard_violations": [
         "이현우의 손가락이 6개로 그려짐 (물리적으로 불가능한 해부학)"
        ],
        "physics": "두 팔 모두 화면 밖의 본체에 의해 지탱되어 허공에 떠 있으며, 물보라가 물리적 충격을 표현함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 구도와 클로즈업 프레이밍을 정확히 따르며, 해부학적 오류 없이 두 손이 미끄러지는 찰나의 순간을 사실적으로 잘 포착했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인간의 손가락이 6개로 그려지는 치명적인 해부학적 오류(Hard Violation)가 있어 사용할 수 없습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "기계 팔은 좌측 상단에서 뻗어 나오고, 인간의 팔은 우측 하단에서 위로 뻗어 서로의 손끝이 방금 분리된 방향을 향하고 있음.",
        "built_space": "갑판 가장자리, 밧줄, 금속 클리트 등 참조 이미지의 선박 구조물들이 카메라 각도에 맞게 정확히 배치되어 있음.",
        "entities": "샌드 베이지 색상의 기계 팔(찰리)과 어두운 색 셔츠를 입고 물에 젖은 인간의 팔(이현우)이 명세와 일치함.",
        "hard_violations": [],
        "physics": "인간의 팔은 파도의 힘에 의해 뒤로 당겨지는 텐션을 보이며 화면 밖 신체에 지탱되어 있고, 기계 팔 역시 안정적으로 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "기계 팔과 인간의 팔이 서로를 향해 뻗어 있으나 맞잡던 궤적이 다소 어긋나 보임.",
        "built_space": "갑판과 난간 구조물이 보이나 디테일의 원근과 배치가 약간 평면적임.",
        "entities": "찰리의 기계 팔과 이현우의 팔이 등장하지만, 이현우의 손가락 개수가 정상 범위를 벗어남.",
        "hard_violations": [
         "이현우의 손가락이 6개로 그려짐 (물리적으로 불가능한 해부학)"
        ],
        "physics": "두 팔 모두 화면 밖의 본체에 의해 지탱되어 허공에 떠 있으며, 물보라가 물리적 충격을 표현함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "좌상단 기계 팔과 우하단 인간 팔의 근접 구도는 맞지만, 손이 이미 벌어진 듯하고 인간 팔이 갑판 안쪽에 겹쳐 보여 붙잡힘이 풀리며 선외로 쓸려가는 관계가 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙에서 마지막 손끝 접촉이 풀리는 순간과 현우의 팔이 뱃전 바깥 우하단으로 이어지는 관계가 명확해, 파도 충격에 의한 손잡음의 실패를 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 기계 전완은 좌상단에서 중앙 아래쪽으로 뻗고, 현우의 팔은 우하단에서 중앙의 기계 손을 향한다. 인간 손가락은 기계 손끝을 향해 굽혀져 있으나 접촉이 대부분 끊어진 듯 보인다. 얼굴과 시선은 크롭 밖이다.",
        "built_space": "손보다 약간 높은 위치에서 내려다본 근접 화면이다. 뒤쪽 뱃전 하나가 좌상단에서 우하단으로 비스듬히 지나며, 왼쪽 가장자리에 굵은 수직 지지대 하나, 하단에 원형 머리의 계선주 하나, 좌하단에 밧줄 뭉치가 보인다. 젖고 녹슨 금속 표면은 이전 장면과 부합한다. 다만 현우의 전완이 뱃전의 갑판 쪽에 겹쳐 보여 배 밖으로 쓸려가는 공간 관계는 덜 선명하다.",
        "entities": "보이는 신체는 찰리의 기계 손과 전완, 현우의 인간 손과 어두운 젖은 소매뿐이다. 찰리의 모래색 각진 장갑과 검은 기계 관절은 참조와 맞는다. 현우의 손은 가늘고 젊어 보이지만 손만으로 성별·연령·민족 정체성을 확정할 수는 없다. 얼굴, 머리, 인이어는 올바르게 크롭 밖이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 화면 밖으로 이어지는 전완에 연결되어 있어 독립적으로 떠 있는 신체 부위는 없다. 손가락이 벌어지는 자세는 놓치는 동작으로 가능하다. 빗물과 파도에서 튄 물이 손 사이와 팔 위를 가로지른다. 찰리의 반대 손과 몸통은 보이지 않아 지지대에 버티는 접촉은 확인할 수 없지만, 이 근접 크롭의 결함은 아니다."
       },
       {
        "label": "B",
        "direction": "찰리의 기계 전완이 좌상단에서 중앙으로 내려오고, 현우의 팔은 중앙 손끝에서 우하단으로 이어진다. 기계 손가락과 인간 손가락의 끝부분이 중앙에서 아직 걸쳐 있어, 붙잡은 접촉이 막 풀리는 방향이 읽힌다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "갑판 가장자리 안쪽에서 손을 비스듬히 내려다보는 근접 시점이다. 뱃전 하나가 왼쪽 위에서 하단 중앙으로 이어져 왼쪽의 젖은 갑판과 오른쪽의 파도를 구분한다. 왼쪽 뒤에 높은 원통형 지지대 하나, 그 앞쪽에 낮은 원형 머리 계선주 하나, 하단 왼쪽에 밧줄 뭉치가 보이며 뱃전에는 세로 보강재가 있다. 현우의 팔이 뱃전 바깥쪽 물 위로 이어져 배와 추락 방향의 구분이 명확하다. 녹슨 도장과 젖은 금속은 장소 참조에 부합한다.",
        "entities": "찰리의 모래색 장갑 전완과 관절식 기계 손, 현우의 맨손과 흠뻑 젖은 어두운 소매가 보인다. 기계의 재질과 마모는 참조에 부합하며 인간 피부가 잘못 추가되지 않았다. 현우의 가는 손은 인물 설정과 모순되지 않지만 구체적인 민족·연령 정체성은 이 크롭만으로 판별할 수 없다. 얼굴이나 다른 인물, 추가 문자는 없다.",
        "hard_violations": [],
        "physics": "기계 손과 인간 손은 각각 전완에 자연스럽게 연결되어 있다. 인간 손가락이 기계 손의 말단에서 빠져나가는 자세이며, 손 사이와 손등을 덮치는 물보라가 접촉을 무너뜨리는 충격을 설명한다. 현우의 팔은 선외 우하단으로 이어져 쓸려가는 동작과 양립한다. 화면 밖 몸통의 지지 상태를 단정할 수는 없지만, 보이는 부분에 무지지 부유나 불가능한 관절은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "좌상단 기계 팔과 우하단 인간 팔의 근접 구도는 맞지만, 손이 이미 벌어진 듯하고 인간 팔이 갑판 안쪽에 겹쳐 보여 붙잡힘이 풀리며 선외로 쓸려가는 관계가 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙에서 마지막 손끝 접촉이 풀리는 순간과 현우의 팔이 뱃전 바깥 우하단으로 이어지는 관계가 명확해, 파도 충격에 의한 손잡음의 실패를 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 기계 전완은 좌상단에서 중앙 아래쪽으로 뻗고, 현우의 팔은 우하단에서 중앙의 기계 손을 향한다. 인간 손가락은 기계 손끝을 향해 굽혀져 있으나 접촉이 대부분 끊어진 듯 보인다. 얼굴과 시선은 크롭 밖이다.",
        "built_space": "손보다 약간 높은 위치에서 내려다본 근접 화면이다. 뒤쪽 뱃전 하나가 좌상단에서 우하단으로 비스듬히 지나며, 왼쪽 가장자리에 굵은 수직 지지대 하나, 하단에 원형 머리의 계선주 하나, 좌하단에 밧줄 뭉치가 보인다. 젖고 녹슨 금속 표면은 이전 장면과 부합한다. 다만 현우의 전완이 뱃전의 갑판 쪽에 겹쳐 보여 배 밖으로 쓸려가는 공간 관계는 덜 선명하다.",
        "entities": "보이는 신체는 찰리의 기계 손과 전완, 현우의 인간 손과 어두운 젖은 소매뿐이다. 찰리의 모래색 각진 장갑과 검은 기계 관절은 참조와 맞는다. 현우의 손은 가늘고 젊어 보이지만 손만으로 성별·연령·민족 정체성을 확정할 수는 없다. 얼굴, 머리, 인이어는 올바르게 크롭 밖이며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 손 모두 화면 밖으로 이어지는 전완에 연결되어 있어 독립적으로 떠 있는 신체 부위는 없다. 손가락이 벌어지는 자세는 놓치는 동작으로 가능하다. 빗물과 파도에서 튄 물이 손 사이와 팔 위를 가로지른다. 찰리의 반대 손과 몸통은 보이지 않아 지지대에 버티는 접촉은 확인할 수 없지만, 이 근접 크롭의 결함은 아니다."
       },
       {
        "label": "A",
        "direction": "찰리의 기계 전완이 좌상단에서 중앙으로 내려오고, 현우의 팔은 중앙 손끝에서 우하단으로 이어진다. 기계 손가락과 인간 손가락의 끝부분이 중앙에서 아직 걸쳐 있어, 붙잡은 접촉이 막 풀리는 방향이 읽힌다. 얼굴과 시선은 보이지 않는다.",
        "built_space": "갑판 가장자리 안쪽에서 손을 비스듬히 내려다보는 근접 시점이다. 뱃전 하나가 왼쪽 위에서 하단 중앙으로 이어져 왼쪽의 젖은 갑판과 오른쪽의 파도를 구분한다. 왼쪽 뒤에 높은 원통형 지지대 하나, 그 앞쪽에 낮은 원형 머리 계선주 하나, 하단 왼쪽에 밧줄 뭉치가 보이며 뱃전에는 세로 보강재가 있다. 현우의 팔이 뱃전 바깥쪽 물 위로 이어져 배와 추락 방향의 구분이 명확하다. 녹슨 도장과 젖은 금속은 장소 참조에 부합한다.",
        "entities": "찰리의 모래색 장갑 전완과 관절식 기계 손, 현우의 맨손과 흠뻑 젖은 어두운 소매가 보인다. 기계의 재질과 마모는 참조에 부합하며 인간 피부가 잘못 추가되지 않았다. 현우의 가는 손은 인물 설정과 모순되지 않지만 구체적인 민족·연령 정체성은 이 크롭만으로 판별할 수 없다. 얼굴이나 다른 인물, 추가 문자는 없다.",
        "hard_violations": [],
        "physics": "기계 손과 인간 손은 각각 전완에 자연스럽게 연결되어 있다. 인간 손가락이 기계 손의 말단에서 빠져나가는 자세이며, 손 사이와 손등을 덮치는 물보라가 접촉을 무너뜨리는 충격을 설명한다. 현우의 팔은 선외 우하단으로 이어져 쓸려가는 동작과 양립한다. 화면 밖 몸통의 지지 상태를 단정할 수는 없지만, 보이는 부분에 무지지 부유나 불가능한 관절은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.317
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.067
   },
   "violations": {
    "B": [
     "[gemini-pro] 이현우의 손가락이 6개로 그려짐 (물리적으로 불가능한 해부학)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1067
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 구도와 클로즈업 프레이밍을 정확히 따르며, 해부학적 오류 없이 두 손이 미끄러지는 찰나의 순간을 사실적으로 잘 포착했습니다."
   },
   {
    "label": "B",
    "score": 1067,
    "verdict_ko": "인간의 손가락이 6개로 그려지는 치명적인 해부학적 오류(Hard Violation)가 있어 사용할 수 없습니다.  ★위반: [gemini-pro] 이현우의 손가락이 6개로 그려짐 (물리적으로 불가능한 해부학)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh3_sel.png",
    "asset_id": "d25e5c48-8314-4a17-9d99-00f681abec0e",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:971162>",
    "asset_id": "74fd6647-055d-4f6f-8fa8-1c687bfa9683",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e24-316e-705c-a245-fb5e6d413965",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S80sh3"
  }
 },
 "S80sh15::signage": {
  "fp": "5679c33bd6f5d491",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::storm_deep_water": {
  "input_fingerprint": "2b5e41b43d255a2c",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "storm_deep_water",
    "tags": [
     "S80sh15"
    ]
   },
   "context_sig": "a14bc760f8f21eee"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 갑판, 폭풍우 속 갑판: 좁은 뱃전 위로 거센 폭우와 파도가 들이치는 혼란스러운 외곽 공간이다. (특징: 장대비와 칠흑 같은 바다 위 거대한 파도의 시각적 묘사; 심하게 기울어지는 갑판 바닥과 금속 기둥; 한 손으로 기둥을 잡고 바다로 떨어지는 현우의 손을 잡은 찰리의 기계 팔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 바다 속\n- 의식을 잃은 현우, 깊게, 더 깊게 아래로 잠기고.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n크리스의 어선 갑판, 폭풍우 속 갑판: 좁은 뱃전 위로 거센 폭우와 파도가 들이치는 혼란스러운 외곽 공간이다. (특징: 장대비와 칠흑 같은 바다 위 거대한 파도의 시각적 묘사; 심하게 기울어지는 갑판 바닥과 금속 기둥; 한 손으로 기둥을 잡고 바다로 떨어지는 현우의 손을 잡은 찰리의 기계 팔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 바다 속\n- 의식을 잃은 현우, 깊게, 더 깊게 아래로 잠기고.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_deep_water_5eb514.png",
  "asset_id": "f48d0265-960a-4203-b955-a256e92761a6",
  "input_asset_ids": [
   "d97b1bff-66a0-47fb-87f6-85df8c1fa6b3"
  ],
  "origin_tag": "S80sh15",
  "place_text": "Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.",
  "origin_inputs": {
   "place_text": "Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.",
   "time_of_day_en": "night, torrential rain and typhoon",
   "conti_asset_id": "d97b1bff-66a0-47fb-87f6-85df8c1fa6b3"
  }
 },
 "S80sh15::bgfirst_bg": {
  "input_fingerprint": "e43afd51eaadaa27",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 눈을 반쯤 감은 채 심해를 향해 몸이 아래로 기울어진 이현우의 평온한 수중 상체.\n\nLOCATION (lock): Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track underwater beside 이현우's upper torso and slightly above his shoulder, looking diagonally downward across his face and chest while matching his descent speed before any further withdrawal. His upper body occupies the center-left with the head below the shoulder line, leaving open water beneath the downward diagonal; his half-closed eyes are unfocused toward the depths, with no conscious object of attention. Hold the lateral three-quarter relationship and let his downward displacement carry the moment, with no imagery added for the mother's voice.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding underwater space (Darkening as he sinks deeper); used as Negative space below the torso makes the continuing descent readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the surrounding underwater visibility grow darker with depth, preserving subdued facial detail without supernatural glow or memory effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 눈을 반쯤 감은 채 심해를 향해 몸이 아래로 기울어진 이현우의 평온한 수중 상체.\n\nLOCATION (lock): Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo.\n\nTIME OF DAY (lock): night, torrential rain and typhoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track underwater beside 이현우's upper torso and slightly above his shoulder, looking diagonally downward across his face and chest while matching his descent speed before any further withdrawal. His upper body occupies the center-left with the head below the shoulder line, leaving open water beneath the downward diagonal; his half-closed eyes are unfocused toward the depths, with no conscious object of attention. Hold the lateral three-quarter relationship and let his downward displacement carry the moment, with no imagery added for the mother's voice.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding underwater space (Darkening as he sinks deeper); used as Negative space below the torso makes the continuing descent readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the surrounding underwater visibility grow darker with depth, preserving subdued facial detail without supernatural glow or memory effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh15__bgfirst_bg.png",
  "asset_id": "7c0d51c8-3794-4812-862e-1bffd9dd819a",
  "input_asset_ids": [
   "d97b1bff-66a0-47fb-87f6-85df8c1fa6b3",
   "f48d0265-960a-4203-b955-a256e92761a6"
  ]
 },
 "S80sh15": {
  "input_fingerprint": "2aeb8c1adc223efe",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 눈을 반쯤 감은 채 심해를 향해 몸이 아래로 기울어진 이현우의 평온한 수중 상체.\n\nLOCATION (lock): Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track underwater beside 이현우's upper torso and slightly above his shoulder, looking diagonally downward across his face and chest while matching his descent speed before any further withdrawal. His upper body occupies the center-left with the head below the shoulder line, leaving open water beneath the downward diagonal; his half-closed eyes are unfocused toward the depths, with no conscious object of attention. Hold the lateral three-quarter relationship and let his downward displacement carry the moment, with no imagery added for the mother's voice.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding underwater space (Darkening as he sinks deeper); used as Negative space below the torso makes the continuing descent readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the surrounding underwater visibility grow darker with depth, preserving subdued facial detail without supernatural glow or memory effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking beneath the sea's surface, his body unsupported in the water. The scene text does not specify the angle of his torso or head, his facing direction, or the positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks through the water, which grows darker with depth. Charlie is also descending underwater. 이현우: He is unconscious and sinking deeper underwater.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 눈을 반쯤 감은 채 심해를 향해 몸이 아래로 기울어진 이현우의 평온한 수중 상체.\n\nLOCATION (lock): Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track underwater beside 이현우's upper torso and slightly above his shoulder, looking diagonally downward across his face and chest while matching his descent speed before any further withdrawal. His upper body occupies the center-left with the head below the shoulder line, leaving open water beneath the downward diagonal; his half-closed eyes are unfocused toward the depths, with no conscious object of attention. Hold the lateral three-quarter relationship and let his downward displacement carry the moment, with no imagery added for the mother's voice.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding underwater space (Darkening as he sinks deeper); used as Negative space below the torso makes the continuing descent readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the surrounding underwater visibility grow darker with depth, preserving subdued facial detail without supernatural glow or memory effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking beneath the sea's surface, his body unsupported in the water. The scene text does not specify the angle of his torso or head, his facing direction, or the positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks through the water, which grows darker with depth. Charlie is also descending underwater. 이현우: He is unconscious and sinking deeper underwater.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night, torrential rain and typhoon.\n\nSHOT TEXT (authoritative, Korean): 눈을 반쯤 감은 채 심해를 향해 몸이 아래로 기울어진 이현우의 평온한 수중 상체.\n\nLOCATION (lock): Deep below the storm-tossed sea surface, in open water beneath the sinking vessel and its falling cargo. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track underwater beside 이현우's upper torso and slightly above his shoulder, looking diagonally downward across his face and chest while matching his descent speed before any further withdrawal. His upper body occupies the center-left with the head below the shoulder line, leaving open water beneath the downward diagonal; his half-closed eyes are unfocused toward the depths, with no conscious object of attention. Hold the lateral three-quarter relationship and let his downward displacement carry the moment, with no imagery added for the mother's voice.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Surrounding underwater space (Darkening as he sinks deeper); used as Negative space below the torso makes the continuing descent readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the surrounding underwater visibility grow darker with depth, preserving subdued facial detail without supernatural glow or memory effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Hyunwoo is unconscious and sinking beneath the sea's surface, his body unsupported in the water. The scene text does not specify the angle of his torso or head, his facing direction, or the positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Heavy cargo from the vessel sinks through the water, which grows darker with depth. Charlie is also descending underwater. 이현우: He is unconscious and sinking deeper underwater.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh15__bgfirst_bg.png",
     "asset_id": "7c0d51c8-3794-4812-862e-1bffd9dd819a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S80sh15.png",
     "asset_id": "d97b1bff-66a0-47fb-87f6-85df8c1fa6b3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1164888>",
     "asset_id": "a6131853-595d-4497-ab83-bbe9e4e4fa3b",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_deep_water_5eb514.png",
     "asset_id": "f48d0265-960a-4203-b955-a256e92761a6",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1164888>",
     "asset_id": "a6131853-595d-4497-ab83-bbe9e4e4fa3b",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 초점 없이 아래를 향하나, 몸의 방향이 대각선 아래보다는 수평에 가깝게 떠 있음.",
    "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
    "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
    "hard_violations": [],
    "physics": "물속에 떠 있는 상태이나, 요구된 하강의 느낌보다는 수중 부유 상태에 가깝게 사지가 퍼져 있음."
   },
   {
    "label": "B",
    "direction": "캐릭터의 몸이 우측 하단의 심해를 향해 대각선으로 기울어져 있으며, 시선은 초점 없이 아래를 향함.",
    "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
    "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
    "hard_violations": [],
    "physics": "의식 잃은 인물이 물속으로 가라앉고 있으며, 수압과 부력에 의해 머리카락과 팔이 자연스럽게 늘어짐."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 카메라 각도(어깨 위에서 얼굴과 가슴을 가로질러 내려다보는 구도)와 하강하는 몸의 기울기(머리가 어깨선 아래에 위치)를 프롬프트에 맞게 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터와 배경 요소는 잘 살렸으나, 몸이 수평에 가깝게 떠 있어 프롬프트가 요구한 하강하는 대각선 구도와 카메라 위치를 만족시키지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "캐릭터의 몸이 우측 하단의 심해를 향해 대각선으로 기울어져 있으며, 시선은 초점 없이 아래를 향함.",
        "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
        "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
        "hard_violations": [],
        "physics": "의식 잃은 인물이 물속으로 가라앉고 있으며, 수압과 부력에 의해 머리카락과 팔이 자연스럽게 늘어짐."
       },
       {
        "label": "A",
        "direction": "시선은 초점 없이 아래를 향하나, 몸의 방향이 대각선 아래보다는 수평에 가깝게 떠 있음.",
        "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
        "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
        "hard_violations": [],
        "physics": "물속에 떠 있는 상태이나, 요구된 하강의 느낌보다는 수중 부유 상태에 가깝게 사지가 퍼져 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 카메라 각도(어깨 위에서 얼굴과 가슴을 가로질러 내려다보는 구도)와 하강하는 몸의 기울기(머리가 어깨선 아래에 위치)를 프롬프트에 맞게 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "캐릭터와 배경 요소는 잘 살렸으나, 몸이 수평에 가깝게 떠 있어 프롬프트가 요구한 하강하는 대각선 구도와 카메라 위치를 만족시키지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "캐릭터의 몸이 우측 하단의 심해를 향해 대각선으로 기울어져 있으며, 시선은 초점 없이 아래를 향함.",
        "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
        "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
        "hard_violations": [],
        "physics": "의식 잃은 인물이 물속으로 가라앉고 있으며, 수압과 부력에 의해 머리카락과 팔이 자연스럽게 늘어짐."
       },
       {
        "label": "A",
        "direction": "시선은 초점 없이 아래를 향하나, 몸의 방향이 대각선 아래보다는 수평에 가깝게 떠 있음.",
        "built_space": "우측 상단에 침몰하는 선박의 실루엣과 컨테이너들이 떨어지고 있으며, 배경 레퍼런스의 공간적 특성과 일치함.",
        "entities": "이현우의 외모, 헤어스타일, 의상(피 묻은 셔츠)이 레퍼런스와 일치함. 추락하는 컨테이너와 선박의 형태도 일치함.",
        "hard_violations": [],
        "physics": "물속에 떠 있는 상태이나, 요구된 하강의 느낌보다는 수중 부유 상태에 가깝게 사지가 퍼져 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "중앙 왼쪽 상체 중심의 미디엄 구도, 반쯤 감긴 눈과 긴소매 의상이 더 충실하지만, 머리가 어깨선 아래로 선행하는 하강 자세는 명확하지 않다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물과 침몰 해역은 맞지만 양팔과 수면을 더 드러낸 넓은 구도이며, 머리가 높은 자세와 거의 감긴 눈, 걷힌 소매가 지정된 순간에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 화면 오른쪽 아래로 숙여져 있고, 가늘게 열린 눈은 특정 물체가 아닌 아래쪽 물을 향한다. 다만 머리는 가슴보다 화면 위에 있으며, 머리가 어깨선 아래에서 심해 쪽으로 몸을 이끄는 하강 대각선은 분명하지 않다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "개방된 수중 공간으로, 오른쪽 위에 선체 하나와 그 아래 오른쪽에 크기가 다른 화물 네 개가 보인다. 참고 장소의 선체·화물·기포 배치는 대체로 유지된다. 상체는 중앙 왼쪽을 크게 차지하고 아래와 오른쪽에 어두운 물이 남는다. 얼굴과 가슴을 위쪽 측면에서 보는 관계는 있으나, 상단의 넓은 수면 노출은 심해를 향해 내려다보는 시선을 약하게 만든다. 고정 설비나 반사상은 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 있으며, 짧고 흐트러진 검은 머리와 얼굴 윤곽은 이현우 참고와 가깝다. 한국계 미국인이라는 국적은 외형으로 확인할 수 없다. 때와 핏자국이 있는 어두운 회녹색 단추 셔츠, 가슴 주머니와 긴소매가 참고에 부합한다. 눈은 가늘게 열려 있고 표정은 평온하다. 추가 인물, 찰리, 어머니의 형상, 자막이나 초자연적 효과는 없다.",
        "hard_violations": [],
        "physics": "몸은 바닥이나 선체에 닿지 않고 물속에 잠겨 있어 부력과 유체 저항을 받는다. 공중에 근거 없이 떠 있는 경우는 아니다. 보이는 팔과 손은 아래로 처지고 머리카락은 물에 퍼져 있어 수중의 수동적인 자세로 가능하다. 다만 정지 화면에서 머리부터 깊이 가라앉는 운동은 뚜렷하지 않다. 화물은 물속 자유 침강, 기포는 부력에 의한 상승으로 설명된다."
       },
       {
        "label": "B",
        "direction": "얼굴은 오른쪽 아래로 기울지만 눈은 거의 완전히 감겨 있어 반쯤 열린 무초점 시선이 잘 드러나지 않는다. 머리는 가슴과 양어깨보다 대체로 높고 몸통은 수평에 가까워, 심해를 향해 머리가 낮아지는 하강 방향이 약하다. 특정 물체를 바라보거나 겨냥하는 동작은 없다.",
        "built_space": "상단 오른쪽의 선체 하나, 오른쪽 아래로 분포한 화물 네 개, 기포와 어두운 개방 수역이 보인다. 참고 장소와의 대응은 좋다. 인물은 중앙 왼쪽에 있으나 A보다 작고 양팔 전체에 가까운 범위와 더 넓은 수면이 포함된다. 아래쪽에 어두운 여백은 있지만, 어깨 옆에서 얼굴과 가슴을 대각선 아래로 따라가는 상체 중심 구도보다 수면 아래 장면 전체를 보여주는 구도에 가깝다. 중복 설비나 불가능한 반사상은 없다.",
        "entities": "젊은 동아시아계 남성 한 명의 얼굴, 체격, 흐트러진 검은 머리는 이현우 참고와 대체로 맞는다. 국적은 영상만으로 판별할 수 없다. 오염된 어두운 단추 셔츠와 가슴 주머니는 맞지만 소매가 팔꿈치 부근까지 걷혀 있어 참고의 긴소매 상태와 다르다. 눈은 거의 감겨 있고 입은 조금 벌어져 있다. 추가 인물이나 어머니의 형상, 글자, 초자연적 눈 표현은 없다.",
        "hard_violations": [],
        "physics": "물속의 몸은 부력과 유체 저항을 받으며, 별도의 고체 받침이 없는 침강 자체는 물리적으로 가능하다. 양팔은 몸통에서 벌어져 있지만 팔꿈치와 손목이 아래로 떨어져 있어 이것만으로 능동적인 수영이나 불가능한 자세라고 단정할 수 없다. 다만 머리가 높고 몸통이 거의 수평이라 아래로 기울어진 무의식적 침강의 운동성이 약하다. 화물과 기포의 위치는 각각 침강과 상승으로 설명할 수 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "중앙 왼쪽 상체 중심의 미디엄 구도, 반쯤 감긴 눈과 긴소매 의상이 더 충실하지만, 머리가 어깨선 아래로 선행하는 하강 자세는 명확하지 않다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "인물과 침몰 해역은 맞지만 양팔과 수면을 더 드러낸 넓은 구도이며, 머리가 높은 자세와 거의 감긴 눈, 걷힌 소매가 지정된 순간에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 화면 오른쪽 아래로 숙여져 있고, 가늘게 열린 눈은 특정 물체가 아닌 아래쪽 물을 향한다. 다만 머리는 가슴보다 화면 위에 있으며, 머리가 어깨선 아래에서 심해 쪽으로 몸을 이끄는 하강 대각선은 분명하지 않다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "개방된 수중 공간으로, 오른쪽 위에 선체 하나와 그 아래 오른쪽에 크기가 다른 화물 네 개가 보인다. 참고 장소의 선체·화물·기포 배치는 대체로 유지된다. 상체는 중앙 왼쪽을 크게 차지하고 아래와 오른쪽에 어두운 물이 남는다. 얼굴과 가슴을 위쪽 측면에서 보는 관계는 있으나, 상단의 넓은 수면 노출은 심해를 향해 내려다보는 시선을 약하게 만든다. 고정 설비나 반사상은 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 있으며, 짧고 흐트러진 검은 머리와 얼굴 윤곽은 이현우 참고와 가깝다. 한국계 미국인이라는 국적은 외형으로 확인할 수 없다. 때와 핏자국이 있는 어두운 회녹색 단추 셔츠, 가슴 주머니와 긴소매가 참고에 부합한다. 눈은 가늘게 열려 있고 표정은 평온하다. 추가 인물, 찰리, 어머니의 형상, 자막이나 초자연적 효과는 없다.",
        "hard_violations": [],
        "physics": "몸은 바닥이나 선체에 닿지 않고 물속에 잠겨 있어 부력과 유체 저항을 받는다. 공중에 근거 없이 떠 있는 경우는 아니다. 보이는 팔과 손은 아래로 처지고 머리카락은 물에 퍼져 있어 수중의 수동적인 자세로 가능하다. 다만 정지 화면에서 머리부터 깊이 가라앉는 운동은 뚜렷하지 않다. 화물은 물속 자유 침강, 기포는 부력에 의한 상승으로 설명된다."
       },
       {
        "label": "A",
        "direction": "얼굴은 오른쪽 아래로 기울지만 눈은 거의 완전히 감겨 있어 반쯤 열린 무초점 시선이 잘 드러나지 않는다. 머리는 가슴과 양어깨보다 대체로 높고 몸통은 수평에 가까워, 심해를 향해 머리가 낮아지는 하강 방향이 약하다. 특정 물체를 바라보거나 겨냥하는 동작은 없다.",
        "built_space": "상단 오른쪽의 선체 하나, 오른쪽 아래로 분포한 화물 네 개, 기포와 어두운 개방 수역이 보인다. 참고 장소와의 대응은 좋다. 인물은 중앙 왼쪽에 있으나 A보다 작고 양팔 전체에 가까운 범위와 더 넓은 수면이 포함된다. 아래쪽에 어두운 여백은 있지만, 어깨 옆에서 얼굴과 가슴을 대각선 아래로 따라가는 상체 중심 구도보다 수면 아래 장면 전체를 보여주는 구도에 가깝다. 중복 설비나 불가능한 반사상은 없다.",
        "entities": "젊은 동아시아계 남성 한 명의 얼굴, 체격, 흐트러진 검은 머리는 이현우 참고와 대체로 맞는다. 국적은 영상만으로 판별할 수 없다. 오염된 어두운 단추 셔츠와 가슴 주머니는 맞지만 소매가 팔꿈치 부근까지 걷혀 있어 참고의 긴소매 상태와 다르다. 눈은 거의 감겨 있고 입은 조금 벌어져 있다. 추가 인물이나 어머니의 형상, 글자, 초자연적 눈 표현은 없다.",
        "hard_violations": [],
        "physics": "물속의 몸은 부력과 유체 저항을 받으며, 별도의 고체 받침이 없는 침강 자체는 물리적으로 가능하다. 양팔은 몸통에서 벌어져 있지만 팔꿈치와 손목이 아래로 떨어져 있어 이것만으로 능동적인 수영이나 불가능한 자세라고 단정할 수 없다. 다만 머리가 높고 몸통이 거의 수평이라 아래로 기울어진 무의식적 침강의 운동성이 약하다. 화물과 기포의 위치는 각각 침강과 상승으로 설명할 수 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.405,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.405,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1405
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 카메라 각도(어깨 위에서 얼굴과 가슴을 가로질러 내려다보는 구도)와 하강하는 몸의 기울기(머리가 어깨선 아래에 위치)를 프롬프트에 맞게 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1405,
    "verdict_ko": "캐릭터와 배경 요소는 잘 살렸으나, 몸이 수평에 가깝게 떠 있어 프롬프트가 요구한 하강하는 대각선 구도와 카메라 위치를 만족시키지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_storm_deep_water_5eb514.png",
    "asset_id": "f48d0265-960a-4203-b955-a256e92761a6",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1164888>",
    "asset_id": "a6131853-595d-4497-ab83-bbe9e4e4fa3b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e29-e3d8-700a-9e7d-8857430c18a3",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S80sh15__bgfirst_bg.png",
   "bg_asset_id": "7c0d51c8-3794-4812-862e-1bffd9dd819a",
   "bg_record_key": "S80sh15::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "storm_deep_water",
   "groupbg_asset_id": "f48d0265-960a-4203-b955-a256e92761a6"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S81sh4::signage": {
  "fp": "6bc0d6161ea58751",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S81sh4": {
  "input_fingerprint": "38d9b07e09f8da2b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 창밖으로 에메랄드빛 푸른 바다가 펼쳐진 병실의 넓은 전경.\n\nLOCATION (lock): Inside the island research institute's temporary-shelter sickroom, beside a white bed and a window overlooking the turquoise sea. Strong fluorescent lighting is on. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the bed-side retreat and rise into a high, diagonally downward view across the room toward the window, retaining 이현우 as the small bedside anchor required by the continuous camera path. Place the white bed across the lower-left area and the window in the upper right, with the emerald-blue sea visible beyond; 이현우 is partway through pushing himself up, his attention still lowered toward his body rather than the lens. Make the increased camera distance the principal change, allowing the room and exterior view to emerge around him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by the newly awakened patient) — Seen diagonally from above in the lower-left portion; used as Connects the room reveal to the preceding bedside view without exceeding a third of the frame; Window (Providing a view of the sea outside) — Its visible opening frames the emerald-blue sea beyond, without an emphasized reflection; used as Upper-right exterior view balancing the bed; Jeju sea (Visible outside in daylight, emerald-blue); used as Establishes the unexpected island setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's strong fluorescent illumination reveals the white bed and clothing, while the emerald-blue sea remains visible outside with controlled highlight contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shelter room has a white bed under bright fluorescent lighting, with an emerald-blue sea visible through the window.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 창밖으로 에메랄드빛 푸른 바다가 펼쳐진 병실의 넓은 전경.\n\nLOCATION (lock): Inside the island research institute's temporary-shelter sickroom, beside a white bed and a window overlooking the turquoise sea. Strong fluorescent lighting is on. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the bed-side retreat and rise into a high, diagonally downward view across the room toward the window, retaining 이현우 as the small bedside anchor required by the continuous camera path. Place the white bed across the lower-left area and the window in the upper right, with the emerald-blue sea visible beyond; 이현우 is partway through pushing himself up, his attention still lowered toward his body rather than the lens. Make the increased camera distance the principal change, allowing the room and exterior view to emerge around him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by the newly awakened patient) — Seen diagonally from above in the lower-left portion; used as Connects the room reveal to the preceding bedside view without exceeding a third of the frame; Window (Providing a view of the sea outside) — Its visible opening frames the emerald-blue sea beyond, without an emphasized reflection; used as Upper-right exterior view balancing the bed; Jeju sea (Visible outside in daylight, emerald-blue); used as Establishes the unexpected island setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's strong fluorescent illumination reveals the white bed and clothing, while the emerald-blue sea remains visible outside with controlled highlight contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shelter room has a white bed under bright fluorescent lighting, with an emerald-blue sea visible through the window.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 창밖으로 에메랄드빛 푸른 바다가 펼쳐진 병실의 넓은 전경.\n\nLOCATION (lock): Inside the island research institute's temporary-shelter sickroom, beside a white bed and a window overlooking the turquoise sea. Strong fluorescent lighting is on. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the bed-side retreat and rise into a high, diagonally downward view across the room toward the window, retaining 이현우 as the small bedside anchor required by the continuous camera path. Place the white bed across the lower-left area and the window in the upper right, with the emerald-blue sea visible beyond; 이현우 is partway through pushing himself up, his attention still lowered toward his body rather than the lens. Make the increased camera distance the principal change, allowing the room and exterior view to emerge around him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: White bed (Occupied by the newly awakened patient) — Seen diagonally from above in the lower-left portion; used as Connects the room reveal to the preceding bedside view without exceeding a third of the frame; Window (Providing a view of the sea outside) — Its visible opening frames the emerald-blue sea beyond, without an emphasized reflection; used as Upper-right exterior view balancing the bed; Jeju sea (Visible outside in daylight, emerald-blue); used as Establishes the unexpected island setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's strong fluorescent illumination reveals the white bed and clothing, while the emerald-blue sea remains visible outside with controlled highlight contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shelter room has a white bed under bright fluorescent lighting, with an emerald-blue sea visible through the window.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
    "built_space": "침대, 벤치, 수납장 등 레퍼런스의 병실 구조를 따르고 있으나, 우측 수납장 뒤에 레퍼런스에 없는 검은색 스툴 의자가 놓여 있음.",
    "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 사물(우측 스툴 의자)을 발명하여 추가함",
     "[gpt-high] 참조 장소와 프롬프트에 없는 의자 한 개를 오른쪽 수납장 옆에 추가했다."
    ],
    "physics": "가구들이 바닥에 정상적으로 지지되어 놓여 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
    "built_space": "침대, 벤치, 벽면 수납장 등 레퍼런스의 병실 공간과 고정 요소들을 임의의 추가 없이 정확하게 재현함.",
    "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
    "hard_violations": [],
    "physics": "모든 사물이 중력에 맞게 바닥에 위치하며 자연스러운 그림자와 반사를 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "레퍼런스와 일치하는 병실 구조와 배경을 정확하게 구현했으며, 불필요한 사물을 추가하지 않아 프롬프트의 지시를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장소의 요소들을 잘 구현했으나, 레퍼런스에 없는 의자를 임의로 추가하여 장소 고정 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
        "built_space": "침대, 벤치, 수납장 등 레퍼런스의 병실 구조를 따르고 있으나, 우측 수납장 뒤에 레퍼런스에 없는 검은색 스툴 의자가 놓여 있음.",
        "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
        "hard_violations": [
         "레퍼런스에 없는 사물(우측 스툴 의자)을 발명하여 추가함"
        ],
        "physics": "가구들이 바닥에 정상적으로 지지되어 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
        "built_space": "침대, 벤치, 벽면 수납장 등 레퍼런스의 병실 공간과 고정 요소들을 임의의 추가 없이 정확하게 재현함.",
        "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
        "hard_violations": [],
        "physics": "모든 사물이 중력에 맞게 바닥에 위치하며 자연스러운 그림자와 반사를 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "레퍼런스와 일치하는 병실 구조와 배경을 정확하게 구현했으며, 불필요한 사물을 추가하지 않아 프롬프트의 지시를 충실히 따랐습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "장소의 요소들을 잘 구현했으나, 레퍼런스에 없는 의자를 임의로 추가하여 장소 고정 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
        "built_space": "침대, 벤치, 수납장 등 레퍼런스의 병실 구조를 따르고 있으나, 우측 수납장 뒤에 레퍼런스에 없는 검은색 스툴 의자가 놓여 있음.",
        "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
        "hard_violations": [
         "레퍼런스에 없는 사물(우측 스툴 의자)을 발명하여 추가함"
        ],
        "physics": "가구들이 바닥에 정상적으로 지지되어 놓여 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 바다가 보이는 정면의 창문을 향하고 있음.",
        "built_space": "침대, 벤치, 벽면 수납장 등 레퍼런스의 병실 공간과 고정 요소들을 임의의 추가 없이 정확하게 재현함.",
        "entities": "하얀 침대, 창문, 에메랄드빛 바다 등 프롬프트가 요구한 요소들이 존재하며 인물은 없음.",
        "hard_violations": [],
        "physics": "모든 사물이 중력에 맞게 바닥에 위치하며 자연스러운 그림자와 반사를 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "높은 사선의 넓은 전경에서 좌하단 흰 침대와 우상단 바다 창을 배치하고, 참조 병실의 설비와 최종 인물 금지 지시를 충실히 유지한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "빈 병실의 넓은 구도와 바다 전망은 맞지만, 참조와 지시에 없는 의자를 오른쪽 벽 옆에 추가해 고정 장소의 충실도를 훼손한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 출입구 쪽 높은 위치에서 침대와 창을 향해 비스듬히 내려다본다. 침대 머리는 창 쪽, 발치는 화면 좌하단을 향한다. 사람이나 조준·이동 물체는 없어 시선과 겨냥 대상은 없다.",
        "built_space": "왼쪽에 흰 침대 한 개, 그 오른쪽에 좁은 보조 탁자 한 개, 뒤쪽 벽에 두 짝으로 나뉜 창 한 개와 블라인드 한 세트가 있다. 왼쪽 벽의 모니터 한 개와 전화기 한 개, 아래 금속 수납장, 창 오른쪽의 작은 수납장 한 개, 켜진 천장 형광등 한 개가 참조 배치와 대응한다. 침대는 화면의 삼분의 일보다 작고 창은 우상단에 있다. 바닥의 밝은 반사는 창과 조명 위치상 가능하며 창에 강조된 반사상은 없다.",
        "entities": "흰 금속 병상, 흰 베개와 구겨진 침구, 낮의 청록빛 바다와 바위 해안이 보인다. 회백색 벽과 바닥, 전화기와 모니터, 벽의 안내물 및 달력도 참조 장소에 대응한다. 인물은 없다. 이는 마지막의 명시적인 인물 금지에는 맞지만, 앞선 카메라 항목의 이현우 등장 요구와는 지시 자체가 상충한다. 새로운 자막이나 그래픽은 보이지 않는다.",
        "hard_violations": [],
        "physics": "침대는 다리와 바퀴로 바닥에 지지되고, 베개와 이불은 매트리스 위에 놓여 있다. 보조 탁자는 바닥에 닿는 받침으로 지지되며, 수납장 위 용기는 상판에 놓여 있다. 모니터와 전화기는 벽 설비에 고정되어 있다. 지지 없이 떠 있는 물체나 불가능한 동작은 없다."
       },
       {
        "label": "B",
        "direction": "카메라는 출입구 쪽에서 방 안의 창을 향해 사선으로 내려다본다. 침대 머리는 창 아래에 있고 발치는 좌하단을 향한다. 사람, 무기, 이동하는 물체는 없어 시선이나 조준 방향은 없다.",
        "built_space": "흰 침대 한 개와 그 오른쪽 보조 탁자 한 개, 뒤쪽의 두 짝 창 한 개와 블라인드 한 세트, 왼쪽 벽의 모니터 한 개와 전화기 한 개가 보인다. 왼쪽 금속 수납장과 창 오른쪽 작은 수납장, 천장 형광등도 참조의 기본 배치를 따른다. 다만 오른쪽 수납장 옆에 참조에 없는 검은 좌판의 의자 한 개가 추가되어 있다. 침대는 좌하단에서 삼분의 일 미만을 차지하고 창은 상단 오른쪽에 놓인다. 바닥 반사는 자연스럽다.",
        "entities": "흰 침대와 침구, 밝은 낮의 청록색 바다, 바위 해안, 회백색 병실 마감이 요청과 일치한다. 모니터·전화기·벽 안내물·달력도 참조와 대응하지만 오른쪽 의자는 추가된 물건이다. 인물은 없어 최종 인물 금지 지시를 따르며, 상충하는 앞선 환자 등장 요구는 실현되지 않았다. 별도의 화면 자막이나 덧씌운 표시는 없다.",
        "hard_violations": [
         "참조 장소와 프롬프트에 없는 의자 한 개를 오른쪽 수납장 옆에 추가했다."
        ],
        "physics": "침대는 다리와 바퀴로, 보조 탁자는 받침으로 바닥에 지지되어 있다. 침구는 매트리스 위에 내려앉고 용기는 수납장 상판에 놓인다. 추가된 의자도 보이는 다리로 바닥에 서 있다. 벽 설비와 천장 조명의 고정도 자연스러우며, 지지 없는 부유 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "높은 사선의 넓은 전경에서 좌하단 흰 침대와 우상단 바다 창을 배치하고, 참조 병실의 설비와 최종 인물 금지 지시를 충실히 유지한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "빈 병실의 넓은 구도와 바다 전망은 맞지만, 참조와 지시에 없는 의자를 오른쪽 벽 옆에 추가해 고정 장소의 충실도를 훼손한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 출입구 쪽 높은 위치에서 침대와 창을 향해 비스듬히 내려다본다. 침대 머리는 창 쪽, 발치는 화면 좌하단을 향한다. 사람이나 조준·이동 물체는 없어 시선과 겨냥 대상은 없다.",
        "built_space": "왼쪽에 흰 침대 한 개, 그 오른쪽에 좁은 보조 탁자 한 개, 뒤쪽 벽에 두 짝으로 나뉜 창 한 개와 블라인드 한 세트가 있다. 왼쪽 벽의 모니터 한 개와 전화기 한 개, 아래 금속 수납장, 창 오른쪽의 작은 수납장 한 개, 켜진 천장 형광등 한 개가 참조 배치와 대응한다. 침대는 화면의 삼분의 일보다 작고 창은 우상단에 있다. 바닥의 밝은 반사는 창과 조명 위치상 가능하며 창에 강조된 반사상은 없다.",
        "entities": "흰 금속 병상, 흰 베개와 구겨진 침구, 낮의 청록빛 바다와 바위 해안이 보인다. 회백색 벽과 바닥, 전화기와 모니터, 벽의 안내물 및 달력도 참조 장소에 대응한다. 인물은 없다. 이는 마지막의 명시적인 인물 금지에는 맞지만, 앞선 카메라 항목의 이현우 등장 요구와는 지시 자체가 상충한다. 새로운 자막이나 그래픽은 보이지 않는다.",
        "hard_violations": [],
        "physics": "침대는 다리와 바퀴로 바닥에 지지되고, 베개와 이불은 매트리스 위에 놓여 있다. 보조 탁자는 바닥에 닿는 받침으로 지지되며, 수납장 위 용기는 상판에 놓여 있다. 모니터와 전화기는 벽 설비에 고정되어 있다. 지지 없이 떠 있는 물체나 불가능한 동작은 없다."
       },
       {
        "label": "A",
        "direction": "카메라는 출입구 쪽에서 방 안의 창을 향해 사선으로 내려다본다. 침대 머리는 창 아래에 있고 발치는 좌하단을 향한다. 사람, 무기, 이동하는 물체는 없어 시선이나 조준 방향은 없다.",
        "built_space": "흰 침대 한 개와 그 오른쪽 보조 탁자 한 개, 뒤쪽의 두 짝 창 한 개와 블라인드 한 세트, 왼쪽 벽의 모니터 한 개와 전화기 한 개가 보인다. 왼쪽 금속 수납장과 창 오른쪽 작은 수납장, 천장 형광등도 참조의 기본 배치를 따른다. 다만 오른쪽 수납장 옆에 참조에 없는 검은 좌판의 의자 한 개가 추가되어 있다. 침대는 좌하단에서 삼분의 일 미만을 차지하고 창은 상단 오른쪽에 놓인다. 바닥 반사는 자연스럽다.",
        "entities": "흰 침대와 침구, 밝은 낮의 청록색 바다, 바위 해안, 회백색 병실 마감이 요청과 일치한다. 모니터·전화기·벽 안내물·달력도 참조와 대응하지만 오른쪽 의자는 추가된 물건이다. 인물은 없어 최종 인물 금지 지시를 따르며, 상충하는 앞선 환자 등장 요구는 실현되지 않았다. 별도의 화면 자막이나 덧씌운 표시는 없다.",
        "hard_violations": [
         "참조 장소와 프롬프트에 없는 의자 한 개를 오른쪽 수납장 옆에 추가했다."
        ],
        "physics": "침대는 다리와 바퀴로, 보조 탁자는 받침으로 바닥에 지지되어 있다. 침구는 매트리스 위에 내려앉고 용기는 수납장 상판에 놓인다. 추가된 의자도 보이는 다리로 바닥에 서 있다. 벽 설비와 천장 조명의 고정도 자연스러우며, 지지 없는 부유 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.056,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.806,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 없는 사물(우측 스툴 의자)을 발명하여 추가함",
     "[gpt-high] 참조 장소와 프롬프트에 없는 의자 한 개를 오른쪽 수납장 옆에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 806
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스와 일치하는 병실 구조와 배경을 정확하게 구현했으며, 불필요한 사물을 추가하지 않아 프롬프트의 지시를 충실히 따랐습니다."
   },
   {
    "label": "A",
    "score": 806,
    "verdict_ko": "장소의 요소들을 잘 구현했으나, 레퍼런스에 없는 의자를 임의로 추가하여 장소 고정 지침을 위반했습니다.  ★위반: [gemini-pro] 레퍼런스에 없는 사물(우측 스툴 의자)을 발명하여 추가함 / [gpt-high] 참조 장소와 프롬프트에 없는 의자 한 개를 오른쪽 수납장 옆에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L139B02.png",
    "asset_id": "b4bdb1e2-9977-4924-a0b7-a115d0c85eb4",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e33-c78c-779f-867b-6b0d049644e1",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S81sh6::signage": {
  "fp": "980a510f7f66105c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S81sh6": {
  "input_fingerprint": "b26283f2d06d4037",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하얀 연구복을 입은 젊은 서지민이 미소 지으며 서 있는 상반신.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter sickroom. Bright fluorescent light and the sea-facing window illuminate the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just behind and beside 이현우's seated shoulder, looking slightly upward toward 서지민 from outside their face-to-face axis in a direct over-shoulder view. Frame the smiling twenty-year-old 서지민 waist-up at center-right in her white research coat, inclining gently toward 이현우 and looking at his off-screen face; his near shoulder forms only a narrow lower-left edge as he turns toward her. Let the reduced distance emphasize her reassuring expression while preserving their conversational orientation.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bed (Supporting the seated patient) — A small bedside portion remains beneath the near shoulder; used as Maintains the patient's position within the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established strong fluorescent room illumination with controlled white-coat highlights and natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, room finishes, fluorescent lighting, and emerald sea beyond the window from the reference. Exclude boat fittings, storm debris, and laboratory machinery.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white bed, bright fluorescent lighting and emerald-blue sea outside the window remain unchanged. 이현우: He remains awake on the bed in white clothes, having only just raised his upper body. 서지민: She is standing in a research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하얀 연구복을 입은 젊은 서지민이 미소 지으며 서 있는 상반신.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter sickroom. Bright fluorescent light and the sea-facing window illuminate the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just behind and beside 이현우's seated shoulder, looking slightly upward toward 서지민 from outside their face-to-face axis in a direct over-shoulder view. Frame the smiling twenty-year-old 서지민 waist-up at center-right in her white research coat, inclining gently toward 이현우 and looking at his off-screen face; his near shoulder forms only a narrow lower-left edge as he turns toward her. Let the reduced distance emphasize her reassuring expression while preserving their conversational orientation.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bed (Supporting the seated patient) — A small bedside portion remains beneath the near shoulder; used as Maintains the patient's position within the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established strong fluorescent room illumination with controlled white-coat highlights and natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, room finishes, fluorescent lighting, and emerald sea beyond the window from the reference. Exclude boat fittings, storm debris, and laboratory machinery.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white bed, bright fluorescent lighting and emerald-blue sea outside the window remain unchanged. 이현우: He remains awake on the bed in white clothes, having only just raised his upper body. 서지민: She is standing in a research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 하얀 연구복을 입은 젊은 서지민이 미소 지으며 서 있는 상반신.\n\nLOCATION (lock): Beside the white bed in the research institute's shelter sickroom. Bright fluorescent light and the sea-facing window illuminate the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the approach just behind and beside 이현우's seated shoulder, looking slightly upward toward 서지민 from outside their face-to-face axis in a direct over-shoulder view. Frame the smiling twenty-year-old 서지민 waist-up at center-right in her white research coat, inclining gently toward 이현우 and looking at his off-screen face; his near shoulder forms only a narrow lower-left edge as he turns toward her. Let the reduced distance emphasize her reassuring expression while preserving their conversational orientation.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Bed (Supporting the seated patient) — A small bedside portion remains beneath the near shoulder; used as Maintains the patient's position within the room.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established strong fluorescent room illumination with controlled white-coat highlights and natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the hospital bed, room finishes, fluorescent lighting, and emerald sea beyond the window from the reference. Exclude boat fittings, storm debris, and laboratory machinery.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white bed, bright fluorescent lighting and emerald-blue sea outside the window remain unchanged. 이현우: He remains awake on the bed in white clothes, having only just raised his upper body. 서지민: She is standing in a research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "지민은 화면 좌측 하단의 현우를 바라보고 있으며, 현우의 어깨 방향도 지민을 향함.",
    "built_space": "병실 창문, 바다, 우측 수납장 등은 일치하나, 환자가 사용 중이어야 할 침대가 배경에 빈 채로 잘못 배치됨.",
    "entities": "지민은 20대 여성, 흰색 연구복, 검은 머리로 레퍼런스와 일치. 현우는 짧은 머리와 흰옷을 입은 뒷모습으로 일치.",
    "hard_violations": [
     "[gemini-pro] 환자가 앉아있는 침대가 전경에 있어야 하나, 배경에 빈 침대로 생성되어 인물 위치와 공간 구조가 물리적으로 모순됨."
    ],
    "physics": "지민은 바닥에 서서 체중을 지탱함. 현우는 앞쪽에 앉아있으나 몸을 받치는 침대 구조가 시각적으로 연결되지 않음."
   },
   {
    "label": "B",
    "direction": "지민과 현우가 서로 마주보며 시선이 올바르게 교차됨.",
    "built_space": "창문과 수납장 등 병실 구조는 나타나나, A와 동일하게 빈 침대가 배경에 나타나 위치 모순이 발생함.",
    "entities": "지민과 현우 모두 주어진 인물 레퍼런스와 의상(흰 연구복, 흰 옷) 조건을 충족함.",
    "hard_violations": [
     "[gemini-pro] 지정된 위치(침대 위)에 인물이 앉아있지 않고 배경에 온전한 빈 침대가 그려지는 구조적 오류가 발생함."
    ],
    "physics": "지민은 올바르게 서 있음. 현우는 하단에 위치하나 그가 앉아있는 지지대가 화면상 명확히 설명되지 않고 배경과 분리됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지민의 외형과 프레이밍은 프롬프트에 부합하나, 현우가 앉아있어야 할 침대가 배경에 빈 상태로 방치되어 공간 묘사에 치명적인 오류가 있음."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 묘사는 양호하지만, A와 마찬가지로 배경에 온전한 빈 침대가 생성되어 환자의 위치와 물리적 구조가 전혀 맞지 않음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지민은 화면 좌측 하단의 현우를 바라보고 있으며, 현우의 어깨 방향도 지민을 향함.",
        "built_space": "병실 창문, 바다, 우측 수납장 등은 일치하나, 환자가 사용 중이어야 할 침대가 배경에 빈 채로 잘못 배치됨.",
        "entities": "지민은 20대 여성, 흰색 연구복, 검은 머리로 레퍼런스와 일치. 현우는 짧은 머리와 흰옷을 입은 뒷모습으로 일치.",
        "hard_violations": [
         "환자가 앉아있는 침대가 전경에 있어야 하나, 배경에 빈 침대로 생성되어 인물 위치와 공간 구조가 물리적으로 모순됨."
        ],
        "physics": "지민은 바닥에 서서 체중을 지탱함. 현우는 앞쪽에 앉아있으나 몸을 받치는 침대 구조가 시각적으로 연결되지 않음."
       },
       {
        "label": "B",
        "direction": "지민과 현우가 서로 마주보며 시선이 올바르게 교차됨.",
        "built_space": "창문과 수납장 등 병실 구조는 나타나나, A와 동일하게 빈 침대가 배경에 나타나 위치 모순이 발생함.",
        "entities": "지민과 현우 모두 주어진 인물 레퍼런스와 의상(흰 연구복, 흰 옷) 조건을 충족함.",
        "hard_violations": [
         "지정된 위치(침대 위)에 인물이 앉아있지 않고 배경에 온전한 빈 침대가 그려지는 구조적 오류가 발생함."
        ],
        "physics": "지민은 올바르게 서 있음. 현우는 하단에 위치하나 그가 앉아있는 지지대가 화면상 명확히 설명되지 않고 배경과 분리됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지민의 외형과 프레이밍은 프롬프트에 부합하나, 현우가 앉아있어야 할 침대가 배경에 빈 상태로 방치되어 공간 묘사에 치명적인 오류가 있음."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물 묘사는 양호하지만, A와 마찬가지로 배경에 온전한 빈 침대가 생성되어 환자의 위치와 물리적 구조가 전혀 맞지 않음."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "지민은 화면 좌측 하단의 현우를 바라보고 있으며, 현우의 어깨 방향도 지민을 향함.",
        "built_space": "병실 창문, 바다, 우측 수납장 등은 일치하나, 환자가 사용 중이어야 할 침대가 배경에 빈 채로 잘못 배치됨.",
        "entities": "지민은 20대 여성, 흰색 연구복, 검은 머리로 레퍼런스와 일치. 현우는 짧은 머리와 흰옷을 입은 뒷모습으로 일치.",
        "hard_violations": [
         "환자가 앉아있는 침대가 전경에 있어야 하나, 배경에 빈 침대로 생성되어 인물 위치와 공간 구조가 물리적으로 모순됨."
        ],
        "physics": "지민은 바닥에 서서 체중을 지탱함. 현우는 앞쪽에 앉아있으나 몸을 받치는 침대 구조가 시각적으로 연결되지 않음."
       },
       {
        "label": "B",
        "direction": "지민과 현우가 서로 마주보며 시선이 올바르게 교차됨.",
        "built_space": "창문과 수납장 등 병실 구조는 나타나나, A와 동일하게 빈 침대가 배경에 나타나 위치 모순이 발생함.",
        "entities": "지민과 현우 모두 주어진 인물 레퍼런스와 의상(흰 연구복, 흰 옷) 조건을 충족함.",
        "hard_violations": [
         "지정된 위치(침대 위)에 인물이 앉아있지 않고 배경에 온전한 빈 침대가 그려지는 구조적 오류가 발생함."
        ],
        "physics": "지민은 올바르게 서 있음. 현우는 하단에 위치하나 그가 앉아있는 지지대가 화면상 명확히 설명되지 않고 배경과 분리됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "중앙 오른쪽의 허리 위 서지민과 환자를 향한 미소·기울임은 잘 맞지만, 이현우의 머리와 어깨가 왼쪽을 크게 차지해 ‘좌하단의 좁은 어깨 가장자리’ 지시를 놓쳤다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물의 미소와 대화 방향, 병실 연속성은 맞지만, 이현우의 뒤통수와 어깨가 A보다 더 크게 들어와 지정된 제한적 전경과 서지민 중심 구도에서 더 멀어졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "서지민은 화면 왼쪽 아래의 이현우 얼굴 쪽으로 눈을 돌리고 몸을 가볍게 기울이며 미소 짓는다. 이현우의 뒤통수와 일부 옆얼굴은 서지민 쪽으로 향한다. 서로 대화하는 방향은 맞으며 렌즈를 직접 보는 모습은 아니다.",
        "built_space": "뒤쪽에 블라인드가 달린 바다 방향 창 하나, 천장 형광등 하나, 왼쪽의 흰 침대 하나와 베개 하나, 오른쪽의 작은 수납장 하나가 보인다. 오른쪽 벽의 달력과 해안 사진도 참조 공간과 대응한다. 서지민은 침대 오른쪽에 서 있고 이현우는 침대 전경 쪽에 위치한다. 고정 설비의 중복이나 불가능한 반사는 보이지 않는다. 다만 침대와 환자 머리·어깨가 지정된 작은 부분보다 넓게 노출된다.",
        "entities": "보이는 사람은 두 명뿐이다. 서지민은 참조와 유사한 젊은 동아시아계 여성 얼굴, 긴 검은 머리, 머리 위 보호안경, 회색 셔츠와 흰 연구 가운을 갖췄다. 가운의 깃과 자수 세부는 참조와 조금 다르다. 이현우는 헝클어진 짧은 검은 머리와 흰 옷을 입었지만 얼굴 대부분이 가려져 정확한 얼굴 일치는 확인하기 어렵다. 흰 침구와 낮의 푸른 청록색 바다는 요구에 부합한다.",
        "hard_violations": [],
        "physics": "서지민의 상체 기울임은 서 있는 사람이 자연스럽게 만들 수 있는 범위이며, 하체와 발은 프레임 밖이다. 이현우의 아래쪽 몸은 가려졌지만 어깨 아래로 이어지는 침대가 있어 앉아 있는 자세와 모순되지 않는다. 베개와 침구는 침대에 놓이고 보호안경은 머리에 걸쳐 있다. 공중에 뜨거나 지지 없이 들린 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "서지민의 시선과 미소는 왼쪽 전경의 이현우 얼굴을 향하고, 상체도 그에게 살짝 기울어 있다. 이현우는 서지민 쪽으로 고개를 돌린 뒷모습이다. 대화 상대를 향하는 시선 관계는 맞는다.",
        "built_space": "블라인드가 있는 창 하나와 천장 형광등 하나, 왼쪽의 침대 하나와 베개 하나, 오른쪽의 수납장 하나가 보인다. 창밖 해안과 오른쪽 벽의 달력은 참조 병실을 이어 간다. 서지민은 침대 옆 오른쪽, 이현우는 침대 전경 쪽에 있다. 설비 중복이나 반사 오류는 보이지 않는다. 그러나 환자의 뒤통수가 왼쪽 높이 대부분을 차지하고 어깨가 하단으로 크게 퍼져, 좁은 좌하단 가장자리만 남기라는 구도와 맞지 않는다.",
        "entities": "추가 인물 없이 서지민과 이현우만 보인다. 서지민의 젊은 동아시아계 여성 외모, 검은 긴 머리, 머리 위 보호안경, 흰 연구 가운과 회색 셔츠는 참조에 대체로 대응한다. 가운의 깃과 로고 형태에는 차이가 있다. 이현우의 검은 헝클어진 머리와 흰 옷은 맞지만 뒤쪽 위주라 얼굴 정체성은 확인할 수 없다. 흰 침대와 주간의 청록색 바다가 존재한다.",
        "hard_violations": [],
        "physics": "서지민의 가벼운 전방 기울임과 내려온 팔은 자연스럽고, 발은 구도 밖이라 접지 여부를 직접 볼 수 없다. 이현우의 골반은 가려져 있지만 몸 아래와 뒤로 침대가 이어져 앉은 자세를 지지할 수 있다. 침구와 베개는 매트리스 위에, 병들은 수납장 위에 놓여 있다. 지지 없는 부유나 불가능한 신체 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "중앙 오른쪽의 허리 위 서지민과 환자를 향한 미소·기울임은 잘 맞지만, 이현우의 머리와 어깨가 왼쪽을 크게 차지해 ‘좌하단의 좁은 어깨 가장자리’ 지시를 놓쳤다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물의 미소와 대화 방향, 병실 연속성은 맞지만, 이현우의 뒤통수와 어깨가 A보다 더 크게 들어와 지정된 제한적 전경과 서지민 중심 구도에서 더 멀어졌다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "서지민은 화면 왼쪽 아래의 이현우 얼굴 쪽으로 눈을 돌리고 몸을 가볍게 기울이며 미소 짓는다. 이현우의 뒤통수와 일부 옆얼굴은 서지민 쪽으로 향한다. 서로 대화하는 방향은 맞으며 렌즈를 직접 보는 모습은 아니다.",
        "built_space": "뒤쪽에 블라인드가 달린 바다 방향 창 하나, 천장 형광등 하나, 왼쪽의 흰 침대 하나와 베개 하나, 오른쪽의 작은 수납장 하나가 보인다. 오른쪽 벽의 달력과 해안 사진도 참조 공간과 대응한다. 서지민은 침대 오른쪽에 서 있고 이현우는 침대 전경 쪽에 위치한다. 고정 설비의 중복이나 불가능한 반사는 보이지 않는다. 다만 침대와 환자 머리·어깨가 지정된 작은 부분보다 넓게 노출된다.",
        "entities": "보이는 사람은 두 명뿐이다. 서지민은 참조와 유사한 젊은 동아시아계 여성 얼굴, 긴 검은 머리, 머리 위 보호안경, 회색 셔츠와 흰 연구 가운을 갖췄다. 가운의 깃과 자수 세부는 참조와 조금 다르다. 이현우는 헝클어진 짧은 검은 머리와 흰 옷을 입었지만 얼굴 대부분이 가려져 정확한 얼굴 일치는 확인하기 어렵다. 흰 침구와 낮의 푸른 청록색 바다는 요구에 부합한다.",
        "hard_violations": [],
        "physics": "서지민의 상체 기울임은 서 있는 사람이 자연스럽게 만들 수 있는 범위이며, 하체와 발은 프레임 밖이다. 이현우의 아래쪽 몸은 가려졌지만 어깨 아래로 이어지는 침대가 있어 앉아 있는 자세와 모순되지 않는다. 베개와 침구는 침대에 놓이고 보호안경은 머리에 걸쳐 있다. 공중에 뜨거나 지지 없이 들린 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "서지민의 시선과 미소는 왼쪽 전경의 이현우 얼굴을 향하고, 상체도 그에게 살짝 기울어 있다. 이현우는 서지민 쪽으로 고개를 돌린 뒷모습이다. 대화 상대를 향하는 시선 관계는 맞는다.",
        "built_space": "블라인드가 있는 창 하나와 천장 형광등 하나, 왼쪽의 침대 하나와 베개 하나, 오른쪽의 수납장 하나가 보인다. 창밖 해안과 오른쪽 벽의 달력은 참조 병실을 이어 간다. 서지민은 침대 옆 오른쪽, 이현우는 침대 전경 쪽에 있다. 설비 중복이나 반사 오류는 보이지 않는다. 그러나 환자의 뒤통수가 왼쪽 높이 대부분을 차지하고 어깨가 하단으로 크게 퍼져, 좁은 좌하단 가장자리만 남기라는 구도와 맞지 않는다.",
        "entities": "추가 인물 없이 서지민과 이현우만 보인다. 서지민의 젊은 동아시아계 여성 외모, 검은 긴 머리, 머리 위 보호안경, 흰 연구 가운과 회색 셔츠는 참조에 대체로 대응한다. 가운의 깃과 로고 형태에는 차이가 있다. 이현우의 검은 헝클어진 머리와 흰 옷은 맞지만 뒤쪽 위주라 얼굴 정체성은 확인할 수 없다. 흰 침대와 주간의 청록색 바다가 존재한다.",
        "hard_violations": [],
        "physics": "서지민의 가벼운 전방 기울임과 내려온 팔은 자연스럽고, 발은 구도 밖이라 접지 여부를 직접 볼 수 없다. 이현우의 골반은 가려져 있지만 몸 아래와 뒤로 침대가 이어져 앉은 자세를 지지할 수 있다. 침구와 베개는 매트리스 위에, 병들은 수납장 위에 놓여 있다. 지지 없는 부유나 불가능한 신체 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.607,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 환자가 앉아있는 침대가 전경에 있어야 하나, 배경에 빈 침대로 생성되어 인물 위치와 공간 구조가 물리적으로 모순됨."
    ],
    "B": [
     "[gemini-pro] 지정된 위치(침대 위)에 인물이 앉아있지 않고 배경에 온전한 빈 침대가 그려지는 구조적 오류가 발생함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1607,
   "B": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1607,
    "verdict_ko": "지민의 외형과 프레이밍은 프롬프트에 부합하나, 현우가 앉아있어야 할 침대가 배경에 빈 상태로 방치되어 공간 묘사에 치명적인 오류가 있음.  ★위반: [gemini-pro] 환자가 앉아있는 침대가 전경에 있어야 하나, 배경에 빈 침대로 생성되어 인물 위치와 공간 구조가 물리적으로 모순됨."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "인물 묘사는 양호하지만, A와 마찬가지로 배경에 온전한 빈 침대가 생성되어 환자의 위치와 물리적 구조가 전혀 맞지 않음.  ★위반: [gemini-pro] 지정된 위치(침대 위)에 인물이 앉아있지 않고 배경에 온전한 빈 침대가 그려지는 구조적 오류가 발생함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S81sh4_sel.png",
    "asset_id": "18c905c0-c790-431b-8613-4af660f9ff2b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869875>",
    "asset_id": "bdc552c6-5e2a-4bae-affd-3dbee764db52",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e37-f7e0-7b2c-a058-f212d4924157",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh4"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S81sh11::signage": {
  "fp": "afdcddf6c612c24d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S81sh11": {
  "input_fingerprint": "4006072d44130ad0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동문이 양옆으로 열리는 중인 그 간격 사이로 눈부신 빛이 쏟아져 나오는 찰나.\n\nLOCATION (lock): At the automatic security doorway in the research institute's corridor. Bright light spills through the widening opening from the center beyond. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the corridor side, continue the low, off-axis advance toward the threshold with a slight upward tilt, directly observing the first separation of the automatic door panels. Keep the doorway around the central third of the image, with adjacent wall and corridor floor retaining scale, and place the narrow opening just right of center while both panels withdraw laterally. Hold the space beyond unrevealed and emphasize the panels' positional change rather than inventing a new interior view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrow gap between separating door panels in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Automatic door panels (Beginning to separate to either side) — Corridor-facing surfaces seen obliquely, with their inner edges defining a narrow gap; used as Central reveal mechanism, kept below forty percent of the total frame; Corridor wall (Visible on either side of the doorway) — Recedes obliquely around the threshold; used as Provides scale and prevents the opening from becoming an abstract close-up; Corridor floor (Visible in front of the threshold); used as Lower-frame depth cue for the advancing camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dazzling light escapes through the first narrow gap, with its source and color unspecified and the corridor exposure held to retain the door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The access-controlled door is opening after a research staff card is presented to the wall reader.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동문이 양옆으로 열리는 중인 그 간격 사이로 눈부신 빛이 쏟아져 나오는 찰나.\n\nLOCATION (lock): At the automatic security doorway in the research institute's corridor. Bright light spills through the widening opening from the center beyond. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the corridor side, continue the low, off-axis advance toward the threshold with a slight upward tilt, directly observing the first separation of the automatic door panels. Keep the doorway around the central third of the image, with adjacent wall and corridor floor retaining scale, and place the narrow opening just right of center while both panels withdraw laterally. Hold the space beyond unrevealed and emphasize the panels' positional change rather than inventing a new interior view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrow gap between separating door panels in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Automatic door panels (Beginning to separate to either side) — Corridor-facing surfaces seen obliquely, with their inner edges defining a narrow gap; used as Central reveal mechanism, kept below forty percent of the total frame; Corridor wall (Visible on either side of the doorway) — Recedes obliquely around the threshold; used as Provides scale and prevents the opening from becoming an abstract close-up; Corridor floor (Visible in front of the threshold); used as Lower-frame depth cue for the advancing camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dazzling light escapes through the first narrow gap, with its source and color unspecified and the corridor exposure held to retain the door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The access-controlled door is opening after a research staff card is presented to the wall reader.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 자동문이 양옆으로 열리는 중인 그 간격 사이로 눈부신 빛이 쏟아져 나오는 찰나.\n\nLOCATION (lock): At the automatic security doorway in the research institute's corridor. Bright light spills through the widening opening from the center beyond. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the corridor side, continue the low, off-axis advance toward the threshold with a slight upward tilt, directly observing the first separation of the automatic door panels. Keep the doorway around the central third of the image, with adjacent wall and corridor floor retaining scale, and place the narrow opening just right of center while both panels withdraw laterally. Hold the space beyond unrevealed and emphasize the panels' positional change rather than inventing a new interior view.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Narrow gap between separating door panels in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Automatic door panels (Beginning to separate to either side) — Corridor-facing surfaces seen obliquely, with their inner edges defining a narrow gap; used as Central reveal mechanism, kept below forty percent of the total frame; Corridor wall (Visible on either side of the doorway) — Recedes obliquely around the threshold; used as Provides scale and prevents the opening from becoming an abstract close-up; Corridor floor (Visible in front of the threshold); used as Lower-frame depth cue for the advancing camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dazzling light escapes through the first narrow gap, with its source and color unspecified and the corridor exposure held to retain the door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The access-controlled door is opening after a research staff card is presented to the wall reader.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "열리는 문 틈 중앙에서부터 밝은 빛이 카메라가 위치한 복도 쪽으로 강하게 쏟아져 나옴.",
    "built_space": "좌우로 복도 벽이 있고 중앙에 금속 재질의 양개형 자동문이 위치함. 좌측 벽에 단말기와 표지판, 우측 벽에 표지판이 레퍼런스와 동일한 위치에 있음.",
    "entities": "지정된 텍스트('연구구역 출입통제', '연구구역 RESEARCH AREA')가 오타 없이 재현된 표지판, 조금 열린 금속 문. 인물은 존재하지 않음.",
    "hard_violations": [],
    "physics": "금속 문이 물리적으로 올바른 궤도로 열리고 있으며, 강한 빛이 바닥 재질의 특성에 맞게 반사됨."
   },
   {
    "label": "B",
    "direction": "문 틈 사이에서 발생한 쏟아지는 빛이 복도를 향해 똑바로 비춤.",
    "built_space": "레퍼런스와 일치하는 복도 구조. 양측 벽면의 표지판과 좌측의 카드 리더기 배치가 정확함.",
    "entities": "프롬프트가 요구한 한글 및 영문 텍스트가 정확히 기재된 두 개의 표지판, 중앙의 금속문. 요구대로 프레임 내에 인물은 없음.",
    "hard_violations": [],
    "physics": "문 패널의 분리와 빛의 확산, 바닥면의 반사 등 물리적 요소들이 자연스럽게 묘사됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간 구조와 표지판 텍스트를 정확히 재현했으며, 문 틈 사이로 새어나오는 빛과 바닥 반사가 프롬프트의 시네마틱한 의도를 훌륭하게 살렸습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 구조, 텍스트, 인물 부재 등의 조건을 모두 정확히 충족했으나, 조명과 렌즈 효과의 극적인 느낌이 A에 비해 다소 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "열리는 문 틈 중앙에서부터 밝은 빛이 카메라가 위치한 복도 쪽으로 강하게 쏟아져 나옴.",
        "built_space": "좌우로 복도 벽이 있고 중앙에 금속 재질의 양개형 자동문이 위치함. 좌측 벽에 단말기와 표지판, 우측 벽에 표지판이 레퍼런스와 동일한 위치에 있음.",
        "entities": "지정된 텍스트('연구구역 출입통제', '연구구역 RESEARCH AREA')가 오타 없이 재현된 표지판, 조금 열린 금속 문. 인물은 존재하지 않음.",
        "hard_violations": [],
        "physics": "금속 문이 물리적으로 올바른 궤도로 열리고 있으며, 강한 빛이 바닥 재질의 특성에 맞게 반사됨."
       },
       {
        "label": "B",
        "direction": "문 틈 사이에서 발생한 쏟아지는 빛이 복도를 향해 똑바로 비춤.",
        "built_space": "레퍼런스와 일치하는 복도 구조. 양측 벽면의 표지판과 좌측의 카드 리더기 배치가 정확함.",
        "entities": "프롬프트가 요구한 한글 및 영문 텍스트가 정확히 기재된 두 개의 표지판, 중앙의 금속문. 요구대로 프레임 내에 인물은 없음.",
        "hard_violations": [],
        "physics": "문 패널의 분리와 빛의 확산, 바닥면의 반사 등 물리적 요소들이 자연스럽게 묘사됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "레퍼런스의 공간 구조와 표지판 텍스트를 정확히 재현했으며, 문 틈 사이로 새어나오는 빛과 바닥 반사가 프롬프트의 시네마틱한 의도를 훌륭하게 살렸습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 구조, 텍스트, 인물 부재 등의 조건을 모두 정확히 충족했으나, 조명과 렌즈 효과의 극적인 느낌이 A에 비해 다소 약합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "열리는 문 틈 중앙에서부터 밝은 빛이 카메라가 위치한 복도 쪽으로 강하게 쏟아져 나옴.",
        "built_space": "좌우로 복도 벽이 있고 중앙에 금속 재질의 양개형 자동문이 위치함. 좌측 벽에 단말기와 표지판, 우측 벽에 표지판이 레퍼런스와 동일한 위치에 있음.",
        "entities": "지정된 텍스트('연구구역 출입통제', '연구구역 RESEARCH AREA')가 오타 없이 재현된 표지판, 조금 열린 금속 문. 인물은 존재하지 않음.",
        "hard_violations": [],
        "physics": "금속 문이 물리적으로 올바른 궤도로 열리고 있으며, 강한 빛이 바닥 재질의 특성에 맞게 반사됨."
       },
       {
        "label": "B",
        "direction": "문 틈 사이에서 발생한 쏟아지는 빛이 복도를 향해 똑바로 비춤.",
        "built_space": "레퍼런스와 일치하는 복도 구조. 양측 벽면의 표지판과 좌측의 카드 리더기 배치가 정확함.",
        "entities": "프롬프트가 요구한 한글 및 영문 텍스트가 정확히 기재된 두 개의 표지판, 중앙의 금속문. 요구대로 프레임 내에 인물은 없음.",
        "hard_violations": [],
        "physics": "문 패널의 분리와 빛의 확산, 바닥면의 반사 등 물리적 요소들이 자연스럽게 묘사됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "낮은 복도 측 시점과 바닥의 깊이감이 지정된 접근 구도에 더 충실하지만, 틈 너머 실내를 완전히 감추지는 못했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "장소와 좁은 개방 틈은 충실하나 참조 사진의 시점을 거의 답습하여 낮은 접근 구도가 약하고, 틈 너머 실내도 드러난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 금속 문짝의 안쪽 세로 모서리가 화면 중심보다 조금 오른쪽에서 좁은 틈을 만든다. 좌우로 분리된 상태는 양옆으로 열리는 동작과 맞지만, 정지 화면에서 이동 자체는 확인되지 않는다. 빛은 틈 안쪽에서 카메라 쪽 복도와 전경 바닥으로 퍼진다. 시선이나 조준 물체는 없다.",
        "built_space": "양개문 한 조, 상부 센서 하나, 왼쪽 벽 카드 판독기 하나, 그 위 출입통제 표지 하나, 아래 배선함 하나, 오른쪽 연구구역 표지 하나가 보인다. 금속 문틀, 노출 배관, 얼룩진 밝은 벽, 검은 걸레받이와 회색 바닥이 참조 장소에 대응한다. 낮은 카메라에서 왼쪽 벽을 비스듬히 보며 문턱에 접근하는 와이드 구도이고, 문은 중경에 있으며 화면 면적의 40퍼센트 미만이다. 다만 틈 안으로 천장 조명과 실내 윤곽이 보여 내부 비공개 조건은 충족하지 못한다.",
        "entities": "사람이나 얼굴은 없다. 문짝은 실제 금속 자동문으로 읽히고, 왼쪽 장치는 참조와 같은 벽 부착형 출입 판독기다. 두 표지의 한국어와 오른쪽 표지의 영문은 참조에 대응하며 별도 자막이나 새 문구는 없다. 직원 카드와 손은 보이지 않지만 카드 제시가 끝난 시점이므로 누락 문제가 아니다.",
        "hard_violations": [],
        "physics": "문짝은 문틀과 문턱의 슬라이딩 구조에 결합되어 있으며 좌우 이동이 가능한 배치다. 판독기, 표지와 배선함은 벽에 고정되어 있다. 틈 아래에서 전경으로 길어지는 바닥 반사는 광원과 카메라 위치상 가능하다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "두 문짝 사이의 좁은 세로 틈이 화면 중심보다 오른쪽에 있고, 문짝은 각각 좌우로 물러난 상태다. 틈의 강한 빛이 복도 방향으로 나오며 바닥에 긴 밝은 띠를 만든다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "양개문 한 조와 상부 센서 하나, 왼쪽 판독기 하나·출입통제 표지 하나·하부 배선함 하나, 오른쪽 연구구역 표지 하나가 참조와 거의 같은 위치에 있다. 벽의 굴곡, 배관, 검은 걸레받이와 회색 바닥도 유지된다. 문은 중경에 있고 양옆 벽과 전경 바닥을 포함하는 와이드 구도지만, 참조 사진과 거의 같은 높이와 구도로 A보다 낮은 시점이 약하다. 틈 너머 천장 조명, 벽과 바닥이 식별되어 내부를 감추라는 조건에 어긋난다.",
        "entities": "사람과 얼굴 없이 금속 자동문, 출입 판독기, 기존 표지들이 보인다. 표지의 한국어 및 영문 표기는 참조의 종류와 대응한다. 별도 문구나 그래픽은 없다. 카드나 손이 없는 것은 카드 제시 후 개방 장면과 양립한다.",
        "hard_violations": [],
        "physics": "문짝은 문틀과 하부 가이드에 지지된 정상적인 슬라이딩 구조로 보인다. 벽 부착물은 모두 고정되어 있고 부유 물체는 없다. 문틈에서 들어온 빛의 바닥 반사와 문 가장자리의 눈부심은 물리적으로 가능한 배치다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "낮은 복도 측 시점과 바닥의 깊이감이 지정된 접근 구도에 더 충실하지만, 틈 너머 실내를 완전히 감추지는 못했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "장소와 좁은 개방 틈은 충실하나 참조 사진의 시점을 거의 답습하여 낮은 접근 구도가 약하고, 틈 너머 실내도 드러난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 금속 문짝의 안쪽 세로 모서리가 화면 중심보다 조금 오른쪽에서 좁은 틈을 만든다. 좌우로 분리된 상태는 양옆으로 열리는 동작과 맞지만, 정지 화면에서 이동 자체는 확인되지 않는다. 빛은 틈 안쪽에서 카메라 쪽 복도와 전경 바닥으로 퍼진다. 시선이나 조준 물체는 없다.",
        "built_space": "양개문 한 조, 상부 센서 하나, 왼쪽 벽 카드 판독기 하나, 그 위 출입통제 표지 하나, 아래 배선함 하나, 오른쪽 연구구역 표지 하나가 보인다. 금속 문틀, 노출 배관, 얼룩진 밝은 벽, 검은 걸레받이와 회색 바닥이 참조 장소에 대응한다. 낮은 카메라에서 왼쪽 벽을 비스듬히 보며 문턱에 접근하는 와이드 구도이고, 문은 중경에 있으며 화면 면적의 40퍼센트 미만이다. 다만 틈 안으로 천장 조명과 실내 윤곽이 보여 내부 비공개 조건은 충족하지 못한다.",
        "entities": "사람이나 얼굴은 없다. 문짝은 실제 금속 자동문으로 읽히고, 왼쪽 장치는 참조와 같은 벽 부착형 출입 판독기다. 두 표지의 한국어와 오른쪽 표지의 영문은 참조에 대응하며 별도 자막이나 새 문구는 없다. 직원 카드와 손은 보이지 않지만 카드 제시가 끝난 시점이므로 누락 문제가 아니다.",
        "hard_violations": [],
        "physics": "문짝은 문틀과 문턱의 슬라이딩 구조에 결합되어 있으며 좌우 이동이 가능한 배치다. 판독기, 표지와 배선함은 벽에 고정되어 있다. 틈 아래에서 전경으로 길어지는 바닥 반사는 광원과 카메라 위치상 가능하다. 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "두 문짝 사이의 좁은 세로 틈이 화면 중심보다 오른쪽에 있고, 문짝은 각각 좌우로 물러난 상태다. 틈의 강한 빛이 복도 방향으로 나오며 바닥에 긴 밝은 띠를 만든다. 사람의 시선이나 겨냥하는 물체는 없다.",
        "built_space": "양개문 한 조와 상부 센서 하나, 왼쪽 판독기 하나·출입통제 표지 하나·하부 배선함 하나, 오른쪽 연구구역 표지 하나가 참조와 거의 같은 위치에 있다. 벽의 굴곡, 배관, 검은 걸레받이와 회색 바닥도 유지된다. 문은 중경에 있고 양옆 벽과 전경 바닥을 포함하는 와이드 구도지만, 참조 사진과 거의 같은 높이와 구도로 A보다 낮은 시점이 약하다. 틈 너머 천장 조명, 벽과 바닥이 식별되어 내부를 감추라는 조건에 어긋난다.",
        "entities": "사람과 얼굴 없이 금속 자동문, 출입 판독기, 기존 표지들이 보인다. 표지의 한국어 및 영문 표기는 참조의 종류와 대응한다. 별도 문구나 그래픽은 없다. 카드나 손이 없는 것은 카드 제시 후 개방 장면과 양립한다.",
        "hard_violations": [],
        "physics": "문짝은 문틀과 하부 가이드에 지지된 정상적인 슬라이딩 구조로 보인다. 벽 부착물은 모두 고정되어 있고 부유 물체는 없다. 문틈에서 들어온 빛의 바닥 반사와 문 가장자리의 눈부심은 물리적으로 가능한 배치다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1857
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 공간 구조와 표지판 텍스트를 정확히 재현했으며, 문 틈 사이로 새어나오는 빛과 바닥 반사가 프롬프트의 시네마틱한 의도를 훌륭하게 살렸습니다."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "요구된 구조, 텍스트, 인물 부재 등의 조건을 모두 정확히 충족했으나, 조명과 렌즈 효과의 극적인 느낌이 A에 비해 다소 약합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L139B01.png",
    "asset_id": "f812f44f-1749-4c6b-9a4d-39a454465c5c",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e3c-dad3-7bc0-a5f8-efbfdac79ab7",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S82sh1::signage": {
  "fp": "46e2a505ae3689e0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S82sh1": {
  "input_fingerprint": "ea3f11c196362dea",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 원자력 발전소 같은 연구 설비와 열대 식물들이 가득한 메인 연구센터의 웅장한 전경.\n\nLOCATION (lock): Inside the island institute's vast main research hall, among large energy-research installations, computers, and tropical plants. Bright interior illumination reveals the expansive scale. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the entrance-side start of the crane path, hold a high, distant, diagonally downward establishing view before beginning the descent, observing the main research space directly. Retain 이현우 and 서지민 as small foreground figures seen from behind, with 이현우 arrested mid-step and 서지민 slightly farther forward, both oriented toward equipment extending beyond the frame. Distribute computers, unfamiliar research apparatus, tropical plants, and flowers through successive depth layers, emphasizing the widened camera distance and monumental layout without letting one machine occupy more than forty percent of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research apparatus (Numerous unfamiliar devices distributed through the main space) — Different oblique faces remain visible across multiple depth layers; used as Creates the vast industrial research scale without asserting a completed reactor; Computers (Present among the research installations) — Seen as oblique workstation forms without legible screen content; used as Smaller scale references among the large apparatus; Tropical plants and flowers (Numerous throughout the main research space); used as Break up the apparatus layers and contrast living forms with the industrial setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use balanced ambient illumination appropriate to the daytime interior, with subdued brightness and controlled contrast rather than invented reactor glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main center contains extensive computers, research equipment, tropical plants and flowers. In the connected laboratory, Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with monitoring wires attached and robotic arms scanning him.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 원자력 발전소 같은 연구 설비와 열대 식물들이 가득한 메인 연구센터의 웅장한 전경.\n\nLOCATION (lock): Inside the island institute's vast main research hall, among large energy-research installations, computers, and tropical plants. Bright interior illumination reveals the expansive scale. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the entrance-side start of the crane path, hold a high, distant, diagonally downward establishing view before beginning the descent, observing the main research space directly. Retain 이현우 and 서지민 as small foreground figures seen from behind, with 이현우 arrested mid-step and 서지민 slightly farther forward, both oriented toward equipment extending beyond the frame. Distribute computers, unfamiliar research apparatus, tropical plants, and flowers through successive depth layers, emphasizing the widened camera distance and monumental layout without letting one machine occupy more than forty percent of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research apparatus (Numerous unfamiliar devices distributed through the main space) — Different oblique faces remain visible across multiple depth layers; used as Creates the vast industrial research scale without asserting a completed reactor; Computers (Present among the research installations) — Seen as oblique workstation forms without legible screen content; used as Smaller scale references among the large apparatus; Tropical plants and flowers (Numerous throughout the main research space); used as Break up the apparatus layers and contrast living forms with the industrial setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use balanced ambient illumination appropriate to the daytime interior, with subdued brightness and controlled contrast rather than invented reactor glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main center contains extensive computers, research equipment, tropical plants and flowers. In the connected laboratory, Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with monitoring wires attached and robotic arms scanning him.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 원자력 발전소 같은 연구 설비와 열대 식물들이 가득한 메인 연구센터의 웅장한 전경.\n\nLOCATION (lock): Inside the island institute's vast main research hall, among large energy-research installations, computers, and tropical plants. Bright interior illumination reveals the expansive scale. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the entrance-side start of the crane path, hold a high, distant, diagonally downward establishing view before beginning the descent, observing the main research space directly. Retain 이현우 and 서지민 as small foreground figures seen from behind, with 이현우 arrested mid-step and 서지민 slightly farther forward, both oriented toward equipment extending beyond the frame. Distribute computers, unfamiliar research apparatus, tropical plants, and flowers through successive depth layers, emphasizing the widened camera distance and monumental layout without letting one machine occupy more than forty percent of the image.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Research apparatus (Numerous unfamiliar devices distributed through the main space) — Different oblique faces remain visible across multiple depth layers; used as Creates the vast industrial research scale without asserting a completed reactor; Computers (Present among the research installations) — Seen as oblique workstation forms without legible screen content; used as Smaller scale references among the large apparatus; Tropical plants and flowers (Numerous throughout the main research space); used as Break up the apparatus layers and contrast living forms with the industrial setting.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use balanced ambient illumination appropriate to the daytime interior, with subdued brightness and controlled contrast rather than invented reactor glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main center contains extensive computers, research equipment, tropical plants and flowers. In the connected laboratory, Charlie lies inactive on a stainless-steel bed inside a circular glass enclosure, with monitoring wires attached and robotic arms scanning him.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 대형 연구소를 대각선 아래로 넓게 내려다봄. 워크스테이션의 모니터들은 사용자의 의자를 향해 올바르게 배치됨. 시선의 주체가 되는 인물은 없음.",
    "built_space": "높은 천장과 창문이 있는 거대한 실내 공간. 좌측에 거대한 원자로 형태의 설비, 우측에 원형 유리 구조물이 안정적으로 배치됨. 층간 깊이와 바닥의 동선 표시가 자연스럽고 구조적으로 일관됨.",
    "entities": "열대 식물, 대형 연구 장비, 컴퓨터 워크스테이션, 원형 유리실 내부의 금속 침대와 로봇 팔이 모두 명확히 존재함. 이현우, 서지민, 찰리는 프레임 내 인물 등장 금지 규칙에 따라 모두 생략됨.",
    "hard_violations": [
     "[gpt-high] 프롬프트에 없는 로고와 작은 글자성 표식이 인쇄된 배너 세 장을 추가하여, 임의 표식 및 문구를 만들지 말라는 지시를 위반했다."
    ],
    "physics": "모든 사물이 바닥이나 책상 위에 안정적으로 지지되어 있으며, 중력에 위배되거나 허공에 떠 있는 객체 없이 물리 법칙에 부합함."
   },
   {
    "label": "B",
    "direction": "카메라가 연구소를 넓게 내려다봄. 중간 좌측 워크스테이션의 일부 모니터 방향이 불규칙하고 의자의 위치와 어긋나 있음.",
    "built_space": "장비와 식물이 공간에 넓게 배치되어 있으나, 좌측의 경사로 구조가 평평한 바닥과 이어지는 지점 및 난간의 연결 부위가 기하학적으로 어긋나 있음.",
    "entities": "식물, 기계 장치, 유리실, 침대, 로봇 팔이 존재하나 세부 형태가 다소 뭉개져 있음. 지정된 인물들은 생략됨.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 건축 구조 (좌측 경사로 바닥면과 난간이 형태 없이 바닥에 융합됨)"
    ],
    "physics": "중간 좌측 책상 주변의 일부 기기가 지지대 없이 융합되거나 떠 있는 것처럼 보이며, 구조물의 결합 부위가 물리적으로 불가능한 형태를 띰."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 웅장한 스케일, 거대 설비, 식물, 원형 유리 연구실 등을 물리적 오류 없이 훌륭하게 구현했으나, 상충하는 지시문('인물 없음' 규칙)으로 인해 지정된 인물들은 생략되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지정된 배경 요소들은 존재하지만 좌측 경사로의 심각한 구조적 오류와 뭉개진 사물 묘사로 인해 사실감이 크게 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 대형 연구소를 대각선 아래로 넓게 내려다봄. 워크스테이션의 모니터들은 사용자의 의자를 향해 올바르게 배치됨. 시선의 주체가 되는 인물은 없음.",
        "built_space": "높은 천장과 창문이 있는 거대한 실내 공간. 좌측에 거대한 원자로 형태의 설비, 우측에 원형 유리 구조물이 안정적으로 배치됨. 층간 깊이와 바닥의 동선 표시가 자연스럽고 구조적으로 일관됨.",
        "entities": "열대 식물, 대형 연구 장비, 컴퓨터 워크스테이션, 원형 유리실 내부의 금속 침대와 로봇 팔이 모두 명확히 존재함. 이현우, 서지민, 찰리는 프레임 내 인물 등장 금지 규칙에 따라 모두 생략됨.",
        "hard_violations": [],
        "physics": "모든 사물이 바닥이나 책상 위에 안정적으로 지지되어 있으며, 중력에 위배되거나 허공에 떠 있는 객체 없이 물리 법칙에 부합함."
       },
       {
        "label": "B",
        "direction": "카메라가 연구소를 넓게 내려다봄. 중간 좌측 워크스테이션의 일부 모니터 방향이 불규칙하고 의자의 위치와 어긋나 있음.",
        "built_space": "장비와 식물이 공간에 넓게 배치되어 있으나, 좌측의 경사로 구조가 평평한 바닥과 이어지는 지점 및 난간의 연결 부위가 기하학적으로 어긋나 있음.",
        "entities": "식물, 기계 장치, 유리실, 침대, 로봇 팔이 존재하나 세부 형태가 다소 뭉개져 있음. 지정된 인물들은 생략됨.",
        "hard_violations": [
         "물리적으로 불가능한 건축 구조 (좌측 경사로 바닥면과 난간이 형태 없이 바닥에 융합됨)"
        ],
        "physics": "중간 좌측 책상 주변의 일부 기기가 지지대 없이 융합되거나 떠 있는 것처럼 보이며, 구조물의 결합 부위가 물리적으로 불가능한 형태를 띰."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 웅장한 스케일, 거대 설비, 식물, 원형 유리 연구실 등을 물리적 오류 없이 훌륭하게 구현했으나, 상충하는 지시문('인물 없음' 규칙)으로 인해 지정된 인물들은 생략되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지정된 배경 요소들은 존재하지만 좌측 경사로의 심각한 구조적 오류와 뭉개진 사물 묘사로 인해 사실감이 크게 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 대형 연구소를 대각선 아래로 넓게 내려다봄. 워크스테이션의 모니터들은 사용자의 의자를 향해 올바르게 배치됨. 시선의 주체가 되는 인물은 없음.",
        "built_space": "높은 천장과 창문이 있는 거대한 실내 공간. 좌측에 거대한 원자로 형태의 설비, 우측에 원형 유리 구조물이 안정적으로 배치됨. 층간 깊이와 바닥의 동선 표시가 자연스럽고 구조적으로 일관됨.",
        "entities": "열대 식물, 대형 연구 장비, 컴퓨터 워크스테이션, 원형 유리실 내부의 금속 침대와 로봇 팔이 모두 명확히 존재함. 이현우, 서지민, 찰리는 프레임 내 인물 등장 금지 규칙에 따라 모두 생략됨.",
        "hard_violations": [],
        "physics": "모든 사물이 바닥이나 책상 위에 안정적으로 지지되어 있으며, 중력에 위배되거나 허공에 떠 있는 객체 없이 물리 법칙에 부합함."
       },
       {
        "label": "B",
        "direction": "카메라가 연구소를 넓게 내려다봄. 중간 좌측 워크스테이션의 일부 모니터 방향이 불규칙하고 의자의 위치와 어긋나 있음.",
        "built_space": "장비와 식물이 공간에 넓게 배치되어 있으나, 좌측의 경사로 구조가 평평한 바닥과 이어지는 지점 및 난간의 연결 부위가 기하학적으로 어긋나 있음.",
        "entities": "식물, 기계 장치, 유리실, 침대, 로봇 팔이 존재하나 세부 형태가 다소 뭉개져 있음. 지정된 인물들은 생략됨.",
        "hard_violations": [
         "물리적으로 불가능한 건축 구조 (좌측 경사로 바닥면과 난간이 형태 없이 바닥에 융합됨)"
        ],
        "physics": "중간 좌측 책상 주변의 일부 기기가 지지대 없이 융합되거나 떠 있는 것처럼 보이며, 구조물의 결합 부위가 물리적으로 불가능한 형태를 띰."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "더 높은 사선 하향 원경과 여러 깊이층의 설비·컴퓨터·열대 식재가 요청한 웅장한 전경에 가깝고, 금지된 인물이나 임의 표식도 보이지 않는다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "연구홀의 규모와 식재는 충실하지만 임의 로고 배너를 추가했으며, A보다 낮고 정면에 가까운 시점이 지정된 높은 사선 하향 전경에 덜 부합한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 전경 난간 너머에서 연구홀 바닥을 뚜렷하게 내려다본다. 통로와 설비 사이로 시선이 뒤쪽 창가까지 이어진다. 사람의 시선이나 무기류는 없다. 오른쪽 유리실의 로봇 팔들은 중앙 침상 쪽을 향하며, 컴퓨터 화면들은 각 작업대의 의자 또는 작업 공간 쪽으로 놓여 있다.",
        "built_space": "전경에 난간과 식재대, 중경에 넓은 바닥 통로와 분산된 작업대, 후경에 높은 창과 대형 설비가 있다. 큰 원통형 설비는 중앙·왼쪽 후방·오른쪽 후방·오른쪽 가장자리에 최소 네 군으로 보이며, 원형 유리실은 오른쪽 전경에 하나 있다. 컴퓨터 작업대는 왼쪽·중앙·오른쪽과 후방에 여러 군으로 분산되어 있다. 단일 장비가 화면의 40%를 넘지는 않는다. 유리와 바닥의 반사에 명백한 광학적 모순은 없다.",
        "entities": "금속 배관과 대형 에너지 연구 설비, 다수의 컴퓨터, 야자류와 넓은 잎의 열대 식물, 분홍색·주황색 꽃이 보인다. 살아 있는 사람이나 얼굴은 없어 최종 인물 금지 지시를 따른다. 오른쪽 원형 유리실에는 금속 침상과 로봇 팔이 있지만 찰리와 몸에 연결된 모니터링 선은 식별되지 않으며, 이 공간이 별도의 연결 실험실인지도 불분명하다. 읽을 수 있는 임의 문구나 로고는 보이지 않는다.",
        "hard_violations": [],
        "physics": "대형 설비와 유리실은 바닥 위 기단으로 지지되고, 컴퓨터는 책상이나 장비 캐비닛 위에 놓여 있다. 식물은 식재대에서 자라며 침상은 하부 프레임으로 받쳐져 있다. 로봇 팔은 유리실 내부의 기계 구조에 연결되어 있고, 지지 없이 떠 있는 물체나 비현실적인 신체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "카메라는 전경 난간 위에서 중앙 통로를 따라 후방을 바라보며 완만하게 내려다본다. A보다 수평에 가까운 관찰 방향이다. 사람의 시선이나 무기류는 없다. 오른쪽 유리실의 로봇 팔은 침상 쪽으로 굽어 있고, 작업대 화면들은 인접한 의자와 작업 위치를 향한다.",
        "built_space": "왼쪽에 큰 원통형 설비 두 기와 상부 정비 통로가 있고, 오른쪽 전경에는 원형 유리실 하나가 있다. 중앙의 길게 열린 통로 양옆과 후방에 여러 컴퓨터 작업대와 소형 장치가 배치되어 있다. 전경 난간과 식재대, 후경의 높은 창이 깊이를 만든다. 기둥에는 문양이 들어간 흰 배너 세 장이 보인다. 설비 하나가 화면의 40%를 넘지는 않으며, 유리실 반사에서 명백히 불가능한 상은 확인되지 않는다.",
        "entities": "원자력 발전 시설을 연상시키는 금속 용기와 배관, 다수의 컴퓨터, 야자류와 열대 식물, 여러 색의 꽃이 보인다. 살아 있는 인물이나 얼굴은 없다. 유리실 안에는 금속 침상과 로봇 팔이 있으나 찰리의 비활성 신체와 부착된 모니터링 선은 명확히 식별되지 않는다. 기둥 배너의 고리 모양 로고와 작은 글자성 표식은 프롬프트가 정하지 않은 추가 요소다.",
        "hard_violations": [
         "프롬프트에 없는 로고와 작은 글자성 표식이 인쇄된 배너 세 장을 추가하여, 임의 표식 및 문구를 만들지 말라는 지시를 위반했다."
        ],
        "physics": "원통형 설비는 바닥에 설치되어 있고 정비 통로에는 기둥과 구조적 지지가 보인다. 모니터는 작업대에, 식물은 화분과 식재대에 놓여 있다. 유리실의 로봇 팔은 고정 기구에 연결되고 침상은 하부 구조로 지지된다. 배너도 기둥 쪽에 부착되어 있으며, 지지 없이 떠 있는 물체나 불가능한 인체 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "더 높은 사선 하향 원경과 여러 깊이층의 설비·컴퓨터·열대 식재가 요청한 웅장한 전경에 가깝고, 금지된 인물이나 임의 표식도 보이지 않는다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "연구홀의 규모와 식재는 충실하지만 임의 로고 배너를 추가했으며, A보다 낮고 정면에 가까운 시점이 지정된 높은 사선 하향 전경에 덜 부합한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 전경 난간 너머에서 연구홀 바닥을 뚜렷하게 내려다본다. 통로와 설비 사이로 시선이 뒤쪽 창가까지 이어진다. 사람의 시선이나 무기류는 없다. 오른쪽 유리실의 로봇 팔들은 중앙 침상 쪽을 향하며, 컴퓨터 화면들은 각 작업대의 의자 또는 작업 공간 쪽으로 놓여 있다.",
        "built_space": "전경에 난간과 식재대, 중경에 넓은 바닥 통로와 분산된 작업대, 후경에 높은 창과 대형 설비가 있다. 큰 원통형 설비는 중앙·왼쪽 후방·오른쪽 후방·오른쪽 가장자리에 최소 네 군으로 보이며, 원형 유리실은 오른쪽 전경에 하나 있다. 컴퓨터 작업대는 왼쪽·중앙·오른쪽과 후방에 여러 군으로 분산되어 있다. 단일 장비가 화면의 40%를 넘지는 않는다. 유리와 바닥의 반사에 명백한 광학적 모순은 없다.",
        "entities": "금속 배관과 대형 에너지 연구 설비, 다수의 컴퓨터, 야자류와 넓은 잎의 열대 식물, 분홍색·주황색 꽃이 보인다. 살아 있는 사람이나 얼굴은 없어 최종 인물 금지 지시를 따른다. 오른쪽 원형 유리실에는 금속 침상과 로봇 팔이 있지만 찰리와 몸에 연결된 모니터링 선은 식별되지 않으며, 이 공간이 별도의 연결 실험실인지도 불분명하다. 읽을 수 있는 임의 문구나 로고는 보이지 않는다.",
        "hard_violations": [],
        "physics": "대형 설비와 유리실은 바닥 위 기단으로 지지되고, 컴퓨터는 책상이나 장비 캐비닛 위에 놓여 있다. 식물은 식재대에서 자라며 침상은 하부 프레임으로 받쳐져 있다. 로봇 팔은 유리실 내부의 기계 구조에 연결되어 있고, 지지 없이 떠 있는 물체나 비현실적인 신체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "카메라는 전경 난간 위에서 중앙 통로를 따라 후방을 바라보며 완만하게 내려다본다. A보다 수평에 가까운 관찰 방향이다. 사람의 시선이나 무기류는 없다. 오른쪽 유리실의 로봇 팔은 침상 쪽으로 굽어 있고, 작업대 화면들은 인접한 의자와 작업 위치를 향한다.",
        "built_space": "왼쪽에 큰 원통형 설비 두 기와 상부 정비 통로가 있고, 오른쪽 전경에는 원형 유리실 하나가 있다. 중앙의 길게 열린 통로 양옆과 후방에 여러 컴퓨터 작업대와 소형 장치가 배치되어 있다. 전경 난간과 식재대, 후경의 높은 창이 깊이를 만든다. 기둥에는 문양이 들어간 흰 배너 세 장이 보인다. 설비 하나가 화면의 40%를 넘지는 않으며, 유리실 반사에서 명백히 불가능한 상은 확인되지 않는다.",
        "entities": "원자력 발전 시설을 연상시키는 금속 용기와 배관, 다수의 컴퓨터, 야자류와 열대 식물, 여러 색의 꽃이 보인다. 살아 있는 인물이나 얼굴은 없다. 유리실 안에는 금속 침상과 로봇 팔이 있으나 찰리의 비활성 신체와 부착된 모니터링 선은 명확히 식별되지 않는다. 기둥 배너의 고리 모양 로고와 작은 글자성 표식은 프롬프트가 정하지 않은 추가 요소다.",
        "hard_violations": [
         "프롬프트에 없는 로고와 작은 글자성 표식이 인쇄된 배너 세 장을 추가하여, 임의 표식 및 문구를 만들지 말라는 지시를 위반했다."
        ],
        "physics": "원통형 설비는 바닥에 설치되어 있고 정비 통로에는 기둥과 구조적 지지가 보인다. 모니터는 작업대에, 식물은 화분과 식재대에 놓여 있다. 유리실의 로봇 팔은 고정 기구에 연결되고 침상은 하부 구조로 지지된다. 배너도 기둥 쪽에 부착되어 있으며, 지지 없이 떠 있는 물체나 불가능한 인체 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 건축 구조 (좌측 경사로 바닥면과 난간이 형태 없이 바닥에 융합됨)"
    ],
    "A": [
     "[gpt-high] 프롬프트에 없는 로고와 작은 글자성 표식이 인쇄된 배너 세 장을 추가하여, 임의 표식 및 문구를 만들지 말라는 지시를 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1250,
   "B": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "요구된 웅장한 스케일, 거대 설비, 식물, 원형 유리 연구실 등을 물리적 오류 없이 훌륭하게 구현했으나, 상충하는 지시문('인물 없음' 규칙)으로 인해 지정된 인물들은 생략되었습니다.  ★위반: [gpt-high] 프롬프트에 없는 로고와 작은 글자성 표식이 인쇄된 배너 세 장을 추가하여, 임의 표식 및 문구를 만들지 말라는 지시를 위반했다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "지정된 배경 요소들은 존재하지만 좌측 경사로의 심각한 구조적 오류와 뭉개진 사물 묘사로 인해 사실감이 크게 떨어집니다.  ★위반: [gemini-pro] 물리적으로 불가능한 건축 구조 (좌측 경사로 바닥면과 난간이 형태 없이 바닥에 융합됨)"
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e41-06e4-713d-b7fe-3b21361aee3e",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S82sh7::signage": {
  "fp": "d715dafe4de790cc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S82sh7": {
  "input_fingerprint": "708140627dc6d4ca",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지소영이 이현우를 향해 환하게 웃으며 악수를 청하듯 오른손을 내민 구도.\n\nLOCATION (lock): In the open visitor circulation area of the brightly illuminated main research hall, beside the large installations and indoor plants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the lateral track settle behind 이현우's right shoulder at shoulder height, observing 지소영 directly from outside their face-to-face axis. Keep his shoulder and partial rear profile at the left edge, with 지소영's waist-up figure on the right and her extended right hand bridging the lower center without exaggerated foreshortening. Her bright smile and eyes address his face just beyond the crop, while his head remains turned toward her; emphasize the offered hand entering the space between them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Main research equipment (Computers and research devices occupy the main center) — Partial side views remain behind 지소영; used as Soft background depth establishes the greeting within the research center; Tropical plants and flowers (Present throughout the main space); used as Separate the human greeting from the equipment without obscuring the extended hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the warmth of her smile without introducing a distinct light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large research installations, computers, tropical plants, flowers, and interior lighting from the reference. Exclude the shelter bed and furnishings from the separate recovery room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Computers, research machinery, tropical plants and flowers fill the main center. Charlie remains inactive on the stainless-steel bed in the connected glass enclosure, with wires attached and robotic scanning arms around him. 이현우: He is standing in the main center, still wearing white clothes. 지소영: She wears a clean research coat and extends a hand for a handshake.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지소영이 이현우를 향해 환하게 웃으며 악수를 청하듯 오른손을 내민 구도.\n\nLOCATION (lock): In the open visitor circulation area of the brightly illuminated main research hall, beside the large installations and indoor plants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the lateral track settle behind 이현우's right shoulder at shoulder height, observing 지소영 directly from outside their face-to-face axis. Keep his shoulder and partial rear profile at the left edge, with 지소영's waist-up figure on the right and her extended right hand bridging the lower center without exaggerated foreshortening. Her bright smile and eyes address his face just beyond the crop, while his head remains turned toward her; emphasize the offered hand entering the space between them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Main research equipment (Computers and research devices occupy the main center) — Partial side views remain behind 지소영; used as Soft background depth establishes the greeting within the research center; Tropical plants and flowers (Present throughout the main space); used as Separate the human greeting from the equipment without obscuring the extended hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the warmth of her smile without introducing a distinct light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large research installations, computers, tropical plants, flowers, and interior lighting from the reference. Exclude the shelter bed and furnishings from the separate recovery room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Computers, research machinery, tropical plants and flowers fill the main center. Charlie remains inactive on the stainless-steel bed in the connected glass enclosure, with wires attached and robotic scanning arms around him. 이현우: He is standing in the main center, still wearing white clothes. 지소영: She wears a clean research coat and extends a hand for a handshake.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지소영이 이현우를 향해 환하게 웃으며 악수를 청하듯 오른손을 내민 구도.\n\nLOCATION (lock): In the open visitor circulation area of the brightly illuminated main research hall, beside the large installations and indoor plants. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Let the lateral track settle behind 이현우's right shoulder at shoulder height, observing 지소영 directly from outside their face-to-face axis. Keep his shoulder and partial rear profile at the left edge, with 지소영's waist-up figure on the right and her extended right hand bridging the lower center without exaggerated foreshortening. Her bright smile and eyes address his face just beyond the crop, while his head remains turned toward her; emphasize the offered hand entering the space between them.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Main research equipment (Computers and research devices occupy the main center) — Partial side views remain behind 지소영; used as Soft background depth establishes the greeting within the research center; Tropical plants and flowers (Present throughout the main space); used as Separate the human greeting from the equipment without obscuring the extended hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve the warmth of her smile without introducing a distinct light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large research installations, computers, tropical plants, flowers, and interior lighting from the reference. Exclude the shelter bed and furnishings from the separate recovery room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Computers, research machinery, tropical plants and flowers fill the main center. Charlie remains inactive on the stainless-steel bed in the connected glass enclosure, with wires attached and robotic scanning arms around him. 이현우: He is standing in the main center, still wearing white clothes. 지소영: She wears a clean research coat and extends a hand for a handshake.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지소영의 시선과 내민 오른손이 화면 밖 이현우의 얼굴과 몸을 정확히 향하고 있음.",
    "built_space": "화면 우측의 원형 유리 부스 및 배경의 연구 기기, 열대 식물들이 지정된 레퍼런스의 스케일에 맞게 올바르게 배치됨.",
    "entities": "지소영과 이현우의 외모 및 지소영의 흰색 연구 가운은 일치하나, 이현우가 지정된 흰색 옷 대신 레퍼런스의 남색 셔츠를 입고 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 지소영이 오른팔을 들어 올린 악수 자세의 무게 중심과 형태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "지소영의 시선과 내민 오른손의 방향이 이현우를 향하고 있음.",
    "built_space": "연구소 배경의 유리 부스, 기기, 식물 등 공간 구성 요소가 적절한 위치에 묘사됨.",
    "entities": "인물들의 외모는 레퍼런스와 일치하나, 이현우가 남색 셔츠를 착용하고 있으며 지소영의 오른손 손가락 묘사가 실패함.",
    "hard_violations": [
     "[gemini-pro] 지소영의 내민 오른손 손가락들이 심하게 융합되고 뭉개져 물리적으로 불가능한 해부학적 형태를 띠고 있음."
    ],
    "physics": "인물들이 서 있는 기본적인 지지 상태는 유지되나, 내민 오른손이 신체 형태로서 정상적으로 기능하지 못하는 구조임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우의 의상이 흰색으로 반영되지 않은 점은 아쉽지만, 요구된 어깨 너머 구도와 지소영의 내민 손을 비교적 온전하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 이현우의 의상 오류와 더불어 핵심적인 요소인 지소영의 오른손 손가락이 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 시선과 내민 오른손이 화면 밖 이현우의 얼굴과 몸을 정확히 향하고 있음.",
        "built_space": "화면 우측의 원형 유리 부스 및 배경의 연구 기기, 열대 식물들이 지정된 레퍼런스의 스케일에 맞게 올바르게 배치됨.",
        "entities": "지소영과 이현우의 외모 및 지소영의 흰색 연구 가운은 일치하나, 이현우가 지정된 흰색 옷 대신 레퍼런스의 남색 셔츠를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 지소영이 오른팔을 들어 올린 악수 자세의 무게 중심과 형태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "지소영의 시선과 내민 오른손의 방향이 이현우를 향하고 있음.",
        "built_space": "연구소 배경의 유리 부스, 기기, 식물 등 공간 구성 요소가 적절한 위치에 묘사됨.",
        "entities": "인물들의 외모는 레퍼런스와 일치하나, 이현우가 남색 셔츠를 착용하고 있으며 지소영의 오른손 손가락 묘사가 실패함.",
        "hard_violations": [
         "지소영의 내민 오른손 손가락들이 심하게 융합되고 뭉개져 물리적으로 불가능한 해부학적 형태를 띠고 있음."
        ],
        "physics": "인물들이 서 있는 기본적인 지지 상태는 유지되나, 내민 오른손이 신체 형태로서 정상적으로 기능하지 못하는 구조임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이현우의 의상이 흰색으로 반영되지 않은 점은 아쉽지만, 요구된 어깨 너머 구도와 지소영의 내민 손을 비교적 온전하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "구도와 배경은 적절하나, 이현우의 의상 오류와 더불어 핵심적인 요소인 지소영의 오른손 손가락이 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 시선과 내민 오른손이 화면 밖 이현우의 얼굴과 몸을 정확히 향하고 있음.",
        "built_space": "화면 우측의 원형 유리 부스 및 배경의 연구 기기, 열대 식물들이 지정된 레퍼런스의 스케일에 맞게 올바르게 배치됨.",
        "entities": "지소영과 이현우의 외모 및 지소영의 흰색 연구 가운은 일치하나, 이현우가 지정된 흰색 옷 대신 레퍼런스의 남색 셔츠를 입고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 지소영이 오른팔을 들어 올린 악수 자세의 무게 중심과 형태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "지소영의 시선과 내민 오른손의 방향이 이현우를 향하고 있음.",
        "built_space": "연구소 배경의 유리 부스, 기기, 식물 등 공간 구성 요소가 적절한 위치에 묘사됨.",
        "entities": "인물들의 외모는 레퍼런스와 일치하나, 이현우가 남색 셔츠를 착용하고 있으며 지소영의 오른손 손가락 묘사가 실패함.",
        "hard_violations": [
         "지소영의 내민 오른손 손가락들이 심하게 융합되고 뭉개져 물리적으로 불가능한 해부학적 형태를 띠고 있음."
        ],
        "physics": "인물들이 서 있는 기본적인 지지 상태는 유지되나, 내민 오른손이 신체 형태로서 정상적으로 기능하지 못하는 구조임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "어깨 너머 미디엄 구도와 밝은 미소는 맞지만, 오른손이 상대보다 카메라 쪽으로 더 뻗어 보이며 이현우의 남색 옷은 흰옷 유지 지시를 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지소영의 시선과 오른손이 이현우에게 향하는 악수 제안이 더 명확하고 인물·연구실도 충실하지만, 이현우의 흰옷 유지 지시는 역시 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영은 왼쪽 전경의 이현우 얼굴을 올려다보며 웃고, 이현우도 고개를 그녀 쪽으로 돌렸다. 오른손은 하단 중앙으로 뻗었지만 팔의 진행 방향은 이현우의 몸보다 카메라 쪽에 가까워 보인다. 악수 제안으로는 읽히며 무기나 이동 중인 물체는 없다.",
        "built_space": "이현우의 뒷머리와 오른쪽 어깨가 왼쪽 전경을 차지하고 지소영은 오른쪽에 허리 부근까지 보인다. 배경에는 큰 원통형 설비의 부분 모습 두 곳, 중앙의 모니터 장착 이동식 장비 한 대, 왼쪽의 세로형 장비함 한 대, 오른쪽의 곡면 유리 enclosure 한 곳이 보인다. 창, 금속 배관, 식재대와 꽃은 참고 연구실의 재료와 낮 조명을 이어가며 두 사람은 장비 사이 열린 통로에 있다. 불가능한 반사나 명백한 시설 중복은 보이지 않는다.",
        "entities": "인물은 두 명뿐이다. 지소영은 참고와 유사한 중년 동아시아계 여성의 얼굴, 정돈된 검은 단발, 흰 연구 가운과 회색 셔츠를 갖췄다. 가운의 가슴 표식은 참고와 유사하나 글자 형태가 달라졌다. 이현우는 젊은 남성의 짧고 헝클어진 검은 머리와 부분 옆얼굴로 보이며, 얼굴 전체의 일치 여부는 확인할 수 없다. 남색 상의는 명시된 흰옷과 다르다. 연구 장비, 컴퓨터, 열대 식물과 꽃은 모두 보인다.",
        "hard_violations": [],
        "physics": "지소영의 오른손은 소매에서 이어지는 손목과 팔로 지지되며, 팔을 앞으로 내밀어 악수를 청하는 동작은 가능하다. 두 사람의 발은 프레임 밖이지만 상체는 자연스러운 기립 자세이고 공중에 떠 있다는 징후는 없다. 이동식 장비는 바퀴로 바닥에 놓이고 식물은 식재대에 심겨 있다."
       },
       {
        "label": "B",
        "direction": "지소영의 눈과 미소는 왼쪽의 이현우 얼굴을 향하고 이현우의 머리도 그녀를 향한다. 오른팔은 하단 중앙에서 왼쪽 상대의 공간으로 뻗고 손바닥은 악수를 받을 수 있도록 비스듬히 열려 있어, 카메라가 아닌 이현우에게 손을 내미는 관계가 더 명확하다.",
        "built_space": "왼쪽 전경에는 이현우의 뒷머리·부분 옆얼굴·오른쪽 어깨가 있고, 오른쪽에는 지소영의 허리 위 모습이 있다. 이현우의 등은 왼쪽 가장자리만 남기라는 지시보다 넓게 들어온다. 배경에는 대형 원통 설비 두 곳의 일부, 왼쪽 작업대의 모니터 두 대, 중앙의 세로형 장비함 한 대, 오른쪽 곡면 유리 enclosure 한 곳과 장비 랙이 보인다. 금속 배관, 큰 창, 흰 식재대와 꽃이 참고 장소를 이어가고 내민 손을 가리지 않는다. 불가능한 반사나 명백한 시설 중복은 보이지 않는다.",
        "entities": "두 인물만 보인다. 지소영의 중년 동아시아계 여성 얼굴, 검은 단발과 체격은 참고에 가깝고 흰 연구 가운, 회색 셔츠, 허리 부분의 어두운 정장 바지가 보인다. 가슴 표식도 참고의 배치를 대체로 따른다. 이현우는 젊은 남성의 머리와 목, 일부 얼굴만 보여 정체성 판단에는 한계가 있지만 머리 모양은 참고와 부합한다. 다만 남색 상의는 흰옷 유지 조건과 명백히 다르다. 컴퓨터와 연구 장치, 열대 식물, 꽃이 존재한다.",
        "hard_violations": [],
        "physics": "내민 오른손은 손목과 팔에 정상적으로 연결되고 팔꿈치를 가볍게 굽힌 악수 제안 자세가 자연스럽다. 두 사람은 기립한 상체로 보이며 발은 촬영 범위 밖이다. 배경 장비와 식재대는 바닥에 놓여 있고 로봇 팔은 유리 공간 안의 장치에 연결되어 있다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "어깨 너머 미디엄 구도와 밝은 미소는 맞지만, 오른손이 상대보다 카메라 쪽으로 더 뻗어 보이며 이현우의 남색 옷은 흰옷 유지 지시를 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지소영의 시선과 오른손이 이현우에게 향하는 악수 제안이 더 명확하고 인물·연구실도 충실하지만, 이현우의 흰옷 유지 지시는 역시 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "지소영은 왼쪽 전경의 이현우 얼굴을 올려다보며 웃고, 이현우도 고개를 그녀 쪽으로 돌렸다. 오른손은 하단 중앙으로 뻗었지만 팔의 진행 방향은 이현우의 몸보다 카메라 쪽에 가까워 보인다. 악수 제안으로는 읽히며 무기나 이동 중인 물체는 없다.",
        "built_space": "이현우의 뒷머리와 오른쪽 어깨가 왼쪽 전경을 차지하고 지소영은 오른쪽에 허리 부근까지 보인다. 배경에는 큰 원통형 설비의 부분 모습 두 곳, 중앙의 모니터 장착 이동식 장비 한 대, 왼쪽의 세로형 장비함 한 대, 오른쪽의 곡면 유리 enclosure 한 곳이 보인다. 창, 금속 배관, 식재대와 꽃은 참고 연구실의 재료와 낮 조명을 이어가며 두 사람은 장비 사이 열린 통로에 있다. 불가능한 반사나 명백한 시설 중복은 보이지 않는다.",
        "entities": "인물은 두 명뿐이다. 지소영은 참고와 유사한 중년 동아시아계 여성의 얼굴, 정돈된 검은 단발, 흰 연구 가운과 회색 셔츠를 갖췄다. 가운의 가슴 표식은 참고와 유사하나 글자 형태가 달라졌다. 이현우는 젊은 남성의 짧고 헝클어진 검은 머리와 부분 옆얼굴로 보이며, 얼굴 전체의 일치 여부는 확인할 수 없다. 남색 상의는 명시된 흰옷과 다르다. 연구 장비, 컴퓨터, 열대 식물과 꽃은 모두 보인다.",
        "hard_violations": [],
        "physics": "지소영의 오른손은 소매에서 이어지는 손목과 팔로 지지되며, 팔을 앞으로 내밀어 악수를 청하는 동작은 가능하다. 두 사람의 발은 프레임 밖이지만 상체는 자연스러운 기립 자세이고 공중에 떠 있다는 징후는 없다. 이동식 장비는 바퀴로 바닥에 놓이고 식물은 식재대에 심겨 있다."
       },
       {
        "label": "A",
        "direction": "지소영의 눈과 미소는 왼쪽의 이현우 얼굴을 향하고 이현우의 머리도 그녀를 향한다. 오른팔은 하단 중앙에서 왼쪽 상대의 공간으로 뻗고 손바닥은 악수를 받을 수 있도록 비스듬히 열려 있어, 카메라가 아닌 이현우에게 손을 내미는 관계가 더 명확하다.",
        "built_space": "왼쪽 전경에는 이현우의 뒷머리·부분 옆얼굴·오른쪽 어깨가 있고, 오른쪽에는 지소영의 허리 위 모습이 있다. 이현우의 등은 왼쪽 가장자리만 남기라는 지시보다 넓게 들어온다. 배경에는 대형 원통 설비 두 곳의 일부, 왼쪽 작업대의 모니터 두 대, 중앙의 세로형 장비함 한 대, 오른쪽 곡면 유리 enclosure 한 곳과 장비 랙이 보인다. 금속 배관, 큰 창, 흰 식재대와 꽃이 참고 장소를 이어가고 내민 손을 가리지 않는다. 불가능한 반사나 명백한 시설 중복은 보이지 않는다.",
        "entities": "두 인물만 보인다. 지소영의 중년 동아시아계 여성 얼굴, 검은 단발과 체격은 참고에 가깝고 흰 연구 가운, 회색 셔츠, 허리 부분의 어두운 정장 바지가 보인다. 가슴 표식도 참고의 배치를 대체로 따른다. 이현우는 젊은 남성의 머리와 목, 일부 얼굴만 보여 정체성 판단에는 한계가 있지만 머리 모양은 참고와 부합한다. 다만 남색 상의는 흰옷 유지 조건과 명백히 다르다. 컴퓨터와 연구 장치, 열대 식물, 꽃이 존재한다.",
        "hard_violations": [],
        "physics": "내민 오른손은 손목과 팔에 정상적으로 연결되고 팔꿈치를 가볍게 굽힌 악수 제안 자세가 자연스럽다. 두 사람은 기립한 상체로 보이며 발은 촬영 범위 밖이다. 배경 장비와 식재대는 바닥에 놓여 있고 로봇 팔은 유리 공간 안의 장치에 연결되어 있다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.446
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.196
   },
   "violations": {
    "B": [
     "[gemini-pro] 지소영의 내민 오른손 손가락들이 심하게 융합되고 뭉개져 물리적으로 불가능한 해부학적 형태를 띠고 있음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1196
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "이현우의 의상이 흰색으로 반영되지 않은 점은 아쉽지만, 요구된 어깨 너머 구도와 지소영의 내민 손을 비교적 온전하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1196,
    "verdict_ko": "구도와 배경은 적절하나, 이현우의 의상 오류와 더불어 핵심적인 요소인 지소영의 오른손 손가락이 심하게 뭉개지는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 지소영의 내민 오른손 손가락들이 심하게 융합되고 뭉개져 물리적으로 불가능한 해부학적 형태를 띠고 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh1_sel.png",
    "asset_id": "c9508968-9d42-4a8b-bca4-331bb522c5e1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:783266>",
    "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e47-6aa1-71db-bd54-c8576cc428f1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh1"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S82sh13::signage": {
  "fp": "cebef3bf86ad1a44",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::ccce42833ea81ab4": {
  "subjects": [],
  "subject_text": "제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀\n열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L184",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::glass_experiment_room": {
  "input_fingerprint": "c25aa35ad4f61325",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "glass_experiment_room",
    "tags": [
     "S82sh13",
     "S84sh1",
     "S85sh12",
     "S88sh11",
     "S88sh37",
     "S88sh45"
    ]
   },
   "context_sig": "877c655344f4772c"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 메인 실험실·유리벽 실험 구역: 원형 유리벽 안에 수술대가 위치하고 컴퓨터 패널이 빙 둘러싼 하이테크 실험 공간이다. (특징: 투명한 대형 원형 유리벽과 내부의 차가운 스테인리스 침대; 침대 위에 누워 전선이 주렁주렁 연결된 찰리; 찰리 주변을 바쁘게 움직이는 인공지능 로봇 스캔 팔; 컴퓨터 스크린에 치솟는 그래프와 'ERROR' 붉은 팝업 글자; 찰리의 링에서 뿜어지는 짙은 흑색 연기; 연구원이 서랍에서 꺼내는 스턴 건(전기 충격기)) / 제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /메인 연구소안의 연결된 실험실\n- 중앙에는 원형으로 된 유리벽이 설치되어 있고, 차가운 스테인리스 재질의 침대 위에 누워있는 찰리.\n- 실험실 유리 벽 안에 누워있는 찰리,\n- 어느새 유리 벽 너머, 여러 전선들을 몸에 붙인 채 앉아있는 찰리.\n- / 메인센터 실험실\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 메인 실험실·유리벽 실험 구역: 원형 유리벽 안에 수술대가 위치하고 컴퓨터 패널이 빙 둘러싼 하이테크 실험 공간이다. (특징: 투명한 대형 원형 유리벽과 내부의 차가운 스테인리스 침대; 침대 위에 누워 전선이 주렁주렁 연결된 찰리; 찰리 주변을 바쁘게 움직이는 인공지능 로봇 스캔 팔; 컴퓨터 스크린에 치솟는 그래프와 'ERROR' 붉은 팝업 글자; 찰리의 링에서 뿜어지는 짙은 흑색 연기; 연구원이 서랍에서 꺼내는 스턴 건(전기 충격기)) / 제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /메인 연구소안의 연결된 실험실\n- 중앙에는 원형으로 된 유리벽이 설치되어 있고, 차가운 스테인리스 재질의 침대 위에 누워있는 찰리.\n- 실험실 유리 벽 안에 누워있는 찰리,\n- 어느새 유리 벽 너머, 여러 전선들을 몸에 붙인 채 앉아있는 찰리.\n- / 메인센터 실험실\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_experiment_room_ea17f9.png",
  "asset_id": "9dc88b2d-de9a-4dc6-88d4-99a8c0dfdee7",
  "input_asset_ids": [
   "544e3944-5241-4703-bc75-804991ca3e7b"
  ],
  "origin_tag": "S82sh13",
  "place_text": "At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.",
  "origin_inputs": {
   "place_text": "At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.",
   "time_of_day_en": "day",
   "conti_asset_id": "544e3944-5241-4703-bc75-804991ca3e7b"
  }
 },
 "S82sh13::bgfirst_bg": {
  "input_fingerprint": "855bd5b80a1b24bb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수술대 앞 유리벽에 두 손을 짚고 안타까운 눈빛으로 찰리를 바라보는 이현우의 뒷모습.\n\nLOCATION (lock): At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the following movement outside the enclosure, at shoulder height behind and slightly left of 이현우, looking directly through the glass rather than at a reflection. His back occupies the left half as both palms meet the wall and his head inclines toward 찰리, whose reclining body remains visible beyond his right shoulder in the middle-right depth. Hold their separation in the same composition, using the arrival of his hands against the glass as the principal change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Glass barrier beneath 이현우's palms in the middle-center of the frame, midground; Treatment bed beyond 이현우's right shoulder in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Separates 이현우 from 찰리 during the procedure) — 찰리 and the bed are visible through the section beneath 이현우's hands; used as Makes the physical barrier between the friends legible; Stainless-steel bed (Supports 찰리 during scanning) — Seen obliquely beyond 이현우's right shoulder; used as Provides a stable horizontal beneath 찰리 and a scale reference; Artificial-intelligence arms (Scanning different parts of 찰리) — Partial articulated profiles flank the bed without covering his head; used as Frames the ongoing treatment inside the enclosure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light retains visibility through the glass and readable contours on 찰리 without theatrical reflections.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수술대 앞 유리벽에 두 손을 짚고 안타까운 눈빛으로 찰리를 바라보는 이현우의 뒷모습.\n\nLOCATION (lock): At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the following movement outside the enclosure, at shoulder height behind and slightly left of 이현우, looking directly through the glass rather than at a reflection. His back occupies the left half as both palms meet the wall and his head inclines toward 찰리, whose reclining body remains visible beyond his right shoulder in the middle-right depth. Hold their separation in the same composition, using the arrival of his hands against the glass as the principal change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Glass barrier beneath 이현우's palms in the middle-center of the frame, midground; Treatment bed beyond 이현우's right shoulder in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Separates 이현우 from 찰리 during the procedure) — 찰리 and the bed are visible through the section beneath 이현우's hands; used as Makes the physical barrier between the friends legible; Stainless-steel bed (Supports 찰리 during scanning) — Seen obliquely beyond 이현우's right shoulder; used as Provides a stable horizontal beneath 찰리 and a scale reference; Artificial-intelligence arms (Scanning different parts of 찰리) — Partial articulated profiles flank the bed without covering his head; used as Frames the ongoing treatment inside the enclosure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light retains visibility through the glass and readable contours on 찰리 without theatrical reflections.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh13__bgfirst_bg.png",
  "asset_id": "2518c02a-4465-4387-8d84-80a1a45d1bd2",
  "input_asset_ids": [
   "544e3944-5241-4703-bc75-804991ca3e7b",
   "9dc88b2d-de9a-4dc6-88d4-99a8c0dfdee7"
  ]
 },
 "S82sh13": {
  "input_fingerprint": "21a0a5312934df4d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수술대 앞 유리벽에 두 손을 짚고 안타까운 눈빛으로 찰리를 바라보는 이현우의 뒷모습.\n\nLOCATION (lock): At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the following movement outside the enclosure, at shoulder height behind and slightly left of 이현우, looking directly through the glass rather than at a reflection. His back occupies the left half as both palms meet the wall and his head inclines toward 찰리, whose reclining body remains visible beyond his right shoulder in the middle-right depth. Hold their separation in the same composition, using the arrival of his hands against the glass as the principal change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Glass barrier beneath 이현우's palms in the middle-center of the frame, midground; Treatment bed beyond 이현우's right shoulder in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Separates 이현우 from 찰리 during the procedure) — 찰리 and the bed are visible through the section beneath 이현우's hands; used as Makes the physical barrier between the friends legible; Stainless-steel bed (Supports 찰리 during scanning) — Seen obliquely beyond 이현우's right shoulder; used as Provides a stable horizontal beneath 찰리 and a scale reference; Artificial-intelligence arms (Scanning different parts of 찰리) — Partial articulated profiles flank the bed without covering his head; used as Frames the ongoing treatment inside the enclosure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light retains visibility through the glass and readable contours on 찰리 without theatrical reflections.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with attached wires and robotic arms scanning his water-damaged hardware. The surrounding laboratory computers are staffed. 이현우: He is at the glass enclosure in white clothes, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수술대 앞 유리벽에 두 손을 짚고 안타까운 눈빛으로 찰리를 바라보는 이현우의 뒷모습.\n\nLOCATION (lock): At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the following movement outside the enclosure, at shoulder height behind and slightly left of 이현우, looking directly through the glass rather than at a reflection. His back occupies the left half as both palms meet the wall and his head inclines toward 찰리, whose reclining body remains visible beyond his right shoulder in the middle-right depth. Hold their separation in the same composition, using the arrival of his hands against the glass as the principal change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Glass barrier beneath 이현우's palms in the middle-center of the frame, midground; Treatment bed beyond 이현우's right shoulder in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Separates 이현우 from 찰리 during the procedure) — 찰리 and the bed are visible through the section beneath 이현우's hands; used as Makes the physical barrier between the friends legible; Stainless-steel bed (Supports 찰리 during scanning) — Seen obliquely beyond 이현우's right shoulder; used as Provides a stable horizontal beneath 찰리 and a scale reference; Artificial-intelligence arms (Scanning different parts of 찰리) — Partial articulated profiles flank the bed without covering his head; used as Frames the ongoing treatment inside the enclosure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light retains visibility through the glass and readable contours on 찰리 without theatrical reflections.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with attached wires and robotic arms scanning his water-damaged hardware. The surrounding laboratory computers are staffed. 이현우: He is at the glass enclosure in white clothes, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수술대 앞 유리벽에 두 손을 짚고 안타까운 눈빛으로 찰리를 바라보는 이현우의 뒷모습.\n\nLOCATION (lock): At the observation side of the circular glass enclosure in the laboratory connected to the main research hall. Laboratory lighting reveals the stainless-steel examination bed and scanning arms beyond. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the following movement outside the enclosure, at shoulder height behind and slightly left of 이현우, looking directly through the glass rather than at a reflection. His back occupies the left half as both palms meet the wall and his head inclines toward 찰리, whose reclining body remains visible beyond his right shoulder in the middle-right depth. Hold their separation in the same composition, using the arrival of his hands against the glass as the principal change.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Glass barrier beneath 이현우's palms in the middle-center of the frame, midground; Treatment bed beyond 이현우's right shoulder in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Circular glass enclosure (Separates 이현우 from 찰리 during the procedure) — 찰리 and the bed are visible through the section beneath 이현우's hands; used as Makes the physical barrier between the friends legible; Stainless-steel bed (Supports 찰리 during scanning) — Seen obliquely beyond 이현우's right shoulder; used as Provides a stable horizontal beneath 찰리 and a scale reference; Artificial-intelligence arms (Scanning different parts of 찰리) — Partial articulated profiles flank the bed without covering his head; used as Frames the ongoing treatment inside the enclosure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light retains visibility through the glass and readable contours on 찰리 without theatrical reflections.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains inactive on the stainless-steel bed inside the circular glass enclosure, with attached wires and robotic arms scanning his water-damaged hardware. The surrounding laboratory computers are staffed. 이현우: He is at the glass enclosure in white clothes, visibly worried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh13__bgfirst_bg.png",
     "asset_id": "2518c02a-4465-4387-8d84-80a1a45d1bd2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S82sh13.png",
     "asset_id": "544e3944-5241-4703-bc75-804991ca3e7b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:766860>",
     "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1299876>",
     "asset_id": "99dc4ad5-1251-4608-89ee-6005d5fc811c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_experiment_room_ea17f9.png",
     "asset_id": "9dc88b2d-de9a-4dc6-88d4-99a8c0dfdee7",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:766860>",
     "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1299876>",
     "asset_id": "99dc4ad5-1251-4608-89ee-6005d5fc811c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 앞쪽의 원형 유리벽 너머 수술대 위에 누워있는 찰리를 향해 시선을 던지고 있다. AI 팔들의 끝부분은 찰리의 몸통과 다리 쪽을 향해 정확히 조준되어 있다.",
    "built_space": "레퍼런스와 동일한 연구실 내부. 이현우 앞을 가로막는 곡면 유리벽과 하단의 조명 띠, 그 안쪽 중앙에 위치한 스테인리스 수술대, 천장에서 내려온 메인 기둥과 로봇 팔들이 정확한 위치에 배치되어 있다. 외부 바닥은 레퍼런스처럼 매끄럽고 단순하게 표현되었다.",
    "entities": "이현우는 검은 짧은 머리에 흰색 상의를 입은 뒷모습이다. 찰리는 레퍼런스 시트와 정확히 일치하는 각진 샌드 베이지 장갑판과 점/선 형태의 이목구비를 가진 흰색 마스크를 장착한 두꺼운 기계 몸체로 묘사되었다.",
    "hard_violations": [
     "[gpt-high] 등장 허용 대상이 아닌 연구원 두 명을 양쪽 작업대에 추가하여, 숏 텍스트에 없는 사람을 보이지 말라는 지침을 위반했다."
    ],
    "physics": "이현우의 양손은 유리벽에 닿아 체중의 일부를 지탱하고 있으며 반사상도 자연스럽다. 찰리는 수술대 위에 중력에 맞게 완전히 누워 있고, 로봇 팔들은 천장 구조물에 단단히 고정되어 있다. 공중에 떠 있거나 물리적으로 불가능한 포즈는 없다."
   },
   {
    "label": "B",
    "direction": "이현우의 몸과 고개가 유리벽 안쪽 수술대의 찰리를 향해 기울어져 있다. 안쪽의 AI 팔들은 찰리의 상체와 머리 쪽을 향해 있다.",
    "built_space": "원형 유리벽과 안쪽의 수술대는 잘 묘사되었으나, 유리벽 바깥쪽(이현우가 서 있는 위치의 왼쪽 하단) 바닥에 레퍼런스 공간에는 존재하지 않는 노란색 곡선 발광 라인이 임의로 추가되어 공간의 구조적 일관성이 깨졌다.",
    "entities": "이현우는 짧은 검은 머리에 흰옷을 입고 있다. 찰리는 기계 몸체이나 마스크에 레퍼런스에 없는 둥근 두 눈이 생겼고, 뚱뚱한 체형보다는 인간의 비율에 가깝게 얇게 표현되었다.",
    "hard_violations": [
     "[gpt-high] 숏에 허용된 이현우와 찰리 외에 여러 연구원을 추가했다. 컴퓨터가 운영 중이라는 상태 설명과 별개로, 인물 가시성 지침은 이들의 등장을 명시적으로 금지한다."
    ],
    "physics": "이현우의 양손이 유리벽을 짚고 있으나 왼손 손가락의 형태가 비정상적으로 두껍게 뭉개져 있다. 찰리는 수술대 위에 안정적으로 눕혀 있으며 다른 물체들의 지지 상태는 정상적이다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "찰리의 고릴라형 기계 몸체와 흰색 마스크, 원형 유리벽의 형태와 바닥 디테일 등 레퍼런스의 요소를 완벽하게 구현하며 제시된 샷의 프레이밍을 정확히 따랐습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스와 달리 찰리의 마스크에 둥근 눈 모양이 추가되었고, 유리벽 외부 바닥에 존재하지 않는 발광 라인이 임의로 생성되었으며 이현우의 왼손 묘사가 다소 어색합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 앞쪽의 원형 유리벽 너머 수술대 위에 누워있는 찰리를 향해 시선을 던지고 있다. AI 팔들의 끝부분은 찰리의 몸통과 다리 쪽을 향해 정확히 조준되어 있다.",
        "built_space": "레퍼런스와 동일한 연구실 내부. 이현우 앞을 가로막는 곡면 유리벽과 하단의 조명 띠, 그 안쪽 중앙에 위치한 스테인리스 수술대, 천장에서 내려온 메인 기둥과 로봇 팔들이 정확한 위치에 배치되어 있다. 외부 바닥은 레퍼런스처럼 매끄럽고 단순하게 표현되었다.",
        "entities": "이현우는 검은 짧은 머리에 흰색 상의를 입은 뒷모습이다. 찰리는 레퍼런스 시트와 정확히 일치하는 각진 샌드 베이지 장갑판과 점/선 형태의 이목구비를 가진 흰색 마스크를 장착한 두꺼운 기계 몸체로 묘사되었다.",
        "hard_violations": [],
        "physics": "이현우의 양손은 유리벽에 닿아 체중의 일부를 지탱하고 있으며 반사상도 자연스럽다. 찰리는 수술대 위에 중력에 맞게 완전히 누워 있고, 로봇 팔들은 천장 구조물에 단단히 고정되어 있다. 공중에 떠 있거나 물리적으로 불가능한 포즈는 없다."
       },
       {
        "label": "B",
        "direction": "이현우의 몸과 고개가 유리벽 안쪽 수술대의 찰리를 향해 기울어져 있다. 안쪽의 AI 팔들은 찰리의 상체와 머리 쪽을 향해 있다.",
        "built_space": "원형 유리벽과 안쪽의 수술대는 잘 묘사되었으나, 유리벽 바깥쪽(이현우가 서 있는 위치의 왼쪽 하단) 바닥에 레퍼런스 공간에는 존재하지 않는 노란색 곡선 발광 라인이 임의로 추가되어 공간의 구조적 일관성이 깨졌다.",
        "entities": "이현우는 짧은 검은 머리에 흰옷을 입고 있다. 찰리는 기계 몸체이나 마스크에 레퍼런스에 없는 둥근 두 눈이 생겼고, 뚱뚱한 체형보다는 인간의 비율에 가깝게 얇게 표현되었다.",
        "hard_violations": [],
        "physics": "이현우의 양손이 유리벽을 짚고 있으나 왼손 손가락의 형태가 비정상적으로 두껍게 뭉개져 있다. 찰리는 수술대 위에 안정적으로 눕혀 있으며 다른 물체들의 지지 상태는 정상적이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "찰리의 고릴라형 기계 몸체와 흰색 마스크, 원형 유리벽의 형태와 바닥 디테일 등 레퍼런스의 요소를 완벽하게 구현하며 제시된 샷의 프레이밍을 정확히 따랐습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "레퍼런스와 달리 찰리의 마스크에 둥근 눈 모양이 추가되었고, 유리벽 외부 바닥에 존재하지 않는 발광 라인이 임의로 생성되었으며 이현우의 왼손 묘사가 다소 어색합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 앞쪽의 원형 유리벽 너머 수술대 위에 누워있는 찰리를 향해 시선을 던지고 있다. AI 팔들의 끝부분은 찰리의 몸통과 다리 쪽을 향해 정확히 조준되어 있다.",
        "built_space": "레퍼런스와 동일한 연구실 내부. 이현우 앞을 가로막는 곡면 유리벽과 하단의 조명 띠, 그 안쪽 중앙에 위치한 스테인리스 수술대, 천장에서 내려온 메인 기둥과 로봇 팔들이 정확한 위치에 배치되어 있다. 외부 바닥은 레퍼런스처럼 매끄럽고 단순하게 표현되었다.",
        "entities": "이현우는 검은 짧은 머리에 흰색 상의를 입은 뒷모습이다. 찰리는 레퍼런스 시트와 정확히 일치하는 각진 샌드 베이지 장갑판과 점/선 형태의 이목구비를 가진 흰색 마스크를 장착한 두꺼운 기계 몸체로 묘사되었다.",
        "hard_violations": [],
        "physics": "이현우의 양손은 유리벽에 닿아 체중의 일부를 지탱하고 있으며 반사상도 자연스럽다. 찰리는 수술대 위에 중력에 맞게 완전히 누워 있고, 로봇 팔들은 천장 구조물에 단단히 고정되어 있다. 공중에 떠 있거나 물리적으로 불가능한 포즈는 없다."
       },
       {
        "label": "B",
        "direction": "이현우의 몸과 고개가 유리벽 안쪽 수술대의 찰리를 향해 기울어져 있다. 안쪽의 AI 팔들은 찰리의 상체와 머리 쪽을 향해 있다.",
        "built_space": "원형 유리벽과 안쪽의 수술대는 잘 묘사되었으나, 유리벽 바깥쪽(이현우가 서 있는 위치의 왼쪽 하단) 바닥에 레퍼런스 공간에는 존재하지 않는 노란색 곡선 발광 라인이 임의로 추가되어 공간의 구조적 일관성이 깨졌다.",
        "entities": "이현우는 짧은 검은 머리에 흰옷을 입고 있다. 찰리는 기계 몸체이나 마스크에 레퍼런스에 없는 둥근 두 눈이 생겼고, 뚱뚱한 체형보다는 인간의 비율에 가깝게 얇게 표현되었다.",
        "hard_violations": [],
        "physics": "이현우의 양손이 유리벽을 짚고 있으나 왼손 손가락의 형태가 비정상적으로 두껍게 뭉개져 있다. 찰리는 수술대 위에 안정적으로 눕혀 있으며 다른 물체들의 지지 상태는 정상적이다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "금지된 배경 인물들로 실격 사유가 있으나, 왼쪽 절반을 채우는 이현우의 등과 오른쪽 어깨 너머 찰리를 담은 미디엄 구도는 B보다 정확하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금지된 연구원 두 명이 등장하며, 장소 구조는 충실하지만 화면을 넓혀 지정된 미디엄 숏보다 연구실 전경을 강조한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 고개를 오른쪽 아래의 침대에 누운 찰리 쪽으로 숙인다. 눈은 뒷모습에 가려 직접 확인되지 않는다. 두 손바닥은 앞쪽 유리를 향한다. 두 스캔 암의 말단은 각각 찰리의 몸통과 상체 쪽을 향하며 얼굴을 가리지 않는다.",
        "built_space": "원형 유리 격실 하나, 중앙 받침대가 있는 금속 침대 하나, 관절형 스캔 암 두 개가 보인다. 곡선 천장 조명과 바닥의 원형 경계, 주변 컴퓨터 설비가 장소 참조와 부합한다. 이현우는 격실 밖 왼쪽 전경에 있고 침대는 오른쪽 깊이에 있다. 어깨 높이의 후방 시점과 큰 등 면적이 지정 구도에 가깝다. 왼손은 중앙이 아니라 화면 왼쪽 가장자리에 가까우며, 찰리를 가리는 강한 반사상은 없다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리, 젊고 마른 체형, 깨끗한 흰 반소매 상의로 나타난다. 얼굴이 가려 정확한 얼굴 일치와 나이는 확인하기 어렵고, 참조의 모자와 마스크는 보이지 않는다. 찰리는 샌드 베이지 장갑판과 흰 기계 얼굴을 가진 비인간 기계이며 침대에 누워 있다. 연결 케이블과 검사 장비도 보인다. 다만 참조의 육중한 고릴라형 비율보다 가늘고 길게 읽힌다. 허용된 두 인물 외에 왼쪽 두 명과 뒤쪽 세 명가량의 연구원이 추가로 보인다.",
        "hard_violations": [
         "숏에 허용된 이현우와 찰리 외에 여러 연구원을 추가했다. 컴퓨터가 운영 중이라는 상태 설명과 별개로, 인물 가시성 지침은 이들의 등장을 명시적으로 금지한다."
        ],
        "physics": "이현우의 팔은 어깨와 굽힌 팔꿈치로 이어지고, 펼친 두 손은 유리면에 닿는 자세다. 발은 프레임 밖이지만 상체가 떠 있다는 징후는 없다. 찰리의 머리와 몸통, 다리는 침대에 놓여 있고, 굽힌 팔은 몸통 위에 기대어 지지되는 것으로 읽힌다. 침대는 금속 기둥과 바닥 받침으로 지지되며, 스캔 암은 고정 구조에 연결되어 있다. 케이블은 침대 가장자리에서 중력 방향으로 늘어진다."
       },
       {
        "label": "B",
        "direction": "이현우는 오른쪽 깊이의 찰리 쪽으로 머리를 기울이고 두 손바닥을 유리 쪽으로 펼친다. 눈 자체는 보이지 않는다. 찰리의 얼굴은 위쪽을 향하고, 두 스캔 암은 몸통과 머리 아래 상체 부위를 향한다. 찰리의 머리를 가리는 장비는 없다.",
        "built_space": "원형 유리 격실 하나와 침대 하나, 스캔 암 두 개가 보이며 천장 원형 조명, 수직 프레임, 바닥 환형 경계가 장소 참조에 매우 가깝다. 이현우는 유리 밖 왼쪽, 찰리는 유리 안 중간 오른쪽에 있다. 다만 인물을 작게 잡고 천장과 바닥, 양쪽 작업대를 넓게 포함해 지정된 미디엄 구도보다 넓다. 두 손은 화면 왼쪽과 중앙으로 벌어져 있다. 유리를 통한 직접 시야이며 불가능한 반사상은 보이지 않는다.",
        "entities": "이현우의 검은 헝클어진 머리와 흰 반소매 상의, 젊은 체형은 설명에 부합한다. 얼굴은 거의 가려져 정체성과 표정을 세밀하게 대조할 수 없다. 찰리는 흰 마스크형 기계 얼굴과 베이지 장갑판을 갖추고 연결선이 있는 금속 침대에 누워 있다. 다만 작은 화면 크기에서는 참조의 긴 팔과 짧은 다리 비율이 뚜렷하지 않다. 왼쪽 작업대에 앉은 연구원 한 명과 오른쪽 작업대에 선 연구원 한 명이 추가로 보인다.",
        "hard_violations": [
         "등장 허용 대상이 아닌 연구원 두 명을 양쪽 작업대에 추가하여, 숏 텍스트에 없는 사람을 보이지 말라는 지침을 위반했다."
        ],
        "physics": "이현우의 팔과 손은 해부학적으로 연결되며 유리에 손을 대는 동작으로 읽힌다. 찰리의 머리, 몸통, 팔과 다리는 침대 위에 놓여 있고 스스로 들고 있는 신체 부위는 보이지 않는다. 침대는 중앙 기둥과 넓은 받침으로, 스캔 암은 고정 지지대와 관절로 지지된다. 연결 케이블은 아래로 처진다. 지지 없이 떠 있는 물체나 신체는 확인되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금지된 배경 인물들로 실격 사유가 있으나, 왼쪽 절반을 채우는 이현우의 등과 오른쪽 어깨 너머 찰리를 담은 미디엄 구도는 B보다 정확하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 연구원 두 명이 등장하며, 장소 구조는 충실하지만 화면을 넓혀 지정된 미디엄 숏보다 연구실 전경을 강조한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 고개를 오른쪽 아래의 침대에 누운 찰리 쪽으로 숙인다. 눈은 뒷모습에 가려 직접 확인되지 않는다. 두 손바닥은 앞쪽 유리를 향한다. 두 스캔 암의 말단은 각각 찰리의 몸통과 상체 쪽을 향하며 얼굴을 가리지 않는다.",
        "built_space": "원형 유리 격실 하나, 중앙 받침대가 있는 금속 침대 하나, 관절형 스캔 암 두 개가 보인다. 곡선 천장 조명과 바닥의 원형 경계, 주변 컴퓨터 설비가 장소 참조와 부합한다. 이현우는 격실 밖 왼쪽 전경에 있고 침대는 오른쪽 깊이에 있다. 어깨 높이의 후방 시점과 큰 등 면적이 지정 구도에 가깝다. 왼손은 중앙이 아니라 화면 왼쪽 가장자리에 가까우며, 찰리를 가리는 강한 반사상은 없다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리, 젊고 마른 체형, 깨끗한 흰 반소매 상의로 나타난다. 얼굴이 가려 정확한 얼굴 일치와 나이는 확인하기 어렵고, 참조의 모자와 마스크는 보이지 않는다. 찰리는 샌드 베이지 장갑판과 흰 기계 얼굴을 가진 비인간 기계이며 침대에 누워 있다. 연결 케이블과 검사 장비도 보인다. 다만 참조의 육중한 고릴라형 비율보다 가늘고 길게 읽힌다. 허용된 두 인물 외에 왼쪽 두 명과 뒤쪽 세 명가량의 연구원이 추가로 보인다.",
        "hard_violations": [
         "숏에 허용된 이현우와 찰리 외에 여러 연구원을 추가했다. 컴퓨터가 운영 중이라는 상태 설명과 별개로, 인물 가시성 지침은 이들의 등장을 명시적으로 금지한다."
        ],
        "physics": "이현우의 팔은 어깨와 굽힌 팔꿈치로 이어지고, 펼친 두 손은 유리면에 닿는 자세다. 발은 프레임 밖이지만 상체가 떠 있다는 징후는 없다. 찰리의 머리와 몸통, 다리는 침대에 놓여 있고, 굽힌 팔은 몸통 위에 기대어 지지되는 것으로 읽힌다. 침대는 금속 기둥과 바닥 받침으로 지지되며, 스캔 암은 고정 구조에 연결되어 있다. 케이블은 침대 가장자리에서 중력 방향으로 늘어진다."
       },
       {
        "label": "A",
        "direction": "이현우는 오른쪽 깊이의 찰리 쪽으로 머리를 기울이고 두 손바닥을 유리 쪽으로 펼친다. 눈 자체는 보이지 않는다. 찰리의 얼굴은 위쪽을 향하고, 두 스캔 암은 몸통과 머리 아래 상체 부위를 향한다. 찰리의 머리를 가리는 장비는 없다.",
        "built_space": "원형 유리 격실 하나와 침대 하나, 스캔 암 두 개가 보이며 천장 원형 조명, 수직 프레임, 바닥 환형 경계가 장소 참조에 매우 가깝다. 이현우는 유리 밖 왼쪽, 찰리는 유리 안 중간 오른쪽에 있다. 다만 인물을 작게 잡고 천장과 바닥, 양쪽 작업대를 넓게 포함해 지정된 미디엄 구도보다 넓다. 두 손은 화면 왼쪽과 중앙으로 벌어져 있다. 유리를 통한 직접 시야이며 불가능한 반사상은 보이지 않는다.",
        "entities": "이현우의 검은 헝클어진 머리와 흰 반소매 상의, 젊은 체형은 설명에 부합한다. 얼굴은 거의 가려져 정체성과 표정을 세밀하게 대조할 수 없다. 찰리는 흰 마스크형 기계 얼굴과 베이지 장갑판을 갖추고 연결선이 있는 금속 침대에 누워 있다. 다만 작은 화면 크기에서는 참조의 긴 팔과 짧은 다리 비율이 뚜렷하지 않다. 왼쪽 작업대에 앉은 연구원 한 명과 오른쪽 작업대에 선 연구원 한 명이 추가로 보인다.",
        "hard_violations": [
         "등장 허용 대상이 아닌 연구원 두 명을 양쪽 작업대에 추가하여, 숏 텍스트에 없는 사람을 보이지 말라는 지침을 위반했다."
        ],
        "physics": "이현우의 팔과 손은 해부학적으로 연결되며 유리에 손을 대는 동작으로 읽힌다. 찰리의 머리, 몸통, 팔과 다리는 침대 위에 놓여 있고 스스로 들고 있는 신체 부위는 보이지 않는다. 침대는 중앙 기둥과 넓은 받침으로, 스캔 암은 고정 지지대와 관절로 지지된다. 연결 케이블은 아래로 처진다. 지지 없이 떠 있는 물체나 신체는 확인되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.528
   },
   "violations": {
    "B": [
     "[gpt-high] 숏에 허용된 이현우와 찰리 외에 여러 연구원을 추가했다. 컴퓨터가 운영 중이라는 상태 설명과 별개로, 인물 가시성 지침은 이들의 등장을 명시적으로 금지한다."
    ],
    "A": [
     "[gpt-high] 등장 허용 대상이 아닌 연구원 두 명을 양쪽 작업대에 추가하여, 숏 텍스트에 없는 사람을 보이지 말라는 지침을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1528
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "찰리의 고릴라형 기계 몸체와 흰색 마스크, 원형 유리벽의 형태와 바닥 디테일 등 레퍼런스의 요소를 완벽하게 구현하며 제시된 샷의 프레이밍을 정확히 따랐습니다.  ★위반: [gpt-high] 등장 허용 대상이 아닌 연구원 두 명을 양쪽 작업대에 추가하여, 숏 텍스트에 없는 사람을 보이지 말라는 지침을 위반했다."
   },
   {
    "label": "B",
    "score": 1528,
    "verdict_ko": "레퍼런스와 달리 찰리의 마스크에 둥근 눈 모양이 추가되었고, 유리벽 외부 바닥에 존재하지 않는 발광 라인이 임의로 생성되었으며 이현우의 왼손 묘사가 다소 어색합니다.  ★위반: [gpt-high] 숏에 허용된 이현우와 찰리 외에 여러 연구원을 추가했다. 컴퓨터가 운영 중이라는 상태 설명과 별개로, 인물 가시성 지침은 이들의 등장을 명시적으로 금지한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_glass_experiment_room_ea17f9.png",
    "asset_id": "9dc88b2d-de9a-4dc6-88d4-99a8c0dfdee7",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1299876>",
    "asset_id": "99dc4ad5-1251-4608-89ee-6005d5fc811c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e4c-bd00-7de4-9435-29d26b445e90",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh13__bgfirst_bg.png",
   "bg_asset_id": "2518c02a-4465-4387-8d84-80a1a45d1bd2",
   "bg_record_key": "S82sh13::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "glass_experiment_room",
   "groupbg_asset_id": "9dc88b2d-de9a-4dc6-88d4-99a8c0dfdee7"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "staged_characters_added": [
   "C01"
  ]
 },
 "S83sh8::signage": {
  "fp": "4cdffc7c0b80169b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::56738a2e383a2aed": {
  "subjects": [],
  "subject_text": "제주도 연구소 임시보호소 병실·숙소\n눈부시게 하얀 조명과 깨끗한 흰색 침구가 있는 모던한 병실이다.",
  "identity": "canonical",
  "scope_id": "L139",
  "scope_role": "location_interior",
  "scope_sha": "6c1ab7cd0b5f3a7a"
 },
 "S83sh8::bgfirst_bg": {
  "input_fingerprint": "52bc67becedb84ac",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 화상통화 모니터 화면 속 앰버의 얼굴 옆으로 라울이 바짝 붙어 있는 구도.\n\nLOCATION (lock): On the active video-call display inside the research institute's temporary-shelter room. The distant callers appear on the lit screen; their surrounding location is not established.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the forward approach beside 이현우's right shoulder position at shoulder height, keeping him outside the crop and retaining an oblique view of the monitor. Place the complete display near center-right at just under two-fifths of the frame, with 앰버 on its left and 라울 pressed close on its right, their adjacent faces occupying most of the transmitted image. Both lean toward the call and attend to 이현우's received image just off their local lens axis; the visible bezel and angled screen distinguish mediated contact from direct presence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Showing 앰버 and 라울 together during the active call) — The image-bearing front is visible obliquely, with its bezel retained and both faces clearly contained within it; used as Defines the screen boundary and the distance implicit in their reunion; 이현우's temporary-shelter room (Surrounds the active call station); used as Unobtrusive surrounding space keeps the monitor a physical object rather than a full-screen replacement image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room subdued and the transmitted faces readable with restrained display brightness and slight signal breakup consistent with the call.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 화상통화 모니터 화면 속 앰버의 얼굴 옆으로 라울이 바짝 붙어 있는 구도.\n\nLOCATION (lock): On the active video-call display inside the research institute's temporary-shelter room. The distant callers appear on the lit screen; their surrounding location is not established.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the forward approach beside 이현우's right shoulder position at shoulder height, keeping him outside the crop and retaining an oblique view of the monitor. Place the complete display near center-right at just under two-fifths of the frame, with 앰버 on its left and 라울 pressed close on its right, their adjacent faces occupying most of the transmitted image. Both lean toward the call and attend to 이현우's received image just off their local lens axis; the visible bezel and angled screen distinguish mediated contact from direct presence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Showing 앰버 and 라울 together during the active call) — The image-bearing front is visible obliquely, with its bezel retained and both faces clearly contained within it; used as Defines the screen boundary and the distance implicit in their reunion; 이현우's temporary-shelter room (Surrounds the active call station); used as Unobtrusive surrounding space keeps the monitor a physical object rather than a full-screen replacement image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room subdued and the transmitted faces readable with restrained display brightness and slight signal breakup consistent with the call.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S83sh8__bgfirst_bg.png",
  "asset_id": "06471ebf-e716-45d7-b55f-cd32b23912b1",
  "input_asset_ids": [
   "a9e86fce-64f3-42df-9b32-8300baa2514e",
   "b4bdb1e2-9977-4924-a0b7-a115d0c85eb4"
  ]
 },
 "S83sh8": {
  "input_fingerprint": "0b795695058e2d98",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 모니터 화면 속 앰버의 얼굴 옆으로 라울이 바짝 붙어 있는 구도.\n\nLOCATION (lock): On the active video-call display inside the research institute's temporary-shelter room. The distant callers appear on the lit screen; their surrounding location is not established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the forward approach beside 이현우's right shoulder position at shoulder height, keeping him outside the crop and retaining an oblique view of the monitor. Place the complete display near center-right at just under two-fifths of the frame, with 앰버 on its left and 라울 pressed close on its right, their adjacent faces occupying most of the transmitted image. Both lean toward the call and attend to 이현우's received image just off their local lens axis; the visible bezel and angled screen distinguish mediated contact from direct presence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Showing 앰버 and 라울 together during the active call) — The image-bearing front is visible obliquely, with its bezel retained and both faces clearly contained within it; used as Defines the screen boundary and the distance implicit in their reunion; 이현우's temporary-shelter room (Surrounds the active call station); used as Unobtrusive surrounding space keeps the monitor a physical object rather than a full-screen replacement image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room subdued and the transmitted faces readable with restrained display brightness and slight signal breakup consistent with the call.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call is active, displaying Amber after her surgery and Raul as he enters the image. The call is taking place in the shelter room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 모니터 화면 속 앰버의 얼굴 옆으로 라울이 바짝 붙어 있는 구도.\n\nLOCATION (lock): On the active video-call display inside the research institute's temporary-shelter room. The distant callers appear on the lit screen; their surrounding location is not established. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the forward approach beside 이현우's right shoulder position at shoulder height, keeping him outside the crop and retaining an oblique view of the monitor. Place the complete display near center-right at just under two-fifths of the frame, with 앰버 on its left and 라울 pressed close on its right, their adjacent faces occupying most of the transmitted image. Both lean toward the call and attend to 이현우's received image just off their local lens axis; the visible bezel and angled screen distinguish mediated contact from direct presence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Showing 앰버 and 라울 together during the active call) — The image-bearing front is visible obliquely, with its bezel retained and both faces clearly contained within it; used as Defines the screen boundary and the distance implicit in their reunion; 이현우's temporary-shelter room (Surrounds the active call station); used as Unobtrusive surrounding space keeps the monitor a physical object rather than a full-screen replacement image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room subdued and the transmitted faces readable with restrained display brightness and slight signal breakup consistent with the call.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call is active, displaying Amber after her surgery and Raul as he enters the image. The call is taking place in the shelter room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 모니터 화면 속 앰버의 얼굴 옆으로 라울이 바짝 붙어 있는 구도.\n\nLOCATION (lock): On the active video-call display inside the research institute's temporary-shelter room. The distant callers appear on the lit screen; their surrounding location is not established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the forward approach beside 이현우's right shoulder position at shoulder height, keeping him outside the crop and retaining an oblique view of the monitor. Place the complete display near center-right at just under two-fifths of the frame, with 앰버 on its left and 라울 pressed close on its right, their adjacent faces occupying most of the transmitted image. Both lean toward the call and attend to 이현우's received image just off their local lens axis; the visible bezel and angled screen distinguish mediated contact from direct presence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Showing 앰버 and 라울 together during the active call) — The image-bearing front is visible obliquely, with its bezel retained and both faces clearly contained within it; used as Defines the screen boundary and the distance implicit in their reunion; 이현우's temporary-shelter room (Surrounds the active call station); used as Unobtrusive surrounding space keeps the monitor a physical object rather than a full-screen replacement image.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the room subdued and the transmitted faces readable with restrained display brightness and slight signal breakup consistent with the call.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call is active, displaying Amber after her surgery and Raul as he enters the image. The call is taking place in the shelter room.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈); 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S83sh8__bgfirst_bg.png",
     "asset_id": "06471ebf-e716-45d7-b55f-cd32b23912b1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S83sh8.png",
     "asset_id": "a9e86fce-64f3-42df-9b32-8300baa2514e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L139B02.png",
     "asset_id": "b4bdb1e2-9977-4924-a0b7-a115d0c85eb4",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1215317>",
     "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1334467>",
     "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "앰버는 송신 화면 왼쪽, 라울은 오른쪽에서 머리를 맞대고 통화 카메라 부근을 바라본다. 수신된 이현우 영상을 보는 시선으로 해석할 수 있으나 렌즈에서 살짝 벗어난 주시점은 뚜렷하지 않다. 오른쪽 전경의 남성은 모니터를 향한다. 화면의 기능면은 관찰자와 남성 쪽을 향해 비스듬히 보인다.",
    "built_space": "왼쪽 통신 설비에 수화기 하나가 있고, 그 위로 기존 벽면 모니터 일부가 보이며, 아래에 별도의 활성 모니터가 추가되어 표시 장치가 두 대다. 뒤에는 침대 하나, 바다를 향한 두 짝 창 하나, 좁은 침상용 탁자 하나가 보인다. 금속 수납장과 벽 재질은 참조와 유사하다. 활성 모니터 전체와 베젤은 보이지만 프레임 중앙 왼쪽에 놓였다.",
    "entities": "화면 속 두 아이는 약 10세로 보이며, 앰버의 금발·큰 눈·밝은 피부·남색 상의와 라울의 짙은 피부·뒤로 모은 곱슬머리·청색 상의가 참조 외형에 대체로 부합한다. 계통 자체는 외형만으로 확정할 수 없다. 오른쪽 전경에는 허용되지 않은 성인 남성의 머리와 어깨가 있으며, 통화 화면 오른쪽 아래에도 성인 남성의 작은 영상이 있다. 수술 이후라는 상태를 별도로 입증하는 시각 단서는 없다.",
    "hard_violations": [
     "화면 밖에 두라고 명시된 이현우로 보이는 성인 남성의 머리와 어깨가 오른쪽 전경에 등장한다.",
     "통화 화면 안에도 허용된 두 아이 외에 성인 남성의 미리보기 영상이 추가되어 있다.",
     "참조의 기존 벽면 모니터를 남긴 채 별도의 통화 모니터를 추가하여 표시 설비가 두 대로 중복된다."
    ],
    "physics": "활성 모니터는 하단 금속 지지부를 통해 수납장 상판에 지지되고, 상단 카메라는 베젤에 고정되어 있다. 두 아이의 머리는 목과 어깨에 자연스럽게 이어지며 서로 기대는 자세도 가능하다. 하체와 좌석은 송신 영상의 범위 밖이므로 지지 여부를 추가로 판단할 수 없다. 전경 남성도 몸통이 프레임 아래로 이어지며 부유 징후는 없다."
   },
   {
    "label": "A",
    "direction": "앰버는 송신 화면 왼쪽에서, 라울은 오른쪽에서 고개를 안쪽으로 기울여 서로 바짝 붙어 있다. 두 아이의 눈은 통화 카메라 부근의 상대 영상을 향하는 것으로 읽힌다. 전경 남성은 오른쪽 모니터를 보고 있고 모니터 앞면도 그와 카메라 쪽을 향해 사용 방향은 자연스럽다.",
    "built_space": "전체 베젤이 보이는 모니터 한 대가 오른쪽 탁자 위에 있다. 뒤에는 침대 하나, 좁은 침상용 탁자 하나, 블라인드가 달린 두 짝 창 하나, 작은 수납장 하나, 천장 형광등 하나가 보인다. 바다 전망과 방의 주요 재료·배치는 참조에 가깝지만, 통화 장치는 참조의 왼쪽 벽면 설비 대신 오른쪽 전경 탁자에 놓였다. 모니터의 오른쪽 배치와 비스듬한 앞면은 A보다 지정 구성에 가깝다.",
    "entities": "화면에는 참조와 유사한 금발의 어린 여자아이와 뒤로 묶은 곱슬머리의 짙은 피부 남자아이가 있고, 각각 남색과 청색 상의를 입었다. 얼굴과 연령대는 대체로 일치하며, 정확한 혼혈 계통은 외형만으로 확정할 수 없다. 두 얼굴이 송신 영상 대부분을 차지하고 약한 신호 잡음도 보인다. 그러나 왼쪽 전경에는 허용되지 않은 성인 남성의 머리와 상체가 크게 등장한다. 수술 이후 상태는 화면만으로 확인되지 않는다.",
    "hard_violations": [
     "이현우를 프레임 밖에 유지하라는 명시적 지시와 달리, 성인 남성의 머리·어깨·등이 왼쪽 전경을 크게 차지한다."
    ],
    "physics": "모니터는 목과 넓은 받침판으로 탁자 위에 안정적으로 지지된다. 아이들의 고개 기울임과 어깨 밀착은 자연스럽고, 목과 몸통의 연결에도 불가능한 자세가 없다. 하체와 좌석은 통화 화면 밖이라 확인 대상이 아니다. 전경 남성의 몸통은 하단으로 이어지고, 공중에 떠 있는 인물이나 물체는 보이지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 아이가 화면 안에서 밀착한 순간은 맞지만, 이현우와 그의 통화 미리보기까지 노출하고 모니터를 중복 배치했으며, 주 화면도 지정된 중앙 오른쪽이 아닌 왼쪽에 있다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "완전한 모니터를 오른쪽에 비스듬히 두고 두 얼굴을 밀착시킨 구성은 더 가깝지만, 제외해야 할 이현우가 전경을 크게 차지해 최종 프레임으로는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버는 송신 화면 왼쪽, 라울은 오른쪽에서 머리를 맞대고 통화 카메라 부근을 바라본다. 수신된 이현우 영상을 보는 시선으로 해석할 수 있으나 렌즈에서 살짝 벗어난 주시점은 뚜렷하지 않다. 오른쪽 전경의 남성은 모니터를 향한다. 화면의 기능면은 관찰자와 남성 쪽을 향해 비스듬히 보인다.",
        "built_space": "왼쪽 통신 설비에 수화기 하나가 있고, 그 위로 기존 벽면 모니터 일부가 보이며, 아래에 별도의 활성 모니터가 추가되어 표시 장치가 두 대다. 뒤에는 침대 하나, 바다를 향한 두 짝 창 하나, 좁은 침상용 탁자 하나가 보인다. 금속 수납장과 벽 재질은 참조와 유사하다. 활성 모니터 전체와 베젤은 보이지만 프레임 중앙 왼쪽에 놓였다.",
        "entities": "화면 속 두 아이는 약 10세로 보이며, 앰버의 금발·큰 눈·밝은 피부·남색 상의와 라울의 짙은 피부·뒤로 모은 곱슬머리·청색 상의가 참조 외형에 대체로 부합한다. 계통 자체는 외형만으로 확정할 수 없다. 오른쪽 전경에는 허용되지 않은 성인 남성의 머리와 어깨가 있으며, 통화 화면 오른쪽 아래에도 성인 남성의 작은 영상이 있다. 수술 이후라는 상태를 별도로 입증하는 시각 단서는 없다.",
        "hard_violations": [
         "화면 밖에 두라고 명시된 이현우로 보이는 성인 남성의 머리와 어깨가 오른쪽 전경에 등장한다.",
         "통화 화면 안에도 허용된 두 아이 외에 성인 남성의 미리보기 영상이 추가되어 있다.",
         "참조의 기존 벽면 모니터를 남긴 채 별도의 통화 모니터를 추가하여 표시 설비가 두 대로 중복된다."
        ],
        "physics": "활성 모니터는 하단 금속 지지부를 통해 수납장 상판에 지지되고, 상단 카메라는 베젤에 고정되어 있다. 두 아이의 머리는 목과 어깨에 자연스럽게 이어지며 서로 기대는 자세도 가능하다. 하체와 좌석은 송신 영상의 범위 밖이므로 지지 여부를 추가로 판단할 수 없다. 전경 남성도 몸통이 프레임 아래로 이어지며 부유 징후는 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 송신 화면 왼쪽에서, 라울은 오른쪽에서 고개를 안쪽으로 기울여 서로 바짝 붙어 있다. 두 아이의 눈은 통화 카메라 부근의 상대 영상을 향하는 것으로 읽힌다. 전경 남성은 오른쪽 모니터를 보고 있고 모니터 앞면도 그와 카메라 쪽을 향해 사용 방향은 자연스럽다.",
        "built_space": "전체 베젤이 보이는 모니터 한 대가 오른쪽 탁자 위에 있다. 뒤에는 침대 하나, 좁은 침상용 탁자 하나, 블라인드가 달린 두 짝 창 하나, 작은 수납장 하나, 천장 형광등 하나가 보인다. 바다 전망과 방의 주요 재료·배치는 참조에 가깝지만, 통화 장치는 참조의 왼쪽 벽면 설비 대신 오른쪽 전경 탁자에 놓였다. 모니터의 오른쪽 배치와 비스듬한 앞면은 A보다 지정 구성에 가깝다.",
        "entities": "화면에는 참조와 유사한 금발의 어린 여자아이와 뒤로 묶은 곱슬머리의 짙은 피부 남자아이가 있고, 각각 남색과 청색 상의를 입었다. 얼굴과 연령대는 대체로 일치하며, 정확한 혼혈 계통은 외형만으로 확정할 수 없다. 두 얼굴이 송신 영상 대부분을 차지하고 약한 신호 잡음도 보인다. 그러나 왼쪽 전경에는 허용되지 않은 성인 남성의 머리와 상체가 크게 등장한다. 수술 이후 상태는 화면만으로 확인되지 않는다.",
        "hard_violations": [
         "이현우를 프레임 밖에 유지하라는 명시적 지시와 달리, 성인 남성의 머리·어깨·등이 왼쪽 전경을 크게 차지한다."
        ],
        "physics": "모니터는 목과 넓은 받침판으로 탁자 위에 안정적으로 지지된다. 아이들의 고개 기울임과 어깨 밀착은 자연스럽고, 목과 몸통의 연결에도 불가능한 자세가 없다. 하체와 좌석은 통화 화면 밖이라 확인 대상이 아니다. 전경 남성의 몸통은 하단으로 이어지고, 공중에 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 아이가 화면 안에서 밀착한 순간은 맞지만, 이현우와 그의 통화 미리보기까지 노출하고 모니터를 중복 배치했으며, 주 화면도 지정된 중앙 오른쪽이 아닌 왼쪽에 있다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "완전한 모니터를 오른쪽에 비스듬히 두고 두 얼굴을 밀착시킨 구성은 더 가깝지만, 제외해야 할 이현우가 전경을 크게 차지해 최종 프레임으로는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "앰버는 송신 화면 왼쪽, 라울은 오른쪽에서 머리를 맞대고 통화 카메라 부근을 바라본다. 수신된 이현우 영상을 보는 시선으로 해석할 수 있으나 렌즈에서 살짝 벗어난 주시점은 뚜렷하지 않다. 오른쪽 전경의 남성은 모니터를 향한다. 화면의 기능면은 관찰자와 남성 쪽을 향해 비스듬히 보인다.",
        "built_space": "왼쪽 통신 설비에 수화기 하나가 있고, 그 위로 기존 벽면 모니터 일부가 보이며, 아래에 별도의 활성 모니터가 추가되어 표시 장치가 두 대다. 뒤에는 침대 하나, 바다를 향한 두 짝 창 하나, 좁은 침상용 탁자 하나가 보인다. 금속 수납장과 벽 재질은 참조와 유사하다. 활성 모니터 전체와 베젤은 보이지만 프레임 중앙 왼쪽에 놓였다.",
        "entities": "화면 속 두 아이는 약 10세로 보이며, 앰버의 금발·큰 눈·밝은 피부·남색 상의와 라울의 짙은 피부·뒤로 모은 곱슬머리·청색 상의가 참조 외형에 대체로 부합한다. 계통 자체는 외형만으로 확정할 수 없다. 오른쪽 전경에는 허용되지 않은 성인 남성의 머리와 어깨가 있으며, 통화 화면 오른쪽 아래에도 성인 남성의 작은 영상이 있다. 수술 이후라는 상태를 별도로 입증하는 시각 단서는 없다.",
        "hard_violations": [
         "화면 밖에 두라고 명시된 이현우로 보이는 성인 남성의 머리와 어깨가 오른쪽 전경에 등장한다.",
         "통화 화면 안에도 허용된 두 아이 외에 성인 남성의 미리보기 영상이 추가되어 있다.",
         "참조의 기존 벽면 모니터를 남긴 채 별도의 통화 모니터를 추가하여 표시 설비가 두 대로 중복된다."
        ],
        "physics": "활성 모니터는 하단 금속 지지부를 통해 수납장 상판에 지지되고, 상단 카메라는 베젤에 고정되어 있다. 두 아이의 머리는 목과 어깨에 자연스럽게 이어지며 서로 기대는 자세도 가능하다. 하체와 좌석은 송신 영상의 범위 밖이므로 지지 여부를 추가로 판단할 수 없다. 전경 남성도 몸통이 프레임 아래로 이어지며 부유 징후는 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 송신 화면 왼쪽에서, 라울은 오른쪽에서 고개를 안쪽으로 기울여 서로 바짝 붙어 있다. 두 아이의 눈은 통화 카메라 부근의 상대 영상을 향하는 것으로 읽힌다. 전경 남성은 오른쪽 모니터를 보고 있고 모니터 앞면도 그와 카메라 쪽을 향해 사용 방향은 자연스럽다.",
        "built_space": "전체 베젤이 보이는 모니터 한 대가 오른쪽 탁자 위에 있다. 뒤에는 침대 하나, 좁은 침상용 탁자 하나, 블라인드가 달린 두 짝 창 하나, 작은 수납장 하나, 천장 형광등 하나가 보인다. 바다 전망과 방의 주요 재료·배치는 참조에 가깝지만, 통화 장치는 참조의 왼쪽 벽면 설비 대신 오른쪽 전경 탁자에 놓였다. 모니터의 오른쪽 배치와 비스듬한 앞면은 A보다 지정 구성에 가깝다.",
        "entities": "화면에는 참조와 유사한 금발의 어린 여자아이와 뒤로 묶은 곱슬머리의 짙은 피부 남자아이가 있고, 각각 남색과 청색 상의를 입었다. 얼굴과 연령대는 대체로 일치하며, 정확한 혼혈 계통은 외형만으로 확정할 수 없다. 두 얼굴이 송신 영상 대부분을 차지하고 약한 신호 잡음도 보인다. 그러나 왼쪽 전경에는 허용되지 않은 성인 남성의 머리와 상체가 크게 등장한다. 수술 이후 상태는 화면만으로 확인되지 않는다.",
        "hard_violations": [
         "이현우를 프레임 밖에 유지하라는 명시적 지시와 달리, 성인 남성의 머리·어깨·등이 왼쪽 전경을 크게 차지한다."
        ],
        "physics": "모니터는 목과 넓은 받침판으로 탁자 위에 안정적으로 지지된다. 아이들의 고개 기울임과 어깨 밀착은 자연스럽고, 목과 몸통의 연결에도 불가능한 자세가 없다. 하체와 좌석은 통화 화면 밖이라 확인 대상이 아니다. 전경 남성의 몸통은 하단으로 이어지고, 공중에 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2,
   "A": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "두 아이가 화면 안에서 밀착한 순간은 맞지만, 이현우와 그의 통화 미리보기까지 노출하고 모니터를 중복 배치했으며, 주 화면도 지정된 중앙 오른쪽이 아닌 왼쪽에 있다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "완전한 모니터를 오른쪽에 비스듬히 두고 두 얼굴을 밀착시킨 구성은 더 가깝지만, 제외해야 할 이현우가 전경을 크게 차지해 최종 프레임으로는 부적합하다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L139B02.png",
    "asset_id": "b4bdb1e2-9977-4924-a0b7-a115d0c85eb4",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1215317>",
    "asset_id": "21ec6f03-01c1-4787-8ea4-a3784e06f866",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1334467>",
    "asset_id": "a557c771-aaac-4090-a251-3139a767186e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e57-f672-7218-89b9-290cef5bd1bd",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S83sh8__bgfirst_bg.png",
   "bg_asset_id": "06471ebf-e716-45d7-b55f-cd32b23912b1",
   "bg_record_key": "S83sh8::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S83sh9::signage": {
  "fp": "17c8177b7a432786",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S83sh9": {
  "input_fingerprint": "34a11b4b37fc5e38",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 화면을 향해 애써 밝은 미소를 짓는 이현우의 씁쓸한 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary-shelter bedroom, with the active monitor providing a local glow in the daytime room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue around 이현우's right side to the monitor-side position, slowing just below eye level and looking slightly upward at his right three-quarter face without crossing the call axis. Place his face on the right two-thirds with breathing room to the left, where he watches the active display outside the crop, his mouth lifting into a smile while his jaw and shoulders resist it. Shift the visual emphasis from the transmitted faces to his answering expression, keeping the call's spatial relationship intact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Temporary-shelter room interior (The call remains in progress); used as A softly resolved background isolates the face without adding new furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued ambient illumination and gentle facial contrast so the strained smile reads without an invented change in light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter bedroom's surfaces, bed, window, and daytime lighting from the reference. Exclude the remote callers as physical occupants and do not import their surroundings into this room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected, now showing Amber after she has taken the phone back; the image has shown intermittent interference. 이현우: He remains on the call, struggling to maintain a reassuring expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 화면을 향해 애써 밝은 미소를 짓는 이현우의 씁쓸한 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary-shelter bedroom, with the active monitor providing a local glow in the daytime room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue around 이현우's right side to the monitor-side position, slowing just below eye level and looking slightly upward at his right three-quarter face without crossing the call axis. Place his face on the right two-thirds with breathing room to the left, where he watches the active display outside the crop, his mouth lifting into a smile while his jaw and shoulders resist it. Shift the visual emphasis from the transmitted faces to his answering expression, keeping the call's spatial relationship intact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Temporary-shelter room interior (The call remains in progress); used as A softly resolved background isolates the face without adding new furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued ambient illumination and gentle facial contrast so the strained smile reads without an invented change in light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter bedroom's surfaces, bed, window, and daytime lighting from the reference. Exclude the remote callers as physical occupants and do not import their surroundings into this room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected, now showing Amber after she has taken the phone back; the image has shown intermittent interference. 이현우: He remains on the call, struggling to maintain a reassuring expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화상통화 화면을 향해 애써 밝은 미소를 짓는 이현우의 씁쓸한 얼굴.\n\nLOCATION (lock): At the video-call station inside the temporary-shelter bedroom, with the active monitor providing a local glow in the daytime room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue around 이현우's right side to the monitor-side position, slowing just below eye level and looking slightly upward at his right three-quarter face without crossing the call axis. Place his face on the right two-thirds with breathing room to the left, where he watches the active display outside the crop, his mouth lifting into a smile while his jaw and shoulders resist it. Shift the visual emphasis from the transmitted faces to his answering expression, keeping the call's spatial relationship intact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Temporary-shelter room interior (The call remains in progress); used as A softly resolved background isolates the face without adding new furnishings.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued ambient illumination and gentle facial contrast so the strained smile reads without an invented change in light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter bedroom's surfaces, bed, window, and daytime lighting from the reference. Exclude the remote callers as physical occupants and do not import their surroundings into this room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The video call remains connected, now showing Amber after she has taken the phone back; the image has shown intermittent interference. 이현우: He remains on the call, struggling to maintain a reassuring expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 프레임 왼쪽 바깥의 모니터를 향하고 있음.",
    "built_space": "참조 이미지와 동일한 창밖 해안 풍경, 침대, 침대 발치의 벤치, 우측 수납장 배치가 부드럽게 아웃포커싱되어 나타남.",
    "entities": "10대 후반 아시아계 남성, 헝클어진 짧은 검은 머리, 흰색 환자복. 억지로 웃는 씁쓸한 표정이 묘사됨.",
    "hard_violations": [],
    "physics": "자연스러운 자세로 앉아 모니터를 응시하며 어깨에 힘이 들어간 상태."
   },
   {
    "label": "B",
    "direction": "시선이 왼쪽 모니터 화면을 향함.",
    "built_space": "창문과 침대, 수납장은 보이나 침대 발치의 흰색 벤치가 누락됨.",
    "entities": "10대 후반 아시아계 남성, 검은 머리, 흰색 환자복. 씁쓸함보다는 평온하고 부드러운 미소에 가까움.",
    "hard_violations": [],
    "physics": "안정적인 자세로 앉아 있음. 물리적 오류 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 프레이밍(우측 2/3 위치, 약간 아래서 올려다보는 앵글)과 억지로 짓는 미소 표정, 배경의 세부 요소(침대 앞 벤치 등)를 모두 정확히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시문에 비해 표정이 평온하며 프레이밍이 다소 중앙에 치우쳐 있고, 배경 레퍼런스의 벤치가 누락됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 프레임 왼쪽 바깥의 모니터를 향하고 있음.",
        "built_space": "참조 이미지와 동일한 창밖 해안 풍경, 침대, 침대 발치의 벤치, 우측 수납장 배치가 부드럽게 아웃포커싱되어 나타남.",
        "entities": "10대 후반 아시아계 남성, 헝클어진 짧은 검은 머리, 흰색 환자복. 억지로 웃는 씁쓸한 표정이 묘사됨.",
        "hard_violations": [],
        "physics": "자연스러운 자세로 앉아 모니터를 응시하며 어깨에 힘이 들어간 상태."
       },
       {
        "label": "B",
        "direction": "시선이 왼쪽 모니터 화면을 향함.",
        "built_space": "창문과 침대, 수납장은 보이나 침대 발치의 흰색 벤치가 누락됨.",
        "entities": "10대 후반 아시아계 남성, 검은 머리, 흰색 환자복. 씁쓸함보다는 평온하고 부드러운 미소에 가까움.",
        "hard_violations": [],
        "physics": "안정적인 자세로 앉아 있음. 물리적 오류 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 프레이밍(우측 2/3 위치, 약간 아래서 올려다보는 앵글)과 억지로 짓는 미소 표정, 배경의 세부 요소(침대 앞 벤치 등)를 모두 정확히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시문에 비해 표정이 평온하며 프레이밍이 다소 중앙에 치우쳐 있고, 배경 레퍼런스의 벤치가 누락됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 프레임 왼쪽 바깥의 모니터를 향하고 있음.",
        "built_space": "참조 이미지와 동일한 창밖 해안 풍경, 침대, 침대 발치의 벤치, 우측 수납장 배치가 부드럽게 아웃포커싱되어 나타남.",
        "entities": "10대 후반 아시아계 남성, 헝클어진 짧은 검은 머리, 흰색 환자복. 억지로 웃는 씁쓸한 표정이 묘사됨.",
        "hard_violations": [],
        "physics": "자연스러운 자세로 앉아 모니터를 응시하며 어깨에 힘이 들어간 상태."
       },
       {
        "label": "B",
        "direction": "시선이 왼쪽 모니터 화면을 향함.",
        "built_space": "창문과 침대, 수납장은 보이나 침대 발치의 흰색 벤치가 누락됨.",
        "entities": "10대 후반 아시아계 남성, 검은 머리, 흰색 환자복. 씁쓸함보다는 평온하고 부드러운 미소에 가까움.",
        "hard_violations": [],
        "physics": "안정적인 자세로 앉아 있음. 물리적 오류 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "화면을 향한 씁쓸한 미소와 방의 연속성은 맞지만, 거의 눈높이인 시점과 상대적으로 밝고 선명한 배경이 지정된 낮은 시점·절제된 얼굴 강조에 덜 충실하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "눈높이 바로 아래에서 본 오른쪽 삼사분면 얼굴, 왼쪽 통화 화면을 향한 시선, 굳은 턱과 억누른 미소를 더 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 두 눈은 왼쪽 전경의 모니터 방향을 향하며 렌즈를 응시하지 않는다. 오른쪽 뺨과 귀가 보이는 삼사분면이다. 모니터의 표시 내용은 보이지 않아 통화 상대는 확인할 수 없지만, 시선과 기기의 위치는 서로 맞는다.",
        "built_space": "뒤쪽에 블라인드가 달린 창 하나, 왼쪽에 흰 금속 침대 하나, 오른쪽에 작은 수납장 하나와 벽 달력이 보인다. 왼쪽 전경에는 모니터 일부가 있다. 참고 장소의 벽 재질과 바다 전망, 침대·창·수납장의 관계를 유지한다. 얼굴은 오른쪽에 크게 놓였으나 카메라는 거의 눈높이로 보이고 배경은 비교적 선명하다. 중복 설비나 반사는 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며, 짧고 헝클어진 검은 머리와 깨끗한 흰색 브이넥 상의가 지정 인물에 부합한다. 참고 인물은 마스크와 모자로 가려져 있어 얼굴 전체의 정확한 동일성은 확인하기 어렵다. 다문 입의 올라간 입꼬리와 촉촉한 눈이 애써 웃는 표정을 만든다. 하의는 프레임 밖이며 원격 통화자가 방 안에 추가되지 않았다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 지지되고 몸통은 화면 아래로 이어진다. 골반과 좌석은 클로즈업 밖이므로 착석 접촉은 확인할 수 없지만 공중에 뜬 자세는 아니다. 침대와 수납장은 바닥에 놓여 있고, 전경 모니터는 하부가 잘려 지지대가 보이지 않는다. 손에 든 물체나 비현실적인 관절은 없다."
       },
       {
        "label": "B",
        "direction": "두 눈과 코는 왼쪽 전경의 모니터 쪽을 향한다. 오른쪽 뺨과 귀가 드러나며 렌즈와 시선이 만나지 않아 통화 축을 유지한다. 실제 표시 면과 상대방은 크롭 밖이어서 앰버나 통신 간섭은 직접 확인할 수 없다.",
        "built_space": "블라인드가 달린 창 하나, 왼쪽 침대 하나, 침대 옆 좁은 벤치 하나, 창 오른쪽 수납장 하나와 천장 조명 하나가 보인다. 참고 방의 설비 수와 상대적 배치에 부합하며 새로운 가구는 보이지 않는다. 왼쪽 모니터 가장자리 너머로 방이 부드럽게 흐려지고, 얼굴은 오른쪽에 크게 배치된다. 약간 올려다보는 시점이 지정된 카메라 높이에 더 가깝다.",
        "entities": "젊은 동아시아계 남성 한 명이며 검은 헝클어진 머리, 마른 체형, 깨끗한 흰색 브이넥 상의가 인물 조건과 맞는다. 참고 사진의 가려진 얼굴 때문에 세부 동일성에는 확인 한계가 있다. 입꼬리는 조금 올라가지만 입술과 턱은 다물려 있고 눈가에는 슬픔이 남아 안심시키려 애쓰는 표정으로 읽힌다. 다른 사람이나 자막은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 받치며 상체는 화면 아래로 이어진다. 좌석과 발은 프레임 밖이므로 접촉점을 판정할 수 없고, 보이는 부분에는 부유나 불가능한 자세가 없다. 침대·벤치·수납장은 바닥에 자리하며 창과 조명은 건물에 고정되어 있다. 모니터의 하부 지지는 크롭 밖이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면을 향한 씁쓸한 미소와 방의 연속성은 맞지만, 거의 눈높이인 시점과 상대적으로 밝고 선명한 배경이 지정된 낮은 시점·절제된 얼굴 강조에 덜 충실하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "눈높이 바로 아래에서 본 오른쪽 삼사분면 얼굴, 왼쪽 통화 화면을 향한 시선, 굳은 턱과 억누른 미소를 더 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 두 눈은 왼쪽 전경의 모니터 방향을 향하며 렌즈를 응시하지 않는다. 오른쪽 뺨과 귀가 보이는 삼사분면이다. 모니터의 표시 내용은 보이지 않아 통화 상대는 확인할 수 없지만, 시선과 기기의 위치는 서로 맞는다.",
        "built_space": "뒤쪽에 블라인드가 달린 창 하나, 왼쪽에 흰 금속 침대 하나, 오른쪽에 작은 수납장 하나와 벽 달력이 보인다. 왼쪽 전경에는 모니터 일부가 있다. 참고 장소의 벽 재질과 바다 전망, 침대·창·수납장의 관계를 유지한다. 얼굴은 오른쪽에 크게 놓였으나 카메라는 거의 눈높이로 보이고 배경은 비교적 선명하다. 중복 설비나 반사는 없다.",
        "entities": "젊은 동아시아계 남성 한 명만 보이며, 짧고 헝클어진 검은 머리와 깨끗한 흰색 브이넥 상의가 지정 인물에 부합한다. 참고 인물은 마스크와 모자로 가려져 있어 얼굴 전체의 정확한 동일성은 확인하기 어렵다. 다문 입의 올라간 입꼬리와 촉촉한 눈이 애써 웃는 표정을 만든다. 하의는 프레임 밖이며 원격 통화자가 방 안에 추가되지 않았다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 지지되고 몸통은 화면 아래로 이어진다. 골반과 좌석은 클로즈업 밖이므로 착석 접촉은 확인할 수 없지만 공중에 뜬 자세는 아니다. 침대와 수납장은 바닥에 놓여 있고, 전경 모니터는 하부가 잘려 지지대가 보이지 않는다. 손에 든 물체나 비현실적인 관절은 없다."
       },
       {
        "label": "A",
        "direction": "두 눈과 코는 왼쪽 전경의 모니터 쪽을 향한다. 오른쪽 뺨과 귀가 드러나며 렌즈와 시선이 만나지 않아 통화 축을 유지한다. 실제 표시 면과 상대방은 크롭 밖이어서 앰버나 통신 간섭은 직접 확인할 수 없다.",
        "built_space": "블라인드가 달린 창 하나, 왼쪽 침대 하나, 침대 옆 좁은 벤치 하나, 창 오른쪽 수납장 하나와 천장 조명 하나가 보인다. 참고 방의 설비 수와 상대적 배치에 부합하며 새로운 가구는 보이지 않는다. 왼쪽 모니터 가장자리 너머로 방이 부드럽게 흐려지고, 얼굴은 오른쪽에 크게 배치된다. 약간 올려다보는 시점이 지정된 카메라 높이에 더 가깝다.",
        "entities": "젊은 동아시아계 남성 한 명이며 검은 헝클어진 머리, 마른 체형, 깨끗한 흰색 브이넥 상의가 인물 조건과 맞는다. 참고 사진의 가려진 얼굴 때문에 세부 동일성에는 확인 한계가 있다. 입꼬리는 조금 올라가지만 입술과 턱은 다물려 있고 눈가에는 슬픔이 남아 안심시키려 애쓰는 표정으로 읽힌다. 다른 사람이나 자막은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 받치며 상체는 화면 아래로 이어진다. 좌석과 발은 프레임 밖이므로 접촉점을 판정할 수 없고, 보이는 부분에는 부유나 불가능한 자세가 없다. 침대·벤치·수납장은 바닥에 자리하며 창과 조명은 건물에 고정되어 있다. 모니터의 하부 지지는 크롭 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.46
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.46
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1460
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 프레이밍(우측 2/3 위치, 약간 아래서 올려다보는 앵글)과 억지로 짓는 미소 표정, 배경의 세부 요소(침대 앞 벤치 등)를 모두 정확히 구현함."
   },
   {
    "label": "B",
    "score": 1460,
    "verdict_ko": "지시문에 비해 표정이 평온하며 프레이밍이 다소 중앙에 치우쳐 있고, 배경 레퍼런스의 벤치가 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S81sh4_sel.png",
    "asset_id": "18c905c0-c790-431b-8613-4af660f9ff2b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e5f-16c6-7ab1-96c1-b0f11cdf2cfa",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh4"
  }
 },
 "S83sh10::signage": {
  "fp": "c469d8d2844fb4e9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S83sh10": {
  "input_fingerprint": "8528cca2ac5a8418",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화면이 꺼진 텅 빈 모니터 앞 이현우의 뒷모습.\n\nLOCATION (lock): At the now-dark video-call monitor inside the temporary-shelter bedroom. The room retains its daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the reverse arc behind and slightly right of 이현우, settling above his head in a downward-looking wide view before he turns toward the bed. Keep his back below center and the blank monitor beyond him toward the upper left, with his head still directed toward its now-empty face and his shoulders releasing after the call. Emphasize the increased camera distance and surrounding vacancy rather than a new lighting cue.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Blank after the call has ended) — Its empty display face is visible beyond 이현우 from a high oblique angle; used as Places the absent callers in the same space as his remaining presence; Temporary-shelter room interior (이현우 is alone after the call); used as Visible space around his figure carries the emotional withdrawal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the room's subdued ambient illumination after the call image disappears, with controlled contrast around his solitary figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter room's furnishings, interior finishes, and daylight appearance from the reference. Exclude the remote callers and any active call image; the monitor is now off.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared from the screen. The shelter bed remains available in the room. 이현우: He is alone in the shelter room after the call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화면이 꺼진 텅 빈 모니터 앞 이현우의 뒷모습.\n\nLOCATION (lock): At the now-dark video-call monitor inside the temporary-shelter bedroom. The room retains its daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the reverse arc behind and slightly right of 이현우, settling above his head in a downward-looking wide view before he turns toward the bed. Keep his back below center and the blank monitor beyond him toward the upper left, with his head still directed toward its now-empty face and his shoulders releasing after the call. Emphasize the increased camera distance and surrounding vacancy rather than a new lighting cue.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Blank after the call has ended) — Its empty display face is visible beyond 이현우 from a high oblique angle; used as Places the absent callers in the same space as his remaining presence; Temporary-shelter room interior (이현우 is alone after the call); used as Visible space around his figure carries the emotional withdrawal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the room's subdued ambient illumination after the call image disappears, with controlled contrast around his solitary figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter room's furnishings, interior finishes, and daylight appearance from the reference. Exclude the remote callers and any active call image; the monitor is now off.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared from the screen. The shelter bed remains available in the room. 이현우: He is alone in the shelter room after the call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 화면이 꺼진 텅 빈 모니터 앞 이현우의 뒷모습.\n\nLOCATION (lock): At the now-dark video-call monitor inside the temporary-shelter bedroom. The room retains its daytime ambient light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the reverse arc behind and slightly right of 이현우, settling above his head in a downward-looking wide view before he turns toward the bed. Keep his back below center and the blank monitor beyond him toward the upper left, with his head still directed toward its now-empty face and his shoulders releasing after the call. Emphasize the increased camera distance and surrounding vacancy rather than a new lighting cue.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Video-call monitor (Blank after the call has ended) — Its empty display face is visible beyond 이현우 from a high oblique angle; used as Places the absent callers in the same space as his remaining presence; Temporary-shelter room interior (이현우 is alone after the call); used as Visible space around his figure carries the emotional withdrawal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the room's subdued ambient illumination after the call image disappears, with controlled contrast around his solitary figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the shelter room's furnishings, interior finishes, and daylight appearance from the reference. Exclude the remote callers and any active call image; the monitor is now off.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The call has ended and the video image has disappeared from the screen. The shelter bed remains available in the room. 이현우: He is alone in the shelter room after the call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 고개가 좌측 상단의 꺼진 모니터를 향해 있음.",
    "built_space": "왼쪽에 책상, 오른쪽에 침대가 배치됨. 참조 이미지의 공간 배치가 왜곡되었고, 침대 측면에 비정상적인 형태의 벤치가 붙어 있음.",
    "entities": "이현우(검은 머리, 뒷모습, 흰색 환자복)와 꺼진 빈 모니터 모두 프롬프트와 일치함.",
    "hard_violations": [
     "[gemini-pro] physically impossible staging (지지대 없이 침대 측면에 허공으로 튀어나온 벤치 구조물)",
     "[gpt-high] 참조에서 확정되지 않은 글자와 줄이 있는 벽면 문서를 모니터 왼쪽 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
    ],
    "physics": "인물은 의자에 앉아 지탱되나, 침대 옆의 흰색 구조물은 아래를 받치는 지지대가 전혀 없어 물리적으로 불가능함."
   },
   {
    "label": "B",
    "direction": "이현우의 시선과 고개가 상단 좌측의 빈 모니터를 명확히 향하고 있음.",
    "built_space": "왼쪽 책상, 오른쪽 침대로 전체 배치가 반전되었으나, 참조 이미지의 철제 침대 프레임, 발치 벤치, 우측 수납장의 형태를 충실히 반영함.",
    "entities": "이현우(뒷모습, 흰색 환자복)와 빈 화면의 모니터 모두 지시사항에 부합함.",
    "hard_violations": [
     "[gpt-high] 참조에서 확정되지 않은 글자와 표가 있는 벽면 문서를 모니터 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
    ],
    "physics": "인물은 의자에 안정적으로 앉아 있으며, 모니터와 침대 등 모든 사물이 정상적인 지지면 위에 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "요구된 카메라 구도와 인물의 자세는 따랐으나, 지지대 없이 허공에 뜬 침대 측면 구조물로 인해 물리적 불가능성 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "방의 좌우 배치가 다소 달라졌으나, 참조 이미지의 가구 디테일을 잘 살렸고 요구된 하향 와이드 샷과 모니터를 향한 인물의 배치를 안정적으로 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 고개가 좌측 상단의 꺼진 모니터를 향해 있음.",
        "built_space": "왼쪽에 책상, 오른쪽에 침대가 배치됨. 참조 이미지의 공간 배치가 왜곡되었고, 침대 측면에 비정상적인 형태의 벤치가 붙어 있음.",
        "entities": "이현우(검은 머리, 뒷모습, 흰색 환자복)와 꺼진 빈 모니터 모두 프롬프트와 일치함.",
        "hard_violations": [
         "physically impossible staging (지지대 없이 침대 측면에 허공으로 튀어나온 벤치 구조물)"
        ],
        "physics": "인물은 의자에 앉아 지탱되나, 침대 옆의 흰색 구조물은 아래를 받치는 지지대가 전혀 없어 물리적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 고개가 상단 좌측의 빈 모니터를 명확히 향하고 있음.",
        "built_space": "왼쪽 책상, 오른쪽 침대로 전체 배치가 반전되었으나, 참조 이미지의 철제 침대 프레임, 발치 벤치, 우측 수납장의 형태를 충실히 반영함.",
        "entities": "이현우(뒷모습, 흰색 환자복)와 빈 화면의 모니터 모두 지시사항에 부합함.",
        "hard_violations": [],
        "physics": "인물은 의자에 안정적으로 앉아 있으며, 모니터와 침대 등 모든 사물이 정상적인 지지면 위에 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "요구된 카메라 구도와 인물의 자세는 따랐으나, 지지대 없이 허공에 뜬 침대 측면 구조물로 인해 물리적 불가능성 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "방의 좌우 배치가 다소 달라졌으나, 참조 이미지의 가구 디테일을 잘 살렸고 요구된 하향 와이드 샷과 모니터를 향한 인물의 배치를 안정적으로 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 고개가 좌측 상단의 꺼진 모니터를 향해 있음.",
        "built_space": "왼쪽에 책상, 오른쪽에 침대가 배치됨. 참조 이미지의 공간 배치가 왜곡되었고, 침대 측면에 비정상적인 형태의 벤치가 붙어 있음.",
        "entities": "이현우(검은 머리, 뒷모습, 흰색 환자복)와 꺼진 빈 모니터 모두 프롬프트와 일치함.",
        "hard_violations": [
         "physically impossible staging (지지대 없이 침대 측면에 허공으로 튀어나온 벤치 구조물)"
        ],
        "physics": "인물은 의자에 앉아 지탱되나, 침대 옆의 흰색 구조물은 아래를 받치는 지지대가 전혀 없어 물리적으로 불가능함."
       },
       {
        "label": "B",
        "direction": "이현우의 시선과 고개가 상단 좌측의 빈 모니터를 명확히 향하고 있음.",
        "built_space": "왼쪽 책상, 오른쪽 침대로 전체 배치가 반전되었으나, 참조 이미지의 철제 침대 프레임, 발치 벤치, 우측 수납장의 형태를 충실히 반영함.",
        "entities": "이현우(뒷모습, 흰색 환자복)와 빈 화면의 모니터 모두 지시사항에 부합함.",
        "hard_violations": [],
        "physics": "인물은 의자에 안정적으로 앉아 있으며, 모니터와 침대 등 모든 사물이 정상적인 지지면 위에 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "뒤쪽 오른편의 높은 와이드 시점, 좌상단의 꺼진 화면, 창가를 향한 침대 배치가 더 충실하지만, 참조에 없는 벽면 문서를 추가했다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "모니터를 보는 뒷모습과 낮의 고독한 공간은 구현했으나, 침대의 방향이 참조와 달라지고 승인되지 않은 벽면 문서까지 추가됐다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 머리는 왼쪽 앞의 꺼진 모니터를 향하며 침대 쪽으로 돌아서지 않았다. 눈은 뒷모습에 가려 직접 확인할 수 없다. 모니터의 표시 면은 인물 쪽을 향하고, 뒤쪽의 높은 카메라에서도 비스듬히 보이는 방향이다.",
        "built_space": "왼쪽 책상에 모니터 한 대, 중앙 아래에 의자 한 개, 뒤쪽에 침대 한 개, 오른쪽 뒤에 수납장 한 개, 뒤 벽에 두 구획의 창 하나가 보인다. 침대는 머리맡이 왼쪽 벽을 향하고 발치가 오른쪽으로 뻗어, 참조에서 창가 쪽 머리맡에서 전경으로 이어지는 배치와 다르다. 흰 벽, 회색 바닥, 바다를 보는 창과 낮의 주변광은 이어진다. 등은 화면 중앙 아래, 모니터는 좌상단에 놓인다.",
        "entities": "검고 헝클어진 짧은 머리와 흰 반소매 환자복 상하의를 입은 남성 한 명만 있다. 드러난 체격과 머리는 이현우의 참조와 대체로 맞지만 얼굴이 보이지 않아 정확한 나이와 얼굴 동일성은 확인할 수 없다. 검은 모니터에는 통화 상대나 영상이 없다. 빈 침대와 흰 침구가 있으며, 책상 조명·책·필기구와 벽면 표 형식 문서도 보인다.",
        "hard_violations": [
         "참조에서 확정되지 않은 글자와 표가 있는 벽면 문서를 모니터 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
        ],
        "physics": "인물의 엉덩이는 의자 좌판에 놓이고 등 뒤에는 등받이가 있다. 다리는 아래로 내려가며 발의 접지는 화면 밖이라 확인할 수 없다. 팔은 무릎 가까이에 자연스럽게 내려와 있고 몸이 공중에 떠 있지는 않다. 모니터는 받침대로 책상에, 침대는 바퀴 달린 프레임으로 바닥에 지지된다."
       },
       {
        "label": "B",
        "direction": "인물의 머리와 상체는 왼쪽 앞 모니터를 향한다. 눈 자체는 보이지 않지만 침대로 몸을 돌리기 전의 방향이 읽힌다. 꺼진 표시 면은 인물을 향하면서 높은 후방 카메라에도 보이므로 사용 방향이 맞다.",
        "built_space": "왼쪽 책상 위 모니터 한 대, 중앙 아래 의자 한 개, 뒤쪽 침대 한 개, 침대 오른쪽 이동식 상판 한 개, 창 오른쪽 수납장 한 개가 보인다. 침대 머리맡은 창가에 있고 발치는 카메라 쪽으로 이어져 참조의 배치와 더 잘 맞는다. 두 구획의 창, 바다 전망, 흰 벽과 회색 바닥도 유지된다. 카메라는 머리보다 높은 뒤쪽에 있고, 등은 중앙 아래, 모니터는 좌상단에 있으며 오른쪽 빈 바닥이 넓게 남는다.",
        "entities": "흰 환자복 상하의와 흰 신발을 착용한 검은 머리의 남성 한 명만 보인다. 머리와 체격은 참조에 부합하며, 얼굴을 억지로 드러내지 않았다. 뒷모습이므로 얼굴의 정확한 동일성이나 한국계 미국인이라는 정체성을 외형만으로 확정할 수 없다. 모니터는 비어 있고 통화 상대는 없다. 빈 침대와 침구, 수납장이 있으며 모니터 위쪽 벽에는 작은 문서와 사진이 추가되어 있다.",
        "hard_violations": [
         "참조에서 확정되지 않은 글자와 줄이 있는 벽면 문서를 모니터 왼쪽 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
        ],
        "physics": "인물은 의자 좌판에 앉아 있으며 등받이는 몸 뒤에 정상적으로 놓인다. 보이는 왼발은 바닥에 닿고, 굽힌 무릎과 내려놓은 팔이 앉은 자세에 맞는다. 모니터는 책상 위 받침대가, 침대와 이동식 상판은 각각 바닥에 닿는 프레임과 바퀴가 지지한다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "뒤쪽 오른편의 높은 와이드 시점, 좌상단의 꺼진 화면, 창가를 향한 침대 배치가 더 충실하지만, 참조에 없는 벽면 문서를 추가했다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "모니터를 보는 뒷모습과 낮의 고독한 공간은 구현했으나, 침대의 방향이 참조와 달라지고 승인되지 않은 벽면 문서까지 추가됐다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 머리는 왼쪽 앞의 꺼진 모니터를 향하며 침대 쪽으로 돌아서지 않았다. 눈은 뒷모습에 가려 직접 확인할 수 없다. 모니터의 표시 면은 인물 쪽을 향하고, 뒤쪽의 높은 카메라에서도 비스듬히 보이는 방향이다.",
        "built_space": "왼쪽 책상에 모니터 한 대, 중앙 아래에 의자 한 개, 뒤쪽에 침대 한 개, 오른쪽 뒤에 수납장 한 개, 뒤 벽에 두 구획의 창 하나가 보인다. 침대는 머리맡이 왼쪽 벽을 향하고 발치가 오른쪽으로 뻗어, 참조에서 창가 쪽 머리맡에서 전경으로 이어지는 배치와 다르다. 흰 벽, 회색 바닥, 바다를 보는 창과 낮의 주변광은 이어진다. 등은 화면 중앙 아래, 모니터는 좌상단에 놓인다.",
        "entities": "검고 헝클어진 짧은 머리와 흰 반소매 환자복 상하의를 입은 남성 한 명만 있다. 드러난 체격과 머리는 이현우의 참조와 대체로 맞지만 얼굴이 보이지 않아 정확한 나이와 얼굴 동일성은 확인할 수 없다. 검은 모니터에는 통화 상대나 영상이 없다. 빈 침대와 흰 침구가 있으며, 책상 조명·책·필기구와 벽면 표 형식 문서도 보인다.",
        "hard_violations": [
         "참조에서 확정되지 않은 글자와 표가 있는 벽면 문서를 모니터 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
        ],
        "physics": "인물의 엉덩이는 의자 좌판에 놓이고 등 뒤에는 등받이가 있다. 다리는 아래로 내려가며 발의 접지는 화면 밖이라 확인할 수 없다. 팔은 무릎 가까이에 자연스럽게 내려와 있고 몸이 공중에 떠 있지는 않다. 모니터는 받침대로 책상에, 침대는 바퀴 달린 프레임으로 바닥에 지지된다."
       },
       {
        "label": "A",
        "direction": "인물의 머리와 상체는 왼쪽 앞 모니터를 향한다. 눈 자체는 보이지 않지만 침대로 몸을 돌리기 전의 방향이 읽힌다. 꺼진 표시 면은 인물을 향하면서 높은 후방 카메라에도 보이므로 사용 방향이 맞다.",
        "built_space": "왼쪽 책상 위 모니터 한 대, 중앙 아래 의자 한 개, 뒤쪽 침대 한 개, 침대 오른쪽 이동식 상판 한 개, 창 오른쪽 수납장 한 개가 보인다. 침대 머리맡은 창가에 있고 발치는 카메라 쪽으로 이어져 참조의 배치와 더 잘 맞는다. 두 구획의 창, 바다 전망, 흰 벽과 회색 바닥도 유지된다. 카메라는 머리보다 높은 뒤쪽에 있고, 등은 중앙 아래, 모니터는 좌상단에 있으며 오른쪽 빈 바닥이 넓게 남는다.",
        "entities": "흰 환자복 상하의와 흰 신발을 착용한 검은 머리의 남성 한 명만 보인다. 머리와 체격은 참조에 부합하며, 얼굴을 억지로 드러내지 않았다. 뒷모습이므로 얼굴의 정확한 동일성이나 한국계 미국인이라는 정체성을 외형만으로 확정할 수 없다. 모니터는 비어 있고 통화 상대는 없다. 빈 침대와 침구, 수납장이 있으며 모니터 위쪽 벽에는 작은 문서와 사진이 추가되어 있다.",
        "hard_violations": [
         "참조에서 확정되지 않은 글자와 줄이 있는 벽면 문서를 모니터 왼쪽 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
        ],
        "physics": "인물은 의자 좌판에 앉아 있으며 등받이는 몸 뒤에 정상적으로 놓인다. 보이는 왼발은 바닥에 닿고, 굽힌 무릎과 내려놓은 팔이 앉은 자세에 맞는다. 모니터는 책상 위 받침대가, 침대와 이동식 상판은 각각 바닥에 닿는 프레임과 바퀴가 지지한다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.321,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible staging (지지대 없이 침대 측면에 허공으로 튀어나온 벤치 구조물)",
     "[gpt-high] 참조에서 확정되지 않은 글자와 줄이 있는 벽면 문서를 모니터 왼쪽 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
    ],
    "B": [
     "[gpt-high] 참조에서 확정되지 않은 글자와 표가 있는 벽면 문서를 모니터 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1321,
   "B": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "요구된 카메라 구도와 인물의 자세는 따랐으나, 지지대 없이 허공에 뜬 침대 측면 구조물로 인해 물리적 불가능성 오류가 발생했습니다.  ★위반: [gemini-pro] physically impossible staging (지지대 없이 침대 측면에 허공으로 튀어나온 벤치 구조물) / [gpt-high] 참조에서 확정되지 않은 글자와 줄이 있는 벽면 문서를 모니터 왼쪽 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "방의 좌우 배치가 다소 달라졌으나, 참조 이미지의 가구 디테일을 잘 살렸고 요구된 하향 와이드 샷과 모니터를 향한 인물의 배치를 안정적으로 구현했습니다.  ★위반: [gpt-high] 참조에서 확정되지 않은 글자와 표가 있는 벽면 문서를 모니터 위에 새로 추가해, 새로운 문구를 만들지 말라는 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S83sh9_sel.png",
    "asset_id": "6d0dd50e-254d-47e1-9456-f2a7beb0813a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e63-47a7-7c39-80e6-a04fc9a3528f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S83sh9"
  }
 },
 "S84sh1::signage": {
  "fp": "c51ca1b47fd01057",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S84sh1": {
  "input_fingerprint": "8b3687de438b6036",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 안 실험대 위에 누워 있는 찰리의 낡은 얼굴 디스플레이에 전원 불빛이 번쩍 켜진 근접 찰나.\n\nLOCATION (lock): On the examination bed inside the main laboratory's glass enclosure. The robot's face display lights up in the nighttime laboratory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the opening camera position inside the enclosure, close above 찰리's reclining face and offset beside the experiment table, looking diagonally downward. His worn face display sits just left of center with the upper torso and attached wires along the lower edge, retaining enough table around him to establish that he is lying down. Observe the display physically powering on as his newly active gaze remains upward; delay the crane's rising retreat until he begins to sit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Experiment table (Supports the reclining 찰리) — A narrow section of its upper surface surrounds his head and shoulders; used as Anchors the close view in a horizontal reclining position; Attached wires (Still connected to 찰리 before he removes them); used as Small visible runs near the lower edge establish the procedure without competing with the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The face display's brief power-on flash punctuates restrained ambient illumination without adding a specified hue or another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is still lying on the laboratory bed inside the glass enclosure, with multiple wires attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 안 실험대 위에 누워 있는 찰리의 낡은 얼굴 디스플레이에 전원 불빛이 번쩍 켜진 근접 찰나.\n\nLOCATION (lock): On the examination bed inside the main laboratory's glass enclosure. The robot's face display lights up in the nighttime laboratory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the opening camera position inside the enclosure, close above 찰리's reclining face and offset beside the experiment table, looking diagonally downward. His worn face display sits just left of center with the upper torso and attached wires along the lower edge, retaining enough table around him to establish that he is lying down. Observe the display physically powering on as his newly active gaze remains upward; delay the crane's rising retreat until he begins to sit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Experiment table (Supports the reclining 찰리) — A narrow section of its upper surface surrounds his head and shoulders; used as Anchors the close view in a horizontal reclining position; Attached wires (Still connected to 찰리 before he removes them); used as Small visible runs near the lower edge establish the procedure without competing with the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The face display's brief power-on flash punctuates restrained ambient illumination without adding a specified hue or another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is still lying on the laboratory bed inside the glass enclosure, with multiple wires attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 유리 벽 안 실험대 위에 누워 있는 찰리의 낡은 얼굴 디스플레이에 전원 불빛이 번쩍 켜진 근접 찰나.\n\nLOCATION (lock): On the examination bed inside the main laboratory's glass enclosure. The robot's face display lights up in the nighttime laboratory. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold the opening camera position inside the enclosure, close above 찰리's reclining face and offset beside the experiment table, looking diagonally downward. His worn face display sits just left of center with the upper torso and attached wires along the lower edge, retaining enough table around him to establish that he is lying down. Observe the display physically powering on as his newly active gaze remains upward; delay the crane's rising retreat until he begins to sit.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Experiment table (Supports the reclining 찰리) — A narrow section of its upper surface surrounds his head and shoulders; used as Anchors the close view in a horizontal reclining position; Attached wires (Still connected to 찰리 before he removes them); used as Small visible runs near the lower edge establish the procedure without competing with the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The face display's brief power-on flash punctuates restrained ambient illumination without adding a specified hue or another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the cold stainless-steel bed inside the circular glass enclosure, his damaged body inactive during examination and with wires attached to it. The bed supports his body, but the exact orientation of his head and torso and the placement of his arms and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is still lying on the laboratory bed inside the glass enclosure, with multiple wires attached across his body. His power has just returned; the wires have not yet been pulled off.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선은 위를 향함.",
    "built_space": "금속 재질의 실험대 위에 누워 있으며 배경에 흐릿한 실험실 환경이 보임.",
    "entities": "찰리의 외형(흰색 마스크, 노란색 빛나는 눈, 샌드 베이지 장갑, 가슴의 파란색 원자로)과 연결된 전선들이 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "찰리의 신체는 금속 실험대 위에 안정적으로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "찰리의 시선은 위를 향함.",
    "built_space": "가운데에 검은색 패드가 깔린 침대 위에 누워 있으며 뒤로 실험실 기기가 보임.",
    "entities": "찰리의 외형과 연결된 전선은 대체로 일치하나, 실험대 표면(검은 패드)이 레퍼런스의 전체 금속 형태와 다름.",
    "hard_violations": [],
    "physics": "찰리의 신체는 침대 표면에 안정적으로 지지되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 근접 구도와 대각선 앵글을 잘 따랐으며, 이전 샷에 명시된 차가운 금속 실험대 재질을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터 구현과 앵글은 무난하나, 이전 샷에서 고정된 금속 실험대 대신 검은색 패드가 깔린 침대를 묘사하여 장소 일관성(Location Lock)을 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선은 위를 향함.",
        "built_space": "금속 재질의 실험대 위에 누워 있으며 배경에 흐릿한 실험실 환경이 보임.",
        "entities": "찰리의 외형(흰색 마스크, 노란색 빛나는 눈, 샌드 베이지 장갑, 가슴의 파란색 원자로)과 연결된 전선들이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "찰리의 신체는 금속 실험대 위에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선은 위를 향함.",
        "built_space": "가운데에 검은색 패드가 깔린 침대 위에 누워 있으며 뒤로 실험실 기기가 보임.",
        "entities": "찰리의 외형과 연결된 전선은 대체로 일치하나, 실험대 표면(검은 패드)이 레퍼런스의 전체 금속 형태와 다름.",
        "hard_violations": [],
        "physics": "찰리의 신체는 침대 표면에 안정적으로 지지되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 근접 구도와 대각선 앵글을 잘 따랐으며, 이전 샷에 명시된 차가운 금속 실험대 재질을 충실히 반영했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "캐릭터 구현과 앵글은 무난하나, 이전 샷에서 고정된 금속 실험대 대신 검은색 패드가 깔린 침대를 묘사하여 장소 일관성(Location Lock)을 어겼습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선은 위를 향함.",
        "built_space": "금속 재질의 실험대 위에 누워 있으며 배경에 흐릿한 실험실 환경이 보임.",
        "entities": "찰리의 외형(흰색 마스크, 노란색 빛나는 눈, 샌드 베이지 장갑, 가슴의 파란색 원자로)과 연결된 전선들이 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "찰리의 신체는 금속 실험대 위에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "찰리의 시선은 위를 향함.",
        "built_space": "가운데에 검은색 패드가 깔린 침대 위에 누워 있으며 뒤로 실험실 기기가 보임.",
        "entities": "찰리의 외형과 연결된 전선은 대체로 일치하나, 실험대 표면(검은 패드)이 레퍼런스의 전체 금속 형태와 다름.",
        "hard_violations": [],
        "physics": "찰리의 신체는 침대 표면에 안정적으로 지지되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 왼쪽 중심의 사선 근접 구도와 위를 향한 시선은 맞지만, 어두운 실험대 받침면이 스테인리스 장소 설정에서 벗어나고 전원 점등 순간의 강조가 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "스테인리스 실험대에 누운 찰리를 비스듬히 내려다보며 얼굴의 점등과 연결된 전선을 보여 주어 구도·장소·행동을 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 흰 얼굴과 두 발광 눈은 누운 몸에서 위쪽, 높은 카메라가 있는 방향을 향한다. 다른 인물이나 겨냥하는 도구는 없으며, 상체를 일으키는 움직임도 보이지 않는다.",
        "built_space": "실험대 한 개의 금속 테두리와 어두운 중앙 받침면이 머리와 어깨 뒤에 보인다. 오른쪽 위에는 흐릿한 장비 일부, 뒤에는 유리벽으로 읽히는 경계가 있다. 얼굴은 화면 중심보다 조금 왼쪽이고 상체와 전선이 아래쪽을 채우지만, 상체가 하단의 좁은 띠보다 크게 나온다. 어두운 받침면은 지정된 차가운 스테인리스 상판과 차이가 있다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 대만 보인다. 낡은 샌드 베이지 장갑판, 흰 마스크형 얼굴, 둥근 눈 두 개와 검은 입 선, 푸른 원형 흉부 원자로가 참조와 부합한다. 인간 피부나 눈은 추가되지 않았다. 여러 전선이 어깨와 흉부 주변에 연결되어 있다. 다리와 팔 전체의 비율은 이 근접 프레임에서 확인할 수 없다. 눈의 점등은 선명하지만 얼굴 전체가 막 켜지는 섬광보다는 지속 발광으로도 읽힌다.",
        "hard_violations": [],
        "physics": "머리 뒤와 어깨·상체는 실험대 받침면에 기대어 있으며 목은 기계 관절로 몸통에 연결되어 있다. 보이는 팔 부분도 몸 옆으로 내려가 있다. 전선은 연결 단자와 몸체·실험대에 걸쳐 지지된다. 공중에 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "누운 찰리의 얼굴과 두 발광 눈이 위쪽을 향하며, 카메라는 그 얼굴을 옆 위에서 비스듬히 내려다본다. 특정 사람을 바라보거나 몸을 일으키는 모습은 없어서 전원만 돌아온 순간의 방향 관계와 맞는다.",
        "built_space": "실험대 한 개의 반사되는 금속 상판이 머리와 어깨 주변에 드러난다. 머리 뒤에는 수직 금속 지지대 하나와 그 기부가 있고, 왼쪽 가장자리에는 별도의 고정구 일부가 보인다. 뒤쪽 유리 경계와 오른쪽의 흐릿한 의료 장비는 실험실 공간을 유지한다. 얼굴은 중심 왼쪽에 놓이고 연결 전선과 상체는 아래쪽에 배치된다. 다만 흉부가 하단에서 차지하는 면적은 지시보다 다소 크다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "찰리 한 대 외에 사람은 없다. 샌드 베이지의 긁힌 장갑판, 흰 얼굴판, 원형 눈 두 개, 입 선 하나, 푸른 에너지의 흉부 원자로가 참조 정체성을 유지한다. 눈과 입 틈의 밝은 빛이 얼굴의 전원 복귀를 강조한다. 입까지 황색으로 빛나는 표현은 참조의 검은 입 선과 조금 다르다. 어깨와 흉부에 여러 전선이 그대로 연결되어 있으며, 프레임 밖 하체의 비율은 평가할 수 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽은 금속 상판의 낮은 받침 부근에 놓이고, 어깨와 등은 실험대에 지지되어 누운 상태로 읽힌다. 보이는 팔은 몸 옆에 내려놓여 있으며 들고 있는 물체는 없다. 전선은 단자에 결합된 채 몸체와 상판 위로 처져 있어 지지가 설명된다. 지지 없이 떠 있는 부분은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 왼쪽 중심의 사선 근접 구도와 위를 향한 시선은 맞지만, 어두운 실험대 받침면이 스테인리스 장소 설정에서 벗어나고 전원 점등 순간의 강조가 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "스테인리스 실험대에 누운 찰리를 비스듬히 내려다보며 얼굴의 점등과 연결된 전선을 보여 주어 구도·장소·행동을 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 흰 얼굴과 두 발광 눈은 누운 몸에서 위쪽, 높은 카메라가 있는 방향을 향한다. 다른 인물이나 겨냥하는 도구는 없으며, 상체를 일으키는 움직임도 보이지 않는다.",
        "built_space": "실험대 한 개의 금속 테두리와 어두운 중앙 받침면이 머리와 어깨 뒤에 보인다. 오른쪽 위에는 흐릿한 장비 일부, 뒤에는 유리벽으로 읽히는 경계가 있다. 얼굴은 화면 중심보다 조금 왼쪽이고 상체와 전선이 아래쪽을 채우지만, 상체가 하단의 좁은 띠보다 크게 나온다. 어두운 받침면은 지정된 차가운 스테인리스 상판과 차이가 있다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "찰리 한 대만 보인다. 낡은 샌드 베이지 장갑판, 흰 마스크형 얼굴, 둥근 눈 두 개와 검은 입 선, 푸른 원형 흉부 원자로가 참조와 부합한다. 인간 피부나 눈은 추가되지 않았다. 여러 전선이 어깨와 흉부 주변에 연결되어 있다. 다리와 팔 전체의 비율은 이 근접 프레임에서 확인할 수 없다. 눈의 점등은 선명하지만 얼굴 전체가 막 켜지는 섬광보다는 지속 발광으로도 읽힌다.",
        "hard_violations": [],
        "physics": "머리 뒤와 어깨·상체는 실험대 받침면에 기대어 있으며 목은 기계 관절로 몸통에 연결되어 있다. 보이는 팔 부분도 몸 옆으로 내려가 있다. 전선은 연결 단자와 몸체·실험대에 걸쳐 지지된다. 공중에 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "누운 찰리의 얼굴과 두 발광 눈이 위쪽을 향하며, 카메라는 그 얼굴을 옆 위에서 비스듬히 내려다본다. 특정 사람을 바라보거나 몸을 일으키는 모습은 없어서 전원만 돌아온 순간의 방향 관계와 맞는다.",
        "built_space": "실험대 한 개의 반사되는 금속 상판이 머리와 어깨 주변에 드러난다. 머리 뒤에는 수직 금속 지지대 하나와 그 기부가 있고, 왼쪽 가장자리에는 별도의 고정구 일부가 보인다. 뒤쪽 유리 경계와 오른쪽의 흐릿한 의료 장비는 실험실 공간을 유지한다. 얼굴은 중심 왼쪽에 놓이고 연결 전선과 상체는 아래쪽에 배치된다. 다만 흉부가 하단에서 차지하는 면적은 지시보다 다소 크다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "찰리 한 대 외에 사람은 없다. 샌드 베이지의 긁힌 장갑판, 흰 얼굴판, 원형 눈 두 개, 입 선 하나, 푸른 에너지의 흉부 원자로가 참조 정체성을 유지한다. 눈과 입 틈의 밝은 빛이 얼굴의 전원 복귀를 강조한다. 입까지 황색으로 빛나는 표현은 참조의 검은 입 선과 조금 다르다. 어깨와 흉부에 여러 전선이 그대로 연결되어 있으며, 프레임 밖 하체의 비율은 평가할 수 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽은 금속 상판의 낮은 받침 부근에 놓이고, 어깨와 등은 실험대에 지지되어 누운 상태로 읽힌다. 보이는 팔은 몸 옆에 내려놓여 있으며 들고 있는 물체는 없다. 전선은 단자에 결합된 채 몸체와 상판 위로 처져 있어 지지가 설명된다. 지지 없이 떠 있는 부분은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.603
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.603
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1603
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 근접 구도와 대각선 앵글을 잘 따랐으며, 이전 샷에 명시된 차가운 금속 실험대 재질을 충실히 반영했습니다."
   },
   {
    "label": "B",
    "score": 1603,
    "verdict_ko": "캐릭터 구현과 앵글은 무난하나, 이전 샷에서 고정된 금속 실험대 대신 검은색 패드가 깔린 침대를 묘사하여 장소 일관성(Location Lock)을 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh13_sel.png",
    "asset_id": "e48c1896-1f82-4804-a304-ec35673941ca",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e6a-b6d9-7c5a-9d09-194abc324d78",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh13"
  },
  "locked_char_refs_kept_body_identity": [
   "찰리(C01)"
  ]
 },
 "S84sh7::signage": {
  "fp": "8cd8211a43591534",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::director_piano_office": {
  "input_fingerprint": "bcadc37cf1fda0c1",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "director_piano_office",
    "tags": [
     "S84sh13",
     "S84sh7",
     "S86sh2",
     "S86sh7"
    ]
   },
   "context_sig": "05a54e3d3f416b2f"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림) / 제주도 연구소 소장실: 모던하고 깔끔한 인테리어에 낡은 피아노가 놓인 지소영의 개인 공간이다. (특징: 최신 가구 사이에 놓인 낡은 목재 피아노; 건반을 두드리는 소영의 뒷모습과 숨어보는 찰리; 책상 위에 놓인 지동현과 소영의 어린 시절 사진 액자)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 센터와 연결된 소장실\n- 깔끔한 인테리어, 모던한 공간.\n중앙 홀에 놓여있는 낡은 피아노.\n- 86. 메인 센터 소장실 - N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림) / 제주도 연구소 소장실: 모던하고 깔끔한 인테리어에 낡은 피아노가 놓인 지소영의 개인 공간이다. (특징: 최신 가구 사이에 놓인 낡은 목재 피아노; 건반을 두드리는 소영의 뒷모습과 숨어보는 찰리; 책상 위에 놓인 지동현과 소영의 어린 시절 사진 액자)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 센터와 연결된 소장실\n- 깔끔한 인테리어, 모던한 공간.\n중앙 홀에 놓여있는 낡은 피아노.\n- 86. 메인 센터 소장실 - N\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_director_piano_office_444ff1.png",
  "asset_id": "e7a34288-7811-4fad-a19d-4dc8f20e85e7",
  "input_asset_ids": [
   "86142bee-74fb-4d76-88b7-ddec73c55c14"
  ],
  "origin_tag": "S84sh7",
  "place_text": "At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.",
  "origin_inputs": {
   "place_text": "At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.",
   "time_of_day_en": "night",
   "conti_asset_id": "86142bee-74fb-4d76-88b7-ddec73c55c14"
  }
 },
 "S84sh7::bgfirst_bg": {
  "input_fingerprint": "5e6b2bb0d1174c26",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모던한 중앙 홀 한가운데 놓인 낡은 피아노 건반 위를 지소영의 손가락이 짚고 있는 전신 구도.\n\nLOCATION (lock): At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a neutral observing position beside the pillar at 지소영's seated head height, looking diagonally across the keyboard rather than assuming 찰리's literal viewpoint. Show her full seated figure on the right, leaning slightly into her fingers on the keys, with the old piano extending toward center-left and occupying less than two-fifths of the frame. Confine the pillar to a narrow left edge and let her downward attention remain on the keyboard before she notices the visitor.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영 in the central hall) — Keyboard and adjacent side are visible obliquely, leaving her hands and seated figure unobstructed; used as Connects her absorbed posture to the source of the familiar melody; Pillar (Present beside the camera's office-side position) — Only a narrow side section is visible at the left edge; used as Provides a restrained foreground boundary for the reveal; Director's office central hall (Clean, modern interior surrounding the old piano); used as Negative space preserves the contrast between the modern room and the familiar instrument.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and measured contrast allow the performance to feel intimate within the clean modern space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모던한 중앙 홀 한가운데 놓인 낡은 피아노 건반 위를 지소영의 손가락이 짚고 있는 전신 구도.\n\nLOCATION (lock): At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a neutral observing position beside the pillar at 지소영's seated head height, looking diagonally across the keyboard rather than assuming 찰리's literal viewpoint. Show her full seated figure on the right, leaning slightly into her fingers on the keys, with the old piano extending toward center-left and occupying less than two-fifths of the frame. Confine the pillar to a narrow left edge and let her downward attention remain on the keyboard before she notices the visitor.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영 in the central hall) — Keyboard and adjacent side are visible obliquely, leaving her hands and seated figure unobstructed; used as Connects her absorbed posture to the source of the familiar melody; Pillar (Present beside the camera's office-side position) — Only a narrow side section is visible at the left edge; used as Provides a restrained foreground boundary for the reveal; Director's office central hall (Clean, modern interior surrounding the old piano); used as Negative space preserves the contrast between the modern room and the familiar instrument.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and measured contrast allow the performance to feel intimate within the clean modern space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh7__bgfirst_bg.png",
  "asset_id": "04610f06-54fc-4285-b5fa-2618df7a7122",
  "input_asset_ids": [
   "86142bee-74fb-4d76-88b7-ddec73c55c14",
   "e7a34288-7811-4fad-a19d-4dc8f20e85e7"
  ]
 },
 "S84sh7": {
  "input_fingerprint": "9564c51bb7cec3c5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모던한 중앙 홀 한가운데 놓인 낡은 피아노 건반 위를 지소영의 손가락이 짚고 있는 전신 구도.\n\nLOCATION (lock): At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a neutral observing position beside the pillar at 지소영's seated head height, looking diagonally across the keyboard rather than assuming 찰리's literal viewpoint. Show her full seated figure on the right, leaning slightly into her fingers on the keys, with the old piano extending toward center-left and occupying less than two-fifths of the frame. Confine the pillar to a narrow left edge and let her downward attention remain on the keyboard before she notices the visitor.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영 in the central hall) — Keyboard and adjacent side are visible obliquely, leaving her hands and seated figure unobstructed; used as Connects her absorbed posture to the source of the familiar melody; Pillar (Present beside the camera's office-side position) — Only a narrow side section is visible at the left edge; used as Provides a restrained foreground boundary for the reveal; Director's office central hall (Clean, modern interior surrounding the old piano); used as Negative space preserves the contrast between the modern room and the familiar instrument.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and measured contrast allow the performance to feel intimate within the clean modern space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office is clean and modern, with an old piano in its central hall. Charlie is awake and mobile after removing the attached wires and opening the glass enclosure's door. 지소영: She is at the piano, playing the familiar melody.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 지소영 right now, so 지소영's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 지소영: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모던한 중앙 홀 한가운데 놓인 낡은 피아노 건반 위를 지소영의 손가락이 짚고 있는 전신 구도.\n\nLOCATION (lock): At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a neutral observing position beside the pillar at 지소영's seated head height, looking diagonally across the keyboard rather than assuming 찰리's literal viewpoint. Show her full seated figure on the right, leaning slightly into her fingers on the keys, with the old piano extending toward center-left and occupying less than two-fifths of the frame. Confine the pillar to a narrow left edge and let her downward attention remain on the keyboard before she notices the visitor.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영 in the central hall) — Keyboard and adjacent side are visible obliquely, leaving her hands and seated figure unobstructed; used as Connects her absorbed posture to the source of the familiar melody; Pillar (Present beside the camera's office-side position) — Only a narrow side section is visible at the left edge; used as Provides a restrained foreground boundary for the reveal; Director's office central hall (Clean, modern interior surrounding the old piano); used as Negative space preserves the contrast between the modern room and the familiar instrument.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and measured contrast allow the performance to feel intimate within the clean modern space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office is clean and modern, with an old piano in its central hall. Charlie is awake and mobile after removing the attached wires and opening the glass enclosure's door. 지소영: She is at the piano, playing the familiar melody.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 지소영 right now, so 지소영's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 지소영: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모던한 중앙 홀 한가운데 놓인 낡은 피아노 건반 위를 지소영의 손가락이 짚고 있는 전신 구도.\n\nLOCATION (lock): At the old piano in the central hall of the director's suite adjoining the research center. The modern interior is illuminated for nighttime use. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a neutral observing position beside the pillar at 지소영's seated head height, looking diagonally across the keyboard rather than assuming 찰리's literal viewpoint. Show her full seated figure on the right, leaning slightly into her fingers on the keys, with the old piano extending toward center-left and occupying less than two-fifths of the frame. Confine the pillar to a narrow left edge and let her downward attention remain on the keyboard before she notices the visitor.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Old piano (Being played by 지소영 in the central hall) — Keyboard and adjacent side are visible obliquely, leaving her hands and seated figure unobstructed; used as Connects her absorbed posture to the source of the familiar melody; Pillar (Present beside the camera's office-side position) — Only a narrow side section is visible at the left edge; used as Provides a restrained foreground boundary for the reveal; Director's office central hall (Clean, modern interior surrounding the old piano); used as Negative space preserves the contrast between the modern room and the familiar instrument.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination and measured contrast allow the performance to feel intimate within the clean modern space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office is clean and modern, with an old piano in its central hall. Charlie is awake and mobile after removing the attached wires and opening the glass enclosure's door. 지소영: She is at the piano, playing the familiar melody.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 지소영 right now, so 지소영's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 지소영: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh7__bgfirst_bg.png",
     "asset_id": "04610f06-54fc-4285-b5fa-2618df7a7122",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S84sh7.png",
     "asset_id": "86142bee-74fb-4d76-88b7-ddec73c55c14",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:783266>",
     "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_director_piano_office_444ff1.png",
     "asset_id": "e7a34288-7811-4fad-a19d-4dc8f20e85e7",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:783266>",
     "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선과 얼굴 방향이 아래쪽 피아노 건반을 정확히 향하고 있습니다.",
    "built_space": "레퍼런스와 동일한 현대적 홀 내부이며, 왼쪽 가장자리에 기둥이 올바르게 배치되어 있습니다.",
    "entities": "지소영의 얼굴, 헤어스타일, 연구 가운 및 낡은 피아노의 외형이 지정된 레퍼런스와 잘 일치합니다.",
    "hard_violations": [],
    "physics": "의자에 앉아 양발을 바닥에 두고 손을 건반에 올린 자세가 물리적으로 안정적입니다."
   },
   {
    "label": "B",
    "direction": "시선이 피아노 건반 쪽을 향하고 있습니다.",
    "built_space": "원본 위치의 인테리어를 잘 반영했으며, 프레임 왼쪽의 기둥 배치도 적절합니다.",
    "entities": "지소영의 복장과 피아노가 등장하지만, 인물의 얼굴 디테일이 레퍼런스와 약간의 차이가 있습니다.",
    "hard_violations": [],
    "physics": "의자에 안착하여 건반을 누르고 있으나, 체중이 실린 느낌이 부족하고 다소 뻣뻣하게 보입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 와이드 샷 구도, 왼쪽 기둥의 배치, 그리고 피아노 건반을 향해 약간 몸을 숙인 인물의 자연스러운 자세를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 공간과 전체적인 앵글은 맞추었으나, 인물의 상체 자세가 다소 경직되어 있고 손가락 묘사의 디테일이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 얼굴 방향이 아래쪽 피아노 건반을 정확히 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 현대적 홀 내부이며, 왼쪽 가장자리에 기둥이 올바르게 배치되어 있습니다.",
        "entities": "지소영의 얼굴, 헤어스타일, 연구 가운 및 낡은 피아노의 외형이 지정된 레퍼런스와 잘 일치합니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 양발을 바닥에 두고 손을 건반에 올린 자세가 물리적으로 안정적입니다."
       },
       {
        "label": "B",
        "direction": "시선이 피아노 건반 쪽을 향하고 있습니다.",
        "built_space": "원본 위치의 인테리어를 잘 반영했으며, 프레임 왼쪽의 기둥 배치도 적절합니다.",
        "entities": "지소영의 복장과 피아노가 등장하지만, 인물의 얼굴 디테일이 레퍼런스와 약간의 차이가 있습니다.",
        "hard_violations": [],
        "physics": "의자에 안착하여 건반을 누르고 있으나, 체중이 실린 느낌이 부족하고 다소 뻣뻣하게 보입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 와이드 샷 구도, 왼쪽 기둥의 배치, 그리고 피아노 건반을 향해 약간 몸을 숙인 인물의 자연스러운 자세를 성공적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "지정된 공간과 전체적인 앵글은 맞추었으나, 인물의 상체 자세가 다소 경직되어 있고 손가락 묘사의 디테일이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 얼굴 방향이 아래쪽 피아노 건반을 정확히 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 현대적 홀 내부이며, 왼쪽 가장자리에 기둥이 올바르게 배치되어 있습니다.",
        "entities": "지소영의 얼굴, 헤어스타일, 연구 가운 및 낡은 피아노의 외형이 지정된 레퍼런스와 잘 일치합니다.",
        "hard_violations": [],
        "physics": "의자에 앉아 양발을 바닥에 두고 손을 건반에 올린 자세가 물리적으로 안정적입니다."
       },
       {
        "label": "B",
        "direction": "시선이 피아노 건반 쪽을 향하고 있습니다.",
        "built_space": "원본 위치의 인테리어를 잘 반영했으며, 프레임 왼쪽의 기둥 배치도 적절합니다.",
        "entities": "지소영의 복장과 피아노가 등장하지만, 인물의 얼굴 디테일이 레퍼런스와 약간의 차이가 있습니다.",
        "hard_violations": [],
        "physics": "의자에 안착하여 건반을 누르고 있으나, 체중이 실린 느낌이 부족하고 다소 뻣뻣하게 보입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽의 앉은 전신과 건반을 향한 시선은 맞지만, 피아노가 조금 더 크게 펼쳐지고 연구 가운이 참조보다 훨씬 길어 B보다 충실도가 낮습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "건반을 비스듬히 보는 전신 와이드 구도, 연주에 몰입한 자세, 참조 공간의 배치와 짧은 연구 가운을 더 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "여성은 고개와 눈을 아래쪽 건반 및 자신의 손으로 향하고 있습니다. 양팔이 건반 쪽으로 뻗어 있고 손가락이 건반 위에 놓여 있어 연주 방향이 맞습니다. 카메라나 방문자를 바라보지 않습니다.",
        "built_space": "왼쪽 가장자리에 원형 기둥 하나의 일부가 보이고, 중앙 왼쪽에는 피아노 한 대, 오른쪽에는 연주용 벤치 하나가 있습니다. 뒤쪽 왼편에는 소파 하나, 낮은 탁자 하나, 플로어 조명 하나와 벽 그림 하나가 있으며, 오른편에는 수납장 하나와 꽃병 하나, 벽 그림 하나 및 유리문이 보입니다. 야경을 향한 통유리, 석재 바닥과 간접조명 천장은 참조 장소에 부합합니다. 여성은 건반 앞 벤치에 앉아 있으며 손과 앉은 전신이 보입니다. 피아노의 가로 범위는 화면의 약 41%로, 5분의 2 미만이라는 지시보다 약간 넓습니다. 바닥의 조명과 가구 반사는 가능한 배치입니다.",
        "entities": "중년의 동아시아계 여성 한 명만 보이며, 정돈된 짙은 단발머리와 얼굴은 지소영 참조에 대체로 부합합니다. 정확한 국적과 나이는 외관만으로 확정할 수 없습니다. 흰 연구 가운, 회색 셔츠, 짙은 정장 바지, 검은 구두와 손목시계가 보입니다. 가운은 참조의 엉덩이 부근 길이보다 길어 무릎 가까이 내려옵니다. 낡은 갈색 그랜드피아노와 검은 누빔 벤치는 참조의 종류와 재질에 맞습니다. 손과 손목은 연주자의 팔에 자연스럽게 이어집니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "골반은 벤치 좌판에 지지되고, 굽힌 다리는 바닥과 페달 쪽으로 자연스럽게 내려옵니다. 한 구두는 바닥에 놓이고 다른 구두는 페달 부근에 있습니다. 양손은 건반에 접촉하며 팔꿈치와 손목의 연결도 연주 가능한 자세입니다. 피아노와 벤치는 다리로 바닥에 지지되어 있고 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "여성의 고개와 시선이 아래쪽 건반으로 향하고, 상체가 손 쪽으로 조금 기울어 있습니다. 양손의 손가락은 자신의 앞에 있는 건반을 짚고 있습니다. 방문자를 알아차리기 전 연주에 집중하는 방향과 일치합니다.",
        "built_space": "왼쪽 가장자리의 기둥 일부, 중앙 왼쪽의 피아노 한 대, 오른쪽의 벤치 하나가 보입니다. 뒤쪽 왼편의 소파 하나, 낮은 탁자 하나, 플로어 조명 하나와 그림 하나, 오른편의 수납장 하나와 꽃병 하나, 그림 하나 및 유리문이 참조와 같은 관계로 배치되어 있습니다. 통유리 너머 야경과 따뜻한 천장 간접조명, 광택 있는 석재 바닥도 일치합니다. 여성의 앉은 전신을 오른쪽에 두고 건반과 피아노 측면을 비스듬히 보여 줍니다. 피아노의 가로 범위는 화면의 약 40%이고 주변 홀의 여백이 유지됩니다. 바닥 반사는 물체와 조명의 위치에 부합합니다.",
        "entities": "중년의 동아시아계 여성 한 명이 등장하며, 짙은 단발머리와 얼굴 윤곽이 지소영 참조에 대체로 맞습니다. 정확한 국적과 나이는 외관만으로 확정할 수 없습니다. 흰 연구 가운의 짧은 길이와 가슴 표식, 회색 셔츠, 짙은 정장 바지, 검은 구두가 참조 복장에 가깝습니다. 닳은 갈색 그랜드피아노와 검은 누빔 벤치도 참조에 부합합니다. 건반 위 손은 인물의 소매와 팔에 이어져 있으며 별도 인물의 손처럼 보이지 않습니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "엉덩이가 벤치 좌판 위에 놓이고 상체는 골반을 중심으로 약간 앞으로 기울어 있습니다. 무릎 아래 다리는 바닥 쪽으로 내려오며, 한 발은 바닥에, 다른 발은 페달 위 또는 바로 인접한 위치에 놓입니다. 손가락과 건반의 접촉, 팔과 손목의 자세는 연주 동작으로 가능합니다. 피아노와 벤치의 다리가 바닥에 닿아 있으며 지지 없이 떠 있는 신체나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽의 앉은 전신과 건반을 향한 시선은 맞지만, 피아노가 조금 더 크게 펼쳐지고 연구 가운이 참조보다 훨씬 길어 B보다 충실도가 낮습니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "건반을 비스듬히 보는 전신 와이드 구도, 연주에 몰입한 자세, 참조 공간의 배치와 짧은 연구 가운을 더 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "여성은 고개와 눈을 아래쪽 건반 및 자신의 손으로 향하고 있습니다. 양팔이 건반 쪽으로 뻗어 있고 손가락이 건반 위에 놓여 있어 연주 방향이 맞습니다. 카메라나 방문자를 바라보지 않습니다.",
        "built_space": "왼쪽 가장자리에 원형 기둥 하나의 일부가 보이고, 중앙 왼쪽에는 피아노 한 대, 오른쪽에는 연주용 벤치 하나가 있습니다. 뒤쪽 왼편에는 소파 하나, 낮은 탁자 하나, 플로어 조명 하나와 벽 그림 하나가 있으며, 오른편에는 수납장 하나와 꽃병 하나, 벽 그림 하나 및 유리문이 보입니다. 야경을 향한 통유리, 석재 바닥과 간접조명 천장은 참조 장소에 부합합니다. 여성은 건반 앞 벤치에 앉아 있으며 손과 앉은 전신이 보입니다. 피아노의 가로 범위는 화면의 약 41%로, 5분의 2 미만이라는 지시보다 약간 넓습니다. 바닥의 조명과 가구 반사는 가능한 배치입니다.",
        "entities": "중년의 동아시아계 여성 한 명만 보이며, 정돈된 짙은 단발머리와 얼굴은 지소영 참조에 대체로 부합합니다. 정확한 국적과 나이는 외관만으로 확정할 수 없습니다. 흰 연구 가운, 회색 셔츠, 짙은 정장 바지, 검은 구두와 손목시계가 보입니다. 가운은 참조의 엉덩이 부근 길이보다 길어 무릎 가까이 내려옵니다. 낡은 갈색 그랜드피아노와 검은 누빔 벤치는 참조의 종류와 재질에 맞습니다. 손과 손목은 연주자의 팔에 자연스럽게 이어집니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "골반은 벤치 좌판에 지지되고, 굽힌 다리는 바닥과 페달 쪽으로 자연스럽게 내려옵니다. 한 구두는 바닥에 놓이고 다른 구두는 페달 부근에 있습니다. 양손은 건반에 접촉하며 팔꿈치와 손목의 연결도 연주 가능한 자세입니다. 피아노와 벤치는 다리로 바닥에 지지되어 있고 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "여성의 고개와 시선이 아래쪽 건반으로 향하고, 상체가 손 쪽으로 조금 기울어 있습니다. 양손의 손가락은 자신의 앞에 있는 건반을 짚고 있습니다. 방문자를 알아차리기 전 연주에 집중하는 방향과 일치합니다.",
        "built_space": "왼쪽 가장자리의 기둥 일부, 중앙 왼쪽의 피아노 한 대, 오른쪽의 벤치 하나가 보입니다. 뒤쪽 왼편의 소파 하나, 낮은 탁자 하나, 플로어 조명 하나와 그림 하나, 오른편의 수납장 하나와 꽃병 하나, 그림 하나 및 유리문이 참조와 같은 관계로 배치되어 있습니다. 통유리 너머 야경과 따뜻한 천장 간접조명, 광택 있는 석재 바닥도 일치합니다. 여성의 앉은 전신을 오른쪽에 두고 건반과 피아노 측면을 비스듬히 보여 줍니다. 피아노의 가로 범위는 화면의 약 40%이고 주변 홀의 여백이 유지됩니다. 바닥 반사는 물체와 조명의 위치에 부합합니다.",
        "entities": "중년의 동아시아계 여성 한 명이 등장하며, 짙은 단발머리와 얼굴 윤곽이 지소영 참조에 대체로 맞습니다. 정확한 국적과 나이는 외관만으로 확정할 수 없습니다. 흰 연구 가운의 짧은 길이와 가슴 표식, 회색 셔츠, 짙은 정장 바지, 검은 구두가 참조 복장에 가깝습니다. 닳은 갈색 그랜드피아노와 검은 누빔 벤치도 참조에 부합합니다. 건반 위 손은 인물의 소매와 팔에 이어져 있으며 별도 인물의 손처럼 보이지 않습니다. 추가 인물이나 화면 위 자막은 없습니다.",
        "hard_violations": [],
        "physics": "엉덩이가 벤치 좌판 위에 놓이고 상체는 골반을 중심으로 약간 앞으로 기울어 있습니다. 무릎 아래 다리는 바닥 쪽으로 내려오며, 한 발은 바닥에, 다른 발은 페달 위 또는 바로 인접한 위치에 놓입니다. 손가락과 건반의 접촉, 팔과 손목의 자세는 연주 동작으로 가능합니다. 피아노와 벤치의 다리가 바닥에 닿아 있으며 지지 없이 떠 있는 신체나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.746
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.746
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1746
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 와이드 샷 구도, 왼쪽 기둥의 배치, 그리고 피아노 건반을 향해 약간 몸을 숙인 인물의 자연스러운 자세를 성공적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 1746,
    "verdict_ko": "지정된 공간과 전체적인 앵글은 맞추었으나, 인물의 상체 자세가 다소 경직되어 있고 손가락 묘사의 디테일이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_director_piano_office_444ff1.png",
    "asset_id": "e7a34288-7811-4fad-a19d-4dc8f20e85e7",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:783266>",
    "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e6f-4f0e-7de5-8a4f-3f1b679a39c6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh7__bgfirst_bg.png",
   "bg_asset_id": "04610f06-54fc-4285-b5fa-2618df7a7122",
   "bg_record_key": "S84sh7::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "director_piano_office",
   "groupbg_asset_id": "e7a34288-7811-4fad-a19d-4dc8f20e85e7"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S84sh13::signage": {
  "fp": "245c252f0e87b30a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S84sh13": {
  "input_fingerprint": "9b100592a040fa03",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 두 기계 팔이 지소영의 상체를 와락 감싸 안고 있는 정지 찰나.\n\nLOCATION (lock): Beside the old piano inside the director's modern suite, under its nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop beside the piano at 지소영's seated head height, observing the embrace obliquely from the side and outside the space between their bodies. Place 찰리 on the left bending toward her and 지소영 on the right yielding into his chest, retaining both mechanical arms completely within the upper-body composition and a small keyboard edge below. Emphasize the closing distance between their bodies: 찰리 inclines his face toward her while she closes her eyes in the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano keyboard (No longer being played as 지소영 receives the embrace) — A small oblique section of the keys remains beneath their upper bodies; used as Preserves the connection between the melody and the reunion without interrupting the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain soft, restrained ambient illumination with precise mechanical contours and gentle facial contrast, letting contact provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the old piano, columns, modern interior finishes, and nighttime lighting from the reference. Exclude the glass laboratory enclosure, surgical bed, and scanning arms from the separate experiment room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the modern central hall. Charlie is free of the monitoring wires and has moved close enough to embrace with both arms. 지소영: She is beside the piano with her eyes gently closed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 두 기계 팔이 지소영의 상체를 와락 감싸 안고 있는 정지 찰나.\n\nLOCATION (lock): Beside the old piano inside the director's modern suite, under its nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop beside the piano at 지소영's seated head height, observing the embrace obliquely from the side and outside the space between their bodies. Place 찰리 on the left bending toward her and 지소영 on the right yielding into his chest, retaining both mechanical arms completely within the upper-body composition and a small keyboard edge below. Emphasize the closing distance between their bodies: 찰리 inclines his face toward her while she closes her eyes in the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano keyboard (No longer being played as 지소영 receives the embrace) — A small oblique section of the keys remains beneath their upper bodies; used as Preserves the connection between the melody and the reunion without interrupting the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain soft, restrained ambient illumination with precise mechanical contours and gentle facial contrast, letting contact provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the old piano, columns, modern interior finishes, and nighttime lighting from the reference. Exclude the glass laboratory enclosure, surgical bed, and scanning arms from the separate experiment room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the modern central hall. Charlie is free of the monitoring wires and has moved close enough to embrace with both arms. 지소영: She is beside the piano with her eyes gently closed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 찰리의 거대한 두 기계 팔이 지소영의 상체를 와락 감싸 안고 있는 정지 찰나.\n\nLOCATION (lock): Beside the old piano inside the director's modern suite, under its nighttime interior lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Stop beside the piano at 지소영's seated head height, observing the embrace obliquely from the side and outside the space between their bodies. Place 찰리 on the left bending toward her and 지소영 on the right yielding into his chest, retaining both mechanical arms completely within the upper-body composition and a small keyboard edge below. Emphasize the closing distance between their bodies: 찰리 inclines his face toward her while she closes her eyes in the embrace.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Old piano keyboard (No longer being played as 지소영 receives the embrace) — A small oblique section of the keys remains beneath their upper bodies; used as Preserves the connection between the melody and the reunion without interrupting the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain soft, restrained ambient illumination with precise mechanical contours and gentle facial contrast, letting contact provide the warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the old piano, columns, modern interior finishes, and nighttime lighting from the reference. Exclude the glass laboratory enclosure, surgical bed, and scanning arms from the separate experiment room.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old piano remains in the modern central hall. Charlie is free of the monitoring wires and has moved close enough to embrace with both arms. 지소영: She is beside the piano with her eyes gently closed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 우측 아래로 고개를 굽혀 지소영을 향하고, 지소영은 눈을 감고 찰리의 가슴에 기대어 있음.",
    "built_space": "화면 좌측에 피아노 건반과 몸체가 보이며, 배경에는 레퍼런스와 동일한 홀의 인테리어 및 야경이 위치함.",
    "entities": "찰리와 지소영의 얼굴, 의상, 기계 부품의 질감 등은 레퍼런스 이미지의 지시사항을 잘 따르고 있으나 찰리의 신체 구조 일부가 훼손됨.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy or staging (화면 좌측 찰리의 오른쪽 어깨 및 위팔이 아래로 향해 있음에도 불구하고, 지소영의 허리를 감싸는 또 다른 오른쪽 팔뚝이 허공/몸통에서 튀어나와 관절이 끊어진 채 중복 생성됨)"
    ],
    "physics": "지소영이 찰리에게 안겨 있으나, 찰리의 오른쪽 팔 구조가 물리적으로 연결되지 않은 상태로 생성되어 팔이 지소영을 지탱하는 구조가 완전히 붕괴됨."
   },
   {
    "label": "B",
    "direction": "찰리는 오른쪽 아래의 지소영을 향해 고개를 숙여 시선을 두고 있으며, 지소영은 눈을 감은 채 찰리의 가슴 쪽으로 얼굴을 기대고 있음.",
    "built_space": "화면 좌측 하단에 오래된 피아노 건반이 위치하고, 배경으로는 레퍼런스 이미지와 동일한 현대적인 홀의 기둥, 창문, 그리고 야경이 정확한 원근감으로 배치됨.",
    "entities": "찰리(고릴라 비율의 기계 몸체, 베이지색 장갑판, 흰색 마스크와 가슴의 푸른 원자로 심볼)와 지소영(50대 한국인 여성, 단정한 짙은 머리, 흰색 연구 가운) 모두 레퍼런스의 디자인과 완벽히 일치함.",
    "hard_violations": [],
    "physics": "찰리의 두 기계 팔이 지소영의 상체와 등, 허리를 감싸 안는 형태가 관절의 방향에 맞게 자연스럽게 연결되어 있으며, 지소영이 찰리에게 체중을 기대는 포옹의 역학이 물리적으로 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 두 기계 팔의 완벽한 포옹 자세, 피아노 건반이 포함된 구도, 인물들의 디테일 및 레퍼런스의 배경을 매우 충실하고 자연스럽게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "캐릭터와 배경의 외형은 잘 묘사되었으나, 찰리의 오른쪽 팔뚝이 어깨 및 팔꿈치 관절과 분리되어 중복 생성되는 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 오른쪽 아래의 지소영을 향해 고개를 숙여 시선을 두고 있으며, 지소영은 눈을 감은 채 찰리의 가슴 쪽으로 얼굴을 기대고 있음.",
        "built_space": "화면 좌측 하단에 오래된 피아노 건반이 위치하고, 배경으로는 레퍼런스 이미지와 동일한 현대적인 홀의 기둥, 창문, 그리고 야경이 정확한 원근감으로 배치됨.",
        "entities": "찰리(고릴라 비율의 기계 몸체, 베이지색 장갑판, 흰색 마스크와 가슴의 푸른 원자로 심볼)와 지소영(50대 한국인 여성, 단정한 짙은 머리, 흰색 연구 가운) 모두 레퍼런스의 디자인과 완벽히 일치함.",
        "hard_violations": [],
        "physics": "찰리의 두 기계 팔이 지소영의 상체와 등, 허리를 감싸 안는 형태가 관절의 방향에 맞게 자연스럽게 연결되어 있으며, 지소영이 찰리에게 체중을 기대는 포옹의 역학이 물리적으로 안정적임."
       },
       {
        "label": "A",
        "direction": "찰리는 우측 아래로 고개를 굽혀 지소영을 향하고, 지소영은 눈을 감고 찰리의 가슴에 기대어 있음.",
        "built_space": "화면 좌측에 피아노 건반과 몸체가 보이며, 배경에는 레퍼런스와 동일한 홀의 인테리어 및 야경이 위치함.",
        "entities": "찰리와 지소영의 얼굴, 의상, 기계 부품의 질감 등은 레퍼런스 이미지의 지시사항을 잘 따르고 있으나 찰리의 신체 구조 일부가 훼손됨.",
        "hard_violations": [
         "physically impossible anatomy or staging (화면 좌측 찰리의 오른쪽 어깨 및 위팔이 아래로 향해 있음에도 불구하고, 지소영의 허리를 감싸는 또 다른 오른쪽 팔뚝이 허공/몸통에서 튀어나와 관절이 끊어진 채 중복 생성됨)"
        ],
        "physics": "지소영이 찰리에게 안겨 있으나, 찰리의 오른쪽 팔 구조가 물리적으로 연결되지 않은 상태로 생성되어 팔이 지소영을 지탱하는 구조가 완전히 붕괴됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 두 기계 팔의 완벽한 포옹 자세, 피아노 건반이 포함된 구도, 인물들의 디테일 및 레퍼런스의 배경을 매우 충실하고 자연스럽게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "캐릭터와 배경의 외형은 잘 묘사되었으나, 찰리의 오른쪽 팔뚝이 어깨 및 팔꿈치 관절과 분리되어 중복 생성되는 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리는 오른쪽 아래의 지소영을 향해 고개를 숙여 시선을 두고 있으며, 지소영은 눈을 감은 채 찰리의 가슴 쪽으로 얼굴을 기대고 있음.",
        "built_space": "화면 좌측 하단에 오래된 피아노 건반이 위치하고, 배경으로는 레퍼런스 이미지와 동일한 현대적인 홀의 기둥, 창문, 그리고 야경이 정확한 원근감으로 배치됨.",
        "entities": "찰리(고릴라 비율의 기계 몸체, 베이지색 장갑판, 흰색 마스크와 가슴의 푸른 원자로 심볼)와 지소영(50대 한국인 여성, 단정한 짙은 머리, 흰색 연구 가운) 모두 레퍼런스의 디자인과 완벽히 일치함.",
        "hard_violations": [],
        "physics": "찰리의 두 기계 팔이 지소영의 상체와 등, 허리를 감싸 안는 형태가 관절의 방향에 맞게 자연스럽게 연결되어 있으며, 지소영이 찰리에게 체중을 기대는 포옹의 역학이 물리적으로 안정적임."
       },
       {
        "label": "A",
        "direction": "찰리는 우측 아래로 고개를 굽혀 지소영을 향하고, 지소영은 눈을 감고 찰리의 가슴에 기대어 있음.",
        "built_space": "화면 좌측에 피아노 건반과 몸체가 보이며, 배경에는 레퍼런스와 동일한 홀의 인테리어 및 야경이 위치함.",
        "entities": "찰리와 지소영의 얼굴, 의상, 기계 부품의 질감 등은 레퍼런스 이미지의 지시사항을 잘 따르고 있으나 찰리의 신체 구조 일부가 훼손됨.",
        "hard_violations": [
         "physically impossible anatomy or staging (화면 좌측 찰리의 오른쪽 어깨 및 위팔이 아래로 향해 있음에도 불구하고, 지소영의 허리를 감싸는 또 다른 오른쪽 팔뚝이 허공/몸통에서 튀어나와 관절이 끊어진 채 중복 생성됨)"
        ],
        "physics": "지소영이 찰리에게 안겨 있으나, 찰리의 오른쪽 팔 구조가 물리적으로 연결되지 않은 상태로 생성되어 팔이 지소영을 지탱하는 구조가 완전히 붕괴됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "좌우 배치와 눈을 감고 가슴에 기대는 포옹은 충실하지만, 뒤쪽 기계 팔의 경로가 대부분 가려지고 건반이 상대적으로 크게 드러나 두 팔 중심의 구도는 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "착석한 인물의 머리 높이에 가까운 비스듬한 상반신 구도에서 두 기계 팔의 감싸는 동작이 더 명확하며, 작은 하단 건반과 밀착한 얼굴 방향도 지시를 잘 따른다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 찰리가 얼굴을 오른쪽 아래 지소영의 정수리 쪽으로 기울인다. 오른쪽 지소영은 눈을 감고 얼굴과 상체를 왼쪽 찰리의 가슴에 붙인다. 앞쪽 기계 팔은 그녀의 몸 앞을 가로질러 등 쪽으로 감기고, 다른 손은 뒤쪽 어깨에 닿는다. 누구도 카메라나 건반을 바라보지 않으며 연주 동작도 없다.",
        "built_space": "왼쪽에 낡은 갈색 그랜드피아노 한 대와 그에 연결된 건반 한 줄이 보인다. 배경에는 왼쪽 소파 한 개, 낮은 탁자 한 개, 플로어램프 한 개, 벽 그림 두 점, 오른쪽 콘솔 한 개가 있으며 유리창과 석재 기둥·벽체, 천장 간접조명이 이전 장소의 구성을 이어간다. 실험실 설비는 없다. 건반은 하단 왼쪽에서 비교적 길게 드러난다. 두 인물의 좌석과 바닥 접촉부는 화면 밖이므로 착석 위치 자체는 확인할 수 없다. 바닥의 조명 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리 한 개체와 지소영 한 명만 보인다. 찰리는 마모된 샌드 베이지 장갑, 육중한 팔, 흰 마스크, 주황색 원형 눈 두 개와 선형 입, 푸른 흉부 원자로를 갖추어 참조와 부합한다. 지소영은 중년 한국인 여성의 외관과 정돈된 검은 단발, 흰 연구 가운과 짙은 안쪽 옷을 유지한다. 얼굴은 참조와 대체로 일치한다. 하반신과 가운 표장은 가려져 평가할 수 없고, 외부 감시선은 보이지 않는다.",
        "hard_violations": [],
        "physics": "앞쪽 팔은 어깨·상완·팔꿈치·전완·손으로 이어지며 손이 지소영의 등과 옆구리를 감싼다. 뒤쪽 손도 어깨에 접촉하지만 해당 팔의 중간 연결부는 몸에 가려진다. 지소영의 머리와 상체는 찰리의 가슴과 두 팔에 기대어 지지된다. 하체가 잘렸다는 이유만으로 공중에 떠 있다고 볼 근거는 없으며, 눈에 보이는 관절과 접촉에는 명백한 물리적 불가능성이 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 찰리의 얼굴과 발광 눈은 오른쪽 아래 지소영의 머리를 향한다. 지소영은 눈을 부드럽게 감고 왼쪽으로 기울어 찰리의 가슴에 뺨을 댄다. 앞쪽 팔은 상체 앞을 가로질러 그녀의 등 쪽을 감싸고, 반대쪽 팔과 손은 뒤에서 어깨와 상체를 둘러싼다. 양팔의 동작 대상이 지소영으로 명확하며 건반을 누르는 손은 없다.",
        "built_space": "낡은 갈색 그랜드피아노 한 대가 왼쪽에 있고, 연결된 건반 한 줄이 화면 하단에 비스듬한 좁은 띠로 남는다. 뒤에는 소파 한 개, 낮은 탁자 한 개, 플로어램프 한 개, 벽 그림 한 점과 오른쪽 가장자리에 일부 보이는 다른 그림, 콘솔 한 개가 보인다. 야간 유리창, 석재 기둥·벽체와 따뜻한 천장 간접조명이 이전 스틸의 현대식 홀을 유지한다. 두 인물 사이가 아닌 바깥쪽에서 보는 비스듬한 상반신 구도다. 좌석과 발은 가려져 정확한 착석 접점은 확인되지 않으며, 금지된 실험실 설비나 불가능한 인물 반사는 없다.",
        "entities": "추가 인물 없이 찰리와 지소영만 등장한다. 찰리의 샌드 베이지 장갑, 긴 육중한 팔, 흰 마스크와 두 주황색 눈, 짧은 입 선, 푸른 가슴 원자로가 참조의 기계 정체성을 유지한다. 지소영은 참조에 가까운 중년 한국인 여성의 얼굴과 검은 단발이며 흰 연구 가운과 짙은 셔츠를 입었다. 정장 바지와 신발은 이 구도에서 평가할 수 없다. 피아노의 오래된 목재 표면도 유지되고 감시용 외부 전선은 없다.",
        "hard_violations": [],
        "physics": "앞쪽 기계 팔의 어깨에서 팔꿈치와 전완으로 이어지는 연결이 읽히며, 손은 지소영의 몸에 밀착한다. 뒤쪽 팔은 일부가 가려져도 오른쪽 어깨 뒤로 돌아오는 장갑과 손의 경로가 A보다 분명하다. 지소영은 찰리의 흉부와 감싼 팔에 기대어 있고, 자신의 팔도 아래쪽에서 찰리에게 밀착한다. 기계 관절을 굽혀 만들 수 있는 포옹이며, 지지 없이 떠 있는 물체나 명백한 신체 관통은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "좌우 배치와 눈을 감고 가슴에 기대는 포옹은 충실하지만, 뒤쪽 기계 팔의 경로가 대부분 가려지고 건반이 상대적으로 크게 드러나 두 팔 중심의 구도는 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "착석한 인물의 머리 높이에 가까운 비스듬한 상반신 구도에서 두 기계 팔의 감싸는 동작이 더 명확하며, 작은 하단 건반과 밀착한 얼굴 방향도 지시를 잘 따른다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 찰리가 얼굴을 오른쪽 아래 지소영의 정수리 쪽으로 기울인다. 오른쪽 지소영은 눈을 감고 얼굴과 상체를 왼쪽 찰리의 가슴에 붙인다. 앞쪽 기계 팔은 그녀의 몸 앞을 가로질러 등 쪽으로 감기고, 다른 손은 뒤쪽 어깨에 닿는다. 누구도 카메라나 건반을 바라보지 않으며 연주 동작도 없다.",
        "built_space": "왼쪽에 낡은 갈색 그랜드피아노 한 대와 그에 연결된 건반 한 줄이 보인다. 배경에는 왼쪽 소파 한 개, 낮은 탁자 한 개, 플로어램프 한 개, 벽 그림 두 점, 오른쪽 콘솔 한 개가 있으며 유리창과 석재 기둥·벽체, 천장 간접조명이 이전 장소의 구성을 이어간다. 실험실 설비는 없다. 건반은 하단 왼쪽에서 비교적 길게 드러난다. 두 인물의 좌석과 바닥 접촉부는 화면 밖이므로 착석 위치 자체는 확인할 수 없다. 바닥의 조명 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리 한 개체와 지소영 한 명만 보인다. 찰리는 마모된 샌드 베이지 장갑, 육중한 팔, 흰 마스크, 주황색 원형 눈 두 개와 선형 입, 푸른 흉부 원자로를 갖추어 참조와 부합한다. 지소영은 중년 한국인 여성의 외관과 정돈된 검은 단발, 흰 연구 가운과 짙은 안쪽 옷을 유지한다. 얼굴은 참조와 대체로 일치한다. 하반신과 가운 표장은 가려져 평가할 수 없고, 외부 감시선은 보이지 않는다.",
        "hard_violations": [],
        "physics": "앞쪽 팔은 어깨·상완·팔꿈치·전완·손으로 이어지며 손이 지소영의 등과 옆구리를 감싼다. 뒤쪽 손도 어깨에 접촉하지만 해당 팔의 중간 연결부는 몸에 가려진다. 지소영의 머리와 상체는 찰리의 가슴과 두 팔에 기대어 지지된다. 하체가 잘렸다는 이유만으로 공중에 떠 있다고 볼 근거는 없으며, 눈에 보이는 관절과 접촉에는 명백한 물리적 불가능성이 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 찰리의 얼굴과 발광 눈은 오른쪽 아래 지소영의 머리를 향한다. 지소영은 눈을 부드럽게 감고 왼쪽으로 기울어 찰리의 가슴에 뺨을 댄다. 앞쪽 팔은 상체 앞을 가로질러 그녀의 등 쪽을 감싸고, 반대쪽 팔과 손은 뒤에서 어깨와 상체를 둘러싼다. 양팔의 동작 대상이 지소영으로 명확하며 건반을 누르는 손은 없다.",
        "built_space": "낡은 갈색 그랜드피아노 한 대가 왼쪽에 있고, 연결된 건반 한 줄이 화면 하단에 비스듬한 좁은 띠로 남는다. 뒤에는 소파 한 개, 낮은 탁자 한 개, 플로어램프 한 개, 벽 그림 한 점과 오른쪽 가장자리에 일부 보이는 다른 그림, 콘솔 한 개가 보인다. 야간 유리창, 석재 기둥·벽체와 따뜻한 천장 간접조명이 이전 스틸의 현대식 홀을 유지한다. 두 인물 사이가 아닌 바깥쪽에서 보는 비스듬한 상반신 구도다. 좌석과 발은 가려져 정확한 착석 접점은 확인되지 않으며, 금지된 실험실 설비나 불가능한 인물 반사는 없다.",
        "entities": "추가 인물 없이 찰리와 지소영만 등장한다. 찰리의 샌드 베이지 장갑, 긴 육중한 팔, 흰 마스크와 두 주황색 눈, 짧은 입 선, 푸른 가슴 원자로가 참조의 기계 정체성을 유지한다. 지소영은 참조에 가까운 중년 한국인 여성의 얼굴과 검은 단발이며 흰 연구 가운과 짙은 셔츠를 입었다. 정장 바지와 신발은 이 구도에서 평가할 수 없다. 피아노의 오래된 목재 표면도 유지되고 감시용 외부 전선은 없다.",
        "hard_violations": [],
        "physics": "앞쪽 기계 팔의 어깨에서 팔꿈치와 전완으로 이어지는 연결이 읽히며, 손은 지소영의 몸에 밀착한다. 뒤쪽 팔은 일부가 가려져도 오른쪽 어깨 뒤로 돌아오는 장갑과 손의 경로가 A보다 분명하다. 지소영은 찰리의 흉부와 감싼 팔에 기대어 있고, 자신의 팔도 아래쪽에서 찰리에게 밀착한다. 기계 관절을 굽혀 만들 수 있는 포옹이며, 지지 없이 떠 있는 물체나 명백한 신체 관통은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.2,
    "B": 1.889
   },
   "adjusted": {
    "A": 0.95,
    "B": 1.889
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy or staging (화면 좌측 찰리의 오른쪽 어깨 및 위팔이 아래로 향해 있음에도 불구하고, 지소영의 허리를 감싸는 또 다른 오른쪽 팔뚝이 허공/몸통에서 튀어나와 관절이 끊어진 채 중복 생성됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1889,
   "A": 950
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "프롬프트가 요구한 두 기계 팔의 완벽한 포옹 자세, 피아노 건반이 포함된 구도, 인물들의 디테일 및 레퍼런스의 배경을 매우 충실하고 자연스럽게 구현한 훌륭한 결과물입니다."
   },
   {
    "label": "A",
    "score": 950,
    "verdict_ko": "캐릭터와 배경의 외형은 잘 묘사되었으나, 찰리의 오른쪽 팔뚝이 어깨 및 팔꿈치 관절과 분리되어 중복 생성되는 심각한 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] physically impossible anatomy or staging (화면 좌측 찰리의 오른쪽 어깨 및 위팔이 아래로 향해 있음에도 불구하고, 지소영의 허리를 감싸는 또 다른 오른쪽 팔뚝이 허공/몸통에서 튀어나와 관절이 끊어진 채 중복 생성됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh7_sel.png",
    "asset_id": "51a1fee3-4efa-450b-a9f2-82655c6d4517",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:783266>",
    "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e77-4d66-7c8f-937d-76d524a3901b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh7"
  }
 },
 "S85sh12::signage": {
  "fp": "2a8998f10fb7d404",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S85sh12": {
  "input_fingerprint": "676a382b9831bc9e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르는 근접 찰나.\n\nLOCATION (lock): Inside the glass-walled test enclosure in the main research laboratory, where cables connect the robot to the experiment. Laboratory lighting and active displays illuminate the setup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the enclosure at 찰리's chest height, close and oblique to his torso, without aligning with his frontal axis. Place the rotating ring slightly below center, occupying less than two-fifths of the image, with both shoulder contours, attached wires, and the lower edge of his bowed face supplying bodily context. Record the first black smoke lifting from the ring while his eyes remain closed in concentration, keeping the camera still until the burst registers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attached experimental wires (Connected to 찰리 during the attempted energy experiment); used as Remain along the torso margins as procedural context; Black smoke from the chest ring (The first burst emerges as the ring begins losing speed); used as Rises across the central chest area while leaving the point of origin readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient contrast separates the black smoke from 찰리's hard-surface contours without adding flames or an unsupported glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with multiple wires reattached; black smoke rises from his chest ring as its rotation slows. The experiment screen displays “ERROR” after the earlier recording display.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르는 근접 찰나.\n\nLOCATION (lock): Inside the glass-walled test enclosure in the main research laboratory, where cables connect the robot to the experiment. Laboratory lighting and active displays illuminate the setup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the enclosure at 찰리's chest height, close and oblique to his torso, without aligning with his frontal axis. Place the rotating ring slightly below center, occupying less than two-fifths of the image, with both shoulder contours, attached wires, and the lower edge of his bowed face supplying bodily context. Record the first black smoke lifting from the ring while his eyes remain closed in concentration, keeping the camera still until the burst registers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attached experimental wires (Connected to 찰리 during the attempted energy experiment); used as Remain along the torso margins as procedural context; Black smoke from the chest ring (The first burst emerges as the ring begins losing speed); used as Rises across the central chest area while leaving the point of origin readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient contrast separates the black smoke from 찰리's hard-surface contours without adding flames or an unsupported glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with multiple wires reattached; black smoke rises from his chest ring as its rotation slows. The experiment screen displays “ERROR” after the earlier recording display.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링에서 시커먼 연기가 훅 피어오르는 근접 찰나.\n\nLOCATION (lock): Inside the glass-walled test enclosure in the main research laboratory, where cables connect the robot to the experiment. Laboratory lighting and active displays illuminate the setup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold inside the enclosure at 찰리's chest height, close and oblique to his torso, without aligning with his frontal axis. Place the rotating ring slightly below center, occupying less than two-fifths of the image, with both shoulder contours, attached wires, and the lower edge of his bowed face supplying bodily context. Record the first black smoke lifting from the ring while his eyes remain closed in concentration, keeping the camera still until the burst registers.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attached experimental wires (Connected to 찰리 during the attempted energy experiment); used as Remain along the torso margins as procedural context; Black smoke from the chest ring (The first burst emerges as the ring begins losing speed); used as Rises across the central chest area while leaving the point of origin readable.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Controlled ambient contrast separates the black smoke from 찰리's hard-surface contours without adding flames or an unsupported glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie is seated inside the glass enclosure with multiple wires reattached; black smoke rises from his chest ring as its rotation slows. The experiment screen displays “ERROR” after the earlier recording display.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 정면 위를 향하고 있으며 가슴의 링에서 검은 연기가 위로 피어오름.",
    "built_space": "이전 샷의 금속 실험 테이블 위에 그대로 누워 있으며 카메라는 위에서 아래로 내려다보는 각도로 위치함.",
    "entities": "찰리가 누워있는 상태이며 레퍼런스와 동일한 둥근 형태의 눈에서 빛이 나고 있음. 가슴에 전선들이 연결되어 있고 연기가 발생함.",
    "hard_violations": [
     "[gemini-pro] 인물이 지시된 자세(앉아있음)가 아닌 누워있는 상태로 배치됨",
     "[gemini-pro] 지시된 카메라 앵글과 위치를 무시하고 이전 샷 레퍼런스의 구도를 그대로 복사함"
    ],
    "physics": "머리와 등이 금속 테이블 표면에 닿아 지지받고 있으며 연기는 중력 반대 방향으로 자연스럽게 상승함."
   },
   {
    "label": "B",
    "direction": "고개를 약간 숙이고 있으며 가슴의 링에서 검은 연기가 위쪽으로 피어오름.",
    "built_space": "유리 벽이 있는 연구소 내부에 앉아 있으며 배경으로 실험실 조명과 구조물이 보임. 카메라는 지시대로 가슴 높이에서 비스듬히 찰리를 향함.",
    "entities": "찰리가 앉아있으며 가슴 링에서 검은 연기가 피어오르고 전선들이 연결되어 있음. 단, 지시된 '눈을 감은' 상태를 표현하기 위해 고정된 금속 마스크의 둥근 눈 구멍을 얇은 가로선으로 임의 변형함.",
    "hard_violations": [],
    "physics": "화면 하단에 가려진 구조물에 앉아 체중을 지지하고 있으며 연기는 공기 중으로 정상적으로 피어오름."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 둥근 눈 형태가 얇은 선으로 임의 변형된 점은 캐릭터 디자인 규정 위반이나, 프롬프트가 지시한 가슴 높이의 비스듬한 카메라 앵글, 고개를 숙이고 앉아있는 자세, 가슴에서 피어오르는 검은 연기 등 핵심 샷 구도를 매우 정확히 구현하여 더 우수합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 명시한 카메라 앵글(가슴 높이, 비스듬한 구도)과 인물의 자세(앉아있음)를 완전히 무시하고 이전 샷 레퍼런스의 누워있는 구도를 그대로 복사하여 심각한 연출 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 정면 위를 향하고 있으며 가슴의 링에서 검은 연기가 위로 피어오름.",
        "built_space": "이전 샷의 금속 실험 테이블 위에 그대로 누워 있으며 카메라는 위에서 아래로 내려다보는 각도로 위치함.",
        "entities": "찰리가 누워있는 상태이며 레퍼런스와 동일한 둥근 형태의 눈에서 빛이 나고 있음. 가슴에 전선들이 연결되어 있고 연기가 발생함.",
        "hard_violations": [
         "인물이 지시된 자세(앉아있음)가 아닌 누워있는 상태로 배치됨",
         "지시된 카메라 앵글과 위치를 무시하고 이전 샷 레퍼런스의 구도를 그대로 복사함"
        ],
        "physics": "머리와 등이 금속 테이블 표면에 닿아 지지받고 있으며 연기는 중력 반대 방향으로 자연스럽게 상승함."
       },
       {
        "label": "B",
        "direction": "고개를 약간 숙이고 있으며 가슴의 링에서 검은 연기가 위쪽으로 피어오름.",
        "built_space": "유리 벽이 있는 연구소 내부에 앉아 있으며 배경으로 실험실 조명과 구조물이 보임. 카메라는 지시대로 가슴 높이에서 비스듬히 찰리를 향함.",
        "entities": "찰리가 앉아있으며 가슴 링에서 검은 연기가 피어오르고 전선들이 연결되어 있음. 단, 지시된 '눈을 감은' 상태를 표현하기 위해 고정된 금속 마스크의 둥근 눈 구멍을 얇은 가로선으로 임의 변형함.",
        "hard_violations": [],
        "physics": "화면 하단에 가려진 구조물에 앉아 체중을 지지하고 있으며 연기는 공기 중으로 정상적으로 피어오름."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 둥근 눈 형태가 얇은 선으로 임의 변형된 점은 캐릭터 디자인 규정 위반이나, 프롬프트가 지시한 가슴 높이의 비스듬한 카메라 앵글, 고개를 숙이고 앉아있는 자세, 가슴에서 피어오르는 검은 연기 등 핵심 샷 구도를 매우 정확히 구현하여 더 우수합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 명시한 카메라 앵글(가슴 높이, 비스듬한 구도)과 인물의 자세(앉아있음)를 완전히 무시하고 이전 샷 레퍼런스의 누워있는 구도를 그대로 복사하여 심각한 연출 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 정면 위를 향하고 있으며 가슴의 링에서 검은 연기가 위로 피어오름.",
        "built_space": "이전 샷의 금속 실험 테이블 위에 그대로 누워 있으며 카메라는 위에서 아래로 내려다보는 각도로 위치함.",
        "entities": "찰리가 누워있는 상태이며 레퍼런스와 동일한 둥근 형태의 눈에서 빛이 나고 있음. 가슴에 전선들이 연결되어 있고 연기가 발생함.",
        "hard_violations": [
         "인물이 지시된 자세(앉아있음)가 아닌 누워있는 상태로 배치됨",
         "지시된 카메라 앵글과 위치를 무시하고 이전 샷 레퍼런스의 구도를 그대로 복사함"
        ],
        "physics": "머리와 등이 금속 테이블 표면에 닿아 지지받고 있으며 연기는 중력 반대 방향으로 자연스럽게 상승함."
       },
       {
        "label": "B",
        "direction": "고개를 약간 숙이고 있으며 가슴의 링에서 검은 연기가 위쪽으로 피어오름.",
        "built_space": "유리 벽이 있는 연구소 내부에 앉아 있으며 배경으로 실험실 조명과 구조물이 보임. 카메라는 지시대로 가슴 높이에서 비스듬히 찰리를 향함.",
        "entities": "찰리가 앉아있으며 가슴 링에서 검은 연기가 피어오르고 전선들이 연결되어 있음. 단, 지시된 '눈을 감은' 상태를 표현하기 위해 고정된 금속 마스크의 둥근 눈 구멍을 얇은 가로선으로 임의 변형함.",
        "hard_violations": [],
        "physics": "화면 하단에 가려진 구조물에 앉아 체중을 지지하고 있으며 연기는 공기 중으로 정상적으로 피어오름."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "가슴 높이의 비스듬한 근접 시점과 검은 연기의 분출은 더 충실하지만, 숙인 얼굴의 아래쪽만 담으라는 지시보다 얼굴이 많이 보이고 링도 다소 낮다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "기존 실험대와 장갑판의 연속성은 좋지만, 이전 사진처럼 위에서 내려다보는 구도와 반쯤 열린 눈이 지정된 가슴 높이 시점·눈을 감고 집중하는 순간에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 화면 왼쪽으로 기울었지만 가슴 링을 향해 명확히 숙여져 있지는 않다. 두 눈은 가느다란 수평 발광 틈으로 표현되어 감은 눈에 가깝다. 검은 연기는 가슴 링에서 목과 화면 위쪽으로 올라가며 발생 부위가 보인다. 카메라는 몸통 정면축에서 옆으로 벗어난 근접 시점이다.",
        "built_space": "양쪽 어깨 윤곽과 몸통 가장자리의 여러 연결 케이블이 보인다. 뒤에는 유리벽의 수직 프레임들과 오른쪽의 부분적으로 잘린 청색 표시 화면 하나가 보인다. 이전 사진의 금속·유리 실험실 재질은 이어지지만 금속 받침 구조는 대부분 프레임 밖이다. 중복 설비나 불가능한 반사는 보이지 않는다. 링은 화면 면적의 5분의 2보다 작지만 중심보다 상당히 아래에 있으며, 얼굴도 아래 가장자리만이 아니라 상당 부분 포함된다.",
        "entities": "인물은 찰리 한 기뿐이다. 마모된 샌드 베이지 장갑판, 흰 마스크형 얼굴, 선형 입, 육중한 어깨, 중앙의 푸른 원형 장치는 참조와 부합한다. 눈은 참조의 원형 구멍 대신 가늘고 긴 틈으로 바뀌었다. 검은 연기와 실험 연결선이 있으며 불꽃이나 추가 인물은 없다. 화면의 'ERROR' 표시는 확인되지 않지만 표시 화면 자체가 가장자리에서 잘려 있다.",
        "hard_violations": [],
        "physics": "머리는 기계 목 관절에, 어깨와 팔은 몸통 관절에 연결되어 있다. 케이블은 커넥터에 결합되어 장갑판을 따라 늘어진다. 연기는 링 주변에서 시작하여 위로 확산하므로 부유하는 고체처럼 보이지 않는다. 좌석과 하체는 근접 구도 밖이어서 착석 접촉점은 확인할 수 없으며, 보이는 부분에 무지지 부유나 불가능한 관절은 없다. 정지 화면만으로 링의 감속 여부는 확정할 수 없다."
       },
       {
        "label": "B",
        "direction": "머리는 뒤로 기대어 얼굴을 카메라 쪽 위로 드러내며, 가슴 쪽으로 숙인 모습이 아니다. 두 눈의 아래쪽 발광 면이 넓게 남아 있어 완전히 감았다기보다 반쯤 뜬 것으로 보인다. 연기는 링 윗부분에서 목 앞을 지나 위로 솟으며 발생 위치를 읽을 수 있다. 카메라는 가슴 높이에서 옆으로 접근하기보다 위에서 몸통을 내려다본다.",
        "built_space": "몸 뒤의 금속 실험대 한 면, 왼쪽의 구멍 난 고정 블록 하나, 머리 뒤 수직 지지봉 하나와 유리벽 프레임들이 보인다. 오른쪽 위에는 일부 잘린 표시 장비가 있다. 금속판과 고정 구조는 이전 사진의 장소를 잘 이어간다. 양쪽 어깨와 연결선도 포함되지만, 실험대 윗면이 넓게 드러나는 높은 시점은 지정된 가슴 높이 카메라와 다르다. 링은 충분히 작으나 아래쪽에 치우치고 얼굴 대부분이 들어온다. 중복 설비나 모순되는 반사는 보이지 않는다.",
        "entities": "찰리 한 기의 베이지색 마모 장갑, 흰 마스크, 원형 눈 테두리와 선형 입, 푸른 가슴 링이 참조와 가깝다. 여러 실험 케이블과 검은 연기가 있으며 추가 인물이나 불꽃은 없다. 눈은 요청한 완전한 감김 상태가 아니다. 가장자리의 표시 장비에서 'ERROR' 문구는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "상체는 뒤쪽 금속 받침에 기대어 지지되는 것으로 보이며 머리와 팔의 기계 관절 연결도 유지된다. 케이블은 몸통 커넥터에 물려 자연스럽게 휘어진다. 연기는 링에서 이어져 상승하며 별도로 떠 있는 물체는 없다. 하체와 좌석 접촉은 프레임 밖이므로 착석 상태 전체를 판정할 수 없다. 링의 푸른 무늬는 보이지만 회전 속도의 감소는 한 장으로 확인하기 어렵다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "가슴 높이의 비스듬한 근접 시점과 검은 연기의 분출은 더 충실하지만, 숙인 얼굴의 아래쪽만 담으라는 지시보다 얼굴이 많이 보이고 링도 다소 낮다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "기존 실험대와 장갑판의 연속성은 좋지만, 이전 사진처럼 위에서 내려다보는 구도와 반쯤 열린 눈이 지정된 가슴 높이 시점·눈을 감고 집중하는 순간에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 화면 왼쪽으로 기울었지만 가슴 링을 향해 명확히 숙여져 있지는 않다. 두 눈은 가느다란 수평 발광 틈으로 표현되어 감은 눈에 가깝다. 검은 연기는 가슴 링에서 목과 화면 위쪽으로 올라가며 발생 부위가 보인다. 카메라는 몸통 정면축에서 옆으로 벗어난 근접 시점이다.",
        "built_space": "양쪽 어깨 윤곽과 몸통 가장자리의 여러 연결 케이블이 보인다. 뒤에는 유리벽의 수직 프레임들과 오른쪽의 부분적으로 잘린 청색 표시 화면 하나가 보인다. 이전 사진의 금속·유리 실험실 재질은 이어지지만 금속 받침 구조는 대부분 프레임 밖이다. 중복 설비나 불가능한 반사는 보이지 않는다. 링은 화면 면적의 5분의 2보다 작지만 중심보다 상당히 아래에 있으며, 얼굴도 아래 가장자리만이 아니라 상당 부분 포함된다.",
        "entities": "인물은 찰리 한 기뿐이다. 마모된 샌드 베이지 장갑판, 흰 마스크형 얼굴, 선형 입, 육중한 어깨, 중앙의 푸른 원형 장치는 참조와 부합한다. 눈은 참조의 원형 구멍 대신 가늘고 긴 틈으로 바뀌었다. 검은 연기와 실험 연결선이 있으며 불꽃이나 추가 인물은 없다. 화면의 'ERROR' 표시는 확인되지 않지만 표시 화면 자체가 가장자리에서 잘려 있다.",
        "hard_violations": [],
        "physics": "머리는 기계 목 관절에, 어깨와 팔은 몸통 관절에 연결되어 있다. 케이블은 커넥터에 결합되어 장갑판을 따라 늘어진다. 연기는 링 주변에서 시작하여 위로 확산하므로 부유하는 고체처럼 보이지 않는다. 좌석과 하체는 근접 구도 밖이어서 착석 접촉점은 확인할 수 없으며, 보이는 부분에 무지지 부유나 불가능한 관절은 없다. 정지 화면만으로 링의 감속 여부는 확정할 수 없다."
       },
       {
        "label": "A",
        "direction": "머리는 뒤로 기대어 얼굴을 카메라 쪽 위로 드러내며, 가슴 쪽으로 숙인 모습이 아니다. 두 눈의 아래쪽 발광 면이 넓게 남아 있어 완전히 감았다기보다 반쯤 뜬 것으로 보인다. 연기는 링 윗부분에서 목 앞을 지나 위로 솟으며 발생 위치를 읽을 수 있다. 카메라는 가슴 높이에서 옆으로 접근하기보다 위에서 몸통을 내려다본다.",
        "built_space": "몸 뒤의 금속 실험대 한 면, 왼쪽의 구멍 난 고정 블록 하나, 머리 뒤 수직 지지봉 하나와 유리벽 프레임들이 보인다. 오른쪽 위에는 일부 잘린 표시 장비가 있다. 금속판과 고정 구조는 이전 사진의 장소를 잘 이어간다. 양쪽 어깨와 연결선도 포함되지만, 실험대 윗면이 넓게 드러나는 높은 시점은 지정된 가슴 높이 카메라와 다르다. 링은 충분히 작으나 아래쪽에 치우치고 얼굴 대부분이 들어온다. 중복 설비나 모순되는 반사는 보이지 않는다.",
        "entities": "찰리 한 기의 베이지색 마모 장갑, 흰 마스크, 원형 눈 테두리와 선형 입, 푸른 가슴 링이 참조와 가깝다. 여러 실험 케이블과 검은 연기가 있으며 추가 인물이나 불꽃은 없다. 눈은 요청한 완전한 감김 상태가 아니다. 가장자리의 표시 장비에서 'ERROR' 문구는 판독되지 않는다.",
        "hard_violations": [],
        "physics": "상체는 뒤쪽 금속 받침에 기대어 지지되는 것으로 보이며 머리와 팔의 기계 관절 연결도 유지된다. 케이블은 몸통 커넥터에 물려 자연스럽게 휘어진다. 연기는 링에서 이어져 상승하며 별도로 떠 있는 물체는 없다. 하체와 좌석 접촉은 프레임 밖이므로 착석 상태 전체를 판정할 수 없다. 링의 푸른 무늬는 보이지만 회전 속도의 감소는 한 장으로 확인하기 어렵다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.964,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.714,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 인물이 지시된 자세(앉아있음)가 아닌 누워있는 상태로 배치됨",
     "[gemini-pro] 지시된 카메라 앵글과 위치를 무시하고 이전 샷 레퍼런스의 구도를 그대로 복사함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "찰리의 둥근 눈 형태가 얇은 선으로 임의 변형된 점은 캐릭터 디자인 규정 위반이나, 프롬프트가 지시한 가슴 높이의 비스듬한 카메라 앵글, 고개를 숙이고 앉아있는 자세, 가슴에서 피어오르는 검은 연기 등 핵심 샷 구도를 매우 정확히 구현하여 더 우수합니다."
   },
   {
    "label": "A",
    "score": 714,
    "verdict_ko": "프롬프트가 명시한 카메라 앵글(가슴 높이, 비스듬한 구도)과 인물의 자세(앉아있음)를 완전히 무시하고 이전 샷 레퍼런스의 누워있는 구도를 그대로 복사하여 심각한 연출 위반이 발생했습니다.  ★위반: [gemini-pro] 인물이 지시된 자세(앉아있음)가 아닌 누워있는 상태로 배치됨 / [gemini-pro] 지시된 카메라 앵글과 위치를 무시하고 이전 샷 레퍼런스의 구도를 그대로 복사함"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh1_sel.png",
    "asset_id": "9fa02819-a001-4a4f-a271-1cd688bc14a5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e7c-4b40-7032-9e07-28ec58f2ca91",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh1"
  }
 },
 "S85sh15::signage": {
  "fp": "a8f5acb244f1303b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::research_dining_hall": {
  "input_fingerprint": "37a76841b017a89d",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "research_dining_hall",
    "tags": [
     "S85sh15",
     "S85sh24"
    ]
   },
   "context_sig": "c5b63d9e573de2b7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 식당. 연구소 안\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 복도·자동문 앞, 메인센터 중앙 홀: 열대 식물 정원과 최첨단 장비가 어우러진 웅장한 로비 및 이동 통로다. (특징: 카드키를 대면 미끄러지듯 열리는 최신식 자동문; 원자력 발전 시설과 열대 식물 온실이 결합된 거대한 스케일의 인테리어; 세련된 50대 중반의 지소영 소장의 흰색 연구복 차림)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- / 식당. 연구소 안\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_dining_hall_a2331c.png",
  "asset_id": "da12b1e2-cc29-4186-b060-6bb1cb26d7f1",
  "input_asset_ids": [
   "efcb4df7-d080-4d19-a38a-e01ccffbc245"
  ],
  "origin_tag": "S85sh15",
  "place_text": "At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.",
  "origin_inputs": {
   "place_text": "At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.",
   "time_of_day_en": "day",
   "conti_asset_id": "efcb4df7-d080-4d19-a38a-e01ccffbc245"
  }
 },
 "S85sh15::bgfirst_bg": {
  "input_fingerprint": "270ba9de465aad1a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 배식을 기다리는 사람들의 긴 줄 옆, 식당 테이블에 마주 앉은 이현우와 서지민의 넓은 구도.\n\nLOCATION (lock): At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut to the dining room and hold the crane's high entry position beside the meal queue, looking diagonally downward across the table from one side of the conversation axis. Place 이현우 on the left and 서지민 opposite on the right, both fully readable in seated posture, with the uneaten bread and food between them; he folds toward his meal while she inclines toward the same food before questioning him. Let the long queue recede beside the table into the upper-left background, its waiting people differing subtly in weight shifts, head angles, arm positions, and spacing rather than repeating a single pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Meal queue receding beside the dining table in the upper-left of the frame, background; Dining table between 이현우 and 서지민 in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table (이현우 and 서지민 sit opposite each other) — Its top is visible diagonally from above, with the open side facing the camera; used as Establishes their conversation axis and keeps the meal between them; Bread and food (Untouched before 이현우); used as A modest foreground detail gives his lowered gaze a concrete subject; Meal queue (Research-center people and residents wait for food in a long line) — The line recedes beside the table toward the serving destination outside the frame; used as Establishes communal life through varied waiting postures and naturally uneven spacing, without adding interactions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Even, subdued ambient illumination and controlled contrast keep the food, the withdrawn diner, and the waiting line equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 배식을 기다리는 사람들의 긴 줄 옆, 식당 테이블에 마주 앉은 이현우와 서지민의 넓은 구도.\n\nLOCATION (lock): At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut to the dining room and hold the crane's high entry position beside the meal queue, looking diagonally downward across the table from one side of the conversation axis. Place 이현우 on the left and 서지민 opposite on the right, both fully readable in seated posture, with the uneaten bread and food between them; he folds toward his meal while she inclines toward the same food before questioning him. Let the long queue recede beside the table into the upper-left background, its waiting people differing subtly in weight shifts, head angles, arm positions, and spacing rather than repeating a single pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Meal queue receding beside the dining table in the upper-left of the frame, background; Dining table between 이현우 and 서지민 in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table (이현우 and 서지민 sit opposite each other) — Its top is visible diagonally from above, with the open side facing the camera; used as Establishes their conversation axis and keeps the meal between them; Bread and food (Untouched before 이현우); used as A modest foreground detail gives his lowered gaze a concrete subject; Meal queue (Research-center people and residents wait for food in a long line) — The line recedes beside the table toward the serving destination outside the frame; used as Establishes communal life through varied waiting postures and naturally uneven spacing, without adding interactions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Even, subdued ambient illumination and controlled contrast keep the food, the withdrawn diner, and the waiting line equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S85sh15__bgfirst_bg.png",
  "asset_id": "7676346d-56c3-4cee-aedd-439ca29ca549",
  "input_asset_ids": [
   "efcb4df7-d080-4d19-a38a-e01ccffbc245",
   "da12b1e2-cc29-4186-b060-6bb1cb26d7f1"
  ]
 },
 "S85sh15": {
  "input_fingerprint": "ad69096adc077209",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배식을 기다리는 사람들의 긴 줄 옆, 식당 테이블에 마주 앉은 이현우와 서지민의 넓은 구도.\n\nLOCATION (lock): At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut to the dining room and hold the crane's high entry position beside the meal queue, looking diagonally downward across the table from one side of the conversation axis. Place 이현우 on the left and 서지민 opposite on the right, both fully readable in seated posture, with the uneaten bread and food between them; he folds toward his meal while she inclines toward the same food before questioning him. Let the long queue recede beside the table into the upper-left background, its waiting people differing subtly in weight shifts, head angles, arm positions, and spacing rather than repeating a single pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Meal queue receding beside the dining table in the upper-left of the frame, background; Dining table between 이현우 and 서지민 in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table (이현우 and 서지민 sit opposite each other) — Its top is visible diagonally from above, with the open side facing the camera; used as Establishes their conversation axis and keeps the meal between them; Bread and food (Untouched before 이현우); used as A modest foreground detail gives his lowered gaze a concrete subject; Meal queue (Research-center people and residents wait for food in a long line) — The line recedes beside the table toward the serving destination outside the frame; used as Establishes communal life through varied waiting postures and naturally uneven spacing, without adding interactions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Even, subdued ambient illumination and controlled contrast keep the food, the withdrawn diner, and the waiting line equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bread and a meal are set on the dining table and remain uneaten. 이현우: He is seated at the table, looking down without eating. 서지민: She is at the dining table, still in her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배식을 기다리는 사람들의 긴 줄 옆, 식당 테이블에 마주 앉은 이현우와 서지민의 넓은 구도.\n\nLOCATION (lock): At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut to the dining room and hold the crane's high entry position beside the meal queue, looking diagonally downward across the table from one side of the conversation axis. Place 이현우 on the left and 서지민 opposite on the right, both fully readable in seated posture, with the uneaten bread and food between them; he folds toward his meal while she inclines toward the same food before questioning him. Let the long queue recede beside the table into the upper-left background, its waiting people differing subtly in weight shifts, head angles, arm positions, and spacing rather than repeating a single pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Meal queue receding beside the dining table in the upper-left of the frame, background; Dining table between 이현우 and 서지민 in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table (이현우 and 서지민 sit opposite each other) — Its top is visible diagonally from above, with the open side facing the camera; used as Establishes their conversation axis and keeps the meal between them; Bread and food (Untouched before 이현우); used as A modest foreground detail gives his lowered gaze a concrete subject; Meal queue (Research-center people and residents wait for food in a long line) — The line recedes beside the table toward the serving destination outside the frame; used as Establishes communal life through varied waiting postures and naturally uneven spacing, without adding interactions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Even, subdued ambient illumination and controlled contrast keep the food, the withdrawn diner, and the waiting line equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bread and a meal are set on the dining table and remain uneaten. 이현우: He is seated at the table, looking down without eating. 서지민: She is at the dining table, still in her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배식을 기다리는 사람들의 긴 줄 옆, 식당 테이블에 마주 앉은 이현우와 서지민의 넓은 구도.\n\nLOCATION (lock): At a dining table beside the serving queue inside the research institute's communal cafeteria. The room is illuminated for daytime meal service. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Make the explicit cut to the dining room and hold the crane's high entry position beside the meal queue, looking diagonally downward across the table from one side of the conversation axis. Place 이현우 on the left and 서지민 opposite on the right, both fully readable in seated posture, with the uneaten bread and food between them; he folds toward his meal while she inclines toward the same food before questioning him. Let the long queue recede beside the table into the upper-left background, its waiting people differing subtly in weight shifts, head angles, arm positions, and spacing rather than repeating a single pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Meal queue receding beside the dining table in the upper-left of the frame, background; Dining table between 이현우 and 서지민 in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Dining table (이현우 and 서지민 sit opposite each other) — Its top is visible diagonally from above, with the open side facing the camera; used as Establishes their conversation axis and keeps the meal between them; Bread and food (Untouched before 이현우); used as A modest foreground detail gives his lowered gaze a concrete subject; Meal queue (Research-center people and residents wait for food in a long line) — The line recedes beside the table toward the serving destination outside the frame; used as Establishes communal life through varied waiting postures and naturally uneven spacing, without adding interactions.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Even, subdued ambient illumination and controlled contrast keep the food, the withdrawn diner, and the waiting line equally legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bread and a meal are set on the dining table and remain uneaten. 이현우: He is seated at the table, looking down without eating. 서지민: She is at the dining table, still in her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S85sh15__bgfirst_bg.png",
     "asset_id": "7676346d-56c3-4cee-aedd-439ca29ca549",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S85sh15.png",
     "asset_id": "efcb4df7-d080-4d19-a38a-e01ccffbc245",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:766860>",
     "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869875>",
     "asset_id": "bdc552c6-5e2a-4bae-affd-3dbee764db52",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_dining_hall_a2331c.png",
     "asset_id": "da12b1e2-cc29-4186-b060-6bb1cb26d7f1",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:766860>",
     "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869875>",
     "asset_id": "bdc552c6-5e2a-4bae-affd-3dbee764db52",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 식판을, 서지민은 이현우를 바라봄. 대기열 인원들은 배식구를 향함.",
    "built_space": "식당 배경은 기준과 일치하나, 카메라는 지시된 높은 시점이 아닌 인물의 눈높이에 위치함.",
    "entities": "이현우(환자복), 서지민(연구 가운 및 고글), 식판 위 음식 모두 조건과 일치함.",
    "hard_violations": [
     "[gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 중앙 뒤편과 오른쪽 테이블에서 별도로 식사하는 인물들을 다수 추가했다."
    ],
    "physics": "모든 인물과 사물이 의자, 바닥, 테이블에 안정적으로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "이현우는 음식을 내려다보고, 서지민은 그를 향함. 대기열은 앞쪽을 주시함.",
    "built_space": "식당 구조가 일치하며, 카메라가 명시된 대로 높은 위치에서 대각선 아래로 테이블을 내려다봄.",
    "entities": "인물의 기본 인상착의와 음식은 일치하나, 기준 이미지에 있는 서지민의 머리 위 고글이 누락됨.",
    "hard_violations": [
     "[gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 뒤쪽과 오른쪽 테이블에서 따로 식사하는 인물 2명을 추가했다."
    ],
    "physics": "착석 및 기립 인물 모두 바닥과 의자에 정상적으로 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "서지민의 고글이 누락되었으나, 지시된 '높은 곳에서 내려다보는' 하이 앵글 카메라 구도를 충실히 구현하여 프레이밍 우선순위에서 우위를 점함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "복장 디테일(고글)은 일치하나, 필수적인 하이 앵글 구도를 무시하고 인물 눈높이에서 촬영되어 연출 지시를 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 식판을, 서지민은 이현우를 바라봄. 대기열 인원들은 배식구를 향함.",
        "built_space": "식당 배경은 기준과 일치하나, 카메라는 지시된 높은 시점이 아닌 인물의 눈높이에 위치함.",
        "entities": "이현우(환자복), 서지민(연구 가운 및 고글), 식판 위 음식 모두 조건과 일치함.",
        "hard_violations": [],
        "physics": "모든 인물과 사물이 의자, 바닥, 테이블에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 음식을 내려다보고, 서지민은 그를 향함. 대기열은 앞쪽을 주시함.",
        "built_space": "식당 구조가 일치하며, 카메라가 명시된 대로 높은 위치에서 대각선 아래로 테이블을 내려다봄.",
        "entities": "인물의 기본 인상착의와 음식은 일치하나, 기준 이미지에 있는 서지민의 머리 위 고글이 누락됨.",
        "hard_violations": [],
        "physics": "착석 및 기립 인물 모두 바닥과 의자에 정상적으로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "서지민의 고글이 누락되었으나, 지시된 '높은 곳에서 내려다보는' 하이 앵글 카메라 구도를 충실히 구현하여 프레이밍 우선순위에서 우위를 점함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "복장 디테일(고글)은 일치하나, 필수적인 하이 앵글 구도를 무시하고 인물 눈높이에서 촬영되어 연출 지시를 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 식판을, 서지민은 이현우를 바라봄. 대기열 인원들은 배식구를 향함.",
        "built_space": "식당 배경은 기준과 일치하나, 카메라는 지시된 높은 시점이 아닌 인물의 눈높이에 위치함.",
        "entities": "이현우(환자복), 서지민(연구 가운 및 고글), 식판 위 음식 모두 조건과 일치함.",
        "hard_violations": [],
        "physics": "모든 인물과 사물이 의자, 바닥, 테이블에 안정적으로 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "이현우는 음식을 내려다보고, 서지민은 그를 향함. 대기열은 앞쪽을 주시함.",
        "built_space": "식당 구조가 일치하며, 카메라가 명시된 대로 높은 위치에서 대각선 아래로 테이블을 내려다봄.",
        "entities": "인물의 기본 인상착의와 음식은 일치하나, 기준 이미지에 있는 서지민의 머리 위 고글이 누락됨.",
        "hard_violations": [],
        "physics": "착석 및 기립 인물 모두 바닥과 의자에 정상적으로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 후보 모두 허용되지 않은 별도 식사객을 추가했지만, A가 높은 하향 시점과 더 작은 인물 규모로 지정된 와이드 구도에 더 가깝다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "별도 식사객을 다수 추가했으며, 낮고 가까운 시점으로 두 주인공을 크게 잡아 크레인의 높은 진입 위치에서 보는 와이드 구도를 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 이현우는 고개를 숙여 자기 앞 음식 접시를 본다. 오른쪽 서지민은 상체를 앞으로 기울이지만 시선은 같은 음식보다 이현우의 얼굴 쪽을 향한다. 대기자들은 왼쪽 배식대 또는 줄의 진행 방향을 보고 있으며 고개와 팔 자세에 차이가 있다. 줄은 왼쪽 가까운 전경부터 화면 위쪽으로 이어져, 상단 왼쪽 배경에 두라는 지정보다 전경 점유가 크다.",
        "built_space": "주 테이블 1개에 왼쪽과 오른쪽의 착석 의자 2개, 카메라 쪽 빈 의자 1개가 보인다. 맞은편 두 사람이 식사를 사이에 두는 배치는 성립하지만 카메라 쪽 의자 등받이가 열린 면 일부를 가린다. 왼쪽 연속 배식대와 유리 가림막, 줄 안내 기둥, 오른쪽 창과 원형 기둥, 녹색 의자와 흰 테이블 행렬, 뒤쪽 화단은 장소 참조와 부합한다. 높은 위치에서 상판을 내려다보며 두 사람의 앉은 자세와 발까지 읽힌다. 테이블은 하단 중앙보다 오른쪽으로 치우쳐 있다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 흰 반소매 환자복 상하의를 입었다. 서지민은 젊은 동아시아계 여성으로 검은 머리, 흰 연구 가운과 흰 바지를 착용했다. 서지민의 머리는 참조보다 길며 참조의 머리 위 보안경은 보이지 않는다. 국적과 정확한 나이는 영상만으로 확정할 수 없다. 두 사람 앞에는 각각 먹지 않은 빵과 음식, 국그릇, 금속 컵이 놓여 있다. 긴 대기 줄은 존재하지만 별도 테이블에 앉은 식사객 2명과 배식 직원들도 보인다.",
        "hard_violations": [
         "숏이 허용한 두 주인공과 배식 대기 줄 외에, 뒤쪽과 오른쪽 테이블에서 따로 식사하는 인물 2명을 추가했다."
        ],
        "physics": "두 주인공의 골반은 각자 의자 좌판에 놓이고 등받이는 몸 뒤에 있다. 다리는 테이블 아래로 내려가며 신발은 바닥에 닿는다. 접은 팔은 상판에 기대고, 식기와 쟁반은 테이블에 지지된다. 대기자들의 발은 바닥에 놓이며 가방은 어깨끈으로 지지된다. 공중에 떠 있거나 지지점 없이 유지되는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽에서 자기 식사 쪽으로 고개와 눈을 내린다. 서지민은 오른쪽에서 앞으로 기울지만 얼굴과 시선은 음식보다 이현우 쪽을 향한다. 대기자들은 배식대와 앞사람 쪽을 보고 자세와 간격에 약간씩 차이가 있다. 줄은 상단 왼쪽으로 멀어지지만 가까운 대기자들이 왼쪽 전경을 크게 차지한다.",
        "built_space": "주 테이블 1개에 좌우 착석 의자 2개와 앞뒤 빈 의자 2개가 보이며, 이는 장소 참조의 네 의자 배치와 가깝다. 앞쪽 빈 의자 등받이는 카메라 쪽 열린 면 일부를 가린다. 왼쪽 배식대, 유리 가림막과 안내 줄, 오른쪽 대형 창과 기둥, 뒤쪽 야자 화단, 천장 루버는 참조 장소를 잘 유지한다. 다만 카메라는 높은 크레인 위치보다 통상적인 눈높이에 가까운 약한 하향 시점이고, 두 주인공과 테이블이 전경을 크게 채운다.",
        "entities": "이현우는 검은 헝클어진 머리와 흰 반소매 환자복을 입은 젊은 동아시아계 남성으로 보인다. 서지민은 검은 머리의 젊은 동아시아계 여성으로 흰 연구 가운, 흰 바지, 머리 위 보안경과 손목시계를 착용해 참조의 소품을 더 많이 유지한다. 정확한 국적과 나이는 확인할 수 없다. 두 사람 사이에 손대지 않은 빵 2개와 식판 2개, 음식과 국그릇, 금속 컵이 있다. 긴 배식 대기 줄 외에 오른쪽과 중앙 뒤편 여러 테이블에서 별도로 식사하는 사람들이 다수 추가되어 있다.",
        "hard_violations": [
         "숏이 허용한 두 주인공과 배식 대기 줄 외에, 중앙 뒤편과 오른쪽 테이블에서 별도로 식사하는 인물들을 다수 추가했다."
        ],
        "physics": "두 주인공은 각각 좌판에 앉고 등받이가 뒤에 있으며, 앞으로 기댄 팔은 테이블에 지지된다. 하체 일부는 테이블과 화면 경계에 가려져 있어 보이지 않는 발의 접촉 상태는 판단하지 않는다. 식판과 빵 접시, 컵은 상판에 안정적으로 놓여 있다. 줄에 선 사람들은 바닥에 발을 딛고 가방은 어깨에 걸었다. 보이는 범위에서 부유나 불가능한 신체 지지는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 후보 모두 허용되지 않은 별도 식사객을 추가했지만, A가 높은 하향 시점과 더 작은 인물 규모로 지정된 와이드 구도에 더 가깝다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "별도 식사객을 다수 추가했으며, 낮고 가까운 시점으로 두 주인공을 크게 잡아 크레인의 높은 진입 위치에서 보는 와이드 구도를 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 이현우는 고개를 숙여 자기 앞 음식 접시를 본다. 오른쪽 서지민은 상체를 앞으로 기울이지만 시선은 같은 음식보다 이현우의 얼굴 쪽을 향한다. 대기자들은 왼쪽 배식대 또는 줄의 진행 방향을 보고 있으며 고개와 팔 자세에 차이가 있다. 줄은 왼쪽 가까운 전경부터 화면 위쪽으로 이어져, 상단 왼쪽 배경에 두라는 지정보다 전경 점유가 크다.",
        "built_space": "주 테이블 1개에 왼쪽과 오른쪽의 착석 의자 2개, 카메라 쪽 빈 의자 1개가 보인다. 맞은편 두 사람이 식사를 사이에 두는 배치는 성립하지만 카메라 쪽 의자 등받이가 열린 면 일부를 가린다. 왼쪽 연속 배식대와 유리 가림막, 줄 안내 기둥, 오른쪽 창과 원형 기둥, 녹색 의자와 흰 테이블 행렬, 뒤쪽 화단은 장소 참조와 부합한다. 높은 위치에서 상판을 내려다보며 두 사람의 앉은 자세와 발까지 읽힌다. 테이블은 하단 중앙보다 오른쪽으로 치우쳐 있다.",
        "entities": "이현우는 짧고 헝클어진 검은 머리의 젊은 동아시아계 남성으로 보이며 흰 반소매 환자복 상하의를 입었다. 서지민은 젊은 동아시아계 여성으로 검은 머리, 흰 연구 가운과 흰 바지를 착용했다. 서지민의 머리는 참조보다 길며 참조의 머리 위 보안경은 보이지 않는다. 국적과 정확한 나이는 영상만으로 확정할 수 없다. 두 사람 앞에는 각각 먹지 않은 빵과 음식, 국그릇, 금속 컵이 놓여 있다. 긴 대기 줄은 존재하지만 별도 테이블에 앉은 식사객 2명과 배식 직원들도 보인다.",
        "hard_violations": [
         "숏이 허용한 두 주인공과 배식 대기 줄 외에, 뒤쪽과 오른쪽 테이블에서 따로 식사하는 인물 2명을 추가했다."
        ],
        "physics": "두 주인공의 골반은 각자 의자 좌판에 놓이고 등받이는 몸 뒤에 있다. 다리는 테이블 아래로 내려가며 신발은 바닥에 닿는다. 접은 팔은 상판에 기대고, 식기와 쟁반은 테이블에 지지된다. 대기자들의 발은 바닥에 놓이며 가방은 어깨끈으로 지지된다. 공중에 떠 있거나 지지점 없이 유지되는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽에서 자기 식사 쪽으로 고개와 눈을 내린다. 서지민은 오른쪽에서 앞으로 기울지만 얼굴과 시선은 음식보다 이현우 쪽을 향한다. 대기자들은 배식대와 앞사람 쪽을 보고 자세와 간격에 약간씩 차이가 있다. 줄은 상단 왼쪽으로 멀어지지만 가까운 대기자들이 왼쪽 전경을 크게 차지한다.",
        "built_space": "주 테이블 1개에 좌우 착석 의자 2개와 앞뒤 빈 의자 2개가 보이며, 이는 장소 참조의 네 의자 배치와 가깝다. 앞쪽 빈 의자 등받이는 카메라 쪽 열린 면 일부를 가린다. 왼쪽 배식대, 유리 가림막과 안내 줄, 오른쪽 대형 창과 기둥, 뒤쪽 야자 화단, 천장 루버는 참조 장소를 잘 유지한다. 다만 카메라는 높은 크레인 위치보다 통상적인 눈높이에 가까운 약한 하향 시점이고, 두 주인공과 테이블이 전경을 크게 채운다.",
        "entities": "이현우는 검은 헝클어진 머리와 흰 반소매 환자복을 입은 젊은 동아시아계 남성으로 보인다. 서지민은 검은 머리의 젊은 동아시아계 여성으로 흰 연구 가운, 흰 바지, 머리 위 보안경과 손목시계를 착용해 참조의 소품을 더 많이 유지한다. 정확한 국적과 나이는 확인할 수 없다. 두 사람 사이에 손대지 않은 빵 2개와 식판 2개, 음식과 국그릇, 금속 컵이 있다. 긴 배식 대기 줄 외에 오른쪽과 중앙 뒤편 여러 테이블에서 별도로 식사하는 사람들이 다수 추가되어 있다.",
        "hard_violations": [
         "숏이 허용한 두 주인공과 배식 대기 줄 외에, 중앙 뒤편과 오른쪽 테이블에서 별도로 식사하는 인물들을 다수 추가했다."
        ],
        "physics": "두 주인공은 각각 좌판에 앉고 등받이가 뒤에 있으며, 앞으로 기댄 팔은 테이블에 지지된다. 하체 일부는 테이블과 화면 경계에 가려져 있어 보이지 않는 발의 접촉 상태는 판단하지 않는다. 식판과 빵 접시, 컵은 상판에 안정적으로 놓여 있다. 줄에 선 사람들은 바닥에 발을 딛고 가방은 어깨에 걸었다. 보이는 범위에서 부유나 불가능한 신체 지지는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 뒤쪽과 오른쪽 테이블에서 따로 식사하는 인물 2명을 추가했다."
    ],
    "A": [
     "[gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 중앙 뒤편과 오른쪽 테이블에서 별도로 식사하는 인물들을 다수 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1750,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "서지민의 고글이 누락되었으나, 지시된 '높은 곳에서 내려다보는' 하이 앵글 카메라 구도를 충실히 구현하여 프레이밍 우선순위에서 우위를 점함.  ★위반: [gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 뒤쪽과 오른쪽 테이블에서 따로 식사하는 인물 2명을 추가했다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "복장 디테일(고글)은 일치하나, 필수적인 하이 앵글 구도를 무시하고 인물 눈높이에서 촬영되어 연출 지시를 위반함.  ★위반: [gpt-high] 숏이 허용한 두 주인공과 배식 대기 줄 외에, 중앙 뒤편과 오른쪽 테이블에서 별도로 식사하는 인물들을 다수 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_dining_hall_a2331c.png",
    "asset_id": "da12b1e2-cc29-4186-b060-6bb1cb26d7f1",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869875>",
    "asset_id": "bdc552c6-5e2a-4bae-affd-3dbee764db52",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e81-b66d-7666-b8f1-914159fba0e7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S85sh15__bgfirst_bg.png",
   "bg_asset_id": "7676346d-56c3-4cee-aedd-439ca29ca549",
   "bg_record_key": "S85sh15::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "research_dining_hall",
   "groupbg_asset_id": "da12b1e2-cc29-4186-b060-6bb1cb26d7f1"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S85sh24::signage": {
  "fp": "199e51a5e97da9b5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S85sh24": {
  "input_fingerprint": "58a6b6b18d93b92b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 식당 테이블을 거칠게 짚어 몸을 반쯤 일으킨, 역동적인 자세의 이현우 상체.\n\nLOCATION (lock): At the same table inside the research institute's communal cafeteria, under daytime dining-room illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the table's open edge, just outside 서지민's shoulder line, looking obliquely upward at 이현우 in a medium frame. Tilt with his half-rise, keeping both forcefully planted hands and a narrow strip of table along the bottom while his torso occupies the center-right; his face remains turned toward 서지민 offscreen left. Let his rising body provide the principal change, without advancing the camera or redirecting his gaze.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's planted hands) — The near edge crosses the lower frame, with a shallow view across its top; used as Provides a fixed horizontal reference against his upward movement; Bread and food (Still uneaten in front of 이현우); used as Remain peripheral beside his hands, preserving the interrupted meal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve facial tension and hand detail without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dining table, bread and meal dishes, serving area, and daytime lighting from the reference. Exclude experimental cables, glass enclosures, and smoke from the separate laboratory.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bread and meal remain on the dining table, with no eating established during the conversation. 이현우: He remains at the dining table, now visibly shocked and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 식당 테이블을 거칠게 짚어 몸을 반쯤 일으킨, 역동적인 자세의 이현우 상체.\n\nLOCATION (lock): At the same table inside the research institute's communal cafeteria, under daytime dining-room illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the table's open edge, just outside 서지민's shoulder line, looking obliquely upward at 이현우 in a medium frame. Tilt with his half-rise, keeping both forcefully planted hands and a narrow strip of table along the bottom while his torso occupies the center-right; his face remains turned toward 서지민 offscreen left. Let his rising body provide the principal change, without advancing the camera or redirecting his gaze.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's planted hands) — The near edge crosses the lower frame, with a shallow view across its top; used as Provides a fixed horizontal reference against his upward movement; Bread and food (Still uneaten in front of 이현우); used as Remain peripheral beside his hands, preserving the interrupted meal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve facial tension and hand detail without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dining table, bread and meal dishes, serving area, and daytime lighting from the reference. Exclude experimental cables, glass enclosures, and smoke from the separate laboratory.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bread and meal remain on the dining table, with no eating established during the conversation. 이현우: He remains at the dining table, now visibly shocked and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 식당 테이블을 거칠게 짚어 몸을 반쯤 일으킨, 역동적인 자세의 이현우 상체.\n\nLOCATION (lock): At the same table inside the research institute's communal cafeteria, under daytime dining-room illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from the table's open edge, just outside 서지민's shoulder line, looking obliquely upward at 이현우 in a medium frame. Tilt with his half-rise, keeping both forcefully planted hands and a narrow strip of table along the bottom while his torso occupies the center-right; his face remains turned toward 서지민 offscreen left. Let his rising body provide the principal change, without advancing the camera or redirecting his gaze.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Dining table (Supporting both of 이현우's planted hands) — The near edge crosses the lower frame, with a shallow view across its top; used as Provides a fixed horizontal reference against his upward movement; Bread and food (Still uneaten in front of 이현우); used as Remain peripheral beside his hands, preserving the interrupted meal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination and controlled contrast preserve facial tension and hand detail without introducing a new lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dining table, bread and meal dishes, serving area, and daytime lighting from the reference. Exclude experimental cables, glass enclosures, and smoke from the separate laboratory.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bread and meal remain on the dining table, with no eating established during the conversation. 이현우: He remains at the dining table, now visibly shocked and angry.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "이현우의 시선과 몸의 방향은 화면 왼쪽 프레임 밖을 향하고 있다.",
    "built_space": "카메라가 레퍼런스와 동일하게 통로(왼쪽 배식구, 오른쪽 창문)를 정면으로 바라보고 있어, 이현우가 원래 자리가 아닌 맞은편 좌석(서지민의 자리)에 위치하게 되는 심각한 공간 축 오류가 발생했다.",
    "entities": "이현우의 외모, 환자복 양식은 일치하며 명찰에 한글 텍스트가 성공적으로 렌더링되었다. 그러나 샷 텍스트에 명시되지 않은 인물(짧은 머리)의 어깨가 왼쪽 전경에 포함되었고, 배경 인물들도 추가되었다.",
    "hard_violations": [
     "[gemini-pro] a person placed where the staging does not put them (이현우가 맞은편 자리로 잘못 배치됨)",
     "[gemini-pro] invented people or objects (전경 및 배경에 샷 텍스트가 지시하지 않은 인물 추가)",
     "[gpt-high] 이현우만 보여야 하는 장면에 다른 인물의 머리와 어깨를 왼쪽 전경에 추가했다.",
     "[gpt-high] 제외 대상으로 지정된 배식대의 이용자·직원과 창가 착석자를 배경에 포함했다."
    ],
    "physics": "양손이 테이블 모서리와 표면에 단단히 밀착되어 반쯤 일어선 상체의 무게 중심을 자연스럽게 지탱하고 있다."
   },
   {
    "label": "B",
    "direction": "이현우의 시선은 화면 왼쪽 밖을 향하고 있다.",
    "built_space": "A와 마찬가지로 카메라가 통로 방향을 그대로 유지하여, 이현우가 자신의 자리가 아닌 반대편 자리에 앉게 되는 구조적 모순을 낳았다.",
    "entities": "이현우의 의상과 외모는 일치하나 명찰의 글씨가 알아볼 수 없게 뭉개졌다. 샷 텍스트에 없는 인물(긴 머리)의 형태가 화면 왼쪽 전경을 차지하며 배경에도 불필요한 인물이 존재한다.",
    "hard_violations": [
     "[gemini-pro] a person placed where the staging does not put them (이현우의 좌석 위치가 반대로 뒤바뀜)",
     "[gemini-pro] invented people or objects (전경 및 배경에 명시되지 않은 인물 다수 포함)",
     "[gpt-high] 이현우만 보여야 하는 장면에 서지민으로 읽히는 전경 여성의 머리와 어깨를 포함했다.",
     "[gpt-high] 이전 장면의 다른 사람을 제외하라는 지시에도 배식 줄의 여러 사람과 창가 착석자를 노출했다."
    ],
    "physics": "양손으로 테이블을 짚고 몸을 일으키고 있으나 화면 왼쪽(캐릭터의 오른손) 손가락 형태가 비정상적으로 길고 평평하게 묘사되어 다소 어색하다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라 축과 좌석 배치를 완전히 잘못 해석하여 캐릭터의 위치가 뒤바뀌고 허가되지 않은 인물을 추가하는 치명적 오류가 있으나, 명찰 텍스트의 한국어 구현과 역동적인 자세 묘사는 긍정적임."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일하게 주인공을 맞은편 자리에 잘못 배치하고 지시되지 않은 인물을 전경에 추가하는 치명적 위반을 범했으며, 명찰 텍스트가 뭉개지고 손 형태가 부자연스러움."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸의 방향은 화면 왼쪽 프레임 밖을 향하고 있다.",
        "built_space": "카메라가 레퍼런스와 동일하게 통로(왼쪽 배식구, 오른쪽 창문)를 정면으로 바라보고 있어, 이현우가 원래 자리가 아닌 맞은편 좌석(서지민의 자리)에 위치하게 되는 심각한 공간 축 오류가 발생했다.",
        "entities": "이현우의 외모, 환자복 양식은 일치하며 명찰에 한글 텍스트가 성공적으로 렌더링되었다. 그러나 샷 텍스트에 명시되지 않은 인물(짧은 머리)의 어깨가 왼쪽 전경에 포함되었고, 배경 인물들도 추가되었다.",
        "hard_violations": [
         "a person placed where the staging does not put them (이현우가 맞은편 자리로 잘못 배치됨)",
         "invented people or objects (전경 및 배경에 샷 텍스트가 지시하지 않은 인물 추가)"
        ],
        "physics": "양손이 테이블 모서리와 표면에 단단히 밀착되어 반쯤 일어선 상체의 무게 중심을 자연스럽게 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 화면 왼쪽 밖을 향하고 있다.",
        "built_space": "A와 마찬가지로 카메라가 통로 방향을 그대로 유지하여, 이현우가 자신의 자리가 아닌 반대편 자리에 앉게 되는 구조적 모순을 낳았다.",
        "entities": "이현우의 의상과 외모는 일치하나 명찰의 글씨가 알아볼 수 없게 뭉개졌다. 샷 텍스트에 없는 인물(긴 머리)의 형태가 화면 왼쪽 전경을 차지하며 배경에도 불필요한 인물이 존재한다.",
        "hard_violations": [
         "a person placed where the staging does not put them (이현우의 좌석 위치가 반대로 뒤바뀜)",
         "invented people or objects (전경 및 배경에 명시되지 않은 인물 다수 포함)"
        ],
        "physics": "양손으로 테이블을 짚고 몸을 일으키고 있으나 화면 왼쪽(캐릭터의 오른손) 손가락 형태가 비정상적으로 길고 평평하게 묘사되어 다소 어색하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "카메라 축과 좌석 배치를 완전히 잘못 해석하여 캐릭터의 위치가 뒤바뀌고 허가되지 않은 인물을 추가하는 치명적 오류가 있으나, 명찰 텍스트의 한국어 구현과 역동적인 자세 묘사는 긍정적임."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "A와 동일하게 주인공을 맞은편 자리에 잘못 배치하고 지시되지 않은 인물을 전경에 추가하는 치명적 위반을 범했으며, 명찰 텍스트가 뭉개지고 손 형태가 부자연스러움."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 시선과 몸의 방향은 화면 왼쪽 프레임 밖을 향하고 있다.",
        "built_space": "카메라가 레퍼런스와 동일하게 통로(왼쪽 배식구, 오른쪽 창문)를 정면으로 바라보고 있어, 이현우가 원래 자리가 아닌 맞은편 좌석(서지민의 자리)에 위치하게 되는 심각한 공간 축 오류가 발생했다.",
        "entities": "이현우의 외모, 환자복 양식은 일치하며 명찰에 한글 텍스트가 성공적으로 렌더링되었다. 그러나 샷 텍스트에 명시되지 않은 인물(짧은 머리)의 어깨가 왼쪽 전경에 포함되었고, 배경 인물들도 추가되었다.",
        "hard_violations": [
         "a person placed where the staging does not put them (이현우가 맞은편 자리로 잘못 배치됨)",
         "invented people or objects (전경 및 배경에 샷 텍스트가 지시하지 않은 인물 추가)"
        ],
        "physics": "양손이 테이블 모서리와 표면에 단단히 밀착되어 반쯤 일어선 상체의 무게 중심을 자연스럽게 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 화면 왼쪽 밖을 향하고 있다.",
        "built_space": "A와 마찬가지로 카메라가 통로 방향을 그대로 유지하여, 이현우가 자신의 자리가 아닌 반대편 자리에 앉게 되는 구조적 모순을 낳았다.",
        "entities": "이현우의 의상과 외모는 일치하나 명찰의 글씨가 알아볼 수 없게 뭉개졌다. 샷 텍스트에 없는 인물(긴 머리)의 형태가 화면 왼쪽 전경을 차지하며 배경에도 불필요한 인물이 존재한다.",
        "hard_violations": [
         "a person placed where the staging does not put them (이현우의 좌석 위치가 반대로 뒤바뀜)",
         "invented people or objects (전경 및 배경에 명시되지 않은 인물 다수 포함)"
        ],
        "physics": "양손으로 테이블을 짚고 몸을 일으키고 있으나 화면 왼쪽(캐릭터의 오른손) 손가락 형태가 비정상적으로 길고 평평하게 묘사되어 다소 어색하다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "양손 지지와 왼쪽 시선은 맞지만, 화면 밖이어야 할 서지민과 다수의 배경 인물을 노출했고 머리 윗부분이 잘려 상승 동작보다 앞으로 숙인 자세가 강조된다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "양손으로 밀며 반쯤 일어나는 상체와 낮은 미디엄 구도는 더 충실하지만, 금지된 전경·배경 인물이 보여 최종 사용 가능한 후보는 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우의 얼굴과 눈은 화면 왼쪽 전경의 긴 머리 인물을 향한다. 왼쪽을 보는 방향은 맞지만, 시선 대상인 서지민은 지시와 달리 화면 안에 들어와 있다. 양팔은 아래로 뻗어 식탁을 짚고 상체는 앞쪽으로 기울어 있다.",
        "built_space": "전경 식탁 한 개와 이현우 뒤의 연녹색 의자 한 개가 보인다. 왼쪽에는 유리 덮개 배식대 한 줄과 차단봉 행렬, 오른쪽에는 창가 식탁·의자 여러 조와 화분이 있다. 흰 기둥, 금속 다리, 낮 채광은 참조 식당과 대체로 일치한다. 식탁은 하단에 놓였지만 전경 인물의 머리와 어깨가 왼쪽 상당 부분을 가리며, 음식이 주변부보다는 하단 중앙에서 두드러진다. 불가능한 반사는 보이지 않는다.",
        "entities": "중앙 오른쪽에는 젊은 동아시아계 남성 한 명이 있으며 짧고 헝클어진 검은 머리, 흰 브이넥 반소매 환자복과 마른 체격은 이현우 참조에 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 눈썹과 벌어진 입은 놀람과 분노를 표현한다. 식판 한 개 위에 빵을 포함한 음식 접시 한 개, 국그릇 한 개, 금속 컵 한 개가 있고 먹는 행동은 없다. 그러나 왼쪽 전경 여성과 배식 줄의 여러 사람, 오른쪽 창가 착석자가 추가되어 있다. 실험실 케이블·연기·유리 격리실은 없다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 서지민으로 읽히는 전경 여성의 머리와 어깨를 포함했다.",
         "이전 장면의 다른 사람을 제외하라는 지시에도 배식 줄의 여러 사람과 창가 착석자를 노출했다."
        ],
        "physics": "벌린 두 손의 손바닥과 손가락이 식탁 상판에 닿아 기울어진 상체를 지지한다. 팔과 어깨의 하중 전달은 가능하며 뒤쪽 의자에서 몸을 들어 올리는 자세로 읽힌다. 발은 프레임 밖이라 접지는 확인할 수 없지만, 몸이 공중에 떠 있다는 증거는 없다. 식판과 식기 모두 상판에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우의 얼굴과 눈은 화면 왼쪽 가장자리의 흐릿한 전경 인물을 향한다. 요구된 왼쪽 시선은 유지하지만 상대를 완전히 화면 밖에 두지 않았다. 두 팔은 좌우로 벌어져 아래쪽 식탁을 누르며, 상체는 앞으로 기울어진 채 위로 올라오는 순간으로 읽힌다.",
        "built_space": "전경 식탁 한 개와 골반 뒤 연녹색 의자 한 개가 보인다. 왼쪽 배식대 한 줄과 차단봉, 오른쪽 창문 열과 식탁·의자 행렬, 기둥 주변 화분이 참조 식당의 재료와 배치를 대체로 잇는다. 카메라는 상체를 약간 올려다보며 양손과 하단의 얕은 상판을 함께 담는다. 머리 전체가 들어오고 몸통이 중앙 오른쪽을 차지해 A보다 요구된 미디엄 구도에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 남성 한 명의 검은 헝클어진 머리, 흰 반소매 환자복과 체격은 이현우 참조에 대체로 맞는다. 국적은 외관만으로 판별할 수 없다. 찌푸린 눈썹과 열린 입에서 분노와 충격이 보인다. 하단에는 식판 한 개, 빵과 음식이 담긴 접시, 국그릇 한 개, 금속 컵 한 개가 남아 있다. 환자복 가슴 표식의 글자 형태는 참조와 차이가 난다. 왼쪽 전경에 다른 인물의 머리·어깨가 있고 배식대 이용자와 직원, 오른쪽 창가 착석자도 보인다. 실험실 요소는 없다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 다른 인물의 머리와 어깨를 왼쪽 전경에 추가했다.",
         "제외 대상으로 지정된 배식대의 이용자·직원과 창가 착석자를 배경에 포함했다."
        ],
        "physics": "양손이 식탁 위에 넓게 펼쳐져 닿아 있고 팔이 상체의 하중을 받는다. 골반이 의자 앞쪽에서 올라오는 모양과 앞으로 기운 몸통이 반기립 동작으로 자연스럽게 연결된다. 발은 보이지 않지만 확인되는 손 지지가 있어 부유 자세는 아니다. 식판·접시·그릇·컵은 모두 상판 또는 식판에 받쳐져 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "양손 지지와 왼쪽 시선은 맞지만, 화면 밖이어야 할 서지민과 다수의 배경 인물을 노출했고 머리 윗부분이 잘려 상승 동작보다 앞으로 숙인 자세가 강조된다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "양손으로 밀며 반쯤 일어나는 상체와 낮은 미디엄 구도는 더 충실하지만, 금지된 전경·배경 인물이 보여 최종 사용 가능한 후보는 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 얼굴과 눈은 화면 왼쪽 전경의 긴 머리 인물을 향한다. 왼쪽을 보는 방향은 맞지만, 시선 대상인 서지민은 지시와 달리 화면 안에 들어와 있다. 양팔은 아래로 뻗어 식탁을 짚고 상체는 앞쪽으로 기울어 있다.",
        "built_space": "전경 식탁 한 개와 이현우 뒤의 연녹색 의자 한 개가 보인다. 왼쪽에는 유리 덮개 배식대 한 줄과 차단봉 행렬, 오른쪽에는 창가 식탁·의자 여러 조와 화분이 있다. 흰 기둥, 금속 다리, 낮 채광은 참조 식당과 대체로 일치한다. 식탁은 하단에 놓였지만 전경 인물의 머리와 어깨가 왼쪽 상당 부분을 가리며, 음식이 주변부보다는 하단 중앙에서 두드러진다. 불가능한 반사는 보이지 않는다.",
        "entities": "중앙 오른쪽에는 젊은 동아시아계 남성 한 명이 있으며 짧고 헝클어진 검은 머리, 흰 브이넥 반소매 환자복과 마른 체격은 이현우 참조에 대체로 부합한다. 국적은 외관만으로 확인할 수 없다. 눈썹과 벌어진 입은 놀람과 분노를 표현한다. 식판 한 개 위에 빵을 포함한 음식 접시 한 개, 국그릇 한 개, 금속 컵 한 개가 있고 먹는 행동은 없다. 그러나 왼쪽 전경 여성과 배식 줄의 여러 사람, 오른쪽 창가 착석자가 추가되어 있다. 실험실 케이블·연기·유리 격리실은 없다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 서지민으로 읽히는 전경 여성의 머리와 어깨를 포함했다.",
         "이전 장면의 다른 사람을 제외하라는 지시에도 배식 줄의 여러 사람과 창가 착석자를 노출했다."
        ],
        "physics": "벌린 두 손의 손바닥과 손가락이 식탁 상판에 닿아 기울어진 상체를 지지한다. 팔과 어깨의 하중 전달은 가능하며 뒤쪽 의자에서 몸을 들어 올리는 자세로 읽힌다. 발은 프레임 밖이라 접지는 확인할 수 없지만, 몸이 공중에 떠 있다는 증거는 없다. 식판과 식기 모두 상판에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우의 얼굴과 눈은 화면 왼쪽 가장자리의 흐릿한 전경 인물을 향한다. 요구된 왼쪽 시선은 유지하지만 상대를 완전히 화면 밖에 두지 않았다. 두 팔은 좌우로 벌어져 아래쪽 식탁을 누르며, 상체는 앞으로 기울어진 채 위로 올라오는 순간으로 읽힌다.",
        "built_space": "전경 식탁 한 개와 골반 뒤 연녹색 의자 한 개가 보인다. 왼쪽 배식대 한 줄과 차단봉, 오른쪽 창문 열과 식탁·의자 행렬, 기둥 주변 화분이 참조 식당의 재료와 배치를 대체로 잇는다. 카메라는 상체를 약간 올려다보며 양손과 하단의 얕은 상판을 함께 담는다. 머리 전체가 들어오고 몸통이 중앙 오른쪽을 차지해 A보다 요구된 미디엄 구도에 가깝다. 불가능한 반사는 보이지 않는다.",
        "entities": "젊은 동아시아계 남성 한 명의 검은 헝클어진 머리, 흰 반소매 환자복과 체격은 이현우 참조에 대체로 맞는다. 국적은 외관만으로 판별할 수 없다. 찌푸린 눈썹과 열린 입에서 분노와 충격이 보인다. 하단에는 식판 한 개, 빵과 음식이 담긴 접시, 국그릇 한 개, 금속 컵 한 개가 남아 있다. 환자복 가슴 표식의 글자 형태는 참조와 차이가 난다. 왼쪽 전경에 다른 인물의 머리·어깨가 있고 배식대 이용자와 직원, 오른쪽 창가 착석자도 보인다. 실험실 요소는 없다.",
        "hard_violations": [
         "이현우만 보여야 하는 장면에 다른 인물의 머리와 어깨를 왼쪽 전경에 추가했다.",
         "제외 대상으로 지정된 배식대의 이용자·직원과 창가 착석자를 배경에 포함했다."
        ],
        "physics": "양손이 식탁 위에 넓게 펼쳐져 닿아 있고 팔이 상체의 하중을 받는다. 골반이 의자 앞쪽에서 올라오는 모양과 앞으로 기운 몸통이 반기립 동작으로 자연스럽게 연결된다. 발은 보이지 않지만 확인되는 손 지지가 있어 부유 자세는 아니다. 식판·접시·그릇·컵은 모두 상판 또는 식판에 받쳐져 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.417
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.167
   },
   "violations": {
    "A": [
     "[gemini-pro] a person placed where the staging does not put them (이현우가 맞은편 자리로 잘못 배치됨)",
     "[gemini-pro] invented people or objects (전경 및 배경에 샷 텍스트가 지시하지 않은 인물 추가)",
     "[gpt-high] 이현우만 보여야 하는 장면에 다른 인물의 머리와 어깨를 왼쪽 전경에 추가했다.",
     "[gpt-high] 제외 대상으로 지정된 배식대의 이용자·직원과 창가 착석자를 배경에 포함했다."
    ],
    "B": [
     "[gemini-pro] a person placed where the staging does not put them (이현우의 좌석 위치가 반대로 뒤바뀜)",
     "[gemini-pro] invented people or objects (전경 및 배경에 명시되지 않은 인물 다수 포함)",
     "[gpt-high] 이현우만 보여야 하는 장면에 서지민으로 읽히는 전경 여성의 머리와 어깨를 포함했다.",
     "[gpt-high] 이전 장면의 다른 사람을 제외하라는 지시에도 배식 줄의 여러 사람과 창가 착석자를 노출했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1750,
   "B": 1167
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "카메라 축과 좌석 배치를 완전히 잘못 해석하여 캐릭터의 위치가 뒤바뀌고 허가되지 않은 인물을 추가하는 치명적 오류가 있으나, 명찰 텍스트의 한국어 구현과 역동적인 자세 묘사는 긍정적임.  ★위반: [gemini-pro] a person placed where the staging does not put them (이현우가 맞은편 자리로 잘못 배치됨) / [gemini-pro] invented people or objects (전경 및 배경에 샷 텍스트가 지시하지 않은 인물 추가) / [gpt-high] 이현우만 보여야 하는 장면에 다른 인물의 머리와 어깨를 왼쪽 전경에 추가했다. / [gpt-high] 제외 대상으로 지정된 배식대의 이용자·직원과 창가 착석자를 배경에 포함했다."
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "A와 동일하게 주인공을 맞은편 자리에 잘못 배치하고 지시되지 않은 인물을 전경에 추가하는 치명적 위반을 범했으며, 명찰 텍스트가 뭉개지고 손 형태가 부자연스러움.  ★위반: [gemini-pro] a person placed where the staging does not put them (이현우의 좌석 위치가 반대로 뒤바뀜) / [gemini-pro] invented people or objects (전경 및 배경에 명시되지 않은 인물 다수 포함) / [gpt-high] 이현우만 보여야 하는 장면에 서지민으로 읽히는 전경 여성의 머리와 어깨를 포함했다. / [gpt-high] 이전 장면의 다른 사람을 제외하라는 지시에도 배식 줄의 여러 사람과 창가 착석자를 노출했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S85sh15_sel.png",
    "asset_id": "9e0ac45a-fe47-4624-b935-bb3f55bde511",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e8a-c68a-7cf3-86df-7bf6f64ebd89",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S85sh15"
  }
 },
 "S86sh2::signage": {
  "fp": "702b65d3acdad265",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S86sh2": {
  "input_fingerprint": "89d41fc061d9a533",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 문가에 선 이현우가 지소영을 향해 핏대 선 얼굴로 고함치는 상체.\n\nLOCATION (lock): Just inside the doorway of the research-center director's office. Interior lighting is on for the nighttime meeting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly inside the open doorway at 이현우's upper-chest height, offset from his path toward 지소영 and tilted slightly upward. Frame his upper body center-left with the doorway behind him, catching his forward-loaded posture and strained face as he shouts at 지소영 offscreen right. Remain static at this entry-stage endpoint, allowing his impending advance—not a camera push—to supply the pressure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office doorway (Open following 이현우's entrance) — Seen obliquely from inside the office, behind his near shoulder; used as Marks the threshold he is about to leave without enclosing him in a symmetrical frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued interior illumination retains readable facial strain with controlled contrast and no added dramatic color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its modern interior and old central piano. 이현우: He has rushed into the office and is standing in an agitated confrontation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 문가에 선 이현우가 지소영을 향해 핏대 선 얼굴로 고함치는 상체.\n\nLOCATION (lock): Just inside the doorway of the research-center director's office. Interior lighting is on for the nighttime meeting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly inside the open doorway at 이현우's upper-chest height, offset from his path toward 지소영 and tilted slightly upward. Frame his upper body center-left with the doorway behind him, catching his forward-loaded posture and strained face as he shouts at 지소영 offscreen right. Remain static at this entry-stage endpoint, allowing his impending advance—not a camera push—to supply the pressure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office doorway (Open following 이현우's entrance) — Seen obliquely from inside the office, behind his near shoulder; used as Marks the threshold he is about to leave without enclosing him in a symmetrical frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued interior illumination retains readable facial strain with controlled contrast and no added dramatic color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its modern interior and old central piano. 이현우: He has rushed into the office and is standing in an agitated confrontation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 문가에 선 이현우가 지소영을 향해 핏대 선 얼굴로 고함치는 상체.\n\nLOCATION (lock): Just inside the doorway of the research-center director's office. Interior lighting is on for the nighttime meeting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly inside the open doorway at 이현우's upper-chest height, offset from his path toward 지소영 and tilted slightly upward. Frame his upper body center-left with the doorway behind him, catching his forward-loaded posture and strained face as he shouts at 지소영 offscreen right. Remain static at this entry-stage endpoint, allowing his impending advance—not a camera push—to supply the pressure.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office doorway (Open following 이현우's entrance) — Seen obliquely from inside the office, behind his near shoulder; used as Marks the threshold he is about to leave without enclosing him in a symmetrical frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued interior illumination retains readable facial strain with controlled contrast and no added dramatic color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The director's office retains its modern interior and old central piano. 이현우: He has rushed into the office and is standing in an agitated confrontation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 우측 화면 밖을 향해 시선을 두고 고함을 치고 있음.",
    "built_space": "레퍼런스와 일치하는 넓은 연구실 내부, 피아노, 야경이 보이는 창문이 배경에 배치됨. 좌측에 출입문 구조물이 보임.",
    "entities": "흰색 환자복을 입은 이현우의 외형이 잘 구현되었으나, 우측 하단에 프롬프트에서 금지한 타 인물(지소영으로 추정)의 어깨가 포함됨.",
    "hard_violations": [
     "[gemini-pro] invented people (화면 밖 우측에 있어야 할 지소영의 어깨가 화면 내에 등장함)",
     "[gpt-high] 오른쪽 전경에 다른 인물의 어깨와 상체 일부를 추가했다. 이현우만 보여야 하고 지소영은 화면 밖에 있어야 한다는 명시적 조건을 위반한다."
    ],
    "physics": "바닥에 단단히 발을 딛고 상체를 앞으로 기울인 자연스러운 서 있는 자세."
   },
   {
    "label": "B",
    "direction": "인물은 우측 화면 밖을 향해 강하게 시선을 두고 소리치고 있음.",
    "built_space": "레퍼런스와 동일한 피아노, 대리석 바닥, 창밖 야경이 정확한 위치에 렌더링되었으며 출입문 구조가 인물을 둘러싸듯 배치됨.",
    "entities": "흰색 환자복을 입고 흐트러진 머리를 한 이현우가 단독으로 정확히 등장함.",
    "hard_violations": [],
    "physics": "바닥을 지지대로 삼아 상체를 앞으로 싣고 있는 역동적이고 자연스러운 자세."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "이현우가 화면 밖을 향해 소리치는 모습을 잘 포착했으며, 지시된 프레이밍과 배경 공간의 디테일을 단독 인물로 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지소영이 화면 밖(offscreen)에 있어야 한다는 명시적 지시와 달리 우측 전경에 다른 인물의 어깨가 노출되어 치명적인 감점 요인이 되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 화면 밖을 향해 시선을 두고 고함을 치고 있음.",
        "built_space": "레퍼런스와 일치하는 넓은 연구실 내부, 피아노, 야경이 보이는 창문이 배경에 배치됨. 좌측에 출입문 구조물이 보임.",
        "entities": "흰색 환자복을 입은 이현우의 외형이 잘 구현되었으나, 우측 하단에 프롬프트에서 금지한 타 인물(지소영으로 추정)의 어깨가 포함됨.",
        "hard_violations": [
         "invented people (화면 밖 우측에 있어야 할 지소영의 어깨가 화면 내에 등장함)"
        ],
        "physics": "바닥에 단단히 발을 딛고 상체를 앞으로 기울인 자연스러운 서 있는 자세."
       },
       {
        "label": "B",
        "direction": "인물은 우측 화면 밖을 향해 강하게 시선을 두고 소리치고 있음.",
        "built_space": "레퍼런스와 동일한 피아노, 대리석 바닥, 창밖 야경이 정확한 위치에 렌더링되었으며 출입문 구조가 인물을 둘러싸듯 배치됨.",
        "entities": "흰색 환자복을 입고 흐트러진 머리를 한 이현우가 단독으로 정확히 등장함.",
        "hard_violations": [],
        "physics": "바닥을 지지대로 삼아 상체를 앞으로 싣고 있는 역동적이고 자연스러운 자세."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "이현우가 화면 밖을 향해 소리치는 모습을 잘 포착했으며, 지시된 프레이밍과 배경 공간의 디테일을 단독 인물로 충실히 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지소영이 화면 밖(offscreen)에 있어야 한다는 명시적 지시와 달리 우측 전경에 다른 인물의 어깨가 노출되어 치명적인 감점 요인이 되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 화면 밖을 향해 시선을 두고 고함을 치고 있음.",
        "built_space": "레퍼런스와 일치하는 넓은 연구실 내부, 피아노, 야경이 보이는 창문이 배경에 배치됨. 좌측에 출입문 구조물이 보임.",
        "entities": "흰색 환자복을 입은 이현우의 외형이 잘 구현되었으나, 우측 하단에 프롬프트에서 금지한 타 인물(지소영으로 추정)의 어깨가 포함됨.",
        "hard_violations": [
         "invented people (화면 밖 우측에 있어야 할 지소영의 어깨가 화면 내에 등장함)"
        ],
        "physics": "바닥에 단단히 발을 딛고 상체를 앞으로 기울인 자연스러운 서 있는 자세."
       },
       {
        "label": "B",
        "direction": "인물은 우측 화면 밖을 향해 강하게 시선을 두고 소리치고 있음.",
        "built_space": "레퍼런스와 동일한 피아노, 대리석 바닥, 창밖 야경이 정확한 위치에 렌더링되었으며 출입문 구조가 인물을 둘러싸듯 배치됨.",
        "entities": "흰색 환자복을 입고 흐트러진 머리를 한 이현우가 단독으로 정확히 등장함.",
        "hard_violations": [],
        "physics": "바닥을 지지대로 삼아 상체를 앞으로 싣고 있는 역동적이고 자연스러운 자세."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "이현우만 보이는 중간 크기의 상체 구도와 오른쪽을 향한 격앙된 고함은 맞지만, 집무실이 인물 뒤로 펼쳐져 지정된 실내측 출입문 카메라 축과 어긋난다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전방으로 쏠린 고함 자세는 맞지만, 오른쪽 전경에 금지된 다른 인물의 어깨를 추가했고 집무실을 등진 배경 배치도 지정 시점과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선이 화면 오른쪽 바깥을 향하고 입을 벌려 고함친다. 화면 밖 오른쪽의 지소영을 향한다는 지시에 부합하며, 렌즈를 직접 바라보지는 않는다. 어깨와 목도 대치 상대 쪽으로 긴장되어 있다.",
        "built_space": "왼쪽에 손잡이가 달린 열린 출입문 한 곳이 있고, 인물 뒤 오른쪽에 그랜드피아노 한 대와 의자 한 개가 보인다. 뒤쪽에는 소파 한 개의 일부, 그림 한 점, 장스탠드 한 개, 야경이 보이는 유리창 벽과 천장 간접조명이 있다. 재료와 야간 조명은 장소 참조와 유사하다. 다만 문턱 너머로 집무실 내부와 피아노가 펼쳐져, 실내에서 출입문을 돌아보는 지정 시점보다는 입구에서 집무실 안쪽을 바라보는 배치로 읽힌다. 바닥의 조명 반사는 자연스럽다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 한 명뿐이다. 짧고 헝클어진 검은 머리, 마른 체격, 깨끗한 흰색 브이넥 반소매 상의가 이현우의 설정과 대체로 맞는다. 참조는 마스크로 얼굴 하부를 가려 정확한 얼굴 전체의 일치 여부는 확인하기 어렵다. 국적은 외형만으로 확인할 수 없다. 참조의 모자와 마스크, 가슴 표식은 없지만 얼굴을 드러내 고함치는 장면은 구현했다. 바지는 프레임 밖이며, 지소영이나 이전 장면의 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "목의 힘줄과 얼굴 근육의 긴장, 약간 앞으로 기운 상체, 아래로 이어지는 양팔은 서서 고함치는 동작으로 가능하다. 하체와 발의 바닥 접촉은 프레임 밖이므로 확인할 수 없지만 부유를 나타내는 단서는 없다. 피아노는 다리와 바퀴로 바닥에 지지되고 의자와 스탠드도 바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우의 눈과 얼굴은 화면 오른쪽을 향하고, 입을 벌린 채 상체를 그쪽으로 내민다. 시선은 오른쪽 전경에 일부 보이는 상대 인물 쪽으로 읽힌다. 대치 방향 자체는 맞지만 지소영을 완전히 화면 밖에 두라는 지시는 지키지 않았다.",
        "built_space": "왼쪽에 경첩이 보이는 열린 출입문 한 곳과 세로 조명이 있고, 뒤 오른쪽에는 그랜드피아노 한 대와 의자 한 개가 있다. 뒤쪽 벽에는 그림 한 점과 장스탠드 한 개, 소파 일부가 보이며 유리창 너머는 밤이다. 참조의 석재·금속 마감과 간접조명은 유지된다. 그러나 집무실과 중앙 피아노가 인물 뒤에 놓여 A와 마찬가지로 지정된 실내측 출입문 카메라 축과 맞지 않는다. 오른쪽 전경의 어깨는 구도를 사실상 상대 인물을 걸친 대치 숏으로 바꾼다.",
        "entities": "주인물은 젊은 동아시아계 남성으로, 헝클어진 검은 머리와 흰색 브이넥 환자복 상의가 설정에 대체로 맞는다. 참조에서 가려진 얼굴 하부의 정확한 동일성은 검증하기 어렵고, 국적도 외형으로 확인할 수 없다. 모자와 마스크, 가슴 표식은 보이지 않는다. 오른쪽 아래에는 별도 인물의 흐릿한 어깨와 상체 일부가 추가되어 있다. 얼굴은 없지만 이현우와 분리된 신체이며, 이 장면에서 허용된 유일한 가시 인물 조건을 위반한다.",
        "hard_violations": [
         "오른쪽 전경에 다른 인물의 어깨와 상체 일부를 추가했다. 이현우만 보여야 하고 지소영은 화면 밖에 있어야 한다는 명시적 조건을 위반한다."
        ],
        "physics": "이현우는 허리와 어깨를 앞으로 기울이고 목에 힘을 준 자세로, 서서 상대에게 고함치는 동작으로 가능하다. 발은 잘려 있어 직접적인 지지점은 확인할 수 없으나 공중에 떠 있는 모습은 아니다. 오른쪽 인물 역시 일부만 잘린 것이므로 부유로 판단할 근거는 없다. 피아노와 의자는 보이는 다리로 바닥에 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "이현우만 보이는 중간 크기의 상체 구도와 오른쪽을 향한 격앙된 고함은 맞지만, 집무실이 인물 뒤로 펼쳐져 지정된 실내측 출입문 카메라 축과 어긋난다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전방으로 쏠린 고함 자세는 맞지만, 오른쪽 전경에 금지된 다른 인물의 어깨를 추가했고 집무실을 등진 배경 배치도 지정 시점과 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선이 화면 오른쪽 바깥을 향하고 입을 벌려 고함친다. 화면 밖 오른쪽의 지소영을 향한다는 지시에 부합하며, 렌즈를 직접 바라보지는 않는다. 어깨와 목도 대치 상대 쪽으로 긴장되어 있다.",
        "built_space": "왼쪽에 손잡이가 달린 열린 출입문 한 곳이 있고, 인물 뒤 오른쪽에 그랜드피아노 한 대와 의자 한 개가 보인다. 뒤쪽에는 소파 한 개의 일부, 그림 한 점, 장스탠드 한 개, 야경이 보이는 유리창 벽과 천장 간접조명이 있다. 재료와 야간 조명은 장소 참조와 유사하다. 다만 문턱 너머로 집무실 내부와 피아노가 펼쳐져, 실내에서 출입문을 돌아보는 지정 시점보다는 입구에서 집무실 안쪽을 바라보는 배치로 읽힌다. 바닥의 조명 반사는 자연스럽다.",
        "entities": "보이는 사람은 젊은 동아시아계 남성 한 명뿐이다. 짧고 헝클어진 검은 머리, 마른 체격, 깨끗한 흰색 브이넥 반소매 상의가 이현우의 설정과 대체로 맞는다. 참조는 마스크로 얼굴 하부를 가려 정확한 얼굴 전체의 일치 여부는 확인하기 어렵다. 국적은 외형만으로 확인할 수 없다. 참조의 모자와 마스크, 가슴 표식은 없지만 얼굴을 드러내 고함치는 장면은 구현했다. 바지는 프레임 밖이며, 지소영이나 이전 장면의 인물은 보이지 않는다.",
        "hard_violations": [],
        "physics": "목의 힘줄과 얼굴 근육의 긴장, 약간 앞으로 기운 상체, 아래로 이어지는 양팔은 서서 고함치는 동작으로 가능하다. 하체와 발의 바닥 접촉은 프레임 밖이므로 확인할 수 없지만 부유를 나타내는 단서는 없다. 피아노는 다리와 바퀴로 바닥에 지지되고 의자와 스탠드도 바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우의 눈과 얼굴은 화면 오른쪽을 향하고, 입을 벌린 채 상체를 그쪽으로 내민다. 시선은 오른쪽 전경에 일부 보이는 상대 인물 쪽으로 읽힌다. 대치 방향 자체는 맞지만 지소영을 완전히 화면 밖에 두라는 지시는 지키지 않았다.",
        "built_space": "왼쪽에 경첩이 보이는 열린 출입문 한 곳과 세로 조명이 있고, 뒤 오른쪽에는 그랜드피아노 한 대와 의자 한 개가 있다. 뒤쪽 벽에는 그림 한 점과 장스탠드 한 개, 소파 일부가 보이며 유리창 너머는 밤이다. 참조의 석재·금속 마감과 간접조명은 유지된다. 그러나 집무실과 중앙 피아노가 인물 뒤에 놓여 A와 마찬가지로 지정된 실내측 출입문 카메라 축과 맞지 않는다. 오른쪽 전경의 어깨는 구도를 사실상 상대 인물을 걸친 대치 숏으로 바꾼다.",
        "entities": "주인물은 젊은 동아시아계 남성으로, 헝클어진 검은 머리와 흰색 브이넥 환자복 상의가 설정에 대체로 맞는다. 참조에서 가려진 얼굴 하부의 정확한 동일성은 검증하기 어렵고, 국적도 외형으로 확인할 수 없다. 모자와 마스크, 가슴 표식은 보이지 않는다. 오른쪽 아래에는 별도 인물의 흐릿한 어깨와 상체 일부가 추가되어 있다. 얼굴은 없지만 이현우와 분리된 신체이며, 이 장면에서 허용된 유일한 가시 인물 조건을 위반한다.",
        "hard_violations": [
         "오른쪽 전경에 다른 인물의 어깨와 상체 일부를 추가했다. 이현우만 보여야 하고 지소영은 화면 밖에 있어야 한다는 명시적 조건을 위반한다."
        ],
        "physics": "이현우는 허리와 어깨를 앞으로 기울이고 목에 힘을 준 자세로, 서서 상대에게 고함치는 동작으로 가능하다. 발은 잘려 있어 직접적인 지지점은 확인할 수 없으나 공중에 떠 있는 모습은 아니다. 오른쪽 인물 역시 일부만 잘린 것이므로 부유로 판단할 근거는 없다. 피아노와 의자는 보이는 다리로 바닥에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.708,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.458,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (화면 밖 우측에 있어야 할 지소영의 어깨가 화면 내에 등장함)",
     "[gpt-high] 오른쪽 전경에 다른 인물의 어깨와 상체 일부를 추가했다. 이현우만 보여야 하고 지소영은 화면 밖에 있어야 한다는 명시적 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 458
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "이현우가 화면 밖을 향해 소리치는 모습을 잘 포착했으며, 지시된 프레이밍과 배경 공간의 디테일을 단독 인물로 충실히 구현했습니다."
   },
   {
    "label": "A",
    "score": 458,
    "verdict_ko": "지소영이 화면 밖(offscreen)에 있어야 한다는 명시적 지시와 달리 우측 전경에 다른 인물의 어깨가 노출되어 치명적인 감점 요인이 되었습니다.  ★위반: [gemini-pro] invented people (화면 밖 우측에 있어야 할 지소영의 어깨가 화면 내에 등장함) / [gpt-high] 오른쪽 전경에 다른 인물의 어깨와 상체 일부를 추가했다. 이현우만 보여야 하고 지소영은 화면 밖에 있어야 한다는 명시적 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S84sh7_sel.png",
    "asset_id": "51a1fee3-4efa-450b-a9f2-82655c6d4517",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e91-ee36-760e-ae2d-82d8d54e01c8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh7"
  }
 },
 "S86sh7::signage": {
  "fp": "00a001d748fe4e0e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S86sh7": {
  "input_fingerprint": "5451bd962be7df31",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 맹렬한 항의에 말문이 막힌 채 입술을 질끈 깨문 지소영의 굳은 얼굴.\n\nLOCATION (lock): Inside the director's office, facing the visitor near the entrance under nighttime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside 이현우's shoulder, finish the established approach slightly above 지소영's eye line with a gentle downward view, observing her directly rather than through his eyes. Her rigid three-quarter face occupies the center-right while his shoulder remains a soft sliver at the extreme left; she bites her lip and looks toward him, and his head remains directed toward her. Ease the dolly to a stop, making reduced distance the final emphasis while preserving their positions and eyelines.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's subdued illumination and controlled contrast, allowing the bitten lip and arrested expression to remain legible without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The modern director's office and its old piano remain undisturbed. 이현우: He remains standing in the office, visibly angry and insistent. 지소영: She remains in her research coat with an uncomfortable, troubled expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 맹렬한 항의에 말문이 막힌 채 입술을 질끈 깨문 지소영의 굳은 얼굴.\n\nLOCATION (lock): Inside the director's office, facing the visitor near the entrance under nighttime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside 이현우's shoulder, finish the established approach slightly above 지소영's eye line with a gentle downward view, observing her directly rather than through his eyes. Her rigid three-quarter face occupies the center-right while his shoulder remains a soft sliver at the extreme left; she bites her lip and looks toward him, and his head remains directed toward her. Ease the dolly to a stop, making reduced distance the final emphasis while preserving their positions and eyelines.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's subdued illumination and controlled contrast, allowing the bitten lip and arrested expression to remain legible without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The modern director's office and its old piano remain undisturbed. 이현우: He remains standing in the office, visibly angry and insistent. 지소영: She remains in her research coat with an uncomfortable, troubled expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 맹렬한 항의에 말문이 막힌 채 입술을 질끈 깨문 지소영의 굳은 얼굴.\n\nLOCATION (lock): Inside the director's office, facing the visitor near the entrance under nighttime office lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just outside 이현우's shoulder, finish the established approach slightly above 지소영's eye line with a gentle downward view, observing her directly rather than through his eyes. Her rigid three-quarter face occupies the center-right while his shoulder remains a soft sliver at the extreme left; she bites her lip and looks toward him, and his head remains directed toward her. Ease the dolly to a stop, making reduced distance the final emphasis while preserving their positions and eyelines.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's subdued illumination and controlled contrast, allowing the bitten lip and arrested expression to remain legible without a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The modern director's office and its old piano remain undisturbed. 이현우: He remains standing in the office, visibly angry and insistent. 지소영: She remains in her research coat with an uncomfortable, troubled expression.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지소영의 시선은 화면 왼쪽의 이현우를 향하고, 이현우의 고개는 지소영 쪽으로 향해 있음.",
    "built_space": "이전 샷과 동일한 야간 사무실 내부. 창밖의 도시 야경과 우측에 놓인 그랜드 피아노가 정확히 배치됨.",
    "entities": "지소영(입술을 굳게 깨문 50대 여성, 레퍼런스와 일치하는 로고가 박힌 흰색 연구복), 이현우(화면 좌측에 걸친 뒷모습, 흰색 브이넥 의상).",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 서서 대치하는 자연스럽고 안정적인 자세임."
   },
   {
    "label": "B",
    "direction": "지소영이 왼쪽의 이현우를 바라보고 있으며, 이현우의 머리 방향도 지소영을 향함.",
    "built_space": "이전 샷의 배경을 잘 유지함. 우측 피아노와 창밖 야경 등의 요소가 자연스럽게 배치됨.",
    "entities": "지소영(입술을 깨문 표정, 흰색 연구복을 입었으나 로고의 하단부가 잘려 있음), 이현우(좌측 전경을 차지하는 뒷모습).",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 서 있는 안정적인 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요청된 숄더샷 구도, 약간 내려다보는 앵글, 입술을 깨문 표정, 그리고 의상의 로고 디테일까지 레퍼런스를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 분위기와 표정은 우수하나, 카메라 앵글이 다소 평면적이고 연구복의 로고가 잘려 묘사됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 시선은 화면 왼쪽의 이현우를 향하고, 이현우의 고개는 지소영 쪽으로 향해 있음.",
        "built_space": "이전 샷과 동일한 야간 사무실 내부. 창밖의 도시 야경과 우측에 놓인 그랜드 피아노가 정확히 배치됨.",
        "entities": "지소영(입술을 굳게 깨문 50대 여성, 레퍼런스와 일치하는 로고가 박힌 흰색 연구복), 이현우(화면 좌측에 걸친 뒷모습, 흰색 브이넥 의상).",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 서서 대치하는 자연스럽고 안정적인 자세임."
       },
       {
        "label": "B",
        "direction": "지소영이 왼쪽의 이현우를 바라보고 있으며, 이현우의 머리 방향도 지소영을 향함.",
        "built_space": "이전 샷의 배경을 잘 유지함. 우측 피아노와 창밖 야경 등의 요소가 자연스럽게 배치됨.",
        "entities": "지소영(입술을 깨문 표정, 흰색 연구복을 입었으나 로고의 하단부가 잘려 있음), 이현우(좌측 전경을 차지하는 뒷모습).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서 있는 안정적인 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요청된 숄더샷 구도, 약간 내려다보는 앵글, 입술을 깨문 표정, 그리고 의상의 로고 디테일까지 레퍼런스를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전반적인 분위기와 표정은 우수하나, 카메라 앵글이 다소 평면적이고 연구복의 로고가 잘려 묘사됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 시선은 화면 왼쪽의 이현우를 향하고, 이현우의 고개는 지소영 쪽으로 향해 있음.",
        "built_space": "이전 샷과 동일한 야간 사무실 내부. 창밖의 도시 야경과 우측에 놓인 그랜드 피아노가 정확히 배치됨.",
        "entities": "지소영(입술을 굳게 깨문 50대 여성, 레퍼런스와 일치하는 로고가 박힌 흰색 연구복), 이현우(화면 좌측에 걸친 뒷모습, 흰색 브이넥 의상).",
        "hard_violations": [],
        "physics": "두 인물 모두 바닥에 서서 대치하는 자연스럽고 안정적인 자세임."
       },
       {
        "label": "B",
        "direction": "지소영이 왼쪽의 이현우를 바라보고 있으며, 이현우의 머리 방향도 지소영을 향함.",
        "built_space": "이전 샷의 배경을 잘 유지함. 우측 피아노와 창밖 야경 등의 요소가 자연스럽게 배치됨.",
        "entities": "지소영(입술을 깨문 표정, 흰색 연구복을 입었으나 로고의 하단부가 잘려 있음), 이현우(좌측 전경을 차지하는 뒷모습).",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 서 있는 안정적인 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지소영의 굳은 얼굴과 입술을 깨문 연기, 중앙 오른쪽 배치와 가까운 얼굴 크기가 더 충실하지만, 이현우가 왼쪽 가장자리의 가느다란 어깨 조각보다 훨씬 크게 나온다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "서로를 향한 시선과 야간 사무실은 맞지만, 가슴까지 드러낸 넓은 구도와 큰 전경 인물이 얼굴 클로즈업 및 좁은 어깨 조각이라는 핵심 지시에서 더 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영의 얼굴은 왼쪽으로 약간 돌아가 있고 눈은 앞에 선 이현우의 얼굴 쪽을 올려다본다. 왼쪽 전경 이현우도 머리를 지소영 쪽으로 향한다. 서로 마주 보는 방향은 맞으며 카메라를 응시하지 않는다.",
        "built_space": "오른쪽 뒤에 오래된 그랜드피아노 한 대와 일부 보이는 벤치 한 개, 왼쪽 뒤에 스탠드 조명 한 개와 액자 한 개가 있다. 창틀로 나뉜 야경 창, 밝은 석재 바닥, 따뜻한 천장 조명이 이전 장면의 공간 재료와 분위기를 따른다. 두 사람 사이 왼쪽에 출입구의 수직 프레임이 보인다. 중복 피아노나 불가능한 거울 반사는 없다. 지소영은 중앙 오른쪽에 있지만 이현우의 머리와 어깨가 화면 왼쪽을 넓게 점유해 지시된 좁은 전경보다 크다.",
        "entities": "두 사람만 보인다. 지소영은 중년 동아시아계 여성으로 짙은 단발, 얼굴 윤곽, 흰 연구 가운과 짙은 회색 셔츠가 인물 참조와 대체로 맞는다. 입술을 안으로 당겨 깨문 듯한 입 모양과 긴장된 눈가가 보인다. 가운의 청록색 문양은 참조보다 크게 보이고 세부 형태도 다르다. 이현우는 검은 짧은 머리와 흰 상의를 입어 이전 장면의 복장을 따른다. 뒷머리와 귀, 목 일부만 보여 얼굴 정체성이나 분노 표정은 확인할 수 없다. 바지와 손은 구도 밖이다.",
        "hard_violations": [],
        "physics": "두 사람의 머리는 목과 상체에 정상적으로 연결되어 있으며 공중에 뜬 자세나 비정상적인 관절은 보이지 않는다. 발과 하체는 잘려 있어 바닥 접촉은 확인할 수 없다. 피아노와 벤치는 다리로 바닥에 지지되고 스탠드 조명도 바닥에 서 있다. 손에 든 물체나 운동 중인 물체는 없다."
       },
       {
        "label": "B",
        "direction": "지소영은 왼쪽 전경 이현우의 얼굴 방향을 바라보며 머리도 그쪽으로 약간 돌렸다. 이현우의 뒷머리와 귀 방향 역시 지소영을 향한다. 두 사람의 시선 관계는 요구와 맞고 다른 대상에 주의를 돌린 모습은 없다.",
        "built_space": "오른쪽 뒤에 그랜드피아노 한 대와 벤치 한 개, 왼쪽 뒤에 스탠드 조명 한 개와 큰 화분 한 개, 일부 가려진 벽 액자가 보인다. 야경 창, 석재 바닥, 출입구 프레임과 따뜻한 간접조명은 참조 공간과 대체로 일치한다. 고정 설비가 중복되거나 불가능한 반사가 나타나지는 않는다. 다만 A보다 상체와 배경을 넓게 보여주며, 지소영의 얼굴은 중앙에 더 가깝고 이현우의 머리와 어깨도 큰 전경으로 남는다.",
        "entities": "지소영과 이현우에 해당하는 두 사람만 있다. 지소영의 중년 여성 얼굴, 정돈된 검은 단발, 흰 가운과 회색 셔츠는 참조와 대체로 맞는다. 다문 입술을 안쪽으로 누르고 미간을 좁힌 불편한 표정이 보인다. 가운에는 산과 물결 모양의 청록색 표장과 '제주도연구소' 글자가 선명하며, 참조의 작고 다른 세부 형태를 가진 표장과 정확히 일치하지 않는다. 이현우의 흰 상의와 검은 머리는 이전 장면과 맞지만 얼굴은 보이지 않아 세부 신원과 분노 연기는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "두 사람은 정상적인 목과 어깨 연결을 가진 곧은 상체로 보이며, 부유하거나 비틀린 신체는 없다. 하체가 화면 밖이므로 발의 지지는 직접 확인할 수 없다. 피아노와 벤치는 바닥에 놓인 다리로, 조명과 화분은 바닥의 받침으로 지지된다. 입술을 누르는 표정은 실제 근육 움직임으로 가능한 범위다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지소영의 굳은 얼굴과 입술을 깨문 연기, 중앙 오른쪽 배치와 가까운 얼굴 크기가 더 충실하지만, 이현우가 왼쪽 가장자리의 가느다란 어깨 조각보다 훨씬 크게 나온다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "서로를 향한 시선과 야간 사무실은 맞지만, 가슴까지 드러낸 넓은 구도와 큰 전경 인물이 얼굴 클로즈업 및 좁은 어깨 조각이라는 핵심 지시에서 더 멀어진다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "지소영의 얼굴은 왼쪽으로 약간 돌아가 있고 눈은 앞에 선 이현우의 얼굴 쪽을 올려다본다. 왼쪽 전경 이현우도 머리를 지소영 쪽으로 향한다. 서로 마주 보는 방향은 맞으며 카메라를 응시하지 않는다.",
        "built_space": "오른쪽 뒤에 오래된 그랜드피아노 한 대와 일부 보이는 벤치 한 개, 왼쪽 뒤에 스탠드 조명 한 개와 액자 한 개가 있다. 창틀로 나뉜 야경 창, 밝은 석재 바닥, 따뜻한 천장 조명이 이전 장면의 공간 재료와 분위기를 따른다. 두 사람 사이 왼쪽에 출입구의 수직 프레임이 보인다. 중복 피아노나 불가능한 거울 반사는 없다. 지소영은 중앙 오른쪽에 있지만 이현우의 머리와 어깨가 화면 왼쪽을 넓게 점유해 지시된 좁은 전경보다 크다.",
        "entities": "두 사람만 보인다. 지소영은 중년 동아시아계 여성으로 짙은 단발, 얼굴 윤곽, 흰 연구 가운과 짙은 회색 셔츠가 인물 참조와 대체로 맞는다. 입술을 안으로 당겨 깨문 듯한 입 모양과 긴장된 눈가가 보인다. 가운의 청록색 문양은 참조보다 크게 보이고 세부 형태도 다르다. 이현우는 검은 짧은 머리와 흰 상의를 입어 이전 장면의 복장을 따른다. 뒷머리와 귀, 목 일부만 보여 얼굴 정체성이나 분노 표정은 확인할 수 없다. 바지와 손은 구도 밖이다.",
        "hard_violations": [],
        "physics": "두 사람의 머리는 목과 상체에 정상적으로 연결되어 있으며 공중에 뜬 자세나 비정상적인 관절은 보이지 않는다. 발과 하체는 잘려 있어 바닥 접촉은 확인할 수 없다. 피아노와 벤치는 다리로 바닥에 지지되고 스탠드 조명도 바닥에 서 있다. 손에 든 물체나 운동 중인 물체는 없다."
       },
       {
        "label": "A",
        "direction": "지소영은 왼쪽 전경 이현우의 얼굴 방향을 바라보며 머리도 그쪽으로 약간 돌렸다. 이현우의 뒷머리와 귀 방향 역시 지소영을 향한다. 두 사람의 시선 관계는 요구와 맞고 다른 대상에 주의를 돌린 모습은 없다.",
        "built_space": "오른쪽 뒤에 그랜드피아노 한 대와 벤치 한 개, 왼쪽 뒤에 스탠드 조명 한 개와 큰 화분 한 개, 일부 가려진 벽 액자가 보인다. 야경 창, 석재 바닥, 출입구 프레임과 따뜻한 간접조명은 참조 공간과 대체로 일치한다. 고정 설비가 중복되거나 불가능한 반사가 나타나지는 않는다. 다만 A보다 상체와 배경을 넓게 보여주며, 지소영의 얼굴은 중앙에 더 가깝고 이현우의 머리와 어깨도 큰 전경으로 남는다.",
        "entities": "지소영과 이현우에 해당하는 두 사람만 있다. 지소영의 중년 여성 얼굴, 정돈된 검은 단발, 흰 가운과 회색 셔츠는 참조와 대체로 맞는다. 다문 입술을 안쪽으로 누르고 미간을 좁힌 불편한 표정이 보인다. 가운에는 산과 물결 모양의 청록색 표장과 '제주도연구소' 글자가 선명하며, 참조의 작고 다른 세부 형태를 가진 표장과 정확히 일치하지 않는다. 이현우의 흰 상의와 검은 머리는 이전 장면과 맞지만 얼굴은 보이지 않아 세부 신원과 분노 연기는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "두 사람은 정상적인 목과 어깨 연결을 가진 곧은 상체로 보이며, 부유하거나 비틀린 신체는 없다. 하체가 화면 밖이므로 발의 지지는 직접 확인할 수 없다. 피아노와 벤치는 바닥에 놓인 다리로, 조명과 화분은 바닥의 받침으로 지지된다. 입술을 누르는 표정은 실제 근육 움직임으로 가능한 범위다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.857
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "요청된 숄더샷 구도, 약간 내려다보는 앵글, 입술을 깨문 표정, 그리고 의상의 로고 디테일까지 레퍼런스를 충실히 구현함."
   },
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "전반적인 분위기와 표정은 우수하나, 카메라 앵글이 다소 평면적이고 연구복의 로고가 잘려 묘사됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이현우 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S86sh2_sel.png",
    "asset_id": "6c5cd4fe-1cdd-4d7e-abf9-39c3ffe1826b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:783266>",
    "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1220012>",
    "asset_id": "19c64d1e-6653-45db-810d-33fcd86e1d51",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e96-b52c-7ffa-a637-34e57c9a14c4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S86sh2"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S87sh5::signage": {
  "fp": "8f93abd9cc7accec",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::5382913fe03bc98e": {
  "subjects": [],
  "subject_text": "유빅사 회장실·지휘 사무실\n대형 화면과 집무용 책상이 있는 사무실. 화면 앞에 마이크와 통신 장비가 배치되어 있다.",
  "identity": "canonical",
  "scope_id": "L95",
  "scope_role": "location_interior",
  "scope_sha": "1657209f0dc0d145"
 },
 "S87sh5::bgfirst_bg": {
  "input_fingerprint": "6bb79016e2ad6a3b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이강준이 턱을 치켜든 채 제주도 출동을 결연하게 명령하는 정면 상체.\n\nLOCATION (lock): In the special-operations commander's military office, at the briefing position. Office lighting is on at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the direct observational approach beside the offscreen reporting soldier, staying below 이강준's eye line and oblique to their shared axis. Place 이강준's upper body center-right, with his raised chin clearly separated from his shoulders as he delivers the order toward the soldier offscreen left, never toward the lens. Hold the document low at the frame boundary and emphasize only his lifted eyeline as the approach settles before the explicit cut to the other office.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Report document (Presented to 이강준) — Only an oblique portion of the document face is visible; its contents are not resolved; used as Provides a small lower-frame trace of the report that prompted the order.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled contrast convey authority without inventing a special source or colored accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이강준이 턱을 치켜든 채 제주도 출동을 결연하게 명령하는 정면 상체.\n\nLOCATION (lock): In the special-operations commander's military office, at the briefing position. Office lighting is on at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the direct observational approach beside the offscreen reporting soldier, staying below 이강준's eye line and oblique to their shared axis. Place 이강준's upper body center-right, with his raised chin clearly separated from his shoulders as he delivers the order toward the soldier offscreen left, never toward the lens. Hold the document low at the frame boundary and emphasize only his lifted eyeline as the approach settles before the explicit cut to the other office.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Report document (Presented to 이강준) — Only an oblique portion of the document face is visible; its contents are not resolved; used as Provides a small lower-frame trace of the report that prompted the order.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled contrast convey authority without inventing a special source or colored accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh5__bgfirst_bg.png",
  "asset_id": "bcf31164-8100-40ee-9f2a-4feaf9f1472e",
  "input_asset_ids": [
   "663f08c8-2257-4e87-9f27-d0e064181563",
   "cfc43618-b868-4ec2-bf22-9433b5630788"
  ]
 },
 "S87sh5": {
  "input_fingerprint": "af4ad17fc3304150",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 턱을 치켜든 채 제주도 출동을 결연하게 명령하는 정면 상체.\n\nLOCATION (lock): In the special-operations commander's military office, at the briefing position. Office lighting is on at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the direct observational approach beside the offscreen reporting soldier, staying below 이강준's eye line and oblique to their shared axis. Place 이강준's upper body center-right, with his raised chin clearly separated from his shoulders as he delivers the order toward the soldier offscreen left, never toward the lens. Hold the document low at the frame boundary and emphasize only his lifted eyeline as the approach settles before the explicit cut to the other office.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Report document (Presented to 이강준) — Only an oblique portion of the document face is visible; its contents are not resolved; used as Provides a small lower-frame trace of the report that prompted the order.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled contrast convey authority without inventing a special source or colored accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The report document concerning the sunken vessel and the butterfly drones has been presented in the special task force commander's office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (한국인 남성, 성숙한 얼굴, 짧게 깎은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 턱을 치켜든 채 제주도 출동을 결연하게 명령하는 정면 상체.\n\nLOCATION (lock): In the special-operations commander's military office, at the briefing position. Office lighting is on at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the direct observational approach beside the offscreen reporting soldier, staying below 이강준's eye line and oblique to their shared axis. Place 이강준's upper body center-right, with his raised chin clearly separated from his shoulders as he delivers the order toward the soldier offscreen left, never toward the lens. Hold the document low at the frame boundary and emphasize only his lifted eyeline as the approach settles before the explicit cut to the other office.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Report document (Presented to 이강준) — Only an oblique portion of the document face is visible; its contents are not resolved; used as Provides a small lower-frame trace of the report that prompted the order.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled contrast convey authority without inventing a special source or colored accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The report document concerning the sunken vessel and the butterfly drones has been presented in the special task force commander's office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (한국인 남성, 성숙한 얼굴, 짧게 깎은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이강준이 턱을 치켜든 채 제주도 출동을 결연하게 명령하는 정면 상체.\n\nLOCATION (lock): In the special-operations commander's military office, at the briefing position. Office lighting is on at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the direct observational approach beside the offscreen reporting soldier, staying below 이강준's eye line and oblique to their shared axis. Place 이강준's upper body center-right, with his raised chin clearly separated from his shoulders as he delivers the order toward the soldier offscreen left, never toward the lens. Hold the document low at the frame boundary and emphasize only his lifted eyeline as the approach settles before the explicit cut to the other office.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Report document (Presented to 이강준) — Only an oblique portion of the document face is visible; its contents are not resolved; used as Provides a small lower-frame trace of the report that prompted the order.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained interior illumination and controlled contrast convey authority without inventing a special source or colored accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The report document concerning the sunken vessel and the butterfly drones has been presented in the special task force commander's office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이강준 (한국인 남성, 성숙한 얼굴, 짧게 깎은 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh5__bgfirst_bg.png",
     "asset_id": "bcf31164-8100-40ee-9f2a-4feaf9f1472e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S87sh5.png",
     "asset_id": "663f08c8-2257-4e87-9f27-d0e064181563",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이강준: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1244522>",
     "asset_id": "3ea5ce75-3b78-4132-a20c-1169884871f0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L95B01.png",
     "asset_id": "cfc43618-b868-4ec2-bf22-9433b5630788",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 이강준: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1244522>",
     "asset_id": "3ea5ce75-3b78-4132-a20c-1169884871f0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "이강준은 왼쪽의 병사를 향해 시선을 던지며 턱을 치켜들고 있습니다.",
    "built_space": "배경에 메인 책상과 임원용 의자가 명확히 보임에도 불구하고, 인물이 전경의 또 다른 책상과 임원용 의자에 앉아 있어 공간 구조가 왜곡되었습니다.",
    "entities": "이강준의 얼굴과 스타일은 레퍼런스와 일치하며, 병사가 들고 있는 클립보드와 문서가 확인됩니다.",
    "hard_violations": [
     "[gemini-pro] 공간 내 유일한 메인 책상과 임원용 의자를 중복으로 배치함 (duplicated fittings)",
     "[gpt-high] 화면 밖이어야 하는 보고 병사의 머리 일부·몸통·손을 노출해, 이강준만 등장해야 한다는 명시적 제한을 위반했다."
    ],
    "physics": "병사의 손이 클립보드를 물리적으로 지지하고 있으며, 인물은 의자에 안정적으로 착석해 있습니다."
   },
   {
    "label": "B",
    "direction": "이강준은 턱을 확연히 치켜든 채 화면 왼쪽 오프스크린의 병사를 올려다보며 명령하는 시선을 보입니다.",
    "built_space": "레퍼런스 사진과 정확히 일치하게 메인 책상 뒤 임원용 의자에 착석해 있으며, 배경의 조명 책장과 우측 창문, 코트 랙의 위치가 완벽합니다.",
    "entities": "이강준의 인물 레퍼런스가 잘 반영되었고, 프레임 하단에 요구된 보고서 문서가 자리하고 있습니다.",
    "hard_violations": [
     "[gpt-high] 화면 밖에 있어야 하는 보고 병사의 몸통을 왼쪽 전경에 추가해, 이강준 외에는 인물을 보이지 말라는 명시적 제한을 위반했다."
    ],
    "physics": "인물의 오른손이 문서를 단단히 쥐고 지지하며, 상체는 의자에 자연스럽게 체중이 실려 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 공간에 하나뿐인 메인 책상과 의자를 복제하여 인물을 가상의 두 번째 책상에 배치한 치명적인 공간 오류가 있어 실격입니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "기준 공간의 정확한 위치에 인물을 배치하였고, 턱을 치켜든 시선, 카메라 앵글, 하단의 문서 배치 등 프롬프트의 요구사항을 훌륭하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이강준은 왼쪽의 병사를 향해 시선을 던지며 턱을 치켜들고 있습니다.",
        "built_space": "배경에 메인 책상과 임원용 의자가 명확히 보임에도 불구하고, 인물이 전경의 또 다른 책상과 임원용 의자에 앉아 있어 공간 구조가 왜곡되었습니다.",
        "entities": "이강준의 얼굴과 스타일은 레퍼런스와 일치하며, 병사가 들고 있는 클립보드와 문서가 확인됩니다.",
        "hard_violations": [
         "공간 내 유일한 메인 책상과 임원용 의자를 중복으로 배치함 (duplicated fittings)"
        ],
        "physics": "병사의 손이 클립보드를 물리적으로 지지하고 있으며, 인물은 의자에 안정적으로 착석해 있습니다."
       },
       {
        "label": "B",
        "direction": "이강준은 턱을 확연히 치켜든 채 화면 왼쪽 오프스크린의 병사를 올려다보며 명령하는 시선을 보입니다.",
        "built_space": "레퍼런스 사진과 정확히 일치하게 메인 책상 뒤 임원용 의자에 착석해 있으며, 배경의 조명 책장과 우측 창문, 코트 랙의 위치가 완벽합니다.",
        "entities": "이강준의 인물 레퍼런스가 잘 반영되었고, 프레임 하단에 요구된 보고서 문서가 자리하고 있습니다.",
        "hard_violations": [],
        "physics": "인물의 오른손이 문서를 단단히 쥐고 지지하며, 상체는 의자에 자연스럽게 체중이 실려 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 공간에 하나뿐인 메인 책상과 의자를 복제하여 인물을 가상의 두 번째 책상에 배치한 치명적인 공간 오류가 있어 실격입니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "기준 공간의 정확한 위치에 인물을 배치하였고, 턱을 치켜든 시선, 카메라 앵글, 하단의 문서 배치 등 프롬프트의 요구사항을 훌륭하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이강준은 왼쪽의 병사를 향해 시선을 던지며 턱을 치켜들고 있습니다.",
        "built_space": "배경에 메인 책상과 임원용 의자가 명확히 보임에도 불구하고, 인물이 전경의 또 다른 책상과 임원용 의자에 앉아 있어 공간 구조가 왜곡되었습니다.",
        "entities": "이강준의 얼굴과 스타일은 레퍼런스와 일치하며, 병사가 들고 있는 클립보드와 문서가 확인됩니다.",
        "hard_violations": [
         "공간 내 유일한 메인 책상과 임원용 의자를 중복으로 배치함 (duplicated fittings)"
        ],
        "physics": "병사의 손이 클립보드를 물리적으로 지지하고 있으며, 인물은 의자에 안정적으로 착석해 있습니다."
       },
       {
        "label": "B",
        "direction": "이강준은 턱을 확연히 치켜든 채 화면 왼쪽 오프스크린의 병사를 올려다보며 명령하는 시선을 보입니다.",
        "built_space": "레퍼런스 사진과 정확히 일치하게 메인 책상 뒤 임원용 의자에 착석해 있으며, 배경의 조명 책장과 우측 창문, 코트 랙의 위치가 완벽합니다.",
        "entities": "이강준의 인물 레퍼런스가 잘 반영되었고, 프레임 하단에 요구된 보고서 문서가 자리하고 있습니다.",
        "hard_violations": [],
        "physics": "인물의 오른손이 문서를 단단히 쥐고 지지하며, 상체는 의자에 자연스럽게 체중이 실려 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "화면 밖이어야 할 보고 병사를 노출한 실격 결함이 있지만, 중앙 오른쪽의 상체 크기와 턱을 든 발화 순간, 낮게 든 문서는 B보다 지시에 가깝다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 보고 병사의 얼굴·몸·손을 노출하며, 넓어진 배경과 크게 드러난 문서, 입을 다문 반응 표정이 요구된 결연한 명령 순간에서 더 멀다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이강준은 턱을 들고 화면 왼쪽 위의 군복 차림 인물을 바라본다. 렌즈를 보지 않고 상대에게 말하는 방향은 맞지만, 그 상대가 화면 밖이 아니라 왼쪽 전경에 보인다. 문서는 아래쪽에서 비스듬히 기울어져 있고 내용은 식별되지 않는다.",
        "built_space": "검은 업무용 의자 하나에 앉아 전경 책상 너머로 상대를 대한다. 뒤쪽에는 두 단의 조명 선반, 오른쪽에는 야경 창, 외투가 걸린 옷걸이 하나와 세로 조명 하나가 보인다. 참조 사무실의 주요 재료와 설비 배치에 대체로 부합하며, 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "성숙한 동아시아계 남성의 얼굴, 짧은 검은 머리, 남색 셔츠와 체격은 이강준 참조와 가깝다. 보고 문서로 읽히는 종이가 하단에 있으며 내용을 읽을 수 없어 침몰선·드론 관련 여부는 확인할 수 없다. 왼쪽의 군복 입은 추가 인물 몸통은 이강준만 보여야 한다는 조건을 위반한다.",
        "hard_violations": [
         "화면 밖에 있어야 하는 보고 병사의 몸통을 왼쪽 전경에 추가해, 이강준 외에는 인물을 보이지 말라는 명시적 제한을 위반했다."
        ],
        "physics": "이강준의 몸은 검은 의자로 지지되고, 아래쪽에 보이는 손이 종이 가장자리를 잡는다. 팔과 책상의 관계도 자연스럽다. 추가 인물은 프레임 밖으로 몸이 이어지는 서 있는 자세이며, 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이강준의 눈과 들린 턱은 왼쪽 위의 전경 인물을 향하고 렌즈는 향하지 않는다. 그러나 보고 병사가 화면 안에 크게 들어와 있다. 문서의 넓은 면과 흐릿한 줄들이 카메라 쪽에 보이며, 이강준에게 제시되는 면의 작은 비스듬한 흔적이라는 지시보다 노출이 크다.",
        "built_space": "이강준은 전경의 검은 업무용 의자 하나에 앉고, 바로 앞에는 별도 탁자처럼 보이는 상판이 있다. 큰 업무용 책상 하나는 그의 뒤에 떨어져 있다. 뒤쪽 조명 선반 두 단, 모니터 하나, 세로 조명 하나, 외투 걸린 옷걸이 하나, 오른쪽 창과 안락의자 하나가 보인다. 사무실 설비는 참조와 유사하지만 업무용 의자와 책상의 관계가 달라졌고, A보다 실내 전체를 넓게 드러낸다.",
        "entities": "이강준의 성숙한 얼굴, 짧은 검은 머리와 남색 셔츠는 참조에 가깝다. 왼쪽에는 다른 남성의 머리 일부와 목, 몸통, 문서를 잡은 손이 보인다. 하단 문서에는 흐릿한 글줄이 있지만 내용은 판독되지 않는다. 추가 남성의 노출은 인물 제한에 어긋난다.",
        "hard_violations": [
         "화면 밖이어야 하는 보고 병사의 머리 일부·몸통·손을 노출해, 이강준만 등장해야 한다는 명시적 제한을 위반했다."
        ],
        "physics": "이강준은 의자 등받이와 좌면에 지지되는 자연스러운 착석 자세이고, 오른팔은 탁자 높이로 내려온다. 전경 문서철은 하단의 손이 잡고 있어 떠 있지 않다. 왼쪽 인물 역시 서 있는 몸이 프레임 밖으로 이어지며, 명백히 불가능한 자세나 무지지 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "화면 밖이어야 할 보고 병사를 노출한 실격 결함이 있지만, 중앙 오른쪽의 상체 크기와 턱을 든 발화 순간, 낮게 든 문서는 B보다 지시에 가깝다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 보고 병사의 얼굴·몸·손을 노출하며, 넓어진 배경과 크게 드러난 문서, 입을 다문 반응 표정이 요구된 결연한 명령 순간에서 더 멀다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이강준은 턱을 들고 화면 왼쪽 위의 군복 차림 인물을 바라본다. 렌즈를 보지 않고 상대에게 말하는 방향은 맞지만, 그 상대가 화면 밖이 아니라 왼쪽 전경에 보인다. 문서는 아래쪽에서 비스듬히 기울어져 있고 내용은 식별되지 않는다.",
        "built_space": "검은 업무용 의자 하나에 앉아 전경 책상 너머로 상대를 대한다. 뒤쪽에는 두 단의 조명 선반, 오른쪽에는 야경 창, 외투가 걸린 옷걸이 하나와 세로 조명 하나가 보인다. 참조 사무실의 주요 재료와 설비 배치에 대체로 부합하며, 불가능한 반사나 명백한 설비 중복은 보이지 않는다.",
        "entities": "성숙한 동아시아계 남성의 얼굴, 짧은 검은 머리, 남색 셔츠와 체격은 이강준 참조와 가깝다. 보고 문서로 읽히는 종이가 하단에 있으며 내용을 읽을 수 없어 침몰선·드론 관련 여부는 확인할 수 없다. 왼쪽의 군복 입은 추가 인물 몸통은 이강준만 보여야 한다는 조건을 위반한다.",
        "hard_violations": [
         "화면 밖에 있어야 하는 보고 병사의 몸통을 왼쪽 전경에 추가해, 이강준 외에는 인물을 보이지 말라는 명시적 제한을 위반했다."
        ],
        "physics": "이강준의 몸은 검은 의자로 지지되고, 아래쪽에 보이는 손이 종이 가장자리를 잡는다. 팔과 책상의 관계도 자연스럽다. 추가 인물은 프레임 밖으로 몸이 이어지는 서 있는 자세이며, 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이강준의 눈과 들린 턱은 왼쪽 위의 전경 인물을 향하고 렌즈는 향하지 않는다. 그러나 보고 병사가 화면 안에 크게 들어와 있다. 문서의 넓은 면과 흐릿한 줄들이 카메라 쪽에 보이며, 이강준에게 제시되는 면의 작은 비스듬한 흔적이라는 지시보다 노출이 크다.",
        "built_space": "이강준은 전경의 검은 업무용 의자 하나에 앉고, 바로 앞에는 별도 탁자처럼 보이는 상판이 있다. 큰 업무용 책상 하나는 그의 뒤에 떨어져 있다. 뒤쪽 조명 선반 두 단, 모니터 하나, 세로 조명 하나, 외투 걸린 옷걸이 하나, 오른쪽 창과 안락의자 하나가 보인다. 사무실 설비는 참조와 유사하지만 업무용 의자와 책상의 관계가 달라졌고, A보다 실내 전체를 넓게 드러낸다.",
        "entities": "이강준의 성숙한 얼굴, 짧은 검은 머리와 남색 셔츠는 참조에 가깝다. 왼쪽에는 다른 남성의 머리 일부와 목, 몸통, 문서를 잡은 손이 보인다. 하단 문서에는 흐릿한 글줄이 있지만 내용은 판독되지 않는다. 추가 남성의 노출은 인물 제한에 어긋난다.",
        "hard_violations": [
         "화면 밖이어야 하는 보고 병사의 머리 일부·몸통·손을 노출해, 이강준만 등장해야 한다는 명시적 제한을 위반했다."
        ],
        "physics": "이강준은 의자 등받이와 좌면에 지지되는 자연스러운 착석 자세이고, 오른팔은 탁자 높이로 내려온다. 전경 문서철은 하단의 손이 잡고 있어 떠 있지 않다. 왼쪽 인물 역시 서 있는 몸이 프레임 밖으로 이어지며, 명백히 불가능한 자세나 무지지 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.875,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.625,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 공간 내 유일한 메인 책상과 임원용 의자를 중복으로 배치함 (duplicated fittings)",
     "[gpt-high] 화면 밖이어야 하는 보고 병사의 머리 일부·몸통·손을 노출해, 이강준만 등장해야 한다는 명시적 제한을 위반했다."
    ],
    "B": [
     "[gpt-high] 화면 밖에 있어야 하는 보고 병사의 몸통을 왼쪽 전경에 추가해, 이강준 외에는 인물을 보이지 말라는 명시적 제한을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 625,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 625,
    "verdict_ko": "지정된 공간에 하나뿐인 메인 책상과 의자를 복제하여 인물을 가상의 두 번째 책상에 배치한 치명적인 공간 오류가 있어 실격입니다.  ★위반: [gemini-pro] 공간 내 유일한 메인 책상과 임원용 의자를 중복으로 배치함 (duplicated fittings) / [gpt-high] 화면 밖이어야 하는 보고 병사의 머리 일부·몸통·손을 노출해, 이강준만 등장해야 한다는 명시적 제한을 위반했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "기준 공간의 정확한 위치에 인물을 배치하였고, 턱을 치켜든 시선, 카메라 앵글, 하단의 문서 배치 등 프롬프트의 요구사항을 훌륭하게 구현했습니다.  ★위반: [gpt-high] 화면 밖에 있어야 하는 보고 병사의 몸통을 왼쪽 전경에 추가해, 이강준 외에는 인물을 보이지 말라는 명시적 제한을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L95B01.png",
    "asset_id": "cfc43618-b868-4ec2-bf22-9433b5630788",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 이강준: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1244522>",
    "asset_id": "3ea5ce75-3b78-4132-a20c-1169884871f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0e9b-6ea2-7563-a13d-4f742b681eb2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh5__bgfirst_bg.png",
   "bg_asset_id": "bcf31164-8100-40ee-9f2a-4feaf9f1472e",
   "bg_record_key": "S87sh5::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S87sh6::signage": {
  "fp": "16309a5a8d3a9733",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S87sh6::bgfirst_bg": {
  "input_fingerprint": "91c57bfee9a02e27",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 유빅사 회장실, 윤성찬이 책상 위 스피커폰을 향해 몸을 기울이고 있는 상체.\n\nLOCATION (lock): At the desk and speakerphone inside a corporate chairman's dark private office. Only subdued nighttime office illumination is established.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After an explicit spatial cut, observe 윤성찬 directly from beside his desk, above his eye line and oblique to his forward body axis, at the opening position of the approach. His upper body occupies the right half while the speakerphone sits small in the lower-left desk area; he inclines his torso and lowers his gaze toward it, absorbed in the intercepted order. Hold this entry composition before beginning the push, withholding his later turn toward the commander.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office desk (Supporting the speakerphone) — Its top is seen diagonally from the side, occupying the lower portion of the frame; used as Connects 윤성찬's leaning posture to the listening device; Speakerphone (On the desk as 윤성찬 listens) — Seen obliquely from above, without requiring readable controls; used as Small secondary focal point beneath his downward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with subdued brightness and enough tonal separation to read 윤성찬's face and the speakerphone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 유빅사 회장실, 윤성찬이 책상 위 스피커폰을 향해 몸을 기울이고 있는 상체.\n\nLOCATION (lock): At the desk and speakerphone inside a corporate chairman's dark private office. Only subdued nighttime office illumination is established.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After an explicit spatial cut, observe 윤성찬 directly from beside his desk, above his eye line and oblique to his forward body axis, at the opening position of the approach. His upper body occupies the right half while the speakerphone sits small in the lower-left desk area; he inclines his torso and lowers his gaze toward it, absorbed in the intercepted order. Hold this entry composition before beginning the push, withholding his later turn toward the commander.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office desk (Supporting the speakerphone) — Its top is seen diagonally from the side, occupying the lower portion of the frame; used as Connects 윤성찬's leaning posture to the listening device; Speakerphone (On the desk as 윤성찬 listens) — Seen obliquely from above, without requiring readable controls; used as Small secondary focal point beneath his downward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with subdued brightness and enough tonal separation to read 윤성찬's face and the speakerphone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh6__bgfirst_bg.png",
  "asset_id": "e67d4cfd-bcde-48f1-92c8-b7f449b43755",
  "input_asset_ids": [
   "851e1c65-9a7b-41ce-9268-bbdfce653ec6",
   "cfc43618-b868-4ec2-bf22-9433b5630788"
  ]
 },
 "S87sh6": {
  "input_fingerprint": "665206f7fba5f901",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 유빅사 회장실, 윤성찬이 책상 위 스피커폰을 향해 몸을 기울이고 있는 상체.\n\nLOCATION (lock): At the desk and speakerphone inside a corporate chairman's dark private office. Only subdued nighttime office illumination is established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After an explicit spatial cut, observe 윤성찬 directly from beside his desk, above his eye line and oblique to his forward body axis, at the opening position of the approach. His upper body occupies the right half while the speakerphone sits small in the lower-left desk area; he inclines his torso and lowers his gaze toward it, absorbed in the intercepted order. Hold this entry composition before beginning the push, withholding his later turn toward the commander.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office desk (Supporting the speakerphone) — Its top is seen diagonally from the side, occupying the lower portion of the frame; used as Connects 윤성찬's leaning posture to the listening device; Speakerphone (On the desk as 윤성찬 listens) — Seen obliquely from above, without requiring readable controls; used as Small secondary focal point beneath his downward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with subdued brightness and enough tonal separation to read 윤성찬's face and the speakerphone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 유빅사 회장실, 윤성찬이 책상 위 스피커폰을 향해 몸을 기울이고 있는 상체.\n\nLOCATION (lock): At the desk and speakerphone inside a corporate chairman's dark private office. Only subdued nighttime office illumination is established. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After an explicit spatial cut, observe 윤성찬 directly from beside his desk, above his eye line and oblique to his forward body axis, at the opening position of the approach. His upper body occupies the right half while the speakerphone sits small in the lower-left desk area; he inclines his torso and lowers his gaze toward it, absorbed in the intercepted order. Hold this entry composition before beginning the push, withholding his later turn toward the commander.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office desk (Supporting the speakerphone) — Its top is seen diagonally from the side, occupying the lower portion of the frame; used as Connects 윤성찬's leaning posture to the listening device; Speakerphone (On the desk as 윤성찬 listens) — Seen obliquely from above, without requiring readable controls; used as Small secondary focal point beneath his downward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with subdued brightness and enough tonal separation to read 윤성찬's face and the speakerphone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 유빅사 회장실, 윤성찬이 책상 위 스피커폰을 향해 몸을 기울이고 있는 상체.\n\nLOCATION (lock): At the desk and speakerphone inside a corporate chairman's dark private office. Only subdued nighttime office illumination is established. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After an explicit spatial cut, observe 윤성찬 directly from beside his desk, above his eye line and oblique to his forward body axis, at the opening position of the approach. His upper body occupies the right half while the speakerphone sits small in the lower-left desk area; he inclines his torso and lowers his gaze toward it, absorbed in the intercepted order. Hold this entry composition before beginning the push, withholding his later turn toward the commander.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Office desk (Supporting the speakerphone) — Its top is seen diagonally from the side, occupying the lower portion of the frame; used as Connects 윤성찬's leaning posture to the listening device; Speakerphone (On the desk as 윤성찬 listens) — Seen obliquely from above, without requiring readable controls; used as Small secondary focal point beneath his downward eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the office dim as described, with subdued brightness and enough tonal separation to read 윤성찬's face and the speakerphone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh6__bgfirst_bg.png",
     "asset_id": "e67d4cfd-bcde-48f1-92c8-b7f449b43755",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S87sh6.png",
     "asset_id": "851e1c65-9a7b-41ce-9268-bbdfce653ec6",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L95B01.png",
     "asset_id": "cfc43618-b868-4ec2-bf22-9433b5630788",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선은 책상 위 스피커폰을 향함.",
    "built_space": "기준 공간의 좌측 벽(문)과 우측 벽(창문) 요소가 인물 뒤편의 단일 벽면에 함께 뒤섞여 배치됨.",
    "entities": "윤성찬의 인상착의와 맞춤 정장 일치. 책상 위 스피커폰 존재.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 무대 배치 (기준 공간의 구조가 엉망으로 섞임)",
     "[gpt-high] 장소 참조에서 막힌 벽으로 이어지는 출입문 쪽과 진열장 사이에 창벽을 만들어, 고정된 창벽의 위치와 실내 구조를 변경했습니다."
    ],
    "physics": "양팔이 책상 위에 얹혀 몸을 지지함."
   },
   {
    "label": "B",
    "direction": "시선은 책상 위 스피커폰을 향해 아래로 향함.",
    "built_space": "책상 좌측에서 우측을 바라보는 카메라 시점에 맞춰 의자, 뒤쪽 책장, 우측 창문의 배치가 논리적으로 구현됨.",
    "entities": "윤성찬의 얼굴 특징, 헤어, 의상이 완벽히 일치. 책상 좌측 하단에 스피커폰 위치.",
    "hard_violations": [],
    "physics": "양손이 책상을 단단히 짚어 기울어진 상체의 체중을 지지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물이 스피커폰을 향해 몸을 기울인 지정된 구도를 잘 살렸으며, 카메라 시점에 따른 기준 공간의 입체적인 구조를 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 동작은 지시사항에 부합하나, 기준 공간의 좌측 문과 우측 창문이 한쪽 벽면에 뒤섞이는 치명적인 공간 왜곡이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 책상 위 스피커폰을 향함.",
        "built_space": "기준 공간의 좌측 벽(문)과 우측 벽(창문) 요소가 인물 뒤편의 단일 벽면에 함께 뒤섞여 배치됨.",
        "entities": "윤성찬의 인상착의와 맞춤 정장 일치. 책상 위 스피커폰 존재.",
        "hard_violations": [
         "물리적으로 불가능한 무대 배치 (기준 공간의 구조가 엉망으로 섞임)"
        ],
        "physics": "양팔이 책상 위에 얹혀 몸을 지지함."
       },
       {
        "label": "B",
        "direction": "시선은 책상 위 스피커폰을 향해 아래로 향함.",
        "built_space": "책상 좌측에서 우측을 바라보는 카메라 시점에 맞춰 의자, 뒤쪽 책장, 우측 창문의 배치가 논리적으로 구현됨.",
        "entities": "윤성찬의 얼굴 특징, 헤어, 의상이 완벽히 일치. 책상 좌측 하단에 스피커폰 위치.",
        "hard_violations": [],
        "physics": "양손이 책상을 단단히 짚어 기울어진 상체의 체중을 지지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물이 스피커폰을 향해 몸을 기울인 지정된 구도를 잘 살렸으며, 카메라 시점에 따른 기준 공간의 입체적인 구조를 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 동작은 지시사항에 부합하나, 기준 공간의 좌측 문과 우측 창문이 한쪽 벽면에 뒤섞이는 치명적인 공간 왜곡이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 책상 위 스피커폰을 향함.",
        "built_space": "기준 공간의 좌측 벽(문)과 우측 벽(창문) 요소가 인물 뒤편의 단일 벽면에 함께 뒤섞여 배치됨.",
        "entities": "윤성찬의 인상착의와 맞춤 정장 일치. 책상 위 스피커폰 존재.",
        "hard_violations": [
         "물리적으로 불가능한 무대 배치 (기준 공간의 구조가 엉망으로 섞임)"
        ],
        "physics": "양팔이 책상 위에 얹혀 몸을 지지함."
       },
       {
        "label": "B",
        "direction": "시선은 책상 위 스피커폰을 향해 아래로 향함.",
        "built_space": "책상 좌측에서 우측을 바라보는 카메라 시점에 맞춰 의자, 뒤쪽 책장, 우측 창문의 배치가 논리적으로 구현됨.",
        "entities": "윤성찬의 얼굴 특징, 헤어, 의상이 완벽히 일치. 책상 좌측 하단에 스피커폰 위치.",
        "hard_violations": [],
        "physics": "양손이 책상을 단단히 짚어 기울어진 상체의 체중을 지지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 상체와 왼쪽 아래의 작은 스피커폰, 이를 향한 기울임과 시선, 야간 회장실의 공간 관계가 잘 맞지만 카메라의 하향각은 다소 약합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "상체 구도와 스피커폰을 향한 청취 자세는 충실하지만, 출입문 쪽 벽과 진열장 사이에 창벽을 배치해 고정된 회장실 구조를 바꾸었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬의 몸통과 얼굴이 화면 왼쪽 아래로 기울고, 내려간 시선은 책상 위 스피커폰 방향을 향합니다. 다른 사람이나 카메라를 돌아보지 않습니다. 스피커폰의 조작부는 인물보다 카메라 쪽을 향하지만, 화면을 읽거나 조작하지 않고 소리를 듣는 순간이므로 청취 동작 자체와는 충돌하지 않습니다.",
        "built_space": "하단에 대각선으로 놓인 목재 책상 하나, 왼쪽 뒤에 회장용 의자 하나와 조명 진열장, 그 오른쪽에 세로 조명 하나와 옷걸이 하나, 오른쪽 배경에 블라인드가 내려온 야간 창벽이 보입니다. 진열장과 창벽의 관계는 장소 참조와 대체로 일치합니다. 인물은 의자에 앉지 않고 책상의 오른쪽 끝 부근에 서서 상판에 몸을 기대고 있습니다. 상체는 오른쪽 절반, 작은 스피커폰은 왼쪽 아래에 놓입니다. 카메라는 책상 옆의 사선 시점이지만 명확히 내려다보는 각도는 강하지 않습니다.",
        "entities": "사람은 노년의 동아시아계 남성 한 명뿐입니다. 짧게 정돈한 은백색 머리, 콧수염, 주름진 얼굴과 체형이 윤성찬 참조에 가깝고, 차콜 그레이 정장과 검은 넥타이, 흰 셔츠 및 포켓스퀘어가 일치합니다. 책상 위 검은 회의용 스피커폰 한 대가 실제 기기로 보입니다. 목재, 가죽, 정장 원단의 재질이 자연스럽고 야간 창밖과 절제된 실내 조명이 확인됩니다. 자막이나 덧씌운 표식은 없습니다.",
        "hard_violations": [],
        "physics": "양손이 책상 상판에 닿아 앞으로 기울인 상체를 지탱합니다. 하체 일부는 책상에 가려져 있지만 서서 책상을 짚는 자세로 자연스럽게 이어지며, 부유하는 몸으로 보이지 않습니다. 스피커폰과 문구류는 상판에 놓여 있고, 의자는 바닥에 놓인 정상적인 가구로 보입니다."
       },
       {
        "label": "B",
        "direction": "윤성찬은 몸을 앞으로 숙이고 고개와 눈을 왼쪽 아래의 스피커폰 쪽으로 내립니다. 손을 모은 채 듣고 있으며 다른 사람 쪽으로 돌아보지 않습니다. 스피커폰은 카메라에 주로 스피커 그릴과 외장을 보이고 있어 청취 상황에 어긋나는 명백한 조작 방향 문제는 없습니다.",
        "built_space": "목재 책상 하나가 하단을 대각선으로 채우고, 인물 뒤 오른쪽에는 회장용 의자 하나와 조명 진열장이 있습니다. 왼쪽에는 출입문 하나, 벽걸이 그림 하나, 낮은 수납장 하나와 큰 화분 하나가 보입니다. 그런데 출입문·그림이 있는 왼쪽 벽과 뒤쪽 진열장 사이에 야간 창벽이 들어갑니다. 참조에서는 창벽이 진열장 오른쪽 벽에 있으며 왼쪽은 막힌 벽이므로, 현재 배치는 동일 공간을 다른 각도에서 본 것으로 설명되지 않습니다. 오른쪽 상체와 왼쪽 아래 스피커폰의 배치는 맞지만 스피커폰의 화면 점유율은 A보다 큽니다.",
        "entities": "노년의 동아시아계 남성 한 명만 등장합니다. 은백색의 짧은 머리, 콧수염, 얼굴 주름, 차콜 정장과 넥타이 및 포켓스퀘어는 윤성찬 참조와 가깝습니다. 책상 위에는 회의용 스피커폰 한 대와 가죽 패드, 문서철, 펜꽂이가 보입니다. 재질과 피부는 실사로 읽히며, 창밖은 밤입니다. 읽을 수 있는 새 문구나 화면 위 오버레이는 보이지 않습니다.",
        "hard_violations": [
         "장소 참조에서 막힌 벽으로 이어지는 출입문 쪽과 진열장 사이에 창벽을 만들어, 고정된 창벽의 위치와 실내 구조를 변경했습니다."
        ],
        "physics": "인물은 등받이가 뒤에 있는 회장용 의자에 앉아 상체를 숙이고, 양팔과 모은 손을 책상에 기대고 있습니다. 골반은 가려졌지만 의자와 몸의 관계가 자연스럽고 팔의 접촉도 정상적입니다. 스피커폰, 패드, 문서철과 펜꽂이는 모두 상판의 지지를 받으며 공중에 뜬 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 상체와 왼쪽 아래의 작은 스피커폰, 이를 향한 기울임과 시선, 야간 회장실의 공간 관계가 잘 맞지만 카메라의 하향각은 다소 약합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "상체 구도와 스피커폰을 향한 청취 자세는 충실하지만, 출입문 쪽 벽과 진열장 사이에 창벽을 배치해 고정된 회장실 구조를 바꾸었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "윤성찬의 몸통과 얼굴이 화면 왼쪽 아래로 기울고, 내려간 시선은 책상 위 스피커폰 방향을 향합니다. 다른 사람이나 카메라를 돌아보지 않습니다. 스피커폰의 조작부는 인물보다 카메라 쪽을 향하지만, 화면을 읽거나 조작하지 않고 소리를 듣는 순간이므로 청취 동작 자체와는 충돌하지 않습니다.",
        "built_space": "하단에 대각선으로 놓인 목재 책상 하나, 왼쪽 뒤에 회장용 의자 하나와 조명 진열장, 그 오른쪽에 세로 조명 하나와 옷걸이 하나, 오른쪽 배경에 블라인드가 내려온 야간 창벽이 보입니다. 진열장과 창벽의 관계는 장소 참조와 대체로 일치합니다. 인물은 의자에 앉지 않고 책상의 오른쪽 끝 부근에 서서 상판에 몸을 기대고 있습니다. 상체는 오른쪽 절반, 작은 스피커폰은 왼쪽 아래에 놓입니다. 카메라는 책상 옆의 사선 시점이지만 명확히 내려다보는 각도는 강하지 않습니다.",
        "entities": "사람은 노년의 동아시아계 남성 한 명뿐입니다. 짧게 정돈한 은백색 머리, 콧수염, 주름진 얼굴과 체형이 윤성찬 참조에 가깝고, 차콜 그레이 정장과 검은 넥타이, 흰 셔츠 및 포켓스퀘어가 일치합니다. 책상 위 검은 회의용 스피커폰 한 대가 실제 기기로 보입니다. 목재, 가죽, 정장 원단의 재질이 자연스럽고 야간 창밖과 절제된 실내 조명이 확인됩니다. 자막이나 덧씌운 표식은 없습니다.",
        "hard_violations": [],
        "physics": "양손이 책상 상판에 닿아 앞으로 기울인 상체를 지탱합니다. 하체 일부는 책상에 가려져 있지만 서서 책상을 짚는 자세로 자연스럽게 이어지며, 부유하는 몸으로 보이지 않습니다. 스피커폰과 문구류는 상판에 놓여 있고, 의자는 바닥에 놓인 정상적인 가구로 보입니다."
       },
       {
        "label": "A",
        "direction": "윤성찬은 몸을 앞으로 숙이고 고개와 눈을 왼쪽 아래의 스피커폰 쪽으로 내립니다. 손을 모은 채 듣고 있으며 다른 사람 쪽으로 돌아보지 않습니다. 스피커폰은 카메라에 주로 스피커 그릴과 외장을 보이고 있어 청취 상황에 어긋나는 명백한 조작 방향 문제는 없습니다.",
        "built_space": "목재 책상 하나가 하단을 대각선으로 채우고, 인물 뒤 오른쪽에는 회장용 의자 하나와 조명 진열장이 있습니다. 왼쪽에는 출입문 하나, 벽걸이 그림 하나, 낮은 수납장 하나와 큰 화분 하나가 보입니다. 그런데 출입문·그림이 있는 왼쪽 벽과 뒤쪽 진열장 사이에 야간 창벽이 들어갑니다. 참조에서는 창벽이 진열장 오른쪽 벽에 있으며 왼쪽은 막힌 벽이므로, 현재 배치는 동일 공간을 다른 각도에서 본 것으로 설명되지 않습니다. 오른쪽 상체와 왼쪽 아래 스피커폰의 배치는 맞지만 스피커폰의 화면 점유율은 A보다 큽니다.",
        "entities": "노년의 동아시아계 남성 한 명만 등장합니다. 은백색의 짧은 머리, 콧수염, 얼굴 주름, 차콜 정장과 넥타이 및 포켓스퀘어는 윤성찬 참조와 가깝습니다. 책상 위에는 회의용 스피커폰 한 대와 가죽 패드, 문서철, 펜꽂이가 보입니다. 재질과 피부는 실사로 읽히며, 창밖은 밤입니다. 읽을 수 있는 새 문구나 화면 위 오버레이는 보이지 않습니다.",
        "hard_violations": [
         "장소 참조에서 막힌 벽으로 이어지는 출입문 쪽과 진열장 사이에 창벽을 만들어, 고정된 창벽의 위치와 실내 구조를 변경했습니다."
        ],
        "physics": "인물은 등받이가 뒤에 있는 회장용 의자에 앉아 상체를 숙이고, 양팔과 모은 손을 책상에 기대고 있습니다. 골반은 가려졌지만 의자와 몸의 관계가 자연스럽고 팔의 접촉도 정상적입니다. 스피커폰, 패드, 문서철과 펜꽂이는 모두 상판의 지지를 받으며 공중에 뜬 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.929,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.679,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 무대 배치 (기준 공간의 구조가 엉망으로 섞임)",
     "[gpt-high] 장소 참조에서 막힌 벽으로 이어지는 출입문 쪽과 진열장 사이에 창벽을 만들어, 고정된 창벽의 위치와 실내 구조를 변경했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 679
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "인물이 스피커폰을 향해 몸을 기울인 지정된 구도를 잘 살렸으며, 카메라 시점에 따른 기준 공간의 입체적인 구조를 정확하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 679,
    "verdict_ko": "인물과 동작은 지시사항에 부합하나, 기준 공간의 좌측 문과 우측 창문이 한쪽 벽면에 뒤섞이는 치명적인 공간 왜곡이 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 무대 배치 (기준 공간의 구조가 엉망으로 섞임) / [gpt-high] 장소 참조에서 막힌 벽으로 이어지는 출입문 쪽과 진열장 사이에 창벽을 만들어, 고정된 창벽의 위치와 실내 구조를 변경했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/episodes/e565764e-23fa-4991-a743-975fd8d061df/images/background_chain/L95B01.png",
    "asset_id": "cfc43618-b868-4ec2-bf22-9433b5630788",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ea2-65b2-7452-998f-0b7649f827ed",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh6__bgfirst_bg.png",
   "bg_asset_id": "e67d4cfd-bcde-48f1-92c8-b7f449b43755",
   "bg_record_key": "S87sh6::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S87sh9::signage": {
  "fp": "346995da69e10cbb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S87sh9": {
  "input_fingerprint": "e50b30057f58145a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 시선을 마주하며 꼿꼿한 자세로 고개를 끄덕이는 유빅 용병대장의 상체.\n\nLOCATION (lock): On the visitor's side of the desk in the corporate chairman's dimly lit office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track just beyond 윤성찬's near shoulder, slightly below the 유빅 용병대장's eye line and on the established side of their dialogue axis. Observe the commander directly in the center-right midground, catching the downward phase of his nod with his torso held firm in acknowledgment; 윤성찬's turned shoulder and partial head remain at the left foreground edge. Their attention stays on each other, with the commander's nod supplying the shot's principal change rather than a further camera move.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim office illumination and controlled contrast, keeping the commander's acknowledgment readable without brightening the setting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the executive office's desk, speakerphone, interior finishes, and subdued nighttime lighting from the reference. Exclude military-office furnishings and the officer from the separate command scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 유빅 용병대장 (한국인 남성, 성인의 얼굴, 짧은 검은 머리); 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 시선을 마주하며 꼿꼿한 자세로 고개를 끄덕이는 유빅 용병대장의 상체.\n\nLOCATION (lock): On the visitor's side of the desk in the corporate chairman's dimly lit office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track just beyond 윤성찬's near shoulder, slightly below the 유빅 용병대장's eye line and on the established side of their dialogue axis. Observe the commander directly in the center-right midground, catching the downward phase of his nod with his torso held firm in acknowledgment; 윤성찬's turned shoulder and partial head remain at the left foreground edge. Their attention stays on each other, with the commander's nod supplying the shot's principal change rather than a further camera move.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim office illumination and controlled contrast, keeping the commander's acknowledgment readable without brightening the setting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the executive office's desk, speakerphone, interior finishes, and subdued nighttime lighting from the reference. Exclude military-office furnishings and the officer from the separate command scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 유빅 용병대장 (한국인 남성, 성인의 얼굴, 짧은 검은 머리); 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 윤성찬의 시선을 마주하며 꼿꼿한 자세로 고개를 끄덕이는 유빅 용병대장의 상체.\n\nLOCATION (lock): On the visitor's side of the desk in the corporate chairman's dimly lit office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the lateral track just beyond 윤성찬's near shoulder, slightly below the 유빅 용병대장's eye line and on the established side of their dialogue axis. Observe the commander directly in the center-right midground, catching the downward phase of his nod with his torso held firm in acknowledgment; 윤성찬's turned shoulder and partial head remain at the left foreground edge. Their attention stays on each other, with the commander's nod supplying the shot's principal change rather than a further camera move.\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the dim office illumination and controlled contrast, keeping the commander's acknowledgment readable without brightening the setting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the executive office's desk, speakerphone, interior finishes, and subdued nighttime lighting from the reference. Exclude military-office furnishings and the officer from the separate command scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 유빅 용병대장 (한국인 남성, 성인의 얼굴, 짧은 검은 머리); 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "용병대장의 시선이 좌측 전경의 윤성찬을 향하며 고개를 끄덕이는 동작을 보임.",
    "built_space": "어두운 사무실 배경, 데스크, 스피커폰이 새로운 앵글에 맞춰 정확한 위치와 방향으로 배치됨.",
    "entities": "용병대장의 인상과 남색 티셔츠 의상이 참조 이미지와 일치하며, 윤성찬의 뒷모습 의상도 이전 샷과 동일함.",
    "hard_violations": [],
    "physics": "인물의 상체와 책상 위의 사물들이 모두 올바르게 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "용병대장의 시선이 좌측의 윤성찬을 향해 아래로 향하고 있음.",
    "built_space": "사무실 배경은 유지되었으나, 스피커폰이 이전 샷과 달리 화면 우측 하단으로 잘못 이동됨.",
    "entities": "용병대장의 얼굴은 일치하나 참조 이미지에 없는 검은색 재킷을 입어 의상 지시를 위반함.",
    "hard_violations": [],
    "physics": "사물과 인물이 물리적 어색함 없이 바닥과 책상에 의해 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "용병대장의 의상, 스피커폰의 위치 및 카메라 구도 등 프롬프트와 참조 이미지의 세부 지시사항을 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라 구도는 적절하나, 용병대장이 지시와 달리 재킷을 입고 있으며 스피커폰의 위치가 이전 샷과 일치하지 않음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "용병대장의 시선이 좌측 전경의 윤성찬을 향하며 고개를 끄덕이는 동작을 보임.",
        "built_space": "어두운 사무실 배경, 데스크, 스피커폰이 새로운 앵글에 맞춰 정확한 위치와 방향으로 배치됨.",
        "entities": "용병대장의 인상과 남색 티셔츠 의상이 참조 이미지와 일치하며, 윤성찬의 뒷모습 의상도 이전 샷과 동일함.",
        "hard_violations": [],
        "physics": "인물의 상체와 책상 위의 사물들이 모두 올바르게 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "용병대장의 시선이 좌측의 윤성찬을 향해 아래로 향하고 있음.",
        "built_space": "사무실 배경은 유지되었으나, 스피커폰이 이전 샷과 달리 화면 우측 하단으로 잘못 이동됨.",
        "entities": "용병대장의 얼굴은 일치하나 참조 이미지에 없는 검은색 재킷을 입어 의상 지시를 위반함.",
        "hard_violations": [],
        "physics": "사물과 인물이 물리적 어색함 없이 바닥과 책상에 의해 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "용병대장의 의상, 스피커폰의 위치 및 카메라 구도 등 프롬프트와 참조 이미지의 세부 지시사항을 충실하게 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라 구도는 적절하나, 용병대장이 지시와 달리 재킷을 입고 있으며 스피커폰의 위치가 이전 샷과 일치하지 않음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "용병대장의 시선이 좌측 전경의 윤성찬을 향하며 고개를 끄덕이는 동작을 보임.",
        "built_space": "어두운 사무실 배경, 데스크, 스피커폰이 새로운 앵글에 맞춰 정확한 위치와 방향으로 배치됨.",
        "entities": "용병대장의 인상과 남색 티셔츠 의상이 참조 이미지와 일치하며, 윤성찬의 뒷모습 의상도 이전 샷과 동일함.",
        "hard_violations": [],
        "physics": "인물의 상체와 책상 위의 사물들이 모두 올바르게 지지되어 있음."
       },
       {
        "label": "B",
        "direction": "용병대장의 시선이 좌측의 윤성찬을 향해 아래로 향하고 있음.",
        "built_space": "사무실 배경은 유지되었으나, 스피커폰이 이전 샷과 달리 화면 우측 하단으로 잘못 이동됨.",
        "entities": "용병대장의 얼굴은 일치하나 참조 이미지에 없는 검은색 재킷을 입어 의상 지시를 위반함.",
        "hard_violations": [],
        "physics": "사물과 인물이 물리적 어색함 없이 바닥과 책상에 의해 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "왼쪽 전경 윤성찬을 향한 눈길과 고개를 낮추는 순간을 더 잘 연결하지만, 참조에 없는 재킷과 약간 앞으로 기운 상체는 불일치한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "참조의 남색 상의와 꼿꼿한 상체는 충실하지만, 눈길이 윤성찬의 눈보다 책상 쪽으로 떨어져 핵심인 시선 교환이 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "용병대장은 턱을 내린 상태에서 눈을 화면 왼쪽으로 돌려 전경 윤성찬의 얼굴 방향을 본다. 윤성찬도 머리를 용병대장 쪽으로 돌렸으며, 눈 자체는 뒤쪽 시점 때문에 보이지 않는다. 고개를 낮추면서 상대를 의식하는 관계가 읽힌다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "목재 책상 하나가 두 사람 사이에 있고, 용병대장은 그 너머 중앙 오른쪽의 검은 의자 하나에 앉아 있다. 윤성찬의 부분 머리와 어깨는 왼쪽 전경에 걸린다. 책상에는 스피커폰 하나, 펜꽂이 하나와 수납함 일부가 보인다. 뒤에는 세로 조명 하나, 외투가 걸린 옷걸이 하나, 큰 화분 하나, 왼쪽 조명 선반과 야경 창문이 있다. 참조의 재료와 야간 조명은 이어지며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 스피커폰의 화면상 위치는 참조와 달라졌지만 카메라도 달라져 이것만으로 공간 모순을 확정할 수 없다.",
        "entities": "등장인물은 두 명뿐이다. 용병대장은 성인 한국인 남성으로 읽히며 짧은 검은 머리, 얼굴 윤곽과 짧은 수염이 인물 참조에 가깝다. 남색 둥근목 상의 위에 참조에는 없는 어두운 재킷을 추가했다. 윤성찬은 은회색의 정돈된 머리와 어두운 정장, 밝은 셔츠가 보여 이전 장면의 외형과 부합한다. 얼굴 대부분이 가려져 세부 인상과 정확한 연령은 확인하기 어렵다. 검은 스피커폰의 형태와 파란 표시등은 참조와 유사하다. 별도의 자막이나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "용병대장의 등 뒤와 오른쪽에 의자 등받이와 팔걸이가 보여 앉은 몸의 지지가 자연스럽다. 골반과 발은 책상에 가려져 있으며, 공중에 떠 있다는 증거는 없다. 상체가 조금 앞으로 기울었지만 목을 굽혀 끄덕이는 동작은 가능하다. 윤성찬의 하체와 지지점은 화면 밖이다. 스피커폰과 문구류는 책상 위에 놓이고, 외투는 옷걸이에 걸려 있어 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "용병대장은 고개를 낮추고 눈도 아래쪽으로 내린다. 눈길은 왼쪽 전경 윤성찬의 눈높이보다는 두 사람 사이의 책상 쪽으로 향해 보인다. 윤성찬의 머리는 용병대장을 향하지만 눈은 보이지 않는다. 끄덕임의 하강 단계는 읽히나 서로 눈을 마주치는 순간은 명확하지 않다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "책상 하나를 사이에 두고 용병대장이 중앙 오른쪽의 검은 의자 하나에 앉아 있으며, 윤성찬의 부분 머리와 어깨가 왼쪽 전경을 차지한다. 스피커폰 하나는 아래쪽 중앙 왼편, 펜꽂이 하나와 칸막이 수납함은 오른쪽에 있다. 배경에는 세로 조명 하나, 외투가 걸린 옷걸이 하나, 큰 화분 하나와 야경 창문이 보이고 오른쪽 뒤에는 소파와 낮은 가구 일부가 드러난다. 참조의 어두운 마감과 조명 분위기를 유지한다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "두 인물만 보인다. 용병대장의 성인 한국인 남성 외형, 짧은 검은 머리와 수염, 남색 둥근목 상의는 인물 참조에 가깝고 재킷을 추가하지 않았다. 윤성찬의 은회색 머리, 어두운 정장과 밝은 셔츠는 이전 장면과 부합하지만 가려진 얼굴의 세부 정체성은 확인하기 어렵다. 스피커폰은 검은 몸체와 파란 표시등을 갖춰 참조의 물건으로 읽힌다. 추가 인물이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "용병대장은 보이는 의자 등받이 앞에 정상적인 방향으로 앉아 있고, 상체는 비교적 곧게 유지한 채 목과 턱만 낮췄다. 끄덕임의 하강 단계로 가능한 자세다. 골반과 발은 가려졌지만 의자가 몸의 지지를 설명한다. 윤성찬의 하체는 화면 밖이다. 책상 소품들은 상판에 놓여 있고 외투는 옷걸이에 걸려 있어, 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "왼쪽 전경 윤성찬을 향한 눈길과 고개를 낮추는 순간을 더 잘 연결하지만, 참조에 없는 재킷과 약간 앞으로 기운 상체는 불일치한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "참조의 남색 상의와 꼿꼿한 상체는 충실하지만, 눈길이 윤성찬의 눈보다 책상 쪽으로 떨어져 핵심인 시선 교환이 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "용병대장은 턱을 내린 상태에서 눈을 화면 왼쪽으로 돌려 전경 윤성찬의 얼굴 방향을 본다. 윤성찬도 머리를 용병대장 쪽으로 돌렸으며, 눈 자체는 뒤쪽 시점 때문에 보이지 않는다. 고개를 낮추면서 상대를 의식하는 관계가 읽힌다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "목재 책상 하나가 두 사람 사이에 있고, 용병대장은 그 너머 중앙 오른쪽의 검은 의자 하나에 앉아 있다. 윤성찬의 부분 머리와 어깨는 왼쪽 전경에 걸린다. 책상에는 스피커폰 하나, 펜꽂이 하나와 수납함 일부가 보인다. 뒤에는 세로 조명 하나, 외투가 걸린 옷걸이 하나, 큰 화분 하나, 왼쪽 조명 선반과 야경 창문이 있다. 참조의 재료와 야간 조명은 이어지며, 중복된 고정 설비나 불가능한 반사는 보이지 않는다. 스피커폰의 화면상 위치는 참조와 달라졌지만 카메라도 달라져 이것만으로 공간 모순을 확정할 수 없다.",
        "entities": "등장인물은 두 명뿐이다. 용병대장은 성인 한국인 남성으로 읽히며 짧은 검은 머리, 얼굴 윤곽과 짧은 수염이 인물 참조에 가깝다. 남색 둥근목 상의 위에 참조에는 없는 어두운 재킷을 추가했다. 윤성찬은 은회색의 정돈된 머리와 어두운 정장, 밝은 셔츠가 보여 이전 장면의 외형과 부합한다. 얼굴 대부분이 가려져 세부 인상과 정확한 연령은 확인하기 어렵다. 검은 스피커폰의 형태와 파란 표시등은 참조와 유사하다. 별도의 자막이나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "용병대장의 등 뒤와 오른쪽에 의자 등받이와 팔걸이가 보여 앉은 몸의 지지가 자연스럽다. 골반과 발은 책상에 가려져 있으며, 공중에 떠 있다는 증거는 없다. 상체가 조금 앞으로 기울었지만 목을 굽혀 끄덕이는 동작은 가능하다. 윤성찬의 하체와 지지점은 화면 밖이다. 스피커폰과 문구류는 책상 위에 놓이고, 외투는 옷걸이에 걸려 있어 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "용병대장은 고개를 낮추고 눈도 아래쪽으로 내린다. 눈길은 왼쪽 전경 윤성찬의 눈높이보다는 두 사람 사이의 책상 쪽으로 향해 보인다. 윤성찬의 머리는 용병대장을 향하지만 눈은 보이지 않는다. 끄덕임의 하강 단계는 읽히나 서로 눈을 마주치는 순간은 명확하지 않다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "책상 하나를 사이에 두고 용병대장이 중앙 오른쪽의 검은 의자 하나에 앉아 있으며, 윤성찬의 부분 머리와 어깨가 왼쪽 전경을 차지한다. 스피커폰 하나는 아래쪽 중앙 왼편, 펜꽂이 하나와 칸막이 수납함은 오른쪽에 있다. 배경에는 세로 조명 하나, 외투가 걸린 옷걸이 하나, 큰 화분 하나와 야경 창문이 보이고 오른쪽 뒤에는 소파와 낮은 가구 일부가 드러난다. 참조의 어두운 마감과 조명 분위기를 유지한다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않는다.",
        "entities": "두 인물만 보인다. 용병대장의 성인 한국인 남성 외형, 짧은 검은 머리와 수염, 남색 둥근목 상의는 인물 참조에 가깝고 재킷을 추가하지 않았다. 윤성찬의 은회색 머리, 어두운 정장과 밝은 셔츠는 이전 장면과 부합하지만 가려진 얼굴의 세부 정체성은 확인하기 어렵다. 스피커폰은 검은 몸체와 파란 표시등을 갖춰 참조의 물건으로 읽힌다. 추가 인물이나 화면 위 문구는 없다.",
        "hard_violations": [],
        "physics": "용병대장은 보이는 의자 등받이 앞에 정상적인 방향으로 앉아 있고, 상체는 비교적 곧게 유지한 채 목과 턱만 낮췄다. 끄덕임의 하강 단계로 가능한 자세다. 골반과 발은 가려졌지만 의자가 몸의 지지를 설명한다. 윤성찬의 하체는 화면 밖이다. 책상 소품들은 상판에 놓여 있고 외투는 옷걸이에 걸려 있어, 지지 없이 떠 있는 몸이나 물건은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.571
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1857,
   "B": 1571
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "용병대장의 의상, 스피커폰의 위치 및 카메라 구도 등 프롬프트와 참조 이미지의 세부 지시사항을 충실하게 구현함."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "카메라 구도는 적절하나, 용병대장이 지시와 달리 재킷을 입고 있으며 스피커폰의 위치가 이전 샷과 일치하지 않음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 윤성찬 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S87sh6_sel.png",
    "asset_id": "4f454b7d-8ecc-47f6-838b-cab0de73ae03",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 유빅 용병대장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1190196>",
    "asset_id": "4c1168d0-bced-42cd-8b81-9986f6c3118c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1456670>",
    "asset_id": "cb968520-53a5-4b55-8260-e357fc83e9f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0eaa-9ae0-7a20-b4b0-26ec1024ab6c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S87sh6"
  },
  "staged_characters_added": [
   "C17"
  ]
 },
 "S88sh11::signage": {
  "fp": "877dc3a3ceb56ebf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh11": {
  "input_fingerprint": "898f5eec35e0ad6b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 기계 팔다리가 단단히 결박된 채 누워 있는 찰리의 전신 구도.\n\nLOCATION (lock): On the stainless-steel examination bed inside the main laboratory's glass-walled test area. The laboratory is lit for the morning procedure. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane above the bed's foot-side corner, looking directly down its diagonal in a wide frame before descending along the same long side. 찰리's complete restrained body extends from lower-left toward upper-right, with every bound limb visible and the bed occupying less than two-fifths of the image; 지소영 bends beside his head at upper-right, stroking it as he looks up toward her. Keep her downward attention on his face, using the elevated distance—not a distorted perspective—to expose his helplessness alongside her tenderness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Bed extending diagonally from feet to head in the middle-center of the frame, midground; 지소영 beside 찰리's head in the upper-right of the frame, midground, looks toward 찰리's face at the head end of the bed.\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — Seen from above its foot-side corner, with its length receding diagonally toward 지소영; used as Establishes the shared spatial anchor and leaves all restraint points visible; Limb restraints (Secured around 찰리's limbs) — Visible at the corresponding limb positions without overlap from 지소영; used as Makes confinement legible within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves precise mechanical contours and restrained stainless-steel highlights while keeping the emotional exchange readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent and restrained on the stainless-steel bed, looking up at Soyoung. The bed supports his body, but the scene text does not specify the individual restraint points, the exact positions of his arms and legs, or the angle of his head and torso.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies restrained on the stainless-steel laboratory bed, with his hands still secured. A stun gun is already concealed in a drawer, as established by its later removal. 지소영: She stands at the laboratory bed in her research coat, looking down sadly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 기계 팔다리가 단단히 결박된 채 누워 있는 찰리의 전신 구도.\n\nLOCATION (lock): On the stainless-steel examination bed inside the main laboratory's glass-walled test area. The laboratory is lit for the morning procedure. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane above the bed's foot-side corner, looking directly down its diagonal in a wide frame before descending along the same long side. 찰리's complete restrained body extends from lower-left toward upper-right, with every bound limb visible and the bed occupying less than two-fifths of the image; 지소영 bends beside his head at upper-right, stroking it as he looks up toward her. Keep her downward attention on his face, using the elevated distance—not a distorted perspective—to expose his helplessness alongside her tenderness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Bed extending diagonally from feet to head in the middle-center of the frame, midground; 지소영 beside 찰리's head in the upper-right of the frame, midground, looks toward 찰리's face at the head end of the bed.\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — Seen from above its foot-side corner, with its length receding diagonally toward 지소영; used as Establishes the shared spatial anchor and leaves all restraint points visible; Limb restraints (Secured around 찰리's limbs) — Visible at the corresponding limb positions without overlap from 지소영; used as Makes confinement legible within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves precise mechanical contours and restrained stainless-steel highlights while keeping the emotional exchange readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent and restrained on the stainless-steel bed, looking up at Soyoung. The bed supports his body, but the scene text does not specify the individual restraint points, the exact positions of his arms and legs, or the angle of his head and torso.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies restrained on the stainless-steel laboratory bed, with his hands still secured. A stun gun is already concealed in a drawer, as established by its later removal. 지소영: She stands at the laboratory bed in her research coat, looking down sadly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 차가운 스테인리스 침대 위에 거대한 기계 팔다리가 단단히 결박된 채 누워 있는 찰리의 전신 구도.\n\nLOCATION (lock): On the stainless-steel examination bed inside the main laboratory's glass-walled test area. The laboratory is lit for the morning procedure. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Pause the crane above the bed's foot-side corner, looking directly down its diagonal in a wide frame before descending along the same long side. 찰리's complete restrained body extends from lower-left toward upper-right, with every bound limb visible and the bed occupying less than two-fifths of the image; 지소영 bends beside his head at upper-right, stroking it as he looks up toward her. Keep her downward attention on his face, using the elevated distance—not a distorted perspective—to expose his helplessness alongside her tenderness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Bed extending diagonally from feet to head in the middle-center of the frame, midground; 지소영 beside 찰리's head in the upper-right of the frame, midground, looks toward 찰리's face at the head end of the bed.\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Supporting 찰리's restrained body) — Seen from above its foot-side corner, with its length receding diagonally toward 지소영; used as Establishes the shared spatial anchor and leaves all restraint points visible; Limb restraints (Secured around 찰리's limbs) — Visible at the corresponding limb positions without overlap from 지소영; used as Makes confinement legible within the full-body composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient illumination preserves precise mechanical contours and restrained stainless-steel highlights while keeping the emotional exchange readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent and restrained on the stainless-steel bed, looking up at Soyoung. The bed supports his body, but the scene text does not specify the individual restraint points, the exact positions of his arms and legs, or the angle of his head and torso.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie lies restrained on the stainless-steel laboratory bed, with his hands still secured. A stun gun is already concealed in a drawer, as established by its later removal. 지소영: She stands at the laboratory bed in her research coat, looking down sadly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지소영은 찰리를 내려다보고, 찰리는 위를 향해 시선을 둠.",
    "built_space": "유리벽 연구실. 대각선 침대가 있으나 레퍼런스와 달리 중앙 기둥이 없는 개방형 뼈대 구조로 잘못 렌더링됨.",
    "entities": "지소영의 외형은 일치하나, 찰리는 레퍼런스의 육중한 고릴라 비율과 흉부 장갑판 디테일이 다소 다르게 나타남.",
    "hard_violations": [],
    "physics": "인물들은 침대 위에 안정적으로 안착되어 물리적 어색함이 없음. 부유물 없음."
   },
   {
    "label": "B",
    "direction": "지소영은 찰리의 얼굴을 내려다보고, 찰리는 위를 응시함.",
    "built_space": "유리벽 연구실. 이전 숏과 동일한 중앙 기둥형 침대와 우측 카트가 묘사됨. 전경 가장자리에 모니터와 기계 팔이 렌더링됨.",
    "entities": "찰리의 육중한 고릴라 비율과 장갑판, 지소영의 50대 여성 외형 및 연구복 모두 레퍼런스와 매우 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 침대와 바닥에 안정적으로 지지되어 있음. 허공에 뜬 부유물 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "전경 확대 제한 지시를 일부 어겼으나, 로봇 외형과 이전 숏의 공간 디테일(침대 기둥 구조 등)을 매우 정확하게 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 프레이밍은 깔끔하나, 찰리의 고릴라 체형 디테일이 부족하고 록(lock)된 침대의 하부 구조를 전혀 다르게 묘사함."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "지소영은 찰리의 얼굴을 내려다보고, 찰리는 위를 응시함.",
        "built_space": "유리벽 연구실. 이전 숏과 동일한 중앙 기둥형 침대와 우측 카트가 묘사됨. 전경 가장자리에 모니터와 기계 팔이 렌더링됨.",
        "entities": "찰리의 육중한 고릴라 비율과 장갑판, 지소영의 50대 여성 외형 및 연구복 모두 레퍼런스와 매우 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 침대와 바닥에 안정적으로 지지되어 있음. 허공에 뜬 부유물 없음."
       },
       {
        "label": "A",
        "direction": "지소영은 찰리를 내려다보고, 찰리는 위를 향해 시선을 둠.",
        "built_space": "유리벽 연구실. 대각선 침대가 있으나 레퍼런스와 달리 중앙 기둥이 없는 개방형 뼈대 구조로 잘못 렌더링됨.",
        "entities": "지소영의 외형은 일치하나, 찰리는 레퍼런스의 육중한 고릴라 비율과 흉부 장갑판 디테일이 다소 다르게 나타남.",
        "hard_violations": [],
        "physics": "인물들은 침대 위에 안정적으로 안착되어 물리적 어색함이 없음. 부유물 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "전경 확대 제한 지시를 일부 어겼으나, 로봇 외형과 이전 숏의 공간 디테일(침대 기둥 구조 등)을 매우 정확하게 구현함."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전반적인 프레이밍은 깔끔하나, 찰리의 고릴라 체형 디테일이 부족하고 록(lock)된 침대의 하부 구조를 전혀 다르게 묘사함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "지소영은 찰리의 얼굴을 내려다보고, 찰리는 위를 응시함.",
        "built_space": "유리벽 연구실. 이전 숏과 동일한 중앙 기둥형 침대와 우측 카트가 묘사됨. 전경 가장자리에 모니터와 기계 팔이 렌더링됨.",
        "entities": "찰리의 육중한 고릴라 비율과 장갑판, 지소영의 50대 여성 외형 및 연구복 모두 레퍼런스와 매우 일치함.",
        "hard_violations": [],
        "physics": "두 인물 모두 침대와 바닥에 안정적으로 지지되어 있음. 허공에 뜬 부유물 없음."
       },
       {
        "label": "A",
        "direction": "지소영은 찰리를 내려다보고, 찰리는 위를 향해 시선을 둠.",
        "built_space": "유리벽 연구실. 대각선 침대가 있으나 레퍼런스와 달리 중앙 기둥이 없는 개방형 뼈대 구조로 잘못 렌더링됨.",
        "entities": "지소영의 외형은 일치하나, 찰리는 레퍼런스의 육중한 고릴라 비율과 흉부 장갑판 디테일이 다소 다르게 나타남.",
        "hard_violations": [],
        "physics": "인물들은 침대 위에 안정적으로 안착되어 물리적 어색함이 없음. 부유물 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "대각선 전신 결박 구도와 머리를 쓰다듬는 동작은 맞지만, 찰리의 얼굴이 소영보다 카메라 쪽 위를 향하고 긴 다리가 고릴라형 체형과 어긋난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "발치 위에서 내려다보는 대각선 전신 구도, 우상단 소영의 손길, 네 팔다리의 명확한 결박과 상대적으로 육중한 찰리의 체형이 더 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 발은 좌하단, 머리는 우상단을 향한다. 소영은 머리 옆에서 찰리의 얼굴을 내려다보며 한 손을 머리에 댄다. 찰리의 흰 얼굴은 소영 쪽으로 돌아가기보다 천장과 카메라 쪽을 향해, 그녀를 올려다보는 관계가 약하다. 왼쪽 시술 도구 끝은 아래쪽 빈 공간을 향하며 찰리를 직접 겨누지 않는다.",
        "built_space": "중앙에 금속 검사 침대 한 대와 중앙 받침대가 있고, 뒤로 곡면 유리벽과 원형 바닥 배수 격자가 이어진다. 왼쪽 가장자리에 시술 팔 두 개의 일부, 오른쪽에 장비 탑 한 대와 모니터 카트 한 대, 우하단에 별도 화면 일부가 보인다. 침대는 화면의 대략 3분의 1을 차지하고 발치 모서리에서 길이 방향을 내려다보는 구도다. 소영은 머리 옆 먼 쪽에 있어 팔다리 결박을 가리지 않는다. 금속과 바닥의 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리와 소영만 등장한다. 찰리의 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개, 선 모양 입과 청색 원자로는 참조와 부합한다. 다만 다리가 길고 몸이 늘씬해져 짧은 다리와 거대한 팔이라는 참조 비율이 약해졌다. 소영은 정돈된 검은 단발의 중년 동아시아계 여성으로 보이며 연구복 안에 짙은 상의를 입었다. 양쪽 팔과 양쪽 하퇴의 결박, 허리 벨트가 보인다. 숨겨진 전기충격기는 노출되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 머리, 몸통과 팔다리는 침대 표면에 놓여 있으며 손도 침대에 내려놓았다. 네 팔다리의 띠는 침대 가장자리 고정 장치로 이어진다. 침대는 금속 받침대가 지지한다. 소영의 하체는 침대 뒤에 가려져 있지만 서서 상체를 숙이고 머리에 손을 얹는 자세에 부유나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리의 전신이 좌하단 발에서 우상단 머리로 이어지고, 소영은 머리의 화면 오른쪽에서 얼굴을 내려다본다. 한 손은 찰리의 정수리에 닿아 있다. 찰리의 얼굴은 위를 향하면서 소영 쪽으로 조금 기울어 A보다 상호 주의 관계가 자연스럽지만, 발광점 눈만으로 정확한 시선 고정까지 확인되지는 않는다. 좌상단 시술 팔의 끝은 침대 왼쪽 빈 공간을 향한다.",
        "built_space": "금속 검사 침대 한 대가 중앙 받침대 위에 놓이고, 곡면 유리벽과 원형 배수 격자가 주변을 둘러싼다. 뒤쪽에 장비 탑 두 대, 왼쪽에 기구 카트 한 대, 오른쪽에 바퀴 달린 의자 한 개, 좌상단에 시술 팔 한 개가 식별된다. 발치 모서리 위의 높은 시점에서 침대 대각선을 내려다보며 침대는 화면의 약 3분의 1을 차지한다. 소영은 우상단 머리 옆에 서 있고 결박 지점을 가리지 않는다. 참조의 유리·금속 실험실 재질과 밝은 아침 절차 조명이 유지되며 불가능한 반사는 보이지 않는다.",
        "entities": "등장 대상은 찰리와 소영뿐이다. 찰리는 베이지 장갑판, 흰 마스크형 얼굴, 주황색 눈 두 개와 입선, 푸른 가슴 원자로를 갖췄다. 참조보다 다리는 여전히 길지만 A보다 몸통과 팔이 육중하고 전체 비율이 압축되어 있다. 소영은 참조와 유사한 중년 동아시아계 여성의 얼굴과 검은 단발이며 흰 연구복과 짙은 상의를 착용한다. 양팔과 양쪽 발목 부근의 결박, 몸통을 가로지르는 띠가 명확하다. 전기충격기는 밖에 드러나지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 몸통과 머리, 양팔과 다리는 침대에 받쳐져 있고 손과 발도 표면에 접촉한다. 팔다리 결박은 침대 측면의 고정 장치에 연결되어 있다. 소영은 침대 옆에 서서 한 손으로 머리를 만지고 다른 손을 침대 가장자리에 얹어 상체를 숙인다. 침대는 중앙 금속 기둥이, 카트와 의자는 바퀴가, 시술 팔은 상부 고정 구조가 지지하며 근거 없이 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "대각선 전신 결박 구도와 머리를 쓰다듬는 동작은 맞지만, 찰리의 얼굴이 소영보다 카메라 쪽 위를 향하고 긴 다리가 고릴라형 체형과 어긋난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "발치 위에서 내려다보는 대각선 전신 구도, 우상단 소영의 손길, 네 팔다리의 명확한 결박과 상대적으로 육중한 찰리의 체형이 더 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 발은 좌하단, 머리는 우상단을 향한다. 소영은 머리 옆에서 찰리의 얼굴을 내려다보며 한 손을 머리에 댄다. 찰리의 흰 얼굴은 소영 쪽으로 돌아가기보다 천장과 카메라 쪽을 향해, 그녀를 올려다보는 관계가 약하다. 왼쪽 시술 도구 끝은 아래쪽 빈 공간을 향하며 찰리를 직접 겨누지 않는다.",
        "built_space": "중앙에 금속 검사 침대 한 대와 중앙 받침대가 있고, 뒤로 곡면 유리벽과 원형 바닥 배수 격자가 이어진다. 왼쪽 가장자리에 시술 팔 두 개의 일부, 오른쪽에 장비 탑 한 대와 모니터 카트 한 대, 우하단에 별도 화면 일부가 보인다. 침대는 화면의 대략 3분의 1을 차지하고 발치 모서리에서 길이 방향을 내려다보는 구도다. 소영은 머리 옆 먼 쪽에 있어 팔다리 결박을 가리지 않는다. 금속과 바닥의 반사에 명백한 광학적 모순은 없다.",
        "entities": "찰리와 소영만 등장한다. 찰리의 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개, 선 모양 입과 청색 원자로는 참조와 부합한다. 다만 다리가 길고 몸이 늘씬해져 짧은 다리와 거대한 팔이라는 참조 비율이 약해졌다. 소영은 정돈된 검은 단발의 중년 동아시아계 여성으로 보이며 연구복 안에 짙은 상의를 입었다. 양쪽 팔과 양쪽 하퇴의 결박, 허리 벨트가 보인다. 숨겨진 전기충격기는 노출되지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 머리, 몸통과 팔다리는 침대 표면에 놓여 있으며 손도 침대에 내려놓았다. 네 팔다리의 띠는 침대 가장자리 고정 장치로 이어진다. 침대는 금속 받침대가 지지한다. 소영의 하체는 침대 뒤에 가려져 있지만 서서 상체를 숙이고 머리에 손을 얹는 자세에 부유나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리의 전신이 좌하단 발에서 우상단 머리로 이어지고, 소영은 머리의 화면 오른쪽에서 얼굴을 내려다본다. 한 손은 찰리의 정수리에 닿아 있다. 찰리의 얼굴은 위를 향하면서 소영 쪽으로 조금 기울어 A보다 상호 주의 관계가 자연스럽지만, 발광점 눈만으로 정확한 시선 고정까지 확인되지는 않는다. 좌상단 시술 팔의 끝은 침대 왼쪽 빈 공간을 향한다.",
        "built_space": "금속 검사 침대 한 대가 중앙 받침대 위에 놓이고, 곡면 유리벽과 원형 배수 격자가 주변을 둘러싼다. 뒤쪽에 장비 탑 두 대, 왼쪽에 기구 카트 한 대, 오른쪽에 바퀴 달린 의자 한 개, 좌상단에 시술 팔 한 개가 식별된다. 발치 모서리 위의 높은 시점에서 침대 대각선을 내려다보며 침대는 화면의 약 3분의 1을 차지한다. 소영은 우상단 머리 옆에 서 있고 결박 지점을 가리지 않는다. 참조의 유리·금속 실험실 재질과 밝은 아침 절차 조명이 유지되며 불가능한 반사는 보이지 않는다.",
        "entities": "등장 대상은 찰리와 소영뿐이다. 찰리는 베이지 장갑판, 흰 마스크형 얼굴, 주황색 눈 두 개와 입선, 푸른 가슴 원자로를 갖췄다. 참조보다 다리는 여전히 길지만 A보다 몸통과 팔이 육중하고 전체 비율이 압축되어 있다. 소영은 참조와 유사한 중년 동아시아계 여성의 얼굴과 검은 단발이며 흰 연구복과 짙은 상의를 착용한다. 양팔과 양쪽 발목 부근의 결박, 몸통을 가로지르는 띠가 명확하다. 전기충격기는 밖에 드러나지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 몸통과 머리, 양팔과 다리는 침대에 받쳐져 있고 손과 발도 표면에 접촉한다. 팔다리 결박은 침대 측면의 고정 장치에 연결되어 있다. 소영은 침대 옆에 서서 한 손으로 머리를 만지고 다른 손을 침대 가장자리에 얹어 상체를 숙인다. 침대는 중앙 금속 기둥이, 카트와 의자는 바퀴가, 시술 팔은 상부 고정 구조가 지지하며 근거 없이 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 1875,
   "A": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "전경 확대 제한 지시를 일부 어겼으나, 로봇 외형과 이전 숏의 공간 디테일(침대 기둥 구조 등)을 매우 정확하게 구현함."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "전반적인 프레이밍은 깔끔하나, 찰리의 고릴라 체형 디테일이 부족하고 록(lock)된 침대의 하부 구조를 전혀 다르게 묘사함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S82sh13_sel.png",
    "asset_id": "e48c1896-1f82-4804-a304-ec35673941ca",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1237799>",
    "asset_id": "3674e762-acc5-4d71-a06e-9b4c0f9b740c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0eaf-ea00-7fcb-8ffa-a908f5505435",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh13"
  },
  "staged_characters_added": [
   "C32"
  ]
 },
 "S88sh37::signage": {
  "fp": "16eb5b9c4a069d85",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh37": {
  "input_fingerprint": "59b494082df01289",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 뭉툭한 기계 손가락이 울고 있는 이현우의 뺨에 닿은 채 멈춘 근접 찰나.\n\nLOCATION (lock): At the side of the examination bed inside the glass-walled laboratory enclosure, under the procedure lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the bed's head-side corner at 이현우's leaning face height, finish the direct observational push across their interaction at an oblique side angle. His tearful face occupies the right half, 찰리's reassuring face remains partly visible lower-left, and the blunt mechanical fingertip touches his cheek near the center without enlarging the hand disproportionately. Stop at contact and prioritize proximity alone, holding their exchanged looks and a narrow bed edge as spatial context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Still supporting 찰리) — A narrow oblique edge remains below the faces; used as Preserves bedside continuity without distracting from the touch.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained ambient illumination with gentle tonal separation across tears and mechanical surfaces, letting tenderness come from contact rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stainless-steel experiment bed, glass enclosure, attached equipment, and daytime laboratory lighting from the reference. Exclude the locked hand restraints in their earlier state; they have been released for this exchange.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains recumbent on the stainless-steel bed after his hand restraints are released, with one hand raised so that his finger touches Hyunwoo's face. His torso remains supported by the bed; which hand is raised and the precise positions of his head, other arm, and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains lying on the stainless-steel bed, but his hands have been released; no release of the remaining restraints has been established. The stand taken up during the confrontation remains displaced, and a stun gun has been drawn from the drawer. 이현우: He is at the bedside, crying heavily.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 뭉툭한 기계 손가락이 울고 있는 이현우의 뺨에 닿은 채 멈춘 근접 찰나.\n\nLOCATION (lock): At the side of the examination bed inside the glass-walled laboratory enclosure, under the procedure lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the bed's head-side corner at 이현우's leaning face height, finish the direct observational push across their interaction at an oblique side angle. His tearful face occupies the right half, 찰리's reassuring face remains partly visible lower-left, and the blunt mechanical fingertip touches his cheek near the center without enlarging the hand disproportionately. Stop at contact and prioritize proximity alone, holding their exchanged looks and a narrow bed edge as spatial context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Still supporting 찰리) — A narrow oblique edge remains below the faces; used as Preserves bedside continuity without distracting from the touch.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained ambient illumination with gentle tonal separation across tears and mechanical surfaces, letting tenderness come from contact rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stainless-steel experiment bed, glass enclosure, attached equipment, and daytime laboratory lighting from the reference. Exclude the locked hand restraints in their earlier state; they have been released for this exchange.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains recumbent on the stainless-steel bed after his hand restraints are released, with one hand raised so that his finger touches Hyunwoo's face. His torso remains supported by the bed; which hand is raised and the precise positions of his head, other arm, and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains lying on the stainless-steel bed, but his hands have been released; no release of the remaining restraints has been established. The stand taken up during the confrontation remains displaced, and a stun gun has been drawn from the drawer. 이현우: He is at the bedside, crying heavily.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 찰리의 뭉툭한 기계 손가락이 울고 있는 이현우의 뺨에 닿은 채 멈춘 근접 찰나.\n\nLOCATION (lock): At the side of the examination bed inside the glass-walled laboratory enclosure, under the procedure lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the bed's head-side corner at 이현우's leaning face height, finish the direct observational push across their interaction at an oblique side angle. His tearful face occupies the right half, 찰리's reassuring face remains partly visible lower-left, and the blunt mechanical fingertip touches his cheek near the center without enlarging the hand disproportionately. Stop at contact and prioritize proximity alone, holding their exchanged looks and a narrow bed edge as spatial context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Stainless-steel bed (Still supporting 찰리) — A narrow oblique edge remains below the faces; used as Preserves bedside continuity without distracting from the touch.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained ambient illumination with gentle tonal separation across tears and mechanical surfaces, letting tenderness come from contact rather than a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stainless-steel experiment bed, glass enclosure, attached equipment, and daytime laboratory lighting from the reference. Exclude the locked hand restraints in their earlier state; they have been released for this exchange.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie remains recumbent on the stainless-steel bed after his hand restraints are released, with one hand raised so that his finger touches Hyunwoo's face. His torso remains supported by the bed; which hand is raised and the precise positions of his head, other arm, and legs are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains lying on the stainless-steel bed, but his hands have been released; no release of the remaining restraints has been established. The stand taken up during the confrontation remains displaced, and a stun gun has been drawn from the drawer. 이현우: He is at the bedside, crying heavily.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 손가락이 이현우의 뺨에 닿아 있고, 두 사람의 시선이 교차함.",
    "built_space": "유리벽 실험실 배경과 하단의 스테인리스 침대 모서리가 보임.",
    "entities": "찰리와 이현우 모두 참조 이미지의 외형 및 복장과 일치함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 머리가 왼쪽에 있는 찰리의 전경 팔이 오른손(엄지가 왼쪽)으로 그려져 신체와 정상적으로 연결될 수 없음."
    ],
    "physics": "침대에 누워 지지받고 있으나, 뺨에 닿은 손이 분리된 것처럼 어색하게 배치됨."
   },
   {
    "label": "B",
    "direction": "찰리의 시선과 손가락이 이현우의 뺨을 향해 정확히 닿아 있음.",
    "built_space": "유리벽 실험실 내부, 하단에 스테인리스 침대 모서리가 위치함.",
    "entities": "찰리(흰색 마스크, 베이지색 기계 몸)와 이현우(검은 머리, 흰색 환자복) 모두 참조와 일치함.",
    "hard_violations": [],
    "physics": "찰리는 침대에 누워 지지받고 있으며, 왼팔이 어깨부터 자연스러운 각도로 들려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 근접 구도와 감정적 교감을 정확히 구현했으며, 기계 팔의 위치와 해부학적 구조가 찰리의 누운 자세와 자연스럽게 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면 전면에 배치된 기계 팔이 오른손으로 그려져, 누워있는 찰리의 몸통과 물리적으로 연결될 수 없는 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 손가락이 이현우의 뺨에 닿아 있고, 두 사람의 시선이 교차함.",
        "built_space": "유리벽 실험실 배경과 하단의 스테인리스 침대 모서리가 보임.",
        "entities": "찰리와 이현우 모두 참조 이미지의 외형 및 복장과 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 머리가 왼쪽에 있는 찰리의 전경 팔이 오른손(엄지가 왼쪽)으로 그려져 신체와 정상적으로 연결될 수 없음."
        ],
        "physics": "침대에 누워 지지받고 있으나, 뺨에 닿은 손이 분리된 것처럼 어색하게 배치됨."
       },
       {
        "label": "B",
        "direction": "찰리의 시선과 손가락이 이현우의 뺨을 향해 정확히 닿아 있음.",
        "built_space": "유리벽 실험실 내부, 하단에 스테인리스 침대 모서리가 위치함.",
        "entities": "찰리(흰색 마스크, 베이지색 기계 몸)와 이현우(검은 머리, 흰색 환자복) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "찰리는 침대에 누워 지지받고 있으며, 왼팔이 어깨부터 자연스러운 각도로 들려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지시된 근접 구도와 감정적 교감을 정확히 구현했으며, 기계 팔의 위치와 해부학적 구조가 찰리의 누운 자세와 자연스럽게 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "화면 전면에 배치된 기계 팔이 오른손으로 그려져, 누워있는 찰리의 몸통과 물리적으로 연결될 수 없는 심각한 해부학적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 손가락이 이현우의 뺨에 닿아 있고, 두 사람의 시선이 교차함.",
        "built_space": "유리벽 실험실 배경과 하단의 스테인리스 침대 모서리가 보임.",
        "entities": "찰리와 이현우 모두 참조 이미지의 외형 및 복장과 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 머리가 왼쪽에 있는 찰리의 전경 팔이 오른손(엄지가 왼쪽)으로 그려져 신체와 정상적으로 연결될 수 없음."
        ],
        "physics": "침대에 누워 지지받고 있으나, 뺨에 닿은 손이 분리된 것처럼 어색하게 배치됨."
       },
       {
        "label": "B",
        "direction": "찰리의 시선과 손가락이 이현우의 뺨을 향해 정확히 닿아 있음.",
        "built_space": "유리벽 실험실 내부, 하단에 스테인리스 침대 모서리가 위치함.",
        "entities": "찰리(흰색 마스크, 베이지색 기계 몸)와 이현우(검은 머리, 흰색 환자복) 모두 참조와 일치함.",
        "hard_violations": [],
        "physics": "찰리는 침대에 누워 지지받고 있으며, 왼팔이 어깨부터 자연스러운 각도로 들려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "눈물 어린 얼굴과 접촉을 더 밀착해 담은 비스듬한 클로즈업이 우세하지만, 찰리의 얼굴이 지정된 왼쪽 아래보다 높고 접촉점도 중앙에서 오른쪽으로 치우친다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "찰리의 왼쪽 아래 배치와 누운 자세는 명확하지만, 상체와 실험실 바닥까지 넓게 보여 근접성과 얼굴 높이의 관찰 시점이 A보다 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우는 왼쪽 아래의 찰리 얼굴을 바라보고, 찰리의 마스크는 현우 쪽으로 돌아가 있다. 세운 기계 손가락의 뭉툭한 끝이 현우의 눈물 흐르는 뺨에 실제로 닿는다. 접촉점은 화면 정중앙보다는 오른쪽이다.",
        "built_space": "두 인물 아래로 스테인리스 검사 침대 한 대의 비스듬한 가장자리가 이어진다. 뒤에는 유리벽과 여러 수직 프레임, 오른쪽 끝에는 부분적으로 잘린 모니터 한 대가 보인다. 현우는 침대 옆에서 몸을 숙이고 찰리는 침대 위에 누워 있어 공간 관계가 성립한다. 중복 침대나 불가능한 인물 반사는 보이지 않는다.",
        "entities": "등장자는 찰리와 현우 두 명뿐이다. 찰리는 마모된 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개와 선 형태의 입, 푸른 가슴 원자로를 갖춰 참조의 기계 정체성과 일치한다. 현우는 10대 후반으로 보이는 동아시아계 남성으로, 헝클어진 검은 머리와 흰 상의를 착용하고 양쪽 뺨에 눈물이 흐른다. 국적은 영상만으로 확인할 수 없다. 하체와 남은 구속구, 이동된 스탠드와 전기충격기는 프레임 밖이므로 판정하지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 아래의 금속 침대에 받쳐져 있고, 들어 올린 손은 손목과 전완의 기계 관절에 연결되어 있다. 손가락은 그 팔의 구동으로 뺨에 접촉하는 자세이며 독립적으로 떠 있지 않다. 현우는 침대 쪽으로 상체를 기울인 자세이고 하체 지지는 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "현우의 시선은 왼쪽 아래에 누운 찰리에게 향하고, 찰리의 얼굴도 현우 쪽 위를 향한다. 기계 손가락 끝은 현우의 뺨에 닿아 있으며 접촉 동작은 분명하다. 접촉점은 중앙보다 오른쪽에 놓인다.",
        "built_space": "스테인리스 침대 한 대가 하단을 비스듬히 가로지르고 찰리를 받친다. 배경에는 유리벽 프레임, 왼쪽의 이동식 장비 카트 한 대, 오른쪽의 상하 화면이 달린 장비 랙 한 대가 보인다. 현우는 침대 옆에 있으며 찰리의 얼굴은 왼쪽 아래에 놓인다. 다만 바닥과 침대 상판이 더 넓게 드러나 카메라가 상호작용을 다소 내려다보는 인상이 강하다. 불가능한 반사나 명백한 설비 복제는 보이지 않는다.",
        "entities": "찰리와 현우 외의 인물은 없다. 찰리의 베이지색 마모 장갑, 흰 마스크, 주황색 눈 두 개, 검은 입선과 푸른 원자로가 참조의 주요 특징을 유지한다. 현우는 젊은 동아시아계 남성의 얼굴, 짧고 헝클어진 검은 머리, 깨끗한 흰 상의를 갖췄고 눈물 자국이 선명하다. 보이지 않는 하의나 구속구 및 장면 밖 소품은 평가할 수 없다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 검사 침대에 누워 지지되고, 올린 전완과 손은 팔 관절로 이어져 있다. 뺨을 만지는 손가락에도 손과 팔이라는 지지가 확인된다. 현우의 상체 기울기는 침대 곁에서 몸을 숙이는 동작으로 가능하며 하체는 프레임 밖이다. 무지지 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "눈물 어린 얼굴과 접촉을 더 밀착해 담은 비스듬한 클로즈업이 우세하지만, 찰리의 얼굴이 지정된 왼쪽 아래보다 높고 접촉점도 중앙에서 오른쪽으로 치우친다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "찰리의 왼쪽 아래 배치와 누운 자세는 명확하지만, 상체와 실험실 바닥까지 넓게 보여 근접성과 얼굴 높이의 관찰 시점이 A보다 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "현우는 왼쪽 아래의 찰리 얼굴을 바라보고, 찰리의 마스크는 현우 쪽으로 돌아가 있다. 세운 기계 손가락의 뭉툭한 끝이 현우의 눈물 흐르는 뺨에 실제로 닿는다. 접촉점은 화면 정중앙보다는 오른쪽이다.",
        "built_space": "두 인물 아래로 스테인리스 검사 침대 한 대의 비스듬한 가장자리가 이어진다. 뒤에는 유리벽과 여러 수직 프레임, 오른쪽 끝에는 부분적으로 잘린 모니터 한 대가 보인다. 현우는 침대 옆에서 몸을 숙이고 찰리는 침대 위에 누워 있어 공간 관계가 성립한다. 중복 침대나 불가능한 인물 반사는 보이지 않는다.",
        "entities": "등장자는 찰리와 현우 두 명뿐이다. 찰리는 마모된 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개와 선 형태의 입, 푸른 가슴 원자로를 갖춰 참조의 기계 정체성과 일치한다. 현우는 10대 후반으로 보이는 동아시아계 남성으로, 헝클어진 검은 머리와 흰 상의를 착용하고 양쪽 뺨에 눈물이 흐른다. 국적은 영상만으로 확인할 수 없다. 하체와 남은 구속구, 이동된 스탠드와 전기충격기는 프레임 밖이므로 판정하지 않는다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 아래의 금속 침대에 받쳐져 있고, 들어 올린 손은 손목과 전완의 기계 관절에 연결되어 있다. 손가락은 그 팔의 구동으로 뺨에 접촉하는 자세이며 독립적으로 떠 있지 않다. 현우는 침대 쪽으로 상체를 기울인 자세이고 하체 지지는 화면 밖이다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "현우의 시선은 왼쪽 아래에 누운 찰리에게 향하고, 찰리의 얼굴도 현우 쪽 위를 향한다. 기계 손가락 끝은 현우의 뺨에 닿아 있으며 접촉 동작은 분명하다. 접촉점은 중앙보다 오른쪽에 놓인다.",
        "built_space": "스테인리스 침대 한 대가 하단을 비스듬히 가로지르고 찰리를 받친다. 배경에는 유리벽 프레임, 왼쪽의 이동식 장비 카트 한 대, 오른쪽의 상하 화면이 달린 장비 랙 한 대가 보인다. 현우는 침대 옆에 있으며 찰리의 얼굴은 왼쪽 아래에 놓인다. 다만 바닥과 침대 상판이 더 넓게 드러나 카메라가 상호작용을 다소 내려다보는 인상이 강하다. 불가능한 반사나 명백한 설비 복제는 보이지 않는다.",
        "entities": "찰리와 현우 외의 인물은 없다. 찰리의 베이지색 마모 장갑, 흰 마스크, 주황색 눈 두 개, 검은 입선과 푸른 원자로가 참조의 주요 특징을 유지한다. 현우는 젊은 동아시아계 남성의 얼굴, 짧고 헝클어진 검은 머리, 깨끗한 흰 상의를 갖췄고 눈물 자국이 선명하다. 보이지 않는 하의나 구속구 및 장면 밖 소품은 평가할 수 없다.",
        "hard_violations": [],
        "physics": "찰리의 몸통은 검사 침대에 누워 지지되고, 올린 전완과 손은 팔 관절로 이어져 있다. 뺨을 만지는 손가락에도 손과 팔이라는 지지가 확인된다. 현우의 상체 기울기는 침대 곁에서 몸을 숙이는 동작으로 가능하며 하체는 프레임 밖이다. 무지지 부유나 불가능한 관절 배치는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.304,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.054,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 머리가 왼쪽에 있는 찰리의 전경 팔이 오른손(엄지가 왼쪽)으로 그려져 신체와 정상적으로 연결될 수 없음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1054
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 근접 구도와 감정적 교감을 정확히 구현했으며, 기계 팔의 위치와 해부학적 구조가 찰리의 누운 자세와 자연스럽게 일치합니다."
   },
   {
    "label": "A",
    "score": 1054,
    "verdict_ko": "화면 전면에 배치된 기계 팔이 오른손으로 그려져, 누워있는 찰리의 몸통과 물리적으로 연결될 수 없는 심각한 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 머리가 왼쪽에 있는 찰리의 전경 팔이 오른손(엄지가 왼쪽)으로 그려져 신체와 정상적으로 연결될 수 없음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S88sh11_sel.png",
    "asset_id": "e64a82d6-b49d-422f-b639-b95001c23f41",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0eb7-f4e6-7fa3-83fa-a492782dfba4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh11"
  }
 },
 "S88sh45::signage": {
  "fp": "9523d5260dcd9b2a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S88sh45": {
  "input_fingerprint": "e6a24789a1dccb42",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 귀를 찢는 듯한 적색경보에 깜짝 놀라 일제히 고개를 든 채 천장을 향해 시선을 고정한 지소영과 서지민의 당황한 구도.\n\nLOCATION (lock): In the main laboratory beside the glass-walled test area. Red warning signals interrupt the room's working illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the bedside lateral track obliquely in front of 지소영's chair at seated chest height, looking directly across both women with a slight upward tilt. Keep 지소영 seated lower-left, her torso recoiling from its slump, and 서지민 higher at right with her shoulders caught in a startled turn; both lift their eyes toward the ceiling with natural neck angles, leaving open space above their faces. Emphasize the upward eyeline change while holding the established spacing, before either redirects attention toward the reporting researcher.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's chair (Occupied by 지소영 after she collapsed into it) — Seen obliquely from the bedside, partially visible beneath her; used as Maintains her seated height and the emotional continuity of her interrupted distress; Laboratory ceiling (Above the women as they react to the alarm); used as A limited upper-frame area supports the upward eyelines without adding an alarm fixture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the red-alert state interrupt the subdued laboratory illumination with a restrained red emphasis while preserving readable faces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the glass enclosure, stainless-steel bed, laboratory equipment, and room materials from the reference. Exclude a calm, alarm-free lighting state; the red emergency warning is now active.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with his hands free and his other restraints not yet released. The laboratory's red alarm has activated. 지소영: She remains at the chair in her research coat, startled out of her distressed posture. 서지민: She remains in the laboratory wearing her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 귀를 찢는 듯한 적색경보에 깜짝 놀라 일제히 고개를 든 채 천장을 향해 시선을 고정한 지소영과 서지민의 당황한 구도.\n\nLOCATION (lock): In the main laboratory beside the glass-walled test area. Red warning signals interrupt the room's working illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the bedside lateral track obliquely in front of 지소영's chair at seated chest height, looking directly across both women with a slight upward tilt. Keep 지소영 seated lower-left, her torso recoiling from its slump, and 서지민 higher at right with her shoulders caught in a startled turn; both lift their eyes toward the ceiling with natural neck angles, leaving open space above their faces. Emphasize the upward eyeline change while holding the established spacing, before either redirects attention toward the reporting researcher.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's chair (Occupied by 지소영 after she collapsed into it) — Seen obliquely from the bedside, partially visible beneath her; used as Maintains her seated height and the emotional continuity of her interrupted distress; Laboratory ceiling (Above the women as they react to the alarm); used as A limited upper-frame area supports the upward eyelines without adding an alarm fixture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the red-alert state interrupt the subdued laboratory illumination with a restrained red emphasis while preserving readable faces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the glass enclosure, stainless-steel bed, laboratory equipment, and room materials from the reference. Exclude a calm, alarm-free lighting state; the red emergency warning is now active.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with his hands free and his other restraints not yet released. The laboratory's red alarm has activated. 지소영: She remains at the chair in her research coat, startled out of her distressed posture. 서지민: She remains in the laboratory wearing her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 귀를 찢는 듯한 적색경보에 깜짝 놀라 일제히 고개를 든 채 천장을 향해 시선을 고정한 지소영과 서지민의 당황한 구도.\n\nLOCATION (lock): In the main laboratory beside the glass-walled test area. Red warning signals interrupt the room's working illumination. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the bedside lateral track obliquely in front of 지소영's chair at seated chest height, looking directly across both women with a slight upward tilt. Keep 지소영 seated lower-left, her torso recoiling from its slump, and 서지민 higher at right with her shoulders caught in a startled turn; both lift their eyes toward the ceiling with natural neck angles, leaving open space above their faces. Emphasize the upward eyeline change while holding the established spacing, before either redirects attention toward the reporting researcher.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지소영's chair (Occupied by 지소영 after she collapsed into it) — Seen obliquely from the bedside, partially visible beneath her; used as Maintains her seated height and the emotional continuity of her interrupted distress; Laboratory ceiling (Above the women as they react to the alarm); used as A limited upper-frame area supports the upward eyelines without adding an alarm fixture.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Let the red-alert state interrupt the subdued laboratory illumination with a restrained red emphasis while preserving readable faces.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the glass enclosure, stainless-steel bed, laboratory equipment, and room materials from the reference. Exclude a calm, alarm-free lighting state; the red emergency warning is now active.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie remains on the stainless-steel bed with his hands free and his other restraints not yet released. The laboratory's red alarm has activated. 지소영: She remains at the chair in her research coat, startled out of her distressed posture. 서지민: She remains in the laboratory wearing her research coat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 지소영 (한국인 여성, 50대 중반의 얼굴, 정돈된 짙은색 머리) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지.; 서지민 (한국인 여성, 20세의 젊은 얼굴, 검은 머리카락) — wearing: 제주도 연구소의 세련되고 깔끔한 미래형 흰색 연구 가운과 정장 바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "지소영과 서지민 모두 고개를 들어 천장 쪽을 바라보고 있으며, 시선이 지시된 대로 상단을 향함.",
    "built_space": "붉은색 비상조명이 켜진 유리 벽면의 실험실. 왼쪽 전경에 로봇 팔 일부가 보이며, 배경 중앙에 스테인리스 침대가 놓여 있음.",
    "entities": "지소영은 참조와 일치하는 짙은 회색 셔츠와 흰색 연구 가운을 입음. 서지민은 연한 회색 셔츠와 가운, 보안경을 착용했으나 바지가 어두운 색임. 로봇은 침대 위에 결박된 상태임.",
    "hard_violations": [
     "[gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
     "[gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 여러 붉은 경보등을 배경에 추가했다."
    ],
    "physics": "지소영은 의자에 앉아 팔걸이에 손을 올린 채 체중을 지탱하고 있으며, 서지민은 바닥에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "두 여성 모두 고개를 위로 젖혀 천장을 향해 시선을 고정하고 있음.",
    "built_space": "적색경보 조명이 켜진 실험실 내부. 중앙에 로봇이 누워 있는 침대가 있으나, 이전 컷에 있던 왼쪽 전경의 기계 장치(로봇 팔)가 보이지 않음.",
    "entities": "지소영의 의상이 참조 이미지(깃이 있는 셔츠)와 달리 둥근 목선의 남색 티셔츠로 변경됨. 서지민은 흰색 바지를 포함해 참조와 대체로 일치함.",
    "hard_violations": [
     "[gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
     "[gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 발광하는 경보등을 배경에 추가했다."
    ],
    "physics": "지소영은 하체를 의자에 지지하고 있으나 오른손은 허공에 떠 있으며, 서지민은 서서 구조물에 손을 얹어 균형을 잡고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 상향 시선과 카메라 구도를 훌륭히 구현했으며 이전 컷에 고정된 지소영의 의상과 주변 환경(로봇 팔 등)을 충실히 유지했으나, 서지민의 바지 색상이 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 인물의 시선 처리와 구도는 적절하나, 이전 컷에서 완전히 고정되어야 할 지소영의 셔츠 형태가 변경되었고 배경의 로봇 팔 등 주요 디테일이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영과 서지민 모두 고개를 들어 천장 쪽을 바라보고 있으며, 시선이 지시된 대로 상단을 향함.",
        "built_space": "붉은색 비상조명이 켜진 유리 벽면의 실험실. 왼쪽 전경에 로봇 팔 일부가 보이며, 배경 중앙에 스테인리스 침대가 놓여 있음.",
        "entities": "지소영은 참조와 일치하는 짙은 회색 셔츠와 흰색 연구 가운을 입음. 서지민은 연한 회색 셔츠와 가운, 보안경을 착용했으나 바지가 어두운 색임. 로봇은 침대 위에 결박된 상태임.",
        "hard_violations": [],
        "physics": "지소영은 의자에 앉아 팔걸이에 손을 올린 채 체중을 지탱하고 있으며, 서지민은 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "두 여성 모두 고개를 위로 젖혀 천장을 향해 시선을 고정하고 있음.",
        "built_space": "적색경보 조명이 켜진 실험실 내부. 중앙에 로봇이 누워 있는 침대가 있으나, 이전 컷에 있던 왼쪽 전경의 기계 장치(로봇 팔)가 보이지 않음.",
        "entities": "지소영의 의상이 참조 이미지(깃이 있는 셔츠)와 달리 둥근 목선의 남색 티셔츠로 변경됨. 서지민은 흰색 바지를 포함해 참조와 대체로 일치함.",
        "hard_violations": [],
        "physics": "지소영은 하체를 의자에 지지하고 있으나 오른손은 허공에 떠 있으며, 서지민은 서서 구조물에 손을 얹어 균형을 잡고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 상향 시선과 카메라 구도를 훌륭히 구현했으며 이전 컷에 고정된 지소영의 의상과 주변 환경(로봇 팔 등)을 충실히 유지했으나, 서지민의 바지 색상이 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 인물의 시선 처리와 구도는 적절하나, 이전 컷에서 완전히 고정되어야 할 지소영의 셔츠 형태가 변경되었고 배경의 로봇 팔 등 주요 디테일이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "지소영과 서지민 모두 고개를 들어 천장 쪽을 바라보고 있으며, 시선이 지시된 대로 상단을 향함.",
        "built_space": "붉은색 비상조명이 켜진 유리 벽면의 실험실. 왼쪽 전경에 로봇 팔 일부가 보이며, 배경 중앙에 스테인리스 침대가 놓여 있음.",
        "entities": "지소영은 참조와 일치하는 짙은 회색 셔츠와 흰색 연구 가운을 입음. 서지민은 연한 회색 셔츠와 가운, 보안경을 착용했으나 바지가 어두운 색임. 로봇은 침대 위에 결박된 상태임.",
        "hard_violations": [],
        "physics": "지소영은 의자에 앉아 팔걸이에 손을 올린 채 체중을 지탱하고 있으며, 서지민은 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "두 여성 모두 고개를 위로 젖혀 천장을 향해 시선을 고정하고 있음.",
        "built_space": "적색경보 조명이 켜진 실험실 내부. 중앙에 로봇이 누워 있는 침대가 있으나, 이전 컷에 있던 왼쪽 전경의 기계 장치(로봇 팔)가 보이지 않음.",
        "entities": "지소영의 의상이 참조 이미지(깃이 있는 셔츠)와 달리 둥근 목선의 남색 티셔츠로 변경됨. 서지민은 흰색 바지를 포함해 참조와 대체로 일치함.",
        "hard_violations": [],
        "physics": "지소영은 하체를 의자에 지지하고 있으나 오른손은 허공에 떠 있으며, 서지민은 서서 구조물에 손을 얹어 균형을 잡고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "낮은 사선 시점과 좌하단 착석·우상단 선 자세, 지소영의 이전 의상을 더 잘 유지하지만, 제외해야 할 찰리와 추가 경보등을 보여 실격이다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 여성의 천장 시선은 맞지만, 찰리와 경보등을 추가하고 시점을 높였으며 지소영의 이전 의상까지 변경했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "지소영은 턱을 들고 화면 위쪽을, 서지민은 위쪽 오른편을 본다. 두 시선 모두 천장 방향으로 향하며 카메라나 다른 연구자를 보지 않는다. 목의 젖힘은 자연스럽고 눈도 정상적인 형태다.",
        "built_space": "왼쪽 아래 의자 1개에 지소영이 앉고 오른쪽에 서지민이 서 있다. 중앙 뒤에는 금속 침대 1개, 오른쪽에는 나란한 모니터 2개와 작업대가 보인다. 유리 칸막이와 금속 설비는 이전 장소의 재질을 따른다. 앉은 인물의 가슴 높이에서 비스듬히 보는 구도에 비교적 가깝지만, 서지민 머리 위 여백은 좁다. 상단 오른쪽과 오른쪽 가장자리에는 명시적으로 추가하지 말라고 한 붉은 경보등이 보인다. 불가능한 반사는 확인되지 않는다.",
        "entities": "중년 여성 지소영과 젊은 여성 서지민의 얼굴·검은 머리·흰 연구 가운은 인물 참조에 대체로 부합한다. 지소영의 짙은 둥근 목 상의와 가운 주머니의 펜은 이전 스틸을 잘 잇는다. 서지민의 긴 머리, 머리 위 보안경과 손목시계도 참조와 맞는다. 그러나 두 여성만 보여야 하는 장면에 찰리의 로봇 몸체가 침대 위로 선명하게 들어왔다. 손목에는 고정 띠로 보이는 부분도 남아 있어 손이 풀린 상태와 어긋난다.",
        "hard_violations": [
         "이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
         "경보 설비를 추가하지 말라는 지시와 달리 발광하는 경보등을 배경에 추가했다."
        ],
        "physics": "지소영의 골반은 의자 좌판에 지지되고 등 뒤에 등받이가 있어, 앉은 채 상체를 갑자기 일으키는 자세가 가능하다. 서지민의 한 손은 오른쪽 작업대 가장자리에 닿고 다른 팔은 팔꿈치를 굽힌 상태다. 발은 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 찰리는 금속 침대에 누워 지지되며, 무지지 물체나 불가능한 신체 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "지소영은 화면 위쪽 오른편으로 눈을 들고, 서지민도 턱과 눈을 위쪽 오른편으로 향한다. 두 사람 모두 천장 쪽을 보는 반응으로 읽힌다. 서지민의 몸통은 옆으로 돌아 있고 얼굴은 위를 향해 놀란 회전 동작을 나타낸다.",
        "built_space": "지소영이 앉은 의자 1개, 중앙 뒤 금속 침대 1개, 오른쪽 모니터 1개와 왼쪽 전경의 부분 모니터 1개가 보인다. 왼쪽 가장자리에는 기계 팔도 일부 보인다. 유리 칸막이와 금속 바닥은 이전 장소와 유사하다. 다만 침대 윗면과 바닥을 넓게 내려다보아 지정된 낮은 가슴 높이의 약한 올려다보기보다 높은 시점으로 읽힌다. 천장 자체는 거의 없고 배경 상부에 적어도 다섯 개의 붉은 발광 설비가 보인다. 불가능한 반사는 확인되지 않는다.",
        "entities": "두 여성의 연령 대비, 검은 머리와 연구 가운은 지정 인물에 대체로 맞고 서지민의 보안경과 시계도 유지된다. 지소영은 이전 스틸의 짙은 둥근 목 상의 대신 단추 달린 회색 셔츠를 입었으며 가운 가슴에도 로고가 생겼다. 중앙 침대에는 이 장면에서 제외해야 하는 찰리가 거의 전신으로 보인다. 찰리의 손목 주변 고정 띠도 남아 있어 손이 자유롭다는 상태가 명확히 구현되지 않았다.",
        "hard_violations": [
         "이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
         "경보 설비를 추가하지 말라는 지시와 달리 여러 붉은 경보등을 배경에 추가했다."
        ],
        "physics": "지소영은 골반을 의자에 두고 상체를 비튼 상태이며 좌판·등받이·팔걸이의 배치가 착석 자세와 양립한다. 서지민은 팔꿈치를 굽히고 몸통을 돌린 서 있는 자세로, 화면 밖 발을 보이지 않는다는 이유만으로 부유했다고 판단할 수 없다. 찰리의 몸은 침대가 지지한다. 해부학적으로 불가능한 동작이나 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "낮은 사선 시점과 좌하단 착석·우상단 선 자세, 지소영의 이전 의상을 더 잘 유지하지만, 제외해야 할 찰리와 추가 경보등을 보여 실격이다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 여성의 천장 시선은 맞지만, 찰리와 경보등을 추가하고 시점을 높였으며 지소영의 이전 의상까지 변경했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "지소영은 턱을 들고 화면 위쪽을, 서지민은 위쪽 오른편을 본다. 두 시선 모두 천장 방향으로 향하며 카메라나 다른 연구자를 보지 않는다. 목의 젖힘은 자연스럽고 눈도 정상적인 형태다.",
        "built_space": "왼쪽 아래 의자 1개에 지소영이 앉고 오른쪽에 서지민이 서 있다. 중앙 뒤에는 금속 침대 1개, 오른쪽에는 나란한 모니터 2개와 작업대가 보인다. 유리 칸막이와 금속 설비는 이전 장소의 재질을 따른다. 앉은 인물의 가슴 높이에서 비스듬히 보는 구도에 비교적 가깝지만, 서지민 머리 위 여백은 좁다. 상단 오른쪽과 오른쪽 가장자리에는 명시적으로 추가하지 말라고 한 붉은 경보등이 보인다. 불가능한 반사는 확인되지 않는다.",
        "entities": "중년 여성 지소영과 젊은 여성 서지민의 얼굴·검은 머리·흰 연구 가운은 인물 참조에 대체로 부합한다. 지소영의 짙은 둥근 목 상의와 가운 주머니의 펜은 이전 스틸을 잘 잇는다. 서지민의 긴 머리, 머리 위 보안경과 손목시계도 참조와 맞는다. 그러나 두 여성만 보여야 하는 장면에 찰리의 로봇 몸체가 침대 위로 선명하게 들어왔다. 손목에는 고정 띠로 보이는 부분도 남아 있어 손이 풀린 상태와 어긋난다.",
        "hard_violations": [
         "이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
         "경보 설비를 추가하지 말라는 지시와 달리 발광하는 경보등을 배경에 추가했다."
        ],
        "physics": "지소영의 골반은 의자 좌판에 지지되고 등 뒤에 등받이가 있어, 앉은 채 상체를 갑자기 일으키는 자세가 가능하다. 서지민의 한 손은 오른쪽 작업대 가장자리에 닿고 다른 팔은 팔꿈치를 굽힌 상태다. 발은 화면 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 찰리는 금속 침대에 누워 지지되며, 무지지 물체나 불가능한 신체 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "지소영은 화면 위쪽 오른편으로 눈을 들고, 서지민도 턱과 눈을 위쪽 오른편으로 향한다. 두 사람 모두 천장 쪽을 보는 반응으로 읽힌다. 서지민의 몸통은 옆으로 돌아 있고 얼굴은 위를 향해 놀란 회전 동작을 나타낸다.",
        "built_space": "지소영이 앉은 의자 1개, 중앙 뒤 금속 침대 1개, 오른쪽 모니터 1개와 왼쪽 전경의 부분 모니터 1개가 보인다. 왼쪽 가장자리에는 기계 팔도 일부 보인다. 유리 칸막이와 금속 바닥은 이전 장소와 유사하다. 다만 침대 윗면과 바닥을 넓게 내려다보아 지정된 낮은 가슴 높이의 약한 올려다보기보다 높은 시점으로 읽힌다. 천장 자체는 거의 없고 배경 상부에 적어도 다섯 개의 붉은 발광 설비가 보인다. 불가능한 반사는 확인되지 않는다.",
        "entities": "두 여성의 연령 대비, 검은 머리와 연구 가운은 지정 인물에 대체로 맞고 서지민의 보안경과 시계도 유지된다. 지소영은 이전 스틸의 짙은 둥근 목 상의 대신 단추 달린 회색 셔츠를 입었으며 가운 가슴에도 로고가 생겼다. 중앙 침대에는 이 장면에서 제외해야 하는 찰리가 거의 전신으로 보인다. 찰리의 손목 주변 고정 띠도 남아 있어 손이 자유롭다는 상태가 명확히 구현되지 않았다.",
        "hard_violations": [
         "이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
         "경보 설비를 추가하지 말라는 지시와 달리 여러 붉은 경보등을 배경에 추가했다."
        ],
        "physics": "지소영은 골반을 의자에 두고 상체를 비튼 상태이며 좌판·등받이·팔걸이의 배치가 착석 자세와 양립한다. 서지민은 팔꿈치를 굽히고 몸통을 돌린 서 있는 자세로, 화면 밖 발을 보이지 않는다는 이유만으로 부유했다고 판단할 수 없다. 찰리의 몸은 침대가 지지한다. 해부학적으로 불가능한 동작이나 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
     "[gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 발광하는 경보등을 배경에 추가했다."
    ],
    "A": [
     "[gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다.",
     "[gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 여러 붉은 경보등을 배경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1500,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "지시된 상향 시선과 카메라 구도를 훌륭히 구현했으며 이전 컷에 고정된 지소영의 의상과 주변 환경(로봇 팔 등)을 충실히 유지했으나, 서지민의 바지 색상이 참조와 다릅니다.  ★위반: [gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다. / [gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 여러 붉은 경보등을 배경에 추가했다."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "두 인물의 시선 처리와 구도는 적절하나, 이전 컷에서 완전히 고정되어야 할 지소영의 셔츠 형태가 변경되었고 배경의 로봇 팔 등 주요 디테일이 누락되었습니다.  ★위반: [gpt-high] 이 장면에서 제외하도록 지정된 찰리의 몸체를 침대 위에 추가로 노출했다. / [gpt-high] 경보 설비를 추가하지 말라는 지시와 달리 발광하는 경보등을 배경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지소영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S88sh11_sel.png",
    "asset_id": "e64a82d6-b49d-422f-b639-b95001c23f41",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 지소영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:783266>",
    "asset_id": "8f1117f8-c2a3-453c-a06d-dbb6f42b564c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 서지민: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869875>",
    "asset_id": "bdc552c6-5e2a-4bae-affd-3dbee764db52",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ebd-5ca0-7653-8385-a02434c50d7e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh11"
  }
 },
 "S89sh45::signage": {
  "fp": "593f0c4a102dc3ce",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::600d52c25a37689e": {
  "subjects": [],
  "subject_text": "제주도 연구소 옥상 잔디밭·안테나·스피커 구역, 연결 계단\n야외 정원이 조성된 옥상 플랫폼으로 거대한 통신 장비가 설치되어 있다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L168",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::research_rooftop": {
  "input_fingerprint": "8bbb262c8a3a4799",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "research_rooftop",
    "tags": [
     "S89sh45",
     "S89sh59"
    ]
   },
   "context_sig": "cee052023da62c12"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 옥상 잔디밭·안테나·스피커 구역, 연결 계단: 야외 정원이 조성된 옥상 플랫폼으로 거대한 통신 장비가 설치되어 있다. (특징: 옥상 잔디밭에 세워진 거대한 파라볼라 안테나와 360도 대형 스피커 모듈; 수신기를 몸에 장착한 현우와 찰리; 착륙하는 헬기에서 레펠로 강하하는 특임대와 인공지능 전투병들; 스피커 진동파로 인해 눈이 뒤집히며 픽픽 쓰러지는 인공지능 병사들; 유탄 파편에 맞아 피 흘리는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 옥상 공원 잔디밭에 거대한 안테나와 360도 사방 스피커\n- 옥상에 윤성찬의 헬기가 착지한다,\n\nTIME OF DAY (lock): day to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 옥상 잔디밭·안테나·스피커 구역, 연결 계단: 야외 정원이 조성된 옥상 플랫폼으로 거대한 통신 장비가 설치되어 있다. (특징: 옥상 잔디밭에 세워진 거대한 파라볼라 안테나와 360도 대형 스피커 모듈; 수신기를 몸에 장착한 현우와 찰리; 착륙하는 헬기에서 레펠로 강하하는 특임대와 인공지능 전투병들; 스피커 진동파로 인해 눈이 뒤집히며 픽픽 쓰러지는 인공지능 병사들; 유탄 파편에 맞아 피 흘리는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 옥상 공원 잔디밭에 거대한 안테나와 360도 사방 스피커\n- 옥상에 윤성찬의 헬기가 착지한다,\n\nTIME OF DAY (lock): day to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_rooftop_9331cf.png",
  "asset_id": "f9fa22f5-21d5-415e-9f40-6d648034d538",
  "input_asset_ids": [
   "689ca5c5-f280-41e4-8ff3-8543eb6a4682"
  ],
  "origin_tag": "S89sh45",
  "place_text": "At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.",
  "origin_inputs": {
   "place_text": "At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.",
   "time_of_day_en": "day to night",
   "conti_asset_id": "689ca5c5-f280-41e4-8ff3-8543eb6a4682"
  }
 },
 "S89sh45::bgfirst_bg": {
  "input_fingerprint": "3bfd19654245172f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활짝 열린 헬기 문밖으로 비릿한 미소를 띠며 한쪽 발을 허공을 향해 내딛고 있는 윤성찬의 거만한 상체.\n\nLOCATION (lock): At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.\n\nTIME OF DAY (lock): day to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from a low rooftop position outside the landed helicopter's open doorway, offset from its outward axis and angled up toward 윤성찬's upper body. Keep him center-right in three-quarter view, smiling toward the rooftop ahead offscreen left as one leg extends outward through the lower frame; show only a bounded portion of the doorway behind him, with rooftop space providing scale. Continue the established lateral track gently, emphasizing his outward weight shift while retaining the helicopter on the right for the subsequent move toward 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 윤성찬 exiting the helicopter in the middle-right of the frame, midground, moves toward Open rooftop space toward screen left; Open helicopter doorway behind 윤성찬 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Helicopter doorway (Fully open on the landed helicopter) — Seen from outside at an oblique angle, behind 윤성찬 on the right; used as Frames his exit without allowing the aircraft to overwhelm the image; Research-center rooftop (The helicopter has landed here); used as Provides lower-frame spatial context and a realistic reference for the outward step.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast keep his smile legible without adding an unsupported weather effect or lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활짝 열린 헬기 문밖으로 비릿한 미소를 띠며 한쪽 발을 허공을 향해 내딛고 있는 윤성찬의 거만한 상체.\n\nLOCATION (lock): At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof.\n\nTIME OF DAY (lock): day to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from a low rooftop position outside the landed helicopter's open doorway, offset from its outward axis and angled up toward 윤성찬's upper body. Keep him center-right in three-quarter view, smiling toward the rooftop ahead offscreen left as one leg extends outward through the lower frame; show only a bounded portion of the doorway behind him, with rooftop space providing scale. Continue the established lateral track gently, emphasizing his outward weight shift while retaining the helicopter on the right for the subsequent move toward 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 윤성찬 exiting the helicopter in the middle-right of the frame, midground, moves toward Open rooftop space toward screen left; Open helicopter doorway behind 윤성찬 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Helicopter doorway (Fully open on the landed helicopter) — Seen from outside at an oblique angle, behind 윤성찬 on the right; used as Frames his exit without allowing the aircraft to overwhelm the image; Research-center rooftop (The helicopter has landed here); used as Provides lower-frame spatial context and a realistic reference for the outward step.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast keep his smile legible without adding an unsupported weather effect or lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh45__bgfirst_bg.png",
  "asset_id": "aa033de8-f3e0-4397-8529-2ff50e103f64",
  "input_asset_ids": [
   "689ca5c5-f280-41e4-8ff3-8543eb6a4682",
   "f9fa22f5-21d5-415e-9f40-6d648034d538"
  ]
 },
 "S89sh45": {
  "input_fingerprint": "e9774b91ce9036b4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 헬기 문밖으로 비릿한 미소를 띠며 한쪽 발을 허공을 향해 내딛고 있는 윤성찬의 거만한 상체.\n\nLOCATION (lock): At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from a low rooftop position outside the landed helicopter's open doorway, offset from its outward axis and angled up toward 윤성찬's upper body. Keep him center-right in three-quarter view, smiling toward the rooftop ahead offscreen left as one leg extends outward through the lower frame; show only a bounded portion of the doorway behind him, with rooftop space providing scale. Continue the established lateral track gently, emphasizing his outward weight shift while retaining the helicopter on the right for the subsequent move toward 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 윤성찬 exiting the helicopter in the middle-right of the frame, midground, moves toward Open rooftop space toward screen left; Open helicopter doorway behind 윤성찬 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Helicopter doorway (Fully open on the landed helicopter) — Seen from outside at an oblique angle, behind 윤성찬 on the right; used as Frames his exit without allowing the aircraft to overwhelm the image; Research-center rooftop (The helicopter has landed here); used as Provides lower-frame spatial context and a realistic reference for the outward step.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast keep his smile legible without adding an unsupported weather effect or lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop has lawn, a large antenna and omnidirectional speakers, with disabled artificial soldiers left from the sonic attack and bombardment damage accumulating. A Yubik helicopter has landed; Charlie is free of the bed restraints but still has experimental sensors attached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 헬기 문밖으로 비릿한 미소를 띠며 한쪽 발을 허공을 향해 내딛고 있는 윤성찬의 거만한 상체.\n\nLOCATION (lock): At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from a low rooftop position outside the landed helicopter's open doorway, offset from its outward axis and angled up toward 윤성찬's upper body. Keep him center-right in three-quarter view, smiling toward the rooftop ahead offscreen left as one leg extends outward through the lower frame; show only a bounded portion of the doorway behind him, with rooftop space providing scale. Continue the established lateral track gently, emphasizing his outward weight shift while retaining the helicopter on the right for the subsequent move toward 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 윤성찬 exiting the helicopter in the middle-right of the frame, midground, moves toward Open rooftop space toward screen left; Open helicopter doorway behind 윤성찬 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Helicopter doorway (Fully open on the landed helicopter) — Seen from outside at an oblique angle, behind 윤성찬 on the right; used as Frames his exit without allowing the aircraft to overwhelm the image; Research-center rooftop (The helicopter has landed here); used as Provides lower-frame spatial context and a realistic reference for the outward step.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast keep his smile legible without adding an unsupported weather effect or lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop has lawn, a large antenna and omnidirectional speakers, with disabled artificial soldiers left from the sonic attack and bombardment damage accumulating. A Yubik helicopter has landed; Charlie is free of the bed restraints but still has experimental sensors attached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 헬기 문밖으로 비릿한 미소를 띠며 한쪽 발을 허공을 향해 내딛고 있는 윤성찬의 거만한 상체.\n\nLOCATION (lock): At the open doorway of a helicopter newly landed on the research institute's rooftop, stepping out onto the exposed roof. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Observe directly from a low rooftop position outside the landed helicopter's open doorway, offset from its outward axis and angled up toward 윤성찬's upper body. Keep him center-right in three-quarter view, smiling toward the rooftop ahead offscreen left as one leg extends outward through the lower frame; show only a bounded portion of the doorway behind him, with rooftop space providing scale. Continue the established lateral track gently, emphasizing his outward weight shift while retaining the helicopter on the right for the subsequent move toward 찰리.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 윤성찬 exiting the helicopter in the middle-right of the frame, midground, moves toward Open rooftop space toward screen left; Open helicopter doorway behind 윤성찬 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Helicopter doorway (Fully open on the landed helicopter) — Seen from outside at an oblique angle, behind 윤성찬 on the right; used as Frames his exit without allowing the aircraft to overwhelm the image; Research-center rooftop (The helicopter has landed here); used as Provides lower-frame spatial context and a realistic reference for the outward step.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime ambient illumination and controlled contrast keep his smile legible without adding an unsupported weather effect or lighting accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop has lawn, a large antenna and omnidirectional speakers, with disabled artificial soldiers left from the sonic attack and bombardment damage accumulating. A Yubik helicopter has landed; Charlie is free of the bed restraints but still has experimental sensors attached.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 윤성찬 (한국인 남성, 70대 후반의 얼굴, 짧게 정돈된 머리) — wearing: 결벽증이 묻어나는 티끌 하나 없는 고급스러운 차콜 그레이 색상의 맞춤 정장과 넥타이. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh45__bgfirst_bg.png",
     "asset_id": "aa033de8-f3e0-4397-8529-2ff50e103f64",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S89sh45.png",
     "asset_id": "689ca5c5-f280-41e4-8ff3-8543eb6a4682",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_rooftop_9331cf.png",
     "asset_id": "f9fa22f5-21d5-415e-9f40-6d648034d538",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:860900>",
     "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "윤성찬은 화면 왼쪽 밖의 옥상을 향해 시선을 던지고 있으며, 한쪽 발을 헬기 밖 허공으로 뻗고 있습니다.",
    "built_space": "참조된 옥상의 안테나와 위성 접시, 바다 배경이 나타나 있으나, 지시되지 않은 여러 인물과 쓰러진 병사들의 몸이 옥상 공간에 어지럽게 배치되어 있습니다.",
    "entities": "윤성찬은 지정된 연령대의 한국인 남성 얼굴과 차콜 그레이 정장 차림을 잘 반영하고 있으나, 샷 텍스트에 언급되지 않은 인물들이 다수 존재합니다.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 추가 인물들(흰옷을 입은 인물, 배경에 서 있는 군인들)이 임의로 배치됨",
     "[gpt-high] 윤성찬만 등장하도록 제한된 숏에 센서 부착 인물 한 명과 서 있는 무장 병사 세 명을 추가했다."
    ],
    "physics": "윤성찬은 왼손으로 헬기 문틀을 잡고 한쪽 발을 허공에 뻗으며 밖으로 나서는 무게 중심 이동을 자연스럽게 보여줍니다."
   },
   {
    "label": "B",
    "direction": "윤성찬은 화면 왼쪽 밖을 향해 여유로운 미소를 지으며 시선을 두고 있으며, 뻗은 다리는 프레임 하단을 향해 내려가고 있습니다.",
    "built_space": "로우 앵글에서 헬기 문과 옥상의 안테나, 위성 접시, 원경의 바다와 섬이 정확한 비율과 위치로 묘사되었으며, 불필요한 요소 없이 깔끔하게 구성되었습니다.",
    "entities": "윤성찬은 70대 후반의 외모, 정돈된 머리, 고급스러운 차콜 그레이 맞춤 정장과 넥타이를 참조 이미지 그대로 완벽하게 재현했습니다.",
    "hard_violations": [],
    "physics": "신체는 헬기 내부에서 바깥으로 나서는 무게 중심을 지니고 있으며, 발은 프레임 밖으로 뻗어 나가며 허공을 딛는 동작을 무리 없이 표현하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 미디엄 샷보다 프레임이 지나치게 넓으며, 샷 텍스트에 없는 다수의 인물(흰옷을 입은 사람과 군인들)이 임의로 추가되어 치명적인 위반이 발생했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "명시된 로우 앵글의 미디엄 샷과 프레임 구도를 정확히 구현하였으며, 인물의 표정, 동작, 단독 등장 조건까지 지시사항을 완벽하게 충족하는 훌륭한 결과물입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 화면 왼쪽 밖의 옥상을 향해 시선을 던지고 있으며, 한쪽 발을 헬기 밖 허공으로 뻗고 있습니다.",
        "built_space": "참조된 옥상의 안테나와 위성 접시, 바다 배경이 나타나 있으나, 지시되지 않은 여러 인물과 쓰러진 병사들의 몸이 옥상 공간에 어지럽게 배치되어 있습니다.",
        "entities": "윤성찬은 지정된 연령대의 한국인 남성 얼굴과 차콜 그레이 정장 차림을 잘 반영하고 있으나, 샷 텍스트에 언급되지 않은 인물들이 다수 존재합니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 추가 인물들(흰옷을 입은 인물, 배경에 서 있는 군인들)이 임의로 배치됨"
        ],
        "physics": "윤성찬은 왼손으로 헬기 문틀을 잡고 한쪽 발을 허공에 뻗으며 밖으로 나서는 무게 중심 이동을 자연스럽게 보여줍니다."
       },
       {
        "label": "B",
        "direction": "윤성찬은 화면 왼쪽 밖을 향해 여유로운 미소를 지으며 시선을 두고 있으며, 뻗은 다리는 프레임 하단을 향해 내려가고 있습니다.",
        "built_space": "로우 앵글에서 헬기 문과 옥상의 안테나, 위성 접시, 원경의 바다와 섬이 정확한 비율과 위치로 묘사되었으며, 불필요한 요소 없이 깔끔하게 구성되었습니다.",
        "entities": "윤성찬은 70대 후반의 외모, 정돈된 머리, 고급스러운 차콜 그레이 맞춤 정장과 넥타이를 참조 이미지 그대로 완벽하게 재현했습니다.",
        "hard_violations": [],
        "physics": "신체는 헬기 내부에서 바깥으로 나서는 무게 중심을 지니고 있으며, 발은 프레임 밖으로 뻗어 나가며 허공을 딛는 동작을 무리 없이 표현하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 미디엄 샷보다 프레임이 지나치게 넓으며, 샷 텍스트에 없는 다수의 인물(흰옷을 입은 사람과 군인들)이 임의로 추가되어 치명적인 위반이 발생했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "명시된 로우 앵글의 미디엄 샷과 프레임 구도를 정확히 구현하였으며, 인물의 표정, 동작, 단독 등장 조건까지 지시사항을 완벽하게 충족하는 훌륭한 결과물입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "윤성찬은 화면 왼쪽 밖의 옥상을 향해 시선을 던지고 있으며, 한쪽 발을 헬기 밖 허공으로 뻗고 있습니다.",
        "built_space": "참조된 옥상의 안테나와 위성 접시, 바다 배경이 나타나 있으나, 지시되지 않은 여러 인물과 쓰러진 병사들의 몸이 옥상 공간에 어지럽게 배치되어 있습니다.",
        "entities": "윤성찬은 지정된 연령대의 한국인 남성 얼굴과 차콜 그레이 정장 차림을 잘 반영하고 있으나, 샷 텍스트에 언급되지 않은 인물들이 다수 존재합니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 추가 인물들(흰옷을 입은 인물, 배경에 서 있는 군인들)이 임의로 배치됨"
        ],
        "physics": "윤성찬은 왼손으로 헬기 문틀을 잡고 한쪽 발을 허공에 뻗으며 밖으로 나서는 무게 중심 이동을 자연스럽게 보여줍니다."
       },
       {
        "label": "B",
        "direction": "윤성찬은 화면 왼쪽 밖을 향해 여유로운 미소를 지으며 시선을 두고 있으며, 뻗은 다리는 프레임 하단을 향해 내려가고 있습니다.",
        "built_space": "로우 앵글에서 헬기 문과 옥상의 안테나, 위성 접시, 원경의 바다와 섬이 정확한 비율과 위치로 묘사되었으며, 불필요한 요소 없이 깔끔하게 구성되었습니다.",
        "entities": "윤성찬은 70대 후반의 외모, 정돈된 머리, 고급스러운 차콜 그레이 맞춤 정장과 넥타이를 참조 이미지 그대로 완벽하게 재현했습니다.",
        "hard_violations": [],
        "physics": "신체는 헬기 내부에서 바깥으로 나서는 무게 중심을 지니고 있으며, 발은 프레임 밖으로 뻗어 나가며 허공을 딛는 동작을 무리 없이 표현하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "윤성찬만을 중앙 오른쪽에서 올려다보는 상체 중심 구도와 출구·옥상 배치는 충실하지만, 시선이 지정된 화면 밖 왼쪽보다 카메라 쪽을 향한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "왼쪽을 향한 미소와 내딛는 동작은 맞지만, 허용되지 않은 인물 네 명을 추가했고 발까지 담는 넓은 구도로 상체 중심 미디엄 숏을 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "몸은 오른쪽 헬기 출입구에서 앞쪽·화면 왼쪽으로 나오며 한쪽 다리가 하단 밖으로 뻗는다. 얼굴은 거의 정면이고 눈길은 카메라 부근에서 약간 화면 오른쪽 위로 향해, 지정된 화면 밖 왼쪽 옥상을 바라보는 모습은 명확하지 않다. 입가에는 자신만만한 미소가 있다. 무기나 손에 든 물건은 없다.",
        "built_space": "오른쪽에 완전히 열린 출입구 하나와 내부 좌석 하나, 하단에 착륙 스키드 일부가 보인다. 왼쪽에는 안테나 탑 하나, 그 주위의 나팔형 스피커 네 개, 접시 안테나 하나가 있으며 잔디·옥상 포장·난간·바다와 섬이 참고 장소와 대응한다. 인물은 문턱 앞 중앙 오른쪽에 있고 카메라는 옥상 쪽 낮은 위치에서 비스듬히 올려다본다. 출입구는 인물 뒤의 제한된 영역을 차지한다. 기체의 하늘 반사도 가능한 방향이다.",
        "entities": "보이는 사람은 윤성찬 한 명뿐이다. 고령의 동아시아계 남성으로 짧고 정돈된 회색 머리, 콧수염, 얼굴 주름이 참고 인물과 대체로 일치한다. 깨끗한 차콜 정장, 흰 셔츠, 어두운 넥타이와 흰 포켓스퀘어도 맞는다. 헬기와 연구소 잔디 옥상, 통신 설비가 보인다. 기체 제작사는 판독할 수 없으며, 보이는 옥상에는 쓰러진 인공 병사나 뚜렷한 폭격 잔해가 없다.",
        "hard_violations": [],
        "physics": "앞으로 내민 다리는 무릎 아래에서 화면 밖으로 이어지고 반대쪽 다리는 뒤에 남아 있다. 두 발의 실제 접촉점은 하단 크롭 때문에 확인할 수 없지만, 문턱에서 뒷다리로 체중을 받으며 나오는 동작으로 읽히며 몸 전체가 공중에 떠 있다는 증거는 없다. 양팔과 손은 자연스럽게 내려와 있다. 헬기는 하단에 보이는 착륙장치로 지지되는 구조다."
       },
       {
        "label": "B",
        "direction": "윤성찬의 얼굴과 눈은 화면 왼쪽 옥상을 향하며 입가에 미소가 있다. 앞다리는 왼쪽 아래의 옥상 쪽으로 뻗고, 화면 오른쪽 손은 문틀을 짚는다. 배경의 센서 부착 인물은 왼쪽을 보고 있으며 서 있는 병사 세 명은 대체로 전경 쪽을 향한다. 병사들의 총구는 주로 아래로 내려가 있고 특정 표적을 겨누지는 않는다.",
        "built_space": "오른쪽에 열린 출입구 하나, 내부 좌석 하나와 하단 착륙장치가 보인다. 왼쪽의 안테나 탑 하나, 나팔형 스피커 네 개, 접시 안테나 하나와 잔디·난간·해안 풍경은 장소 참고와 대응한다. 인물은 출입구 중앙 오른쪽에 있지만 앞발 전체와 뒷다리 대부분까지 보여 미디엄 숏보다 넓다. 상단에는 회전날개 일부가 걸리고, 넓게 드러난 옥상은 추가 인물과 쓰러진 병사들로 채워져 있다.",
        "entities": "주인공의 고령 동아시아계 남성 외형, 회색 머리와 콧수염, 차콜 정장과 넥타이는 참고와 대체로 일치한다. 그러나 센서와 케이블을 부착한 젊은 인물 한 명, 무장한 채 서 있는 병사 세 명이 추가로 등장한다. 지면에는 여러 인공 병사형 몸체와 파편이 있어 잔존 상태 설정은 표현하지만, 이 숏에서 윤성찬만 보이게 하라는 인물 제한을 어겼다. 추가 문구나 그래픽은 보이지 않는다.",
        "hard_violations": [
         "윤성찬만 등장하도록 제한된 숏에 센서 부착 인물 한 명과 서 있는 무장 병사 세 명을 추가했다."
        ],
        "physics": "주인공은 화면 오른쪽 손으로 문틀을 붙잡고, 뒤쪽 다리를 출입구·착륙장치 쪽에 남긴 채 앞발을 허공으로 내민다. 뒤쪽 발의 접촉점은 하단에서 잘리지만 손의 지지와 다리의 체중 이동이 보여 가능한 하차 동작이다. 배경의 서 있는 인물들은 발로 옥상을 딛고 있고 쓰러진 몸체와 잔해는 지면에 놓여 있다. 회전날개는 화면 밖 기체 상부로 이어져 지지 없는 부유 물체로 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "윤성찬만을 중앙 오른쪽에서 올려다보는 상체 중심 구도와 출구·옥상 배치는 충실하지만, 시선이 지정된 화면 밖 왼쪽보다 카메라 쪽을 향한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "왼쪽을 향한 미소와 내딛는 동작은 맞지만, 허용되지 않은 인물 네 명을 추가했고 발까지 담는 넓은 구도로 상체 중심 미디엄 숏을 벗어났다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "몸은 오른쪽 헬기 출입구에서 앞쪽·화면 왼쪽으로 나오며 한쪽 다리가 하단 밖으로 뻗는다. 얼굴은 거의 정면이고 눈길은 카메라 부근에서 약간 화면 오른쪽 위로 향해, 지정된 화면 밖 왼쪽 옥상을 바라보는 모습은 명확하지 않다. 입가에는 자신만만한 미소가 있다. 무기나 손에 든 물건은 없다.",
        "built_space": "오른쪽에 완전히 열린 출입구 하나와 내부 좌석 하나, 하단에 착륙 스키드 일부가 보인다. 왼쪽에는 안테나 탑 하나, 그 주위의 나팔형 스피커 네 개, 접시 안테나 하나가 있으며 잔디·옥상 포장·난간·바다와 섬이 참고 장소와 대응한다. 인물은 문턱 앞 중앙 오른쪽에 있고 카메라는 옥상 쪽 낮은 위치에서 비스듬히 올려다본다. 출입구는 인물 뒤의 제한된 영역을 차지한다. 기체의 하늘 반사도 가능한 방향이다.",
        "entities": "보이는 사람은 윤성찬 한 명뿐이다. 고령의 동아시아계 남성으로 짧고 정돈된 회색 머리, 콧수염, 얼굴 주름이 참고 인물과 대체로 일치한다. 깨끗한 차콜 정장, 흰 셔츠, 어두운 넥타이와 흰 포켓스퀘어도 맞는다. 헬기와 연구소 잔디 옥상, 통신 설비가 보인다. 기체 제작사는 판독할 수 없으며, 보이는 옥상에는 쓰러진 인공 병사나 뚜렷한 폭격 잔해가 없다.",
        "hard_violations": [],
        "physics": "앞으로 내민 다리는 무릎 아래에서 화면 밖으로 이어지고 반대쪽 다리는 뒤에 남아 있다. 두 발의 실제 접촉점은 하단 크롭 때문에 확인할 수 없지만, 문턱에서 뒷다리로 체중을 받으며 나오는 동작으로 읽히며 몸 전체가 공중에 떠 있다는 증거는 없다. 양팔과 손은 자연스럽게 내려와 있다. 헬기는 하단에 보이는 착륙장치로 지지되는 구조다."
       },
       {
        "label": "A",
        "direction": "윤성찬의 얼굴과 눈은 화면 왼쪽 옥상을 향하며 입가에 미소가 있다. 앞다리는 왼쪽 아래의 옥상 쪽으로 뻗고, 화면 오른쪽 손은 문틀을 짚는다. 배경의 센서 부착 인물은 왼쪽을 보고 있으며 서 있는 병사 세 명은 대체로 전경 쪽을 향한다. 병사들의 총구는 주로 아래로 내려가 있고 특정 표적을 겨누지는 않는다.",
        "built_space": "오른쪽에 열린 출입구 하나, 내부 좌석 하나와 하단 착륙장치가 보인다. 왼쪽의 안테나 탑 하나, 나팔형 스피커 네 개, 접시 안테나 하나와 잔디·난간·해안 풍경은 장소 참고와 대응한다. 인물은 출입구 중앙 오른쪽에 있지만 앞발 전체와 뒷다리 대부분까지 보여 미디엄 숏보다 넓다. 상단에는 회전날개 일부가 걸리고, 넓게 드러난 옥상은 추가 인물과 쓰러진 병사들로 채워져 있다.",
        "entities": "주인공의 고령 동아시아계 남성 외형, 회색 머리와 콧수염, 차콜 정장과 넥타이는 참고와 대체로 일치한다. 그러나 센서와 케이블을 부착한 젊은 인물 한 명, 무장한 채 서 있는 병사 세 명이 추가로 등장한다. 지면에는 여러 인공 병사형 몸체와 파편이 있어 잔존 상태 설정은 표현하지만, 이 숏에서 윤성찬만 보이게 하라는 인물 제한을 어겼다. 추가 문구나 그래픽은 보이지 않는다.",
        "hard_violations": [
         "윤성찬만 등장하도록 제한된 숏에 센서 부착 인물 한 명과 서 있는 무장 병사 세 명을 추가했다."
        ],
        "physics": "주인공은 화면 오른쪽 손으로 문틀을 붙잡고, 뒤쪽 다리를 출입구·착륙장치 쪽에 남긴 채 앞발을 허공으로 내민다. 뒤쪽 발의 접촉점은 하단에서 잘리지만 손의 지지와 다리의 체중 이동이 보여 가능한 하차 동작이다. 배경의 서 있는 인물들은 발로 옥상을 딛고 있고 쓰러진 몸체와 잔해는 지면에 놓여 있다. 회전날개는 화면 밖 기체 상부로 이어져 지지 없는 부유 물체로 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.486,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.236,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 추가 인물들(흰옷을 입은 인물, 배경에 서 있는 군인들)이 임의로 배치됨",
     "[gpt-high] 윤성찬만 등장하도록 제한된 숏에 센서 부착 인물 한 명과 서 있는 무장 병사 세 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 236,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 236,
    "verdict_ko": "지정된 미디엄 샷보다 프레임이 지나치게 넓으며, 샷 텍스트에 없는 다수의 인물(흰옷을 입은 사람과 군인들)이 임의로 추가되어 치명적인 위반이 발생했습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 추가 인물들(흰옷을 입은 인물, 배경에 서 있는 군인들)이 임의로 배치됨 / [gpt-high] 윤성찬만 등장하도록 제한된 숏에 센서 부착 인물 한 명과 서 있는 무장 병사 세 명을 추가했다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "명시된 로우 앵글의 미디엄 샷과 프레임 구도를 정확히 구현하였으며, 인물의 표정, 동작, 단독 등장 조건까지 지시사항을 완벽하게 충족하는 훌륭한 결과물입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_rooftop_9331cf.png",
    "asset_id": "f9fa22f5-21d5-415e-9f40-6d648034d538",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 윤성찬: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:860900>",
    "asset_id": "f51a7e69-64ab-4e45-a7c5-a7f8a2d1a69d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ec2-8bf0-762d-a8cf-db5721046c3c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh45__bgfirst_bg.png",
   "bg_asset_id": "aa033de8-f3e0-4397-8529-2ff50e103f64",
   "bg_record_key": "S89sh45::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "research_rooftop",
   "groupbg_asset_id": "f9fa22f5-21d5-415e-9f40-6d648034d538"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S89sh59::signage": {
  "fp": "d9df77028db11f44",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S89sh59": {
  "input_fingerprint": "ace4ee27df31a021",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 거대한 빛기둥 중심에 서서 고개를 꺾은 채 바닥에 쓰러진 이현우를 부드럽게 내려다보고 있는 찰리의 구도.\n\nLOCATION (lock): On the research institute's exposed rooftop lawn, near the antenna and all-direction speaker array, at the center of the rising energy beam. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Maintain the slow inward dolly from beside 이현우's fallen shoulder, low above the rooftop and tilted upward toward 찰리's three-quarter profile, observing their exchange directly rather than adopting either character's eyes. Place 이현우's partial profile along the lower-left foreground and 찰리's upper body in the center-right midground, with his inclined head and gentle downward gaze answering 이현우's upward look. Hold their positions and eyeline steady so that decreasing camera distance alone intensifies this middle phase of the farewell, while the rising beam remains behind and around 찰리.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rooftop park lawn (Beneath the fallen 이현우) — Seen at a shallow angle along the lower frame edge; used as Provides a narrow grounding plane for the low camera and the injured figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brilliant white beam provides the scene's overwhelming illumination, with exposure preserving the tenderness of the faces and the definition of 찰리's hard surfaces before the eventual whiteout.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A brilliant, nearly white column of light rises from Charlie's rapidly spinning chest ring; the attached experimental sensors remain in place. The laboratory is already heavily damaged by bombardment, and the Ubik command helicopter has landed on the roof. 이현우: He has been wounded by grenade fragments and still wears the wireless receiver attached earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 거대한 빛기둥 중심에 서서 고개를 꺾은 채 바닥에 쓰러진 이현우를 부드럽게 내려다보고 있는 찰리의 구도.\n\nLOCATION (lock): On the research institute's exposed rooftop lawn, near the antenna and all-direction speaker array, at the center of the rising energy beam. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Maintain the slow inward dolly from beside 이현우's fallen shoulder, low above the rooftop and tilted upward toward 찰리's three-quarter profile, observing their exchange directly rather than adopting either character's eyes. Place 이현우's partial profile along the lower-left foreground and 찰리's upper body in the center-right midground, with his inclined head and gentle downward gaze answering 이현우's upward look. Hold their positions and eyeline steady so that decreasing camera distance alone intensifies this middle phase of the farewell, while the rising beam remains behind and around 찰리.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rooftop park lawn (Beneath the fallen 이현우) — Seen at a shallow angle along the lower frame edge; used as Provides a narrow grounding plane for the low camera and the injured figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brilliant white beam provides the scene's overwhelming illumination, with exposure preserving the tenderness of the faces and the definition of 찰리's hard surfaces before the eventual whiteout.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A brilliant, nearly white column of light rises from Charlie's rapidly spinning chest ring; the attached experimental sensors remain in place. The laboratory is already heavily damaged by bombardment, and the Ubik command helicopter has landed on the roof. 이현우: He has been wounded by grenade fragments and still wears the wireless receiver attached earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 거대한 빛기둥 중심에 서서 고개를 꺾은 채 바닥에 쓰러진 이현우를 부드럽게 내려다보고 있는 찰리의 구도.\n\nLOCATION (lock): On the research institute's exposed rooftop lawn, near the antenna and all-direction speaker array, at the center of the rising energy beam. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Maintain the slow inward dolly from beside 이현우's fallen shoulder, low above the rooftop and tilted upward toward 찰리's three-quarter profile, observing their exchange directly rather than adopting either character's eyes. Place 이현우's partial profile along the lower-left foreground and 찰리's upper body in the center-right midground, with his inclined head and gentle downward gaze answering 이현우's upward look. Hold their positions and eyeline steady so that decreasing camera distance alone intensifies this middle phase of the farewell, while the rising beam remains behind and around 찰리.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rooftop park lawn (Beneath the fallen 이현우) — Seen at a shallow angle along the lower frame edge; used as Provides a narrow grounding plane for the low camera and the injured figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brilliant white beam provides the scene's overwhelming illumination, with exposure preserving the tenderness of the faces and the definition of 찰리's hard surfaces before the eventual whiteout.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A brilliant, nearly white column of light rises from Charlie's rapidly spinning chest ring; the attached experimental sensors remain in place. The laboratory is already heavily damaged by bombardment, and the Ubik command helicopter has landed on the roof. 이현우: He has been wounded by grenade fragments and still wears the wireless receiver attached earlier.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우는 위를 쳐다보고 있으나, 찰리는 고개를 숙이지 않고 정면 허공을 응시하고 있어 두 인물의 시선이 교차하지 않음.",
    "built_space": "옥상 잔디밭, 안테나, 배경의 헬리콥터 등이 지시된 위치와 샷 비율에 맞게 잘 배치되어 있음.",
    "entities": "찰리와 이현우의 외형 묘사는 레퍼런스와 대체로 일치하며, 이현우의 부상 상태와 옷깃의 무선 수신기도 잘 표현됨.",
    "hard_violations": [
     "[gemini-pro] 찰리의 시선이 이현우를 향하지 않고 정면을 응시하여 '고개를 꺾은 채 쓰러진 이현우를 부드럽게 내려다보고 있는(inclined head and gentle downward gaze)'이라는 핵심 행동 및 구도 지시를 위반함."
    ],
    "physics": "이현우의 등과 머리가 지면에 지탱되어 있으며, 찰리 역시 땅에 두 발을 딛고 자연스럽게 서 있음."
   },
   {
    "label": "B",
    "direction": "바닥에 쓰러진 이현우가 위를 올려다보고, 찰리가 고개를 꺾어 그를 부드럽게 내려다보며 두 캐릭터 간의 시선이 정확히 맞닿아 있음.",
    "built_space": "연구소 옥상 잔디밭 지면, 뒤편의 안테나 구조물이 로우 앵글 카메라 구도에 맞게 적절히 묘사됨(크롭으로 인해 헬기 미노출).",
    "entities": "찰리의 기계 외형과 얼굴 마스크, 가슴의 푸른빛 심볼이 명확하며, 이현우의 상처 입은 얼굴과 가슴에 위치한 무선 수신기가 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "이현우는 바닥에 안정적으로 누워 지탱되고 있고, 찰리는 잔디밭 위에 무게 중심을 잘 잡고 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "찰리가 고개를 숙여 이현우를 내려다보는 시선 교환을 완벽하게 포착하였으며, 카메라 구도와 캐릭터들의 세부 디테일을 프롬프트에 맞게 훌륭히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "배경과 인물의 질감은 사실적으로 묘사되었으나, 찰리가 이현우를 내려다보지 않고 정면을 응시하여 핵심적인 시선 및 고개 각도 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 위를 쳐다보고 있으나, 찰리는 고개를 숙이지 않고 정면 허공을 응시하고 있어 두 인물의 시선이 교차하지 않음.",
        "built_space": "옥상 잔디밭, 안테나, 배경의 헬리콥터 등이 지시된 위치와 샷 비율에 맞게 잘 배치되어 있음.",
        "entities": "찰리와 이현우의 외형 묘사는 레퍼런스와 대체로 일치하며, 이현우의 부상 상태와 옷깃의 무선 수신기도 잘 표현됨.",
        "hard_violations": [
         "찰리의 시선이 이현우를 향하지 않고 정면을 응시하여 '고개를 꺾은 채 쓰러진 이현우를 부드럽게 내려다보고 있는(inclined head and gentle downward gaze)'이라는 핵심 행동 및 구도 지시를 위반함."
        ],
        "physics": "이현우의 등과 머리가 지면에 지탱되어 있으며, 찰리 역시 땅에 두 발을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "바닥에 쓰러진 이현우가 위를 올려다보고, 찰리가 고개를 꺾어 그를 부드럽게 내려다보며 두 캐릭터 간의 시선이 정확히 맞닿아 있음.",
        "built_space": "연구소 옥상 잔디밭 지면, 뒤편의 안테나 구조물이 로우 앵글 카메라 구도에 맞게 적절히 묘사됨(크롭으로 인해 헬기 미노출).",
        "entities": "찰리의 기계 외형과 얼굴 마스크, 가슴의 푸른빛 심볼이 명확하며, 이현우의 상처 입은 얼굴과 가슴에 위치한 무선 수신기가 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 누워 지탱되고 있고, 찰리는 잔디밭 위에 무게 중심을 잘 잡고 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "찰리가 고개를 숙여 이현우를 내려다보는 시선 교환을 완벽하게 포착하였으며, 카메라 구도와 캐릭터들의 세부 디테일을 프롬프트에 맞게 훌륭히 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "배경과 인물의 질감은 사실적으로 묘사되었으나, 찰리가 이현우를 내려다보지 않고 정면을 응시하여 핵심적인 시선 및 고개 각도 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 위를 쳐다보고 있으나, 찰리는 고개를 숙이지 않고 정면 허공을 응시하고 있어 두 인물의 시선이 교차하지 않음.",
        "built_space": "옥상 잔디밭, 안테나, 배경의 헬리콥터 등이 지시된 위치와 샷 비율에 맞게 잘 배치되어 있음.",
        "entities": "찰리와 이현우의 외형 묘사는 레퍼런스와 대체로 일치하며, 이현우의 부상 상태와 옷깃의 무선 수신기도 잘 표현됨.",
        "hard_violations": [
         "찰리의 시선이 이현우를 향하지 않고 정면을 응시하여 '고개를 꺾은 채 쓰러진 이현우를 부드럽게 내려다보고 있는(inclined head and gentle downward gaze)'이라는 핵심 행동 및 구도 지시를 위반함."
        ],
        "physics": "이현우의 등과 머리가 지면에 지탱되어 있으며, 찰리 역시 땅에 두 발을 딛고 자연스럽게 서 있음."
       },
       {
        "label": "B",
        "direction": "바닥에 쓰러진 이현우가 위를 올려다보고, 찰리가 고개를 꺾어 그를 부드럽게 내려다보며 두 캐릭터 간의 시선이 정확히 맞닿아 있음.",
        "built_space": "연구소 옥상 잔디밭 지면, 뒤편의 안테나 구조물이 로우 앵글 카메라 구도에 맞게 적절히 묘사됨(크롭으로 인해 헬기 미노출).",
        "entities": "찰리의 기계 외형과 얼굴 마스크, 가슴의 푸른빛 심볼이 명확하며, 이현우의 상처 입은 얼굴과 가슴에 위치한 무선 수신기가 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "이현우는 바닥에 안정적으로 누워 지탱되고 있고, 찰리는 잔디밭 위에 무게 중심을 잘 잡고 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "쓰러진 이현우의 어깨 곁에서 올려다보는 밀착 구도와 찰리의 기울어진 삼사분면 얼굴, 서로 응답하는 시선이 B보다 지정된 미디엄 숏에 가깝다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "인물과 장소는 대체로 맞지만, 찰리의 하체와 배경 설비까지 더 넓게 보여 주고 얼굴도 정면에 가까워 지정된 상반신 중심의 이별 구도가 약해진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 왼쪽 아래에 누워 눈을 오른쪽 위의 찰리 얼굴로 향한다. 찰리는 고개를 왼쪽 아래로 기울여 이현우 쪽을 내려다보며, 두 얼굴의 방향이 서로 응답한다. 흰 빛줄기는 찰리 뒤와 주변에서 수직으로 솟지만 가슴 고리에서 직접 이어지는 발광 경로는 분명하지 않다.",
        "built_space": "화면 아래의 좁은 잔디 면에 이현우가 누워 있고, 찰리는 중앙 오른쪽에 있다. 왼쪽 배경에는 안테나 탑 한 기와 같은 탑에 달린 혼 스피커 네 개가 보이며, 아래로 옥상 설비와 난간, 멀리 바다와 섬이 이어진다. 접시 안테나와 헬기는 이 구도에서 확인되지 않지만, 이를 보여 주기 위해 화면을 넓힐 필요는 없다. 어깨 가까이의 낮은 상향 시점과 크게 잡힌 찰리가 요청한 접근 구도에 가깝다.",
        "entities": "등장 인물은 이현우와 찰리 둘뿐이다. 이현우는 젊은 동아시아계 남성으로 보이며, 헝클어진 짧은 검은 머리, 얼굴의 상처, 피 묻은 흰옷이 참고와 부합한다. 목 아래에는 검은 수신기처럼 보이는 장치와 선이 있다. 찰리는 마모된 샌드 베이지 장갑, 육중한 긴 팔, 흰 마스크, 주황색 원형 눈 두 개, 선형 입, 푸른 가슴 고리를 유지한다. 별도로 부착된 실험 센서는 식별하기 어렵다. 하늘은 해 질 무렵이며 흰 빛기둥이 강한 조명을 제공한다.",
        "hard_violations": [],
        "physics": "이현우의 뒤통수와 등 쪽은 잔디 면에 놓여 있으며, 얼굴만 위로 돌린 자세가 가능하다. 찰리의 팔과 손은 관절에 연결되어 아래로 내려와 있고 다리는 지면 방향으로 이어진다. 발의 접촉점은 전경과 화면 끝에 가려져 확인되지 않지만, 몸이 떠 있다는 징후는 없다. 수신기처럼 보이는 장치는 가슴의 옷 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "이현우는 왼쪽 아래에서 오른쪽 위의 찰리를 바라본다. 찰리도 머리를 아래로 숙였으나 얼굴 면이 카메라 쪽에 더 정면으로 열려 있어, 이현우를 향한 삼사분면 시선은 A보다 덜 뚜렷하다. 빛기둥은 찰리 뒤에서 위로 뻗으며, 가슴 고리와 기둥의 직접적인 연결은 보이지 않는다.",
        "built_space": "잔디에 누운 이현우가 하단 전경을 차지하고 찰리가 중앙 오른쪽에 선다. 왼쪽에는 안테나 탑 한 기와 혼 스피커 네 개, 접시 안테나 한 기가 있고, 오른쪽에는 착륙한 헬기 한 대가 보인다. 난간과 해안 배경도 참고 장소와 연결된다. 다만 찰리의 무릎 아래와 넓은 배경까지 포함해, 요청한 상반신 중심 미디엄 숏보다 넓은 구도다.",
        "entities": "추가 인물 없이 이현우와 찰리만 보인다. 이현우의 동아시아계 젊은 남성 외형, 검은 머리, 얼굴 상처와 피 묻은 흰옷은 참고의 주요 특징을 유지한다. 목 아래에는 파란 표시등이 있는 작은 검은 수신기가 보인다. 찰리의 베이지 장갑, 흰 마스크와 두 발광 눈, 선형 입, 푸른 가슴 고리, 긴 팔은 참고와 부합한다. 실험 센서는 명확하게 식별되지 않는다. 어두워지는 하늘과 흰 기둥은 낮에서 밤으로 넘어가는 분위기를 만든다.",
        "hard_violations": [],
        "physics": "이현우의 머리와 몸통은 잔디에 받쳐져 있으며 누운 자세에 무리가 없다. 찰리는 다리를 벌리고 팔을 내리고 있고, 발은 전경에 가려져 접촉점을 직접 볼 수 없다. 다리가 잔디 방향으로 자연스럽게 이어져 공중에 떠 있는 몸으로 보이지 않는다. 수신기는 옷 위에 붙어 있고, 헬기는 옥상 착륙면에 놓인 것으로 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "쓰러진 이현우의 어깨 곁에서 올려다보는 밀착 구도와 찰리의 기울어진 삼사분면 얼굴, 서로 응답하는 시선이 B보다 지정된 미디엄 숏에 가깝다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "인물과 장소는 대체로 맞지만, 찰리의 하체와 배경 설비까지 더 넓게 보여 주고 얼굴도 정면에 가까워 지정된 상반신 중심의 이별 구도가 약해진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 왼쪽 아래에 누워 눈을 오른쪽 위의 찰리 얼굴로 향한다. 찰리는 고개를 왼쪽 아래로 기울여 이현우 쪽을 내려다보며, 두 얼굴의 방향이 서로 응답한다. 흰 빛줄기는 찰리 뒤와 주변에서 수직으로 솟지만 가슴 고리에서 직접 이어지는 발광 경로는 분명하지 않다.",
        "built_space": "화면 아래의 좁은 잔디 면에 이현우가 누워 있고, 찰리는 중앙 오른쪽에 있다. 왼쪽 배경에는 안테나 탑 한 기와 같은 탑에 달린 혼 스피커 네 개가 보이며, 아래로 옥상 설비와 난간, 멀리 바다와 섬이 이어진다. 접시 안테나와 헬기는 이 구도에서 확인되지 않지만, 이를 보여 주기 위해 화면을 넓힐 필요는 없다. 어깨 가까이의 낮은 상향 시점과 크게 잡힌 찰리가 요청한 접근 구도에 가깝다.",
        "entities": "등장 인물은 이현우와 찰리 둘뿐이다. 이현우는 젊은 동아시아계 남성으로 보이며, 헝클어진 짧은 검은 머리, 얼굴의 상처, 피 묻은 흰옷이 참고와 부합한다. 목 아래에는 검은 수신기처럼 보이는 장치와 선이 있다. 찰리는 마모된 샌드 베이지 장갑, 육중한 긴 팔, 흰 마스크, 주황색 원형 눈 두 개, 선형 입, 푸른 가슴 고리를 유지한다. 별도로 부착된 실험 센서는 식별하기 어렵다. 하늘은 해 질 무렵이며 흰 빛기둥이 강한 조명을 제공한다.",
        "hard_violations": [],
        "physics": "이현우의 뒤통수와 등 쪽은 잔디 면에 놓여 있으며, 얼굴만 위로 돌린 자세가 가능하다. 찰리의 팔과 손은 관절에 연결되어 아래로 내려와 있고 다리는 지면 방향으로 이어진다. 발의 접촉점은 전경과 화면 끝에 가려져 확인되지 않지만, 몸이 떠 있다는 징후는 없다. 수신기처럼 보이는 장치는 가슴의 옷 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "이현우는 왼쪽 아래에서 오른쪽 위의 찰리를 바라본다. 찰리도 머리를 아래로 숙였으나 얼굴 면이 카메라 쪽에 더 정면으로 열려 있어, 이현우를 향한 삼사분면 시선은 A보다 덜 뚜렷하다. 빛기둥은 찰리 뒤에서 위로 뻗으며, 가슴 고리와 기둥의 직접적인 연결은 보이지 않는다.",
        "built_space": "잔디에 누운 이현우가 하단 전경을 차지하고 찰리가 중앙 오른쪽에 선다. 왼쪽에는 안테나 탑 한 기와 혼 스피커 네 개, 접시 안테나 한 기가 있고, 오른쪽에는 착륙한 헬기 한 대가 보인다. 난간과 해안 배경도 참고 장소와 연결된다. 다만 찰리의 무릎 아래와 넓은 배경까지 포함해, 요청한 상반신 중심 미디엄 숏보다 넓은 구도다.",
        "entities": "추가 인물 없이 이현우와 찰리만 보인다. 이현우의 동아시아계 젊은 남성 외형, 검은 머리, 얼굴 상처와 피 묻은 흰옷은 참고의 주요 특징을 유지한다. 목 아래에는 파란 표시등이 있는 작은 검은 수신기가 보인다. 찰리의 베이지 장갑, 흰 마스크와 두 발광 눈, 선형 입, 푸른 가슴 고리, 긴 팔은 참고와 부합한다. 실험 센서는 명확하게 식별되지 않는다. 어두워지는 하늘과 흰 기둥은 낮에서 밤으로 넘어가는 분위기를 만든다.",
        "hard_violations": [],
        "physics": "이현우의 머리와 몸통은 잔디에 받쳐져 있으며 누운 자세에 무리가 없다. 찰리는 다리를 벌리고 팔을 내리고 있고, 발은 전경에 가려져 접촉점을 직접 볼 수 없다. 다리가 잔디 방향으로 자연스럽게 이어져 공중에 떠 있는 몸으로 보이지 않는다. 수신기는 옷 위에 붙어 있고, 헬기는 옥상 착륙면에 놓인 것으로 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.319,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.069,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 찰리의 시선이 이현우를 향하지 않고 정면을 응시하여 '고개를 꺾은 채 쓰러진 이현우를 부드럽게 내려다보고 있는(inclined head and gentle downward gaze)'이라는 핵심 행동 및 구도 지시를 위반함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1069
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "찰리가 고개를 숙여 이현우를 내려다보는 시선 교환을 완벽하게 포착하였으며, 카메라 구도와 캐릭터들의 세부 디테일을 프롬프트에 맞게 훌륭히 구현했습니다."
   },
   {
    "label": "A",
    "score": 1069,
    "verdict_ko": "배경과 인물의 질감은 사실적으로 묘사되었으나, 찰리가 이현우를 내려다보지 않고 정면을 응시하여 핵심적인 시선 및 고개 각도 지시를 위반했습니다.  ★위반: [gemini-pro] 찰리의 시선이 이현우를 향하지 않고 정면을 응시하여 '고개를 꺾은 채 쓰러진 이현우를 부드럽게 내려다보고 있는(inclined head and gentle downward gaze)'이라는 핵심 행동 및 구도 지시를 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh45_sel.png",
    "asset_id": "9bd26dbc-0a76-4d9c-bd40-6d136ae60665",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1177379>",
    "asset_id": "39ede2f0-a73c-4159-a2cd-6ca09f52b2f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ecc-d365-79c0-bb00-a0a67d4ae51e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S89sh45"
  },
  "staged_characters_added": [
   "C03"
  ]
 },
 "S89sh75::signage": {
  "fp": "a64e020afc769644",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::research_exterior_aftermath": {
  "input_fingerprint": "10d0b46b2746daca",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "research_exterior_aftermath",
    "tags": [
     "S89sh75"
    ]
   },
   "context_sig": "c5c8600f95dae2d2"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 옥상 잔디밭·안테나·스피커 구역, 연결 계단: 야외 정원이 조성된 옥상 플랫폼으로 거대한 통신 장비가 설치되어 있다. (특징: 옥상 잔디밭에 세워진 거대한 파라볼라 안테나와 360도 대형 스피커 모듈; 수신기를 몸에 장착한 현우와 찰리; 착륙하는 헬기에서 레펠로 강하하는 특임대와 인공지능 전투병들; 스피커 진동파로 인해 눈이 뒤집히며 픽픽 쓰러지는 인공지능 병사들; 유탄 파편에 맞아 피 흘리는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /제주도 연구소 밖 - N\n\nTIME OF DAY (lock): day to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 옥상 잔디밭·안테나·스피커 구역, 연결 계단: 야외 정원이 조성된 옥상 플랫폼으로 거대한 통신 장비가 설치되어 있다. (특징: 옥상 잔디밭에 세워진 거대한 파라볼라 안테나와 360도 대형 스피커 모듈; 수신기를 몸에 장착한 현우와 찰리; 착륙하는 헬기에서 레펠로 강하하는 특임대와 인공지능 전투병들; 스피커 진동파로 인해 눈이 뒤집히며 픽픽 쓰러지는 인공지능 병사들; 유탄 파편에 맞아 피 흘리는 현우)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- /제주도 연구소 밖 - N\n\nTIME OF DAY (lock): day to night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_exterior_aftermath_113eed.png",
  "asset_id": "87360af2-3c8c-4720-a9dd-269c990c14ce",
  "input_asset_ids": [
   "718aaa5e-115f-4d74-8289-87ad019344c2"
  ],
  "origin_tag": "S89sh75",
  "place_text": "At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.",
  "origin_inputs": {
   "place_text": "At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.",
   "time_of_day_en": "day to night",
   "conti_asset_id": "718aaa5e-115f-4d74-8289-87ad019344c2"
  }
 },
 "S89sh75::bgfirst_bg": {
  "input_fingerprint": "ed9ed2a9af23d4eb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 두 팔이 전원이 꺼진 찰리의 차가운 금속 잔해를 자신의 품에 꽉 껴안고 있는 애처로운 구도.\n\nLOCATION (lock): At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.\n\nTIME OF DAY (lock): day to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the final static composition close behind 이현우's side, slightly above his shoulder and angled downward into the space between his forearms and chest. His hunched shoulder borders the upper-left frame, his arms enclose the center, and 찰리's powered-down torso remains occupy less than a third of the image, held firmly against him rather than enlarged by foreground perspective. Observe the embrace directly, with 이현우's bowed head directed toward the remains and his hands tightening around them; the camera has stopped, leaving that compression of his arms as the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ground outside the laboratory (Visible beneath the embrace after the beam has disappeared) — A small downward-viewed area remains at the lower-right edge; used as Retains exterior spatial context without competing with the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination and controlled contrast keep the clasped arms and cold metal remains legible without continuing the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이현우의 두 팔이 전원이 꺼진 찰리의 차가운 금속 잔해를 자신의 품에 꽉 껴안고 있는 애처로운 구도.\n\nLOCATION (lock): At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies.\n\nTIME OF DAY (lock): day to night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the final static composition close behind 이현우's side, slightly above his shoulder and angled downward into the space between his forearms and chest. His hunched shoulder borders the upper-left frame, his arms enclose the center, and 찰리's powered-down torso remains occupy less than a third of the image, held firmly against him rather than enlarged by foreground perspective. Observe the embrace directly, with 이현우's bowed head directed toward the remains and his hands tightening around them; the camera has stopped, leaving that compression of his arms as the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ground outside the laboratory (Visible beneath the embrace after the beam has disappeared) — A small downward-viewed area remains at the lower-right edge; used as Retains exterior spatial context without competing with the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination and controlled contrast keep the clasped arms and cold metal remains legible without continuing the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh75__bgfirst_bg.png",
  "asset_id": "185b8f96-594a-43dc-8a4d-9fc3f4fe2079",
  "input_asset_ids": [
   "718aaa5e-115f-4d74-8289-87ad019344c2",
   "87360af2-3c8c-4720-a9dd-269c990c14ce"
  ]
 },
 "S89sh75": {
  "input_fingerprint": "268d5cd1ad4fae62",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 두 팔이 전원이 꺼진 찰리의 차가운 금속 잔해를 자신의 품에 꽉 껴안고 있는 애처로운 구도.\n\nLOCATION (lock): At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the final static composition close behind 이현우's side, slightly above his shoulder and angled downward into the space between his forearms and chest. His hunched shoulder borders the upper-left frame, his arms enclose the center, and 찰리's powered-down torso remains occupy less than a third of the image, held firmly against him rather than enlarged by foreground perspective. Observe the embrace directly, with 이현우's bowed head directed toward the remains and his hands tightening around them; the camera has stopped, leaving that compression of his arms as the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ground outside the laboratory (Visible beneath the embrace after the beam has disappeared) — A small downward-viewed area remains at the lower-right edge; used as Retains exterior spatial context without competing with the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination and controlled contrast keep the clasped arms and cold metal remains legible without continuing the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains are gathered tightly against Hyunwoo's chest in both of Hyunwoo's arms. Only a surviving portion of his destroyed body remains, and the scene text does not establish an intact arrangement of head and limbs or a facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Outside the laboratory at night, the column of light has vanished. Charlie is reduced to a surviving portion of his torso and other remains, with his power now completely off. 이현우: His grenade-fragment injuries remain untreated, and the wireless receiver remains attached. He holds the powerless metal remains tightly and sobs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 두 팔이 전원이 꺼진 찰리의 차가운 금속 잔해를 자신의 품에 꽉 껴안고 있는 애처로운 구도.\n\nLOCATION (lock): At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the final static composition close behind 이현우's side, slightly above his shoulder and angled downward into the space between his forearms and chest. His hunched shoulder borders the upper-left frame, his arms enclose the center, and 찰리's powered-down torso remains occupy less than a third of the image, held firmly against him rather than enlarged by foreground perspective. Observe the embrace directly, with 이현우's bowed head directed toward the remains and his hands tightening around them; the camera has stopped, leaving that compression of his arms as the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ground outside the laboratory (Visible beneath the embrace after the beam has disappeared) — A small downward-viewed area remains at the lower-right edge; used as Retains exterior spatial context without competing with the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination and controlled contrast keep the clasped arms and cold metal remains legible without continuing the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains are gathered tightly against Hyunwoo's chest in both of Hyunwoo's arms. Only a surviving portion of his destroyed body remains, and the scene text does not establish an intact arrangement of head and limbs or a facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Outside the laboratory at night, the column of light has vanished. Charlie is reduced to a surviving portion of his torso and other remains, with his power now completely off. 이현우: His grenade-fragment injuries remain untreated, and the wireless receiver remains attached. He holds the powerless metal remains tightly and sobs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day to night.\n\nSHOT TEXT (authoritative, Korean): 이현우의 두 팔이 전원이 꺼진 찰리의 차가운 금속 잔해를 자신의 품에 꽉 껴안고 있는 애처로운 구도.\n\nLOCATION (lock): At the devastated outdoor blast site of the research institute at night, where the robot's remaining wreckage lies. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin the final static composition close behind 이현우's side, slightly above his shoulder and angled downward into the space between his forearms and chest. His hunched shoulder borders the upper-left frame, his arms enclose the center, and 찰리's powered-down torso remains occupy less than a third of the image, held firmly against him rather than enlarged by foreground perspective. Observe the embrace directly, with 이현우's bowed head directed toward the remains and his hands tightening around them; the camera has stopped, leaving that compression of his arms as the sole emphasis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Ground outside the laboratory (Visible beneath the embrace after the beam has disappeared) — A small downward-viewed area remains at the lower-right edge; used as Retains exterior spatial context without competing with the arms.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination and controlled contrast keep the clasped arms and cold metal remains legible without continuing the vanished beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's powered-off remains are gathered tightly against Hyunwoo's chest in both of Hyunwoo's arms. Only a surviving portion of his destroyed body remains, and the scene text does not establish an intact arrangement of head and limbs or a facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Outside the laboratory at night, the column of light has vanished. Charlie is reduced to a surviving portion of his torso and other remains, with his power now completely off. 이현우: His grenade-fragment injuries remain untreated, and the wireless receiver remains attached. He holds the powerless metal remains tightly and sobs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이현우 right now, so 이현우's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이현우: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.); 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh75__bgfirst_bg.png",
     "asset_id": "185b8f96-594a-43dc-8a4d-9fc3f4fe2079",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S89sh75.png",
     "asset_id": "718aaa5e-115f-4d74-8289-87ad019344c2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1230529>",
     "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1177379>",
     "asset_id": "39ede2f0-a73c-4159-a2cd-6ca09f52b2f2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_exterior_aftermath_113eed.png",
     "asset_id": "87360af2-3c8c-4720-a9dd-269c990c14ce",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1230529>",
     "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1177379>",
     "asset_id": "39ede2f0-a73c-4159-a2cd-6ca09f52b2f2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 약간 위쪽에서 아래를 향해 내려다보는 각도를 취하고 있으며, 이현우의 시선과 고개는 품에 안은 로봇의 잔해를 향하고 있습니다.",
    "built_space": "배경에는 야간의 파괴된 야외 연구소 폭발 현장이 펼쳐져 있으며, 부서진 콘크리트 구조물, 풀밭, 멀리 보이는 불빛들이 레퍼런스 이미지의 장소와 일치합니다. 화면 우측 하단에 바닥면이 잘 배치되어 있습니다.",
    "entities": "피 묻은 흰색 유니폼을 입고 짧은 검은 머리를 한 이현우가 등장합니다. 그가 안고 있는 찰리의 잔해는 모래색(샌드 베이지) 장갑을 띠고 있으나, 몸통보다는 어깨나 관절 부품에 가까워 보입니다.",
    "hard_violations": [],
    "physics": "이현우의 팔과 손이 로봇의 금속 잔해를 단단히 감싸 안고 지탱하고 있으며, 중력과 무게감이 자연스럽게 표현되었습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 아래로 향하는 각도를 유지하고 있으며, 이현우는 고개를 숙인 채 품 안의 금속 잔해를 바라보고 있습니다.",
    "built_space": "야간의 파괴된 폭발 현장 배경이 잘 묘사되어 있으며, 부서진 구조물과 잔해, 화면 우측 하단의 포장된 바닥 타일이 위치 레퍼런스의 디테일과 정확히 일치합니다.",
    "entities": "피투성이가 된 흰색 유니폼과 검은 머리를 한 이현우의 모습이 레퍼런스와 일치합니다. 품에 안긴 찰리의 잔해는 레퍼런스에서 볼 수 있는 특유의 원형 가슴 코어와 모래색 장갑판, 내부 전선 등이 드러나 있어 '몸통 잔해'라는 지시를 완벽히 따릅니다.",
    "hard_violations": [],
    "physics": "이현우의 왼팔이 로봇의 무거운 몸통 잔해를 아래에서부터 감싸 안아 확실하게 지지하고 있으며, 자세와 그립이 매우 물리적으로 타당합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "카메라 앵글과 피사체의 배치 비율을 완벽하게 구현했으며, 레퍼런스의 가슴 코어 디자인이 명확히 반영된 찰리의 몸통 잔해를 디테일하게 묘사하여 지시사항을 훌륭하게 충족합니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도와 배경의 스케일은 잘 따랐으나, 로봇의 잔해가 찰리의 몸통(가슴 부분)이라기보다는 단순한 기계 관절 부위로 보여 디테일 면에서 다소 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 약간 위쪽에서 아래를 향해 내려다보는 각도를 취하고 있으며, 이현우의 시선과 고개는 품에 안은 로봇의 잔해를 향하고 있습니다.",
        "built_space": "배경에는 야간의 파괴된 야외 연구소 폭발 현장이 펼쳐져 있으며, 부서진 콘크리트 구조물, 풀밭, 멀리 보이는 불빛들이 레퍼런스 이미지의 장소와 일치합니다. 화면 우측 하단에 바닥면이 잘 배치되어 있습니다.",
        "entities": "피 묻은 흰색 유니폼을 입고 짧은 검은 머리를 한 이현우가 등장합니다. 그가 안고 있는 찰리의 잔해는 모래색(샌드 베이지) 장갑을 띠고 있으나, 몸통보다는 어깨나 관절 부품에 가까워 보입니다.",
        "hard_violations": [],
        "physics": "이현우의 팔과 손이 로봇의 금속 잔해를 단단히 감싸 안고 지탱하고 있으며, 중력과 무게감이 자연스럽게 표현되었습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 아래로 향하는 각도를 유지하고 있으며, 이현우는 고개를 숙인 채 품 안의 금속 잔해를 바라보고 있습니다.",
        "built_space": "야간의 파괴된 폭발 현장 배경이 잘 묘사되어 있으며, 부서진 구조물과 잔해, 화면 우측 하단의 포장된 바닥 타일이 위치 레퍼런스의 디테일과 정확히 일치합니다.",
        "entities": "피투성이가 된 흰색 유니폼과 검은 머리를 한 이현우의 모습이 레퍼런스와 일치합니다. 품에 안긴 찰리의 잔해는 레퍼런스에서 볼 수 있는 특유의 원형 가슴 코어와 모래색 장갑판, 내부 전선 등이 드러나 있어 '몸통 잔해'라는 지시를 완벽히 따릅니다.",
        "hard_violations": [],
        "physics": "이현우의 왼팔이 로봇의 무거운 몸통 잔해를 아래에서부터 감싸 안아 확실하게 지지하고 있으며, 자세와 그립이 매우 물리적으로 타당합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "카메라 앵글과 피사체의 배치 비율을 완벽하게 구현했으며, 레퍼런스의 가슴 코어 디자인이 명확히 반영된 찰리의 몸통 잔해를 디테일하게 묘사하여 지시사항을 훌륭하게 충족합니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도와 배경의 스케일은 잘 따랐으나, 로봇의 잔해가 찰리의 몸통(가슴 부분)이라기보다는 단순한 기계 관절 부위로 보여 디테일 면에서 다소 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 약간 위쪽에서 아래를 향해 내려다보는 각도를 취하고 있으며, 이현우의 시선과 고개는 품에 안은 로봇의 잔해를 향하고 있습니다.",
        "built_space": "배경에는 야간의 파괴된 야외 연구소 폭발 현장이 펼쳐져 있으며, 부서진 콘크리트 구조물, 풀밭, 멀리 보이는 불빛들이 레퍼런스 이미지의 장소와 일치합니다. 화면 우측 하단에 바닥면이 잘 배치되어 있습니다.",
        "entities": "피 묻은 흰색 유니폼을 입고 짧은 검은 머리를 한 이현우가 등장합니다. 그가 안고 있는 찰리의 잔해는 모래색(샌드 베이지) 장갑을 띠고 있으나, 몸통보다는 어깨나 관절 부품에 가까워 보입니다.",
        "hard_violations": [],
        "physics": "이현우의 팔과 손이 로봇의 금속 잔해를 단단히 감싸 안고 지탱하고 있으며, 중력과 무게감이 자연스럽게 표현되었습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 이현우의 왼쪽 어깨 뒤에서 아래로 향하는 각도를 유지하고 있으며, 이현우는 고개를 숙인 채 품 안의 금속 잔해를 바라보고 있습니다.",
        "built_space": "야간의 파괴된 폭발 현장 배경이 잘 묘사되어 있으며, 부서진 구조물과 잔해, 화면 우측 하단의 포장된 바닥 타일이 위치 레퍼런스의 디테일과 정확히 일치합니다.",
        "entities": "피투성이가 된 흰색 유니폼과 검은 머리를 한 이현우의 모습이 레퍼런스와 일치합니다. 품에 안긴 찰리의 잔해는 레퍼런스에서 볼 수 있는 특유의 원형 가슴 코어와 모래색 장갑판, 내부 전선 등이 드러나 있어 '몸통 잔해'라는 지시를 완벽히 따릅니다.",
        "hard_violations": [],
        "physics": "이현우의 왼팔이 로봇의 무거운 몸통 잔해를 아래에서부터 감싸 안아 확실하게 지지하고 있으며, 자세와 그립이 매우 물리적으로 타당합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "어깨 뒤에서 품 안을 내려다보는 시점과 잔해에 깊이 숙인 머리가 지시에 더 가깝지만, 배경이 우하단의 작은 영역을 넘어 넓게 노출됩니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 손으로 잔해를 조이는 행동은 명확하지만, 수평에 가까운 시점으로 하늘과 연구소 전경까지 보여 주어 지정된 하향 클로즈업에서 더 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "현우의 머리와 얼굴은 가슴에 붙인 금속 잔해를 향해 깊이 숙여져 있습니다. 눈은 가려져 직접적인 시선은 확인할 수 없습니다. 보이는 손은 잔해 오른쪽 가장자리를 안쪽으로 감싸며, 로봇에는 별도의 시선이나 진행 방향이 없습니다.",
        "built_space": "왼쪽 전경의 어깨 너머로 품 안을 비스듬히 내려다봅니다. 오른쪽에는 젖은 포장면, 풀, 콘크리트 파편이 있고, 상단에는 무너진 건물 한 동과 난간, 복수의 작은 조명이 보입니다. 참조 장소의 재료와 야간 환경에 부합하지만, 배경이 오른쪽과 상단에 넓게 남아 우하단의 작은 지면만 보여 달라는 지시에는 미달합니다. 복제된 고정 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "인물은 현우 한 명이며, 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성의 옆얼굴, 피 묻은 흰옷과 치료되지 않은 얼굴·팔 상처가 참조에 부합합니다. 정확한 얼굴 일치와 나이는 가려진 부분 때문에 제한적으로만 판단됩니다. 찰리는 샌드 베이지 장갑판, 원형 흉부 기구, 노출된 내부 구조를 가진 파괴된 몸통 일부로 표현되며 화면의 약 3분의 1 미만입니다. 사람 피부나 눈은 부여되지 않았고 발광도 없습니다. 무선 수신기는 명확히 식별되지 않으며, 흐느낌보다 숙인 자세로 슬픔을 전달합니다.",
        "hard_violations": [],
        "physics": "잔해는 현우의 가슴과 앞쪽 팔 사이에 끼워져 있고, 오른쪽에 보이는 손이 장갑판을 잡습니다. 반대쪽 팔과 손의 접촉은 대부분 가려져 두 팔의 압박을 모두 읽기는 어렵지만, 보이는 지지만으로도 잔해가 떠 있는 상태는 아닙니다. 부서진 부품은 몸통에 붙거나 품 안에 모여 있으며 독립적으로 공중에 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "현우는 품 안의 잔해를 향해 고개를 숙이고 있습니다. 눈 자체는 머리카락과 얼굴 각도로 가려집니다. 한 손은 잔해 위쪽 측면을 잡고 다른 손은 아래쪽 앞면을 감싸 두 손의 힘이 가슴 쪽으로 모입니다. 찰리의 잔해에는 능동적인 시선이나 움직임이 없습니다.",
        "built_space": "왼쪽 어깨 뒤의 가까운 시점이지만, 카메라가 품 안을 내려다보기보다 바깥을 향해 수평에 가깝게 열려 있습니다. 중앙 뒤에 무너진 건물 한 동, 왼쪽에 다른 건물 일부, 여러 안테나와 난간, 지면 조명, 멀리 바다와 불빛이 보입니다. 장소의 주요 재료와 배치는 참조와 대체로 맞지만, 하늘과 원경까지 크게 포함하여 배경을 우하단의 작은 지면으로 제한하는 구도와 어긋납니다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않습니다.",
        "entities": "현우 한 명과 찰리의 몸통 잔해만 보입니다. 현우의 검은 머리, 젊은 동아시아계 남성 외형, 찢어지고 피 묻은 흰옷, 얼굴과 팔의 상처는 참조에 부합합니다. 얼굴 대부분이 가려져 정확한 동일인 여부는 제한적으로 판단됩니다. 찰리는 베이지색 각진 금속 외장과 검은 관절·내부 구조가 있는 파손된 몸통 일부이며, 화면의 3분의 1 미만을 차지합니다. 완전한 머리와 팔다리를 억지로 복원하지 않았고 전원 발광도 없습니다. 무선 수신기는 식별되지 않으며, 눈물이나 흐느낌 자체는 뚜렷하지 않습니다.",
        "hard_violations": [],
        "physics": "두 손이 각각 잔해의 측면과 아래쪽 앞면에 닿아 있고, 양팔이 이를 가슴에 압착합니다. 무게를 팔과 몸통으로 받치는 지지가 분명합니다. 로봇 잔해가 스스로 자세를 유지하거나 부품이 지지 없이 떠 있는 모습은 없으며, 보이는 손과 팔의 연결도 물리적으로 가능합니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "어깨 뒤에서 품 안을 내려다보는 시점과 잔해에 깊이 숙인 머리가 지시에 더 가깝지만, 배경이 우하단의 작은 영역을 넘어 넓게 노출됩니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 손으로 잔해를 조이는 행동은 명확하지만, 수평에 가까운 시점으로 하늘과 연구소 전경까지 보여 주어 지정된 하향 클로즈업에서 더 벗어납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "현우의 머리와 얼굴은 가슴에 붙인 금속 잔해를 향해 깊이 숙여져 있습니다. 눈은 가려져 직접적인 시선은 확인할 수 없습니다. 보이는 손은 잔해 오른쪽 가장자리를 안쪽으로 감싸며, 로봇에는 별도의 시선이나 진행 방향이 없습니다.",
        "built_space": "왼쪽 전경의 어깨 너머로 품 안을 비스듬히 내려다봅니다. 오른쪽에는 젖은 포장면, 풀, 콘크리트 파편이 있고, 상단에는 무너진 건물 한 동과 난간, 복수의 작은 조명이 보입니다. 참조 장소의 재료와 야간 환경에 부합하지만, 배경이 오른쪽과 상단에 넓게 남아 우하단의 작은 지면만 보여 달라는 지시에는 미달합니다. 복제된 고정 설비나 불가능한 반사는 보이지 않습니다.",
        "entities": "인물은 현우 한 명이며, 짧고 헝클어진 검은 머리, 젊은 동아시아계 남성의 옆얼굴, 피 묻은 흰옷과 치료되지 않은 얼굴·팔 상처가 참조에 부합합니다. 정확한 얼굴 일치와 나이는 가려진 부분 때문에 제한적으로만 판단됩니다. 찰리는 샌드 베이지 장갑판, 원형 흉부 기구, 노출된 내부 구조를 가진 파괴된 몸통 일부로 표현되며 화면의 약 3분의 1 미만입니다. 사람 피부나 눈은 부여되지 않았고 발광도 없습니다. 무선 수신기는 명확히 식별되지 않으며, 흐느낌보다 숙인 자세로 슬픔을 전달합니다.",
        "hard_violations": [],
        "physics": "잔해는 현우의 가슴과 앞쪽 팔 사이에 끼워져 있고, 오른쪽에 보이는 손이 장갑판을 잡습니다. 반대쪽 팔과 손의 접촉은 대부분 가려져 두 팔의 압박을 모두 읽기는 어렵지만, 보이는 지지만으로도 잔해가 떠 있는 상태는 아닙니다. 부서진 부품은 몸통에 붙거나 품 안에 모여 있으며 독립적으로 공중에 떠 있는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "현우는 품 안의 잔해를 향해 고개를 숙이고 있습니다. 눈 자체는 머리카락과 얼굴 각도로 가려집니다. 한 손은 잔해 위쪽 측면을 잡고 다른 손은 아래쪽 앞면을 감싸 두 손의 힘이 가슴 쪽으로 모입니다. 찰리의 잔해에는 능동적인 시선이나 움직임이 없습니다.",
        "built_space": "왼쪽 어깨 뒤의 가까운 시점이지만, 카메라가 품 안을 내려다보기보다 바깥을 향해 수평에 가깝게 열려 있습니다. 중앙 뒤에 무너진 건물 한 동, 왼쪽에 다른 건물 일부, 여러 안테나와 난간, 지면 조명, 멀리 바다와 불빛이 보입니다. 장소의 주요 재료와 배치는 참조와 대체로 맞지만, 하늘과 원경까지 크게 포함하여 배경을 우하단의 작은 지면으로 제한하는 구도와 어긋납니다. 설비 중복이나 광학적으로 불가능한 반사는 보이지 않습니다.",
        "entities": "현우 한 명과 찰리의 몸통 잔해만 보입니다. 현우의 검은 머리, 젊은 동아시아계 남성 외형, 찢어지고 피 묻은 흰옷, 얼굴과 팔의 상처는 참조에 부합합니다. 얼굴 대부분이 가려져 정확한 동일인 여부는 제한적으로 판단됩니다. 찰리는 베이지색 각진 금속 외장과 검은 관절·내부 구조가 있는 파손된 몸통 일부이며, 화면의 3분의 1 미만을 차지합니다. 완전한 머리와 팔다리를 억지로 복원하지 않았고 전원 발광도 없습니다. 무선 수신기는 식별되지 않으며, 눈물이나 흐느낌 자체는 뚜렷하지 않습니다.",
        "hard_violations": [],
        "physics": "두 손이 각각 잔해의 측면과 아래쪽 앞면에 닿아 있고, 양팔이 이를 가슴에 압착합니다. 무게를 팔과 몸통으로 받치는 지지가 분명합니다. 로봇 잔해가 스스로 자세를 유지하거나 부품이 지지 없이 떠 있는 모습은 없으며, 보이는 손과 팔의 연결도 물리적으로 가능합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.657,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.657,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1657
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "카메라 앵글과 피사체의 배치 비율을 완벽하게 구현했으며, 레퍼런스의 가슴 코어 디자인이 명확히 반영된 찰리의 몸통 잔해를 디테일하게 묘사하여 지시사항을 훌륭하게 충족합니다."
   },
   {
    "label": "A",
    "score": 1657,
    "verdict_ko": "지시된 카메라 구도와 배경의 스케일은 잘 따랐으나, 로봇의 잔해가 찰리의 몸통(가슴 부분)이라기보다는 단순한 기계 관절 부위로 보여 디테일 면에서 다소 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_research_exterior_aftermath_113eed.png",
    "asset_id": "87360af2-3c8c-4720-a9dd-269c990c14ce",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1230529>",
    "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1177379>",
    "asset_id": "39ede2f0-a73c-4159-a2cd-6ca09f52b2f2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ed2-0b34-7c7f-a176-b2f14c92cd11",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S89sh75__bgfirst_bg.png",
   "bg_asset_id": "185b8f96-594a-43dc-8a4d-9fc3f4fe2079",
   "bg_record_key": "S89sh75::bgfirst_bg",
   "chain_winner": false,
   "winner_origin": "B",
   "authority": "groupbg",
   "group_key": "research_exterior_aftermath",
   "groupbg_asset_id": "87360af2-3c8c-4720-a9dd-269c990c14ce"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S90sh2::signage": {
  "fp": "20ac5ba357693bcc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::f77e0a36999163e3": {
  "subjects": [],
  "subject_text": "제주도 연구소 부품 보관·홀로그램 상영실, 찰리 재조립 작업실\n어둡고 넓은 공간에 찰리의 잔해가 제단처럼 보관된 애도 및 재건 공간이다.",
  "identity_fallback": true,
  "reason": "scope_incomplete",
  "scope_id": "L163",
  "scope_role": null,
  "scope_sha": null
 },
 "groupbg::reconstruction_hall": {
  "input_fingerprint": "0ffdf7c2f358fd74",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "reconstruction_hall",
    "tags": [
     "S90sh16",
     "S90sh19",
     "S90sh2"
    ]
   },
   "context_sig": "d9560882e481f0d1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 부품 보관·홀로그램 상영실, 찰리 재조립 작업실: 어둡고 넓은 공간에 찰리의 잔해가 제단처럼 보관된 애도 및 재건 공간이다. (특징: 전 세계 각지에 배치된 찰리의 형상을 띄우는 초대형 모니터 벽면; 테이블에 고이 놓인 찰리의 그을린 기계 부품 잔해들; 소영이 현우에게 건네는 작은 반도체 칩 조각; 책상 버튼을 누르자 공중에 떠오르는 찰리와 앰버 형상의 입체 홀로그램 광선; 로봇 팔에 의해 다시 맞춰지는 부품 조각들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 넓은 공간. 전체적으로 어두운.\n- 그 아래, 제단처럼 보이는 테이블 위에 분해된 찰리의 부품들이 놓여있다.\n- / 좌석의 현우도 일어나 찰리의 손을 만지려하며 미소를 짓는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n제주도 연구소 부품 보관·홀로그램 상영실, 찰리 재조립 작업실: 어둡고 넓은 공간에 찰리의 잔해가 제단처럼 보관된 애도 및 재건 공간이다. (특징: 전 세계 각지에 배치된 찰리의 형상을 띄우는 초대형 모니터 벽면; 테이블에 고이 놓인 찰리의 그을린 기계 부품 잔해들; 소영이 현우에게 건네는 작은 반도체 칩 조각; 책상 버튼을 누르자 공중에 떠오르는 찰리와 앰버 형상의 입체 홀로그램 광선; 로봇 팔에 의해 다시 맞춰지는 부품 조각들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 넓은 공간. 전체적으로 어두운.\n- 그 아래, 제단처럼 보이는 테이블 위에 분해된 찰리의 부품들이 놓여있다.\n- / 좌석의 현우도 일어나 찰리의 손을 만지려하며 미소를 짓는다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reconstruction_hall_c365e6.png",
  "asset_id": "cead43ee-15ec-4696-a2d0-d3f4e8e2edc9",
  "input_asset_ids": [
   "8660d680-0c6a-4575-9790-c7eeb6feb6aa"
  ],
  "origin_tag": "S90sh2",
  "place_text": "At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.",
  "origin_inputs": {
   "place_text": "At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.",
   "time_of_day_en": "night",
   "conti_asset_id": "8660d680-0c6a-4575-9790-c7eeb6feb6aa"
  }
 },
 "S90sh2::bgfirst_bg": {
  "input_fingerprint": "b6b2b8d23bb098d3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모니터 아래 제단 같은 테이블 위에 찰리의 낡고 분해된 금속 부품들이 흩어져 놓인 구도.\n\nLOCATION (lock): At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane descent at the near corner of the altar-like table, retaining an oblique downward view rather than flattening the arrangement into an overhead diagram. Observe 찰리's worn, disassembled metal parts directly across the central band of the frame, keeping individual pieces modest in scale and readable against the tabletop, its near edge, and the surrounding room. The monitor has passed beyond the upper edge and 이현우 remains outside the crop beyond the table; the final reduction in camera distance concentrates attention on the interrupted body without introducing another visual axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Altar-like table (Supporting 찰리's separated parts beneath the off-frame monitor) — The upper surface and near corner are visible obliquely, with the near edge crossing the lower frame; used as Organizes the parts into a restrained central arrangement while preserving their scale; Laboratory interior (Open space visible beyond the table); used as Leaves quiet peripheral space around the close observation of the parts.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the laboratory broadly dark, with subdued neutral illumination sufficient to distinguish the separated metal parts without assigning a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모니터 아래 제단 같은 테이블 위에 찰리의 낡고 분해된 금속 부품들이 흩어져 놓인 구도.\n\nLOCATION (lock): At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane descent at the near corner of the altar-like table, retaining an oblique downward view rather than flattening the arrangement into an overhead diagram. Observe 찰리's worn, disassembled metal parts directly across the central band of the frame, keeping individual pieces modest in scale and readable against the tabletop, its near edge, and the surrounding room. The monitor has passed beyond the upper edge and 이현우 remains outside the crop beyond the table; the final reduction in camera distance concentrates attention on the interrupted body without introducing another visual axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Altar-like table (Supporting 찰리's separated parts beneath the off-frame monitor) — The upper surface and near corner are visible obliquely, with the near edge crossing the lower frame; used as Organizes the parts into a restrained central arrangement while preserving their scale; Laboratory interior (Open space visible beyond the table); used as Leaves quiet peripheral space around the close observation of the parts.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the laboratory broadly dark, with subdued neutral illumination sufficient to distinguish the separated metal parts without assigning a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S90sh2__bgfirst_bg.png",
  "asset_id": "3180a031-cf36-49a4-a3fd-6b690d429dd9",
  "input_asset_ids": [
   "8660d680-0c6a-4575-9790-c7eeb6feb6aa",
   "cead43ee-15ec-4696-a2d0-d3f4e8e2edc9"
  ]
 },
 "S90sh2": {
  "input_fingerprint": "1b64034d3b461ea5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 아래 제단 같은 테이블 위에 찰리의 낡고 분해된 금속 부품들이 흩어져 놓인 구도.\n\nLOCATION (lock): At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane descent at the near corner of the altar-like table, retaining an oblique downward view rather than flattening the arrangement into an overhead diagram. Observe 찰리's worn, disassembled metal parts directly across the central band of the frame, keeping individual pieces modest in scale and readable against the tabletop, its near edge, and the surrounding room. The monitor has passed beyond the upper edge and 이현우 remains outside the crop beyond the table; the final reduction in camera distance concentrates attention on the interrupted body without introducing another visual axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Altar-like table (Supporting 찰리's separated parts beneath the off-frame monitor) — The upper surface and near corner are visible obliquely, with the near edge crossing the lower frame; used as Organizes the parts into a restrained central arrangement while preserving their scale; Laboratory interior (Open space visible beyond the table); used as Leaves quiet peripheral space around the close observation of the parts.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the laboratory broadly dark, with subdued neutral illumination sufficient to distinguish the separated metal parts without assigning a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's disassembled metal components rest on the altar-like table beneath the large monitor. His body is not assembled, so there is no intact torso, head-and-limb arrangement, or whole-body facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dim laboratory, Charlie's disassembled components lie on an altar-like table beneath a large monitor. The monitor displays energy-source Charlies distributed across Dubai, Europe, China and Africa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 아래 제단 같은 테이블 위에 찰리의 낡고 분해된 금속 부품들이 흩어져 놓인 구도.\n\nLOCATION (lock): At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane descent at the near corner of the altar-like table, retaining an oblique downward view rather than flattening the arrangement into an overhead diagram. Observe 찰리's worn, disassembled metal parts directly across the central band of the frame, keeping individual pieces modest in scale and readable against the tabletop, its near edge, and the surrounding room. The monitor has passed beyond the upper edge and 이현우 remains outside the crop beyond the table; the final reduction in camera distance concentrates attention on the interrupted body without introducing another visual axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Altar-like table (Supporting 찰리's separated parts beneath the off-frame monitor) — The upper surface and near corner are visible obliquely, with the near edge crossing the lower frame; used as Organizes the parts into a restrained central arrangement while preserving their scale; Laboratory interior (Open space visible beyond the table); used as Leaves quiet peripheral space around the close observation of the parts.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the laboratory broadly dark, with subdued neutral illumination sufficient to distinguish the separated metal parts without assigning a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's disassembled metal components rest on the altar-like table beneath the large monitor. His body is not assembled, so there is no intact torso, head-and-limb arrangement, or whole-body facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dim laboratory, Charlie's disassembled components lie on an altar-like table beneath a large monitor. The monitor displays energy-source Charlies distributed across Dubai, Europe, China and Africa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 아래 제단 같은 테이블 위에 찰리의 낡고 분해된 금속 부품들이 흩어져 놓인 구도.\n\nLOCATION (lock): At the altar-like parts table beneath a large active monitor in a spacious, dark research room. The screen provides the prominent light in the room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the crane descent at the near corner of the altar-like table, retaining an oblique downward view rather than flattening the arrangement into an overhead diagram. Observe 찰리's worn, disassembled metal parts directly across the central band of the frame, keeping individual pieces modest in scale and readable against the tabletop, its near edge, and the surrounding room. The monitor has passed beyond the upper edge and 이현우 remains outside the crop beyond the table; the final reduction in camera distance concentrates attention on the interrupted body without introducing another visual axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Altar-like table (Supporting 찰리's separated parts beneath the off-frame monitor) — The upper surface and near corner are visible obliquely, with the near edge crossing the lower frame; used as Organizes the parts into a restrained central arrangement while preserving their scale; Laboratory interior (Open space visible beyond the table); used as Leaves quiet peripheral space around the close observation of the parts.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the laboratory broadly dark, with subdued neutral illumination sufficient to distinguish the separated metal parts without assigning a new light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie's disassembled metal components rest on the altar-like table beneath the large monitor. His body is not assembled, so there is no intact torso, head-and-limb arrangement, or whole-body facing direction.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): In the dim laboratory, Charlie's disassembled components lie on an altar-like table beneath a large monitor. The monitor displays energy-source Charlies distributed across Dubai, Europe, China and Africa.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S90sh2__bgfirst_bg.png",
     "asset_id": "3180a031-cf36-49a4-a3fd-6b690d429dd9",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S90sh2.png",
     "asset_id": "8660d680-0c6a-4575-9790-c7eeb6feb6aa",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1230529>",
     "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reconstruction_hall_c365e6.png",
     "asset_id": "cead43ee-15ec-4696-a2d0-d3f4e8e2edc9",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1230529>",
     "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
    "built_space": "어두운 연구실 내부. 대형 테이블이 전면에 있고 뒤쪽 벽에 켜진 모니터가 위치함(모니터가 프레임에 포함되어 프레이밍 지시 위반).",
    "entities": "찰리의 분해된 금속 부품들. 머리 부품이 레퍼런스와 일치하는 흰색 마스크 형태로 정확히 표현됨.",
    "hard_violations": [],
    "physics": "모든 기계 부품이 테이블 표면에 흩어져 안정적으로 지지받고 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
    "built_space": "어두운 연구실 내부. 테이블과 배경의 대형 모니터가 모두 프레임에 포함됨.",
    "entities": "분해된 기계 부품들. 머리 부품이 흰색 마스크 형태가 아닌 일반적인 로봇 헬멧 디자인으로 잘못 그려짐.",
    "hard_violations": [],
    "physics": "모든 부품이 중력에 따라 테이블 위에 놓여 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "모니터가 프레임 밖으로 벗어나는 클로즈업 지시를 어기고 와이드 샷으로 연출되었으나, 찰리의 머리 부품(흰색 마스크와 검은 눈)을 레퍼런스와 정확하게 일치시켰습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프레이밍 지시(클로즈업)를 위반했을 뿐만 아니라, 찰리의 머리가 레퍼런스의 마스크가 아닌 임의의 기계 헬멧으로 묘사되어 캐릭터 일치도에서 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
        "built_space": "어두운 연구실 내부. 대형 테이블이 전면에 있고 뒤쪽 벽에 켜진 모니터가 위치함(모니터가 프레임에 포함되어 프레이밍 지시 위반).",
        "entities": "찰리의 분해된 금속 부품들. 머리 부품이 레퍼런스와 일치하는 흰색 마스크 형태로 정확히 표현됨.",
        "hard_violations": [],
        "physics": "모든 기계 부품이 테이블 표면에 흩어져 안정적으로 지지받고 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
        "built_space": "어두운 연구실 내부. 테이블과 배경의 대형 모니터가 모두 프레임에 포함됨.",
        "entities": "분해된 기계 부품들. 머리 부품이 흰색 마스크 형태가 아닌 일반적인 로봇 헬멧 디자인으로 잘못 그려짐.",
        "hard_violations": [],
        "physics": "모든 부품이 중력에 따라 테이블 위에 놓여 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "모니터가 프레임 밖으로 벗어나는 클로즈업 지시를 어기고 와이드 샷으로 연출되었으나, 찰리의 머리 부품(흰색 마스크와 검은 눈)을 레퍼런스와 정확하게 일치시켰습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프레이밍 지시(클로즈업)를 위반했을 뿐만 아니라, 찰리의 머리가 레퍼런스의 마스크가 아닌 임의의 기계 헬멧으로 묘사되어 캐릭터 일치도에서 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
        "built_space": "어두운 연구실 내부. 대형 테이블이 전면에 있고 뒤쪽 벽에 켜진 모니터가 위치함(모니터가 프레임에 포함되어 프레이밍 지시 위반).",
        "entities": "찰리의 분해된 금속 부품들. 머리 부품이 레퍼런스와 일치하는 흰색 마스크 형태로 정확히 표현됨.",
        "hard_violations": [],
        "physics": "모든 기계 부품이 테이블 표면에 흩어져 안정적으로 지지받고 있음."
       },
       {
        "label": "B",
        "direction": "카메라가 테이블 위를 비스듬히 내려다보는 시점.",
        "built_space": "어두운 연구실 내부. 테이블과 배경의 대형 모니터가 모두 프레임에 포함됨.",
        "entities": "분해된 기계 부품들. 머리 부품이 흰색 마스크 형태가 아닌 일반적인 로봇 헬멧 디자인으로 잘못 그려짐.",
        "hard_violations": [],
        "physics": "모든 부품이 중력에 따라 테이블 위에 놓여 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "B보다 테이블에 가까운 관찰 구도라 우세하지만, 모니터를 화면 밖으로 제외하지 못했고 찰리의 점·선형 얼굴도 재현하지 못했다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찰리의 얼굴과 장소는 더 충실하지만, 장소 참고 사진에 가까운 넓은 구도로 모니터와 주변 설비까지 보여 주어 지정된 클로즈업에서 더 멀어졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "분리된 머리의 마스크 면은 카메라 쪽에서 약간 오른쪽을 향한다. 왼쪽 팔과 손은 화면 아래쪽으로, 오른쪽 팔과 손은 오른쪽 아래로 뻗어 있지만 특정 대상을 가리키는 행동은 아니다. 인간의 시선이나 무기는 없다. 대형 모니터의 표시 면은 테이블 쪽을 향한다.",
        "built_space": "금속 제단형 테이블 한 개를 비스듬히 내려다보며 상판과 하단을 가로지르는 앞 테두리가 보인다. 가까운 모서리의 꼭짓점은 하단 밖으로 잘린다. 뒤에는 대형 모니터 한 개, 오른쪽 서버 랙 한 개, 소형 화면들이 놓인 작업대와 의자 한 개, 야간 창문이 보인다. 참고 장소의 재료와 배치는 대체로 유지했지만, 프레임 밖이어야 할 대형 모니터가 상단을 크게 차지한다.",
        "entities": "사람 없이 기계 머리 한 개, 손이 붙은 긴 팔 두 개, 중앙의 큰 몸통형 모듈, 여러 관절 하우징과 장갑판·체결 부품이 놓여 있다. 금속의 마모와 밝은 장갑판은 부합하지만, 머리는 캐릭터 참고의 둥근 검은 눈 두 개와 선형 입 대신 각진 투구형 얼굴이다. 중앙 몸통도 상당히 조립된 덩어리로 남아 있다. 모니터에는 세계 지도와 연결선이 보이며, 지역별 에너지원 찰리의 분포라는 세부 내용은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "머리와 큰 모듈들은 하부 외피로 상판에 닿아 있고, 팔과 손도 상판에 내려놓여 있다. 판재와 볼트는 눕거나 평평한 끝면으로 서 있다. 지지 없이 떠 있는 부품이나 움직이는 몸체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "분리된 머리의 흰 얼굴은 카메라 쪽에서 약간 오른쪽을 향하며, 검은 원형 눈에는 특정 대상을 보는 행동이 없다. 왼쪽 손은 화면 아래쪽, 오른쪽 손은 오른쪽 아래를 향해 상판에 놓였다. 조준이나 이동은 없고, 대형 모니터는 테이블을 향한다.",
        "built_space": "금속 제단형 테이블 한 개의 상판, 가까운 모서리, 앞면까지 넓게 보인다. 뒤쪽 대형 모니터 한 개와 그 받침, 오른쪽 서버 랙 한 개, 소형 화면들이 있는 작업대와 의자 한 개, 창문과 넓은 바닥이 드러난다. 왼쪽 전경에는 참고에도 있는 장비 일부가 보인다. 장소 재현은 충실하지만, 모니터를 제외한 근접 관찰 대신 주변 공간을 상당히 포함한 넓은 구도다.",
        "entities": "기계 머리 한 개, 긴 팔 두 개와 각각의 손, 큰 중앙 몸통형 모듈, 분리된 관절·장갑판·볼트가 있으며 인간은 없다. 흰 마스크에 검은 원형 눈 두 개와 짧은 선형 입이 있어 A보다 찰리의 얼굴 정체성에 가깝다. 장갑은 참고 캐릭터의 샌드 베이지보다 희게 보이고, 중앙 몸통은 여전히 큰 조립체로 남아 있다. 모니터의 세계 지도는 보이지만 지역별 찰리 표시는 명확하지 않다.",
        "hard_violations": [],
        "physics": "머리는 하단으로, 몸통형 모듈과 팔은 외피의 접촉면으로 테이블에 지지된다. 손가락도 상판 가까이에 내려앉아 있다. 작은 판재와 체결 부품에도 상판 지지가 있으며, 공중에 떠 있거나 불가능하게 버티는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "B보다 테이블에 가까운 관찰 구도라 우세하지만, 모니터를 화면 밖으로 제외하지 못했고 찰리의 점·선형 얼굴도 재현하지 못했다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "찰리의 얼굴과 장소는 더 충실하지만, 장소 참고 사진에 가까운 넓은 구도로 모니터와 주변 설비까지 보여 주어 지정된 클로즈업에서 더 멀어졌다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "분리된 머리의 마스크 면은 카메라 쪽에서 약간 오른쪽을 향한다. 왼쪽 팔과 손은 화면 아래쪽으로, 오른쪽 팔과 손은 오른쪽 아래로 뻗어 있지만 특정 대상을 가리키는 행동은 아니다. 인간의 시선이나 무기는 없다. 대형 모니터의 표시 면은 테이블 쪽을 향한다.",
        "built_space": "금속 제단형 테이블 한 개를 비스듬히 내려다보며 상판과 하단을 가로지르는 앞 테두리가 보인다. 가까운 모서리의 꼭짓점은 하단 밖으로 잘린다. 뒤에는 대형 모니터 한 개, 오른쪽 서버 랙 한 개, 소형 화면들이 놓인 작업대와 의자 한 개, 야간 창문이 보인다. 참고 장소의 재료와 배치는 대체로 유지했지만, 프레임 밖이어야 할 대형 모니터가 상단을 크게 차지한다.",
        "entities": "사람 없이 기계 머리 한 개, 손이 붙은 긴 팔 두 개, 중앙의 큰 몸통형 모듈, 여러 관절 하우징과 장갑판·체결 부품이 놓여 있다. 금속의 마모와 밝은 장갑판은 부합하지만, 머리는 캐릭터 참고의 둥근 검은 눈 두 개와 선형 입 대신 각진 투구형 얼굴이다. 중앙 몸통도 상당히 조립된 덩어리로 남아 있다. 모니터에는 세계 지도와 연결선이 보이며, 지역별 에너지원 찰리의 분포라는 세부 내용은 확인하기 어렵다.",
        "hard_violations": [],
        "physics": "머리와 큰 모듈들은 하부 외피로 상판에 닿아 있고, 팔과 손도 상판에 내려놓여 있다. 판재와 볼트는 눕거나 평평한 끝면으로 서 있다. 지지 없이 떠 있는 부품이나 움직이는 몸체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "분리된 머리의 흰 얼굴은 카메라 쪽에서 약간 오른쪽을 향하며, 검은 원형 눈에는 특정 대상을 보는 행동이 없다. 왼쪽 손은 화면 아래쪽, 오른쪽 손은 오른쪽 아래를 향해 상판에 놓였다. 조준이나 이동은 없고, 대형 모니터는 테이블을 향한다.",
        "built_space": "금속 제단형 테이블 한 개의 상판, 가까운 모서리, 앞면까지 넓게 보인다. 뒤쪽 대형 모니터 한 개와 그 받침, 오른쪽 서버 랙 한 개, 소형 화면들이 있는 작업대와 의자 한 개, 창문과 넓은 바닥이 드러난다. 왼쪽 전경에는 참고에도 있는 장비 일부가 보인다. 장소 재현은 충실하지만, 모니터를 제외한 근접 관찰 대신 주변 공간을 상당히 포함한 넓은 구도다.",
        "entities": "기계 머리 한 개, 긴 팔 두 개와 각각의 손, 큰 중앙 몸통형 모듈, 분리된 관절·장갑판·볼트가 있으며 인간은 없다. 흰 마스크에 검은 원형 눈 두 개와 짧은 선형 입이 있어 A보다 찰리의 얼굴 정체성에 가깝다. 장갑은 참고 캐릭터의 샌드 베이지보다 희게 보이고, 중앙 몸통은 여전히 큰 조립체로 남아 있다. 모니터의 세계 지도는 보이지만 지역별 찰리 표시는 명확하지 않다.",
        "hard_violations": [],
        "physics": "머리는 하단으로, 몸통형 모듈과 팔은 외피의 접촉면으로 테이블에 지지된다. 손가락도 상판 가까이에 내려앉아 있다. 작은 판재와 체결 부품에도 상판 지지가 있으며, 공중에 떠 있거나 불가능하게 버티는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.8,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1800,
   "B": 1667
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1800,
    "verdict_ko": "모니터가 프레임 밖으로 벗어나는 클로즈업 지시를 어기고 와이드 샷으로 연출되었으나, 찰리의 머리 부품(흰색 마스크와 검은 눈)을 레퍼런스와 정확하게 일치시켰습니다."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "프레이밍 지시(클로즈업)를 위반했을 뿐만 아니라, 찰리의 머리가 레퍼런스의 마스크가 아닌 임의의 기계 헬멧으로 묘사되어 캐릭터 일치도에서 크게 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_reconstruction_hall_c365e6.png",
    "asset_id": "cead43ee-15ec-4696-a2d0-d3f4e8e2edc9",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1230529>",
    "asset_id": "067e469f-9f75-4785-8fa9-66e54ef5eb89",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0edc-3e80-7592-9420-63ea728a7bff",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S90sh2__bgfirst_bg.png",
   "bg_asset_id": "3180a031-cf36-49a4-a3fd-6b690d429dd9",
   "bg_record_key": "S90sh2::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "reconstruction_hall",
   "groupbg_asset_id": "cead43ee-15ec-4696-a2d0-d3f4e8e2edc9"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S90sh16::signage": {
  "fp": "ecf7579caa4dc472",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S90sh16": {
  "input_fingerprint": "07416c19e6912b1d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 미소 지은 채 홀로그램 찰리의 손을 향해 자신의 손을 조심스럽게 뻗은 근접 찰나.\n\nLOCATION (lock): In the hologram viewing area before the parts table in the spacious research room. The active projection illuminates the otherwise dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track beside 이현우 at shoulder height, retaining the established oblique view of his smiling three-quarter face and the space between the approaching hands. Place his face in the upper-left and his carefully extended hand across the lower center, while 홀로그램 찰리's partial torso and hand enter from the right with his head outside the crop; 이현우 looks toward that off-frame face, not toward the lens. Observe the real figure and the holographic image together in the laboratory, emphasizing only the narrowing hand gap and preserving a visible separation rather than implying solid contact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory interior (Visible as limited spatial context behind the reaching figures); used as Keeps the encounter grounded in the room rather than presenting the hologram as a full-frame alternate reality.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's subdued darkness and restrained contrast, distinguishing the holographic appearance without adding an unsupported colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large monitor, altar-like table, scattered mechanical parts, and dim room lighting from the reference. Exclude the previous worldwide display content; replace it with the current holographic presentation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 홀로그램 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large display has changed to Jeju, and a holographic Charlie appears with images of Amber and Raul approaching him. The physical, disassembled Charlie components remain on the table in the dim laboratory. 이현우: He remains covered in wounds and has received the memory chip. He has risen from his seat and reaches out with a smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 홀로그램 찰리 (로봇 형태의 입체 영상, 고릴라형 비율, 긴 팔과 짧은 다리, 샌드 베이지 외장, 흰 마스크형 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 미소 지은 채 홀로그램 찰리의 손을 향해 자신의 손을 조심스럽게 뻗은 근접 찰나.\n\nLOCATION (lock): In the hologram viewing area before the parts table in the spacious research room. The active projection illuminates the otherwise dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track beside 이현우 at shoulder height, retaining the established oblique view of his smiling three-quarter face and the space between the approaching hands. Place his face in the upper-left and his carefully extended hand across the lower center, while 홀로그램 찰리's partial torso and hand enter from the right with his head outside the crop; 이현우 looks toward that off-frame face, not toward the lens. Observe the real figure and the holographic image together in the laboratory, emphasizing only the narrowing hand gap and preserving a visible separation rather than implying solid contact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory interior (Visible as limited spatial context behind the reaching figures); used as Keeps the encounter grounded in the room rather than presenting the hologram as a full-frame alternate reality.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's subdued darkness and restrained contrast, distinguishing the holographic appearance without adding an unsupported colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large monitor, altar-like table, scattered mechanical parts, and dim room lighting from the reference. Exclude the previous worldwide display content; replace it with the current holographic presentation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 홀로그램 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large display has changed to Jeju, and a holographic Charlie appears with images of Amber and Raul approaching him. The physical, disassembled Charlie components remain on the table in the dim laboratory. 이현우: He remains covered in wounds and has received the memory chip. He has risen from his seat and reaches out with a smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 홀로그램 찰리 (로봇 형태의 입체 영상, 고릴라형 비율, 긴 팔과 짧은 다리, 샌드 베이지 외장, 흰 마스크형 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 이현우가 환하게 미소 지은 채 홀로그램 찰리의 손을 향해 자신의 손을 조심스럽게 뻗은 근접 찰나.\n\nLOCATION (lock): In the hologram viewing area before the parts table in the spacious research room. The active projection illuminates the otherwise dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the lateral track beside 이현우 at shoulder height, retaining the established oblique view of his smiling three-quarter face and the space between the approaching hands. Place his face in the upper-left and his carefully extended hand across the lower center, while 홀로그램 찰리's partial torso and hand enter from the right with his head outside the crop; 이현우 looks toward that off-frame face, not toward the lens. Observe the real figure and the holographic image together in the laboratory, emphasizing only the narrowing hand gap and preserving a visible separation rather than implying solid contact.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Laboratory interior (Visible as limited spatial context behind the reaching figures); used as Keeps the encounter grounded in the room rather than presenting the hologram as a full-frame alternate reality.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the laboratory's subdued darkness and restrained contrast, distinguishing the holographic appearance without adding an unsupported colored glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large monitor, altar-like table, scattered mechanical parts, and dim room lighting from the reference. Exclude the previous worldwide display content; replace it with the current holographic presentation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 홀로그램 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large display has changed to Jeju, and a holographic Charlie appears with images of Amber and Raul approaching him. The physical, disassembled Charlie components remain on the table in the dim laboratory. 이현우: He remains covered in wounds and has received the memory chip. He has risen from his seat and reaches out with a smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이현우 (한국계 미국인 남성, 10대 후반의 얼굴, 짧은 검은 머리, 헝클어진 머리칼) — wearing: 제주도 연구소의 무균 상태를 보여주는 새하얗고 깔끔한 상하의 환자복.; 홀로그램 찰리 (로봇 형태의 입체 영상, 고릴라형 비율, 긴 팔과 짧은 다리, 샌드 베이지 외장, 흰 마스크형 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이현우가 우측 상단을 바라보며 손을 뻗고 있으며, 로봇의 손 역시 그를 향해 마주 뻗음.",
    "built_space": "연구실 배경에 부품 테이블과 대형 모니터가 있음. 단, 모니터에는 제주도가 아닌 고릴라형 로봇과 두 인물의 모습이 띄워져 있음.",
    "entities": "이현우는 상처를 입고 미소 짓는 모습이 잘 표현됨. 홀로그램 찰리는 우측에서 등장하나, 지시와 달리 프레임 상단에 턱과 얼굴 일부가 포함됨.",
    "hard_violations": [],
    "physics": "손과 팔이 서로를 향해 뻗어 있으며 자연스러운 근육의 긴장과 자세로 지탱됨."
   },
   {
    "label": "B",
    "direction": "이현우의 시선이 화면 밖 찰리의 얼굴 위치를 향하며, 두 사람의 손이 서로를 향해 뻗어 있음.",
    "built_space": "어두운 연구실 내부로, 대형 모니터와 기계 부품이 놓인 테이블이 배치됨. 모니터 화면에는 제주도 지도와 두 인물의 이미지가 표시됨.",
    "entities": "이현우는 흰 환자복을 입고 얼굴과 팔에 상처가 있는 채로 미소 짓고 있음. 홀로그램 찰리는 스캔라인 질감의 입체 영상으로 우측에서 등장하며, 머리는 프레임 밖에 위치함.",
    "hard_violations": [],
    "physics": "뻗은 두 팔은 어깨와 몸통에 의해 자연스럽게 지탱되며 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "대형 모니터에 제주도 지도를 정확히 반영했고, 찰리의 머리를 프레임 밖으로 배치하라는 구도 지시를 훌륭하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "모니터 화면에 제주도 지도가 아닌 로봇을 표시했으며, 프레임에서 제외되어야 할 찰리의 얼굴 일부가 노출되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 밖 찰리의 얼굴 위치를 향하며, 두 사람의 손이 서로를 향해 뻗어 있음.",
        "built_space": "어두운 연구실 내부로, 대형 모니터와 기계 부품이 놓인 테이블이 배치됨. 모니터 화면에는 제주도 지도와 두 인물의 이미지가 표시됨.",
        "entities": "이현우는 흰 환자복을 입고 얼굴과 팔에 상처가 있는 채로 미소 짓고 있음. 홀로그램 찰리는 스캔라인 질감의 입체 영상으로 우측에서 등장하며, 머리는 프레임 밖에 위치함.",
        "hard_violations": [],
        "physics": "뻗은 두 팔은 어깨와 몸통에 의해 자연스럽게 지탱되며 허공에 떠 있음."
       },
       {
        "label": "A",
        "direction": "이현우가 우측 상단을 바라보며 손을 뻗고 있으며, 로봇의 손 역시 그를 향해 마주 뻗음.",
        "built_space": "연구실 배경에 부품 테이블과 대형 모니터가 있음. 단, 모니터에는 제주도가 아닌 고릴라형 로봇과 두 인물의 모습이 띄워져 있음.",
        "entities": "이현우는 상처를 입고 미소 짓는 모습이 잘 표현됨. 홀로그램 찰리는 우측에서 등장하나, 지시와 달리 프레임 상단에 턱과 얼굴 일부가 포함됨.",
        "hard_violations": [],
        "physics": "손과 팔이 서로를 향해 뻗어 있으며 자연스러운 근육의 긴장과 자세로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "대형 모니터에 제주도 지도를 정확히 반영했고, 찰리의 머리를 프레임 밖으로 배치하라는 구도 지시를 훌륭하게 따랐습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "모니터 화면에 제주도 지도가 아닌 로봇을 표시했으며, 프레임에서 제외되어야 할 찰리의 얼굴 일부가 노출되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이현우의 시선이 화면 밖 찰리의 얼굴 위치를 향하며, 두 사람의 손이 서로를 향해 뻗어 있음.",
        "built_space": "어두운 연구실 내부로, 대형 모니터와 기계 부품이 놓인 테이블이 배치됨. 모니터 화면에는 제주도 지도와 두 인물의 이미지가 표시됨.",
        "entities": "이현우는 흰 환자복을 입고 얼굴과 팔에 상처가 있는 채로 미소 짓고 있음. 홀로그램 찰리는 스캔라인 질감의 입체 영상으로 우측에서 등장하며, 머리는 프레임 밖에 위치함.",
        "hard_violations": [],
        "physics": "뻗은 두 팔은 어깨와 몸통에 의해 자연스럽게 지탱되며 허공에 떠 있음."
       },
       {
        "label": "A",
        "direction": "이현우가 우측 상단을 바라보며 손을 뻗고 있으며, 로봇의 손 역시 그를 향해 마주 뻗음.",
        "built_space": "연구실 배경에 부품 테이블과 대형 모니터가 있음. 단, 모니터에는 제주도가 아닌 고릴라형 로봇과 두 인물의 모습이 띄워져 있음.",
        "entities": "이현우는 상처를 입고 미소 짓는 모습이 잘 표현됨. 홀로그램 찰리는 우측에서 등장하나, 지시와 달리 프레임 상단에 턱과 얼굴 일부가 포함됨.",
        "hard_violations": [],
        "physics": "손과 팔이 서로를 향해 뻗어 있으며 자연스러운 근육의 긴장과 자세로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "더 밀착된 얼굴 중심 구도와 제주로 바뀐 화면, 분명한 손 사이 간격이 우세하지만, 찰리의 머리 일부가 들어오고 환자복이 오염되어 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "미소와 손 뻗기 방향은 맞지만, 상체와 작업대를 더 넓게 보여 근접 구도에서 멀어지고 찰리의 얼굴이 들어오며 대형 화면의 제주 전환도 보이지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이현우는 렌즈가 아니라 오른쪽 위 찰리의 얼굴 방향을 바라본다. 자신의 손은 오른쪽 찰리의 기계 손을 향하고, 찰리의 손가락도 왼쪽 이현우의 손을 향한다. 양쪽 손끝 사이에는 명확한 빈틈이 있다. 무기나 다른 지향성 소품은 없다.",
        "built_space": "뒤쪽 대형 모니터 1개와 하단의 금속 부품 작업대 1개가 보인다. 작업대 위에는 분리된 로봇 머리 1개, 팔과 손 조립체 및 여러 외장 부품이 놓여 있다. 어두운 기둥과 따뜻한 선형 조명이 이전 장소의 재질과 분위기를 이어간다. 이현우와 홀로그램은 작업대 앞 좌우에 배치된다. 얼굴이 왼쪽 위에 크게 있지만 손은 하단 중앙보다 조금 높고, 오른쪽 위에는 찰리의 머리 하단 일부가 들어온다. 불가능한 반사는 보이지 않는다.",
        "entities": "짧고 헝클어진 검은 머리의 동아시아계 청소년 남성이 밝게 웃으며 얼굴·목·팔에 상처를 지니고 있어 이현우의 주요 조건에 부합한다. 미국 국적은 외관으로 확인할 수 없다. 흰 브이넥 환자복은 참조와 유사하나 피와 얼룩이 있어 깨끗한 복장 지시와 다르다. 찰리는 사람 피부가 아닌 기계식 외장과 관절 손을 가진 입체 영상이지만, 청색 테두리와 발광 때문에 샌드 베이지색이 약해졌다. 모니터에는 제주 형태의 지도와 남녀 인물 영상이 보인다. 메모리 칩과 하체는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 손은 손목과 전완, 굽힌 팔꿈치로 몸에 자연스럽게 연결되어 있어 조심스럽게 뻗는 동작이 가능하다. 발은 화면 밖이므로 접지 상태를 판정할 수 없다. 찰리의 손도 기계식 손목과 팔에 연결되며, 명시된 홀로그램이므로 실물의 공중 부유로 볼 근거는 없다. 분해 부품은 작업대 표면에 지지되어 있고 두 손은 접촉하지 않는다."
       },
       {
        "label": "B",
        "direction": "이현우의 시선은 오른쪽 위 찰리의 얼굴 쪽으로 향하며 렌즈를 보지 않는다. 이현우의 손은 찰리의 펼친 손바닥과 손가락을 향해 오른쪽으로 뻗고, 찰리의 손은 왼쪽으로 내밀어져 있다. 두 손 사이의 간격이 유지된다. 찰리의 눈은 잘려 있어 자체 시선은 확인하기 어렵다.",
        "built_space": "대형 모니터 1개, 넓은 금속 작업대 1개, 왼쪽 후경의 작은 작업 구역이 보인다. 작업대에는 분리된 팔·손 조립체와 여러 외장 부품이 놓여 있다. 창밖의 어두운 산과 불빛, 검은 기둥, 선형 조명은 야간 연구실의 연속성을 지킨다. 두 인물은 작업대 앞에 있지만, 이현우의 허리 가까이와 작업대 넓은 면적까지 보여 요구된 근접 구도보다 넓다. 찰리의 마스크형 얼굴 하단도 오른쪽 위에 상당 부분 들어온다. 중복된 주요 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 검은 머리의 젊은 동아시아계 남성으로, 밝은 미소와 얼굴·목·팔의 상처가 표현되어 있다. 흰 환자복은 참조 계열이지만 어깨의 혈흔과 여러 얼룩은 깨끗한 복장 지시와 어긋난다. 찰리의 베이지 외장, 두꺼운 로봇 팔, 기계식 손과 흰 마스크 일부는 참조의 특징을 따른다. 다만 청백색 윤곽 발광이 강하다. 모니터는 로봇 전신과 남녀 초상 영상을 보여주며 제주로 바뀐 표시 내용은 확인되지 않는다. 칩과 화면 밖 하체는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 뻗은 팔과 손은 어깨부터 연속적으로 연결되고 팔꿈치 굽힘도 자연스럽다. 하체 접지는 화면 밖이다. 찰리의 열린 손은 손목과 전완에 연결된 홀로그램으로 표현되어 지지 없는 실물 손이 아니다. 작업대의 부품들은 상판에 놓여 있으며, 두 손은 겹치거나 물리적으로 맞잡지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "더 밀착된 얼굴 중심 구도와 제주로 바뀐 화면, 분명한 손 사이 간격이 우세하지만, 찰리의 머리 일부가 들어오고 환자복이 오염되어 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "미소와 손 뻗기 방향은 맞지만, 상체와 작업대를 더 넓게 보여 근접 구도에서 멀어지고 찰리의 얼굴이 들어오며 대형 화면의 제주 전환도 보이지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이현우는 렌즈가 아니라 오른쪽 위 찰리의 얼굴 방향을 바라본다. 자신의 손은 오른쪽 찰리의 기계 손을 향하고, 찰리의 손가락도 왼쪽 이현우의 손을 향한다. 양쪽 손끝 사이에는 명확한 빈틈이 있다. 무기나 다른 지향성 소품은 없다.",
        "built_space": "뒤쪽 대형 모니터 1개와 하단의 금속 부품 작업대 1개가 보인다. 작업대 위에는 분리된 로봇 머리 1개, 팔과 손 조립체 및 여러 외장 부품이 놓여 있다. 어두운 기둥과 따뜻한 선형 조명이 이전 장소의 재질과 분위기를 이어간다. 이현우와 홀로그램은 작업대 앞 좌우에 배치된다. 얼굴이 왼쪽 위에 크게 있지만 손은 하단 중앙보다 조금 높고, 오른쪽 위에는 찰리의 머리 하단 일부가 들어온다. 불가능한 반사는 보이지 않는다.",
        "entities": "짧고 헝클어진 검은 머리의 동아시아계 청소년 남성이 밝게 웃으며 얼굴·목·팔에 상처를 지니고 있어 이현우의 주요 조건에 부합한다. 미국 국적은 외관으로 확인할 수 없다. 흰 브이넥 환자복은 참조와 유사하나 피와 얼룩이 있어 깨끗한 복장 지시와 다르다. 찰리는 사람 피부가 아닌 기계식 외장과 관절 손을 가진 입체 영상이지만, 청색 테두리와 발광 때문에 샌드 베이지색이 약해졌다. 모니터에는 제주 형태의 지도와 남녀 인물 영상이 보인다. 메모리 칩과 하체는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "이현우의 손은 손목과 전완, 굽힌 팔꿈치로 몸에 자연스럽게 연결되어 있어 조심스럽게 뻗는 동작이 가능하다. 발은 화면 밖이므로 접지 상태를 판정할 수 없다. 찰리의 손도 기계식 손목과 팔에 연결되며, 명시된 홀로그램이므로 실물의 공중 부유로 볼 근거는 없다. 분해 부품은 작업대 표면에 지지되어 있고 두 손은 접촉하지 않는다."
       },
       {
        "label": "A",
        "direction": "이현우의 시선은 오른쪽 위 찰리의 얼굴 쪽으로 향하며 렌즈를 보지 않는다. 이현우의 손은 찰리의 펼친 손바닥과 손가락을 향해 오른쪽으로 뻗고, 찰리의 손은 왼쪽으로 내밀어져 있다. 두 손 사이의 간격이 유지된다. 찰리의 눈은 잘려 있어 자체 시선은 확인하기 어렵다.",
        "built_space": "대형 모니터 1개, 넓은 금속 작업대 1개, 왼쪽 후경의 작은 작업 구역이 보인다. 작업대에는 분리된 팔·손 조립체와 여러 외장 부품이 놓여 있다. 창밖의 어두운 산과 불빛, 검은 기둥, 선형 조명은 야간 연구실의 연속성을 지킨다. 두 인물은 작업대 앞에 있지만, 이현우의 허리 가까이와 작업대 넓은 면적까지 보여 요구된 근접 구도보다 넓다. 찰리의 마스크형 얼굴 하단도 오른쪽 위에 상당 부분 들어온다. 중복된 주요 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "이현우는 검은 머리의 젊은 동아시아계 남성으로, 밝은 미소와 얼굴·목·팔의 상처가 표현되어 있다. 흰 환자복은 참조 계열이지만 어깨의 혈흔과 여러 얼룩은 깨끗한 복장 지시와 어긋난다. 찰리의 베이지 외장, 두꺼운 로봇 팔, 기계식 손과 흰 마스크 일부는 참조의 특징을 따른다. 다만 청백색 윤곽 발광이 강하다. 모니터는 로봇 전신과 남녀 초상 영상을 보여주며 제주로 바뀐 표시 내용은 확인되지 않는다. 칩과 화면 밖 하체는 판단할 수 없다.",
        "hard_violations": [],
        "physics": "이현우의 뻗은 팔과 손은 어깨부터 연속적으로 연결되고 팔꿈치 굽힘도 자연스럽다. 하체 접지는 화면 밖이다. 찰리의 열린 손은 손목과 전완에 연결된 홀로그램으로 표현되어 지지 없는 실물 손이 아니다. 작업대의 부품들은 상판에 놓여 있으며, 두 손은 겹치거나 물리적으로 맞잡지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.482,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.482,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 2000,
   "A": 1482
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "대형 모니터에 제주도 지도를 정확히 반영했고, 찰리의 머리를 프레임 밖으로 배치하라는 구도 지시를 훌륭하게 따랐습니다."
   },
   {
    "label": "A",
    "score": 1482,
    "verdict_ko": "모니터 화면에 제주도 지도가 아닌 로봇을 표시했으며, 프레임에서 제외되어야 할 찰리의 얼굴 일부가 노출되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S90sh2_sel.png",
    "asset_id": "6214d07f-85e5-43d8-be91-e9cc8b2fd588",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이현우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:766860>",
    "asset_id": "9169e9a9-8769-48e3-8394-9e002b18f080",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 홀로그램 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1209867>",
    "asset_id": "71c835fb-2f85-4373-880b-cac5dc515f78",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ee5-a58a-70d3-86ad-4b7ed8a32577",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S90sh2"
  }
 },
 "S90sh19::signage": {
  "fp": "c9f8bc42ae473570",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S90sh19": {
  "input_fingerprint": "b23f824e80038209",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조각이 맞춰진 찰리의 낡은 금속 얼굴이 눈을 감은 채 평온하게 누워 있는 얼굴 클로즈업.\n\nLOCATION (lock): On the reassembly table in the spacious research room, where the robot's broken components are being fitted together. The room remains dim around the active display and projection. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward dolly beside the head end of the operating table, above and laterally offset from 찰리's face, holding the inherited oblique downward angle. His assembled face occupies the central half of the image on a gentle diagonal, with closed eyes clearly readable and portions of the head, upper body, and table retained at the edges. This is direct observation of his peaceful, unresponsive repose, not entry into a dream; camera distance is the only intensified axis as the movement resolves into stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting the reassembled 찰리) — A narrow portion of the supporting upper surface is visible beside and beneath his head; used as Establishes physical support and continuity with the assembly passage without overwhelming the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued laboratory illumination gently separates the worn metal facial planes while preserving the scene's darkness and quiet, sleep-like tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the operating table as his broken components are fitted back together, with his eyes closed. The table supports the reconstructed body, but the exact orientation of his head and torso, the placement of his limbs, and the extent of completed reassembly are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken components are being fitted together on the operating table, with his eyes closed as though asleep. The Jeju holographic display remains active in the otherwise dim laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조각이 맞춰진 찰리의 낡은 금속 얼굴이 눈을 감은 채 평온하게 누워 있는 얼굴 클로즈업.\n\nLOCATION (lock): On the reassembly table in the spacious research room, where the robot's broken components are being fitted together. The room remains dim around the active display and projection. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward dolly beside the head end of the operating table, above and laterally offset from 찰리's face, holding the inherited oblique downward angle. His assembled face occupies the central half of the image on a gentle diagonal, with closed eyes clearly readable and portions of the head, upper body, and table retained at the edges. This is direct observation of his peaceful, unresponsive repose, not entry into a dream; camera distance is the only intensified axis as the movement resolves into stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting the reassembled 찰리) — A narrow portion of the supporting upper surface is visible beside and beneath his head; used as Establishes physical support and continuity with the assembly passage without overwhelming the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued laboratory illumination gently separates the worn metal facial planes while preserving the scene's darkness and quiet, sleep-like tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the operating table as his broken components are fitted back together, with his eyes closed. The table supports the reconstructed body, but the exact orientation of his head and torso, the placement of his limbs, and the extent of completed reassembly are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken components are being fitted together on the operating table, with his eyes closed as though asleep. The Jeju holographic display remains active in the otherwise dim laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조각이 맞춰진 찰리의 낡은 금속 얼굴이 눈을 감은 채 평온하게 누워 있는 얼굴 클로즈업.\n\nLOCATION (lock): On the reassembly table in the spacious research room, where the robot's broken components are being fitted together. The room remains dim around the active display and projection. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the forward dolly beside the head end of the operating table, above and laterally offset from 찰리's face, holding the inherited oblique downward angle. His assembled face occupies the central half of the image on a gentle diagonal, with closed eyes clearly readable and portions of the head, upper body, and table retained at the edges. This is direct observation of his peaceful, unresponsive repose, not entry into a dream; camera distance is the only intensified axis as the movement resolves into stillness.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Operating table (Supporting the reassembled 찰리) — A narrow portion of the supporting upper surface is visible beside and beneath his head; used as Establishes physical support and continuity with the assembly passage without overwhelming the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued laboratory illumination gently separates the worn metal facial planes while preserving the scene's darkness and quiet, sleep-like tenderness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Charlie is recumbent on the operating table as his broken components are fitted back together, with his eyes closed. The table supports the reconstructed body, but the exact orientation of his head and torso, the placement of his limbs, and the extent of completed reassembly are not specified.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's broken components are being fitted together on the operating table, with his eyes closed as though asleep. The Jeju holographic display remains active in the otherwise dim laboratory.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
    "built_space": "이전 샷과 동일한 어두운 금속 질감의 수술대와 조명 띠가 보이며, 흩어진 볼트 부품들이 놓여 있음.",
    "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
    "hard_violations": [],
    "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
   },
   {
    "label": "B",
    "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
    "built_space": "어두운 금속 수술대와 조명 띠가 보이나, 주변 부품 묘사가 비교적 적음.",
    "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
    "hard_violations": [],
    "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요청된 클로즈업 구도와 카메라 앵글을 정확히 구현하였으며, 이전 샷의 수술대 디테일과 찰리의 낡은 금속 질감을 매우 훌륭하게 살려냈습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "구도와 찰리의 디테일이 훌륭하나, 주변 부품 등의 묘사에서 A에 비해 디테일이 약간 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
        "built_space": "이전 샷과 동일한 어두운 금속 질감의 수술대와 조명 띠가 보이며, 흩어진 볼트 부품들이 놓여 있음.",
        "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
       },
       {
        "label": "B",
        "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
        "built_space": "어두운 금속 수술대와 조명 띠가 보이나, 주변 부품 묘사가 비교적 적음.",
        "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요청된 클로즈업 구도와 카메라 앵글을 정확히 구현하였으며, 이전 샷의 수술대 디테일과 찰리의 낡은 금속 질감을 매우 훌륭하게 살려냈습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "구도와 찰리의 디테일이 훌륭하나, 주변 부품 등의 묘사에서 A에 비해 디테일이 약간 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
        "built_space": "이전 샷과 동일한 어두운 금속 질감의 수술대와 조명 띠가 보이며, 흩어진 볼트 부품들이 놓여 있음.",
        "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
       },
       {
        "label": "B",
        "direction": "찰리는 눈을 감고 누워 있어 특정 방향을 응시하지 않음.",
        "built_space": "어두운 금속 수술대와 조명 띠가 보이나, 주변 부품 묘사가 비교적 적음.",
        "entities": "찰리(흰색 마스크형 얼굴, 샌드 베이지 장갑판, 감은 눈)가 레퍼런스와 일치하게 재현됨.",
        "hard_violations": [],
        "physics": "찰리의 머리와 상체가 수술대 표면 위에 자연스럽게 눕혀져 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "머리 위 측면에서 내려다보는 사선 클로즈업과 절제된 조명이, 눈을 감고 평온하게 누운 찰리의 낡은 금속 얼굴을 더 충실하게 강조한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "찰리의 정체성과 감긴 눈, 조립대의 지지는 잘 맞지만, A보다 정면성이 강하고 밝은 가슴 장갑과 주변 부품이 얼굴의 주목도를 나눈다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 얼굴은 위쪽을 향하고 머리에서 가슴으로 이어지는 축은 화면 왼쪽 위에서 오른쪽 아래로 비스듬히 놓인다. 두 원형 눈에는 닫힌 기계식 눈꺼풀과 가는 틈이 보이며, 특정 대상을 응시하지 않는다. 카메라는 얼굴 위의 측면에서 비스듬히 내려다본다. 겨누는 물체나 이동하는 신체는 없다.",
        "built_space": "긁힌 금속 상판을 가진 조립대 하나가 머리 뒤와 왼쪽에 이어지고, 가장자리의 선형 조명이 보인다. 왼쪽 상판에는 원통형 체결 부품 두 개가 놓여 있다. 얼굴이 중앙부를 차지하고 양어깨와 가슴 일부는 아래쪽과 오른쪽 가장자리에서 잘린다. 이전 장면의 금속 작업대와 어두운 청색·따뜻한 선형 조명의 재료 및 조명 계열이 유지된다. 대형 지도 화면은 이 클로즈업 밖에 있어 작동 여부를 확인할 수 없다. 중복 설비나 불가능한 거울 반사는 보이지 않는다.",
        "entities": "등장 개체는 기계 몸체의 찰리 하나뿐이다. 흰 마스크형 얼굴, 원형 눈 두 개, 선형 입, 이마의 작은 직사각형 부품과 체결점, 원형 귀 관절, 마모된 샌드 베이지 장갑이 캐릭터 참조와 대응한다. 가장자리에 보이는 가슴의 푸른 원형 부품도 캐릭터 참조와 맞는다. 인간의 피부·머리카락·치아는 추가되지 않았다. 긴 팔과 짧은 다리의 전체 비례는 프레임 밖이므로 평가할 수 없다. 화면 위에 추가된 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽은 상판 가까이에 놓이고 목의 기계 연결부가 머리와 몸통을 연결한다. 몸통과 어깨는 조립대 위에 누워 있으며, 보이는 부분에서 능동적으로 몸을 들어 올리는 자세는 없다. 왼쪽의 체결 부품 두 개는 상판에 밑면을 대고 서 있다. 지지 없이 떠 있는 신체나 부품은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "찰리는 얼굴을 위로 향하고 눈을 감고 있으며 시선의 대상은 없다. 머리와 몸통은 화면 왼쪽 위에서 오른쪽 아래로 이어진다. 카메라도 위에서 내려다보지만 A보다 얼굴과 상체를 정면에 가깝게 보여준다. 겨누는 물체나 이동 동작은 없다.",
        "built_space": "조립대 하나의 금속 상판이 머리 뒤와 왼쪽에 보이고, 오른쪽 가장자리를 따라 밝은 선형 조명이 이어진다. 왼쪽에는 납작한 장갑판 한 개와 화면에 일부만 들어온 것을 포함한 체결 부품 네 개가 보인다. 얼굴은 중앙부에 있으나 양어깨와 가슴 장갑도 아래쪽 화면을 크게 차지한다. 작업대의 흠집과 차갑고 어두운 연구실 배경은 이전 장면과 부합한다. 지도 화면은 프레임 밖이라 확인할 수 없다. 설비 중복이나 광학적으로 불가능한 반사는 없다.",
        "entities": "찰리 한 개체만 등장한다. 흰 마스크, 닫힌 기계식 원형 눈, 선형 입, 이마의 직사각형 부품, 원형 귀 관절과 샌드 베이지 장갑이 참조의 정체성을 유지한다. 푸른 가슴 원형 부품도 보인다. 얼굴과 장갑에는 물리적인 긁힘과 마모가 있으며 인간 피부나 얼굴 근육은 없다. 주변 장갑판과 체결 부품은 이전 장면의 재조립 상황에 해당한다. 추가 인물이나 삽입 문자는 없고, 전신 비례는 크롭 때문에 확인할 수 없다.",
        "hard_violations": [],
        "physics": "몸통과 어깨는 조립대 위에 누워 있고 머리는 목의 기계 연결부를 통해 몸통에 연결되어 있다. 뒤통수와 상판의 정확한 접촉선은 가려져 있지만, 머리가 연결부 없이 떠 있지는 않다. 왼쪽 장갑판과 체결 부품은 모두 상판에 놓여 있다. 노출된 신체 부분에서 무지지 부유나 움직임을 요구하는 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "머리 위 측면에서 내려다보는 사선 클로즈업과 절제된 조명이, 눈을 감고 평온하게 누운 찰리의 낡은 금속 얼굴을 더 충실하게 강조한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "찰리의 정체성과 감긴 눈, 조립대의 지지는 잘 맞지만, A보다 정면성이 강하고 밝은 가슴 장갑과 주변 부품이 얼굴의 주목도를 나눈다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "찰리의 얼굴은 위쪽을 향하고 머리에서 가슴으로 이어지는 축은 화면 왼쪽 위에서 오른쪽 아래로 비스듬히 놓인다. 두 원형 눈에는 닫힌 기계식 눈꺼풀과 가는 틈이 보이며, 특정 대상을 응시하지 않는다. 카메라는 얼굴 위의 측면에서 비스듬히 내려다본다. 겨누는 물체나 이동하는 신체는 없다.",
        "built_space": "긁힌 금속 상판을 가진 조립대 하나가 머리 뒤와 왼쪽에 이어지고, 가장자리의 선형 조명이 보인다. 왼쪽 상판에는 원통형 체결 부품 두 개가 놓여 있다. 얼굴이 중앙부를 차지하고 양어깨와 가슴 일부는 아래쪽과 오른쪽 가장자리에서 잘린다. 이전 장면의 금속 작업대와 어두운 청색·따뜻한 선형 조명의 재료 및 조명 계열이 유지된다. 대형 지도 화면은 이 클로즈업 밖에 있어 작동 여부를 확인할 수 없다. 중복 설비나 불가능한 거울 반사는 보이지 않는다.",
        "entities": "등장 개체는 기계 몸체의 찰리 하나뿐이다. 흰 마스크형 얼굴, 원형 눈 두 개, 선형 입, 이마의 작은 직사각형 부품과 체결점, 원형 귀 관절, 마모된 샌드 베이지 장갑이 캐릭터 참조와 대응한다. 가장자리에 보이는 가슴의 푸른 원형 부품도 캐릭터 참조와 맞는다. 인간의 피부·머리카락·치아는 추가되지 않았다. 긴 팔과 짧은 다리의 전체 비례는 프레임 밖이므로 평가할 수 없다. 화면 위에 추가된 글자는 없다.",
        "hard_violations": [],
        "physics": "머리 뒤쪽은 상판 가까이에 놓이고 목의 기계 연결부가 머리와 몸통을 연결한다. 몸통과 어깨는 조립대 위에 누워 있으며, 보이는 부분에서 능동적으로 몸을 들어 올리는 자세는 없다. 왼쪽의 체결 부품 두 개는 상판에 밑면을 대고 서 있다. 지지 없이 떠 있는 신체나 부품은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "찰리는 얼굴을 위로 향하고 눈을 감고 있으며 시선의 대상은 없다. 머리와 몸통은 화면 왼쪽 위에서 오른쪽 아래로 이어진다. 카메라도 위에서 내려다보지만 A보다 얼굴과 상체를 정면에 가깝게 보여준다. 겨누는 물체나 이동 동작은 없다.",
        "built_space": "조립대 하나의 금속 상판이 머리 뒤와 왼쪽에 보이고, 오른쪽 가장자리를 따라 밝은 선형 조명이 이어진다. 왼쪽에는 납작한 장갑판 한 개와 화면에 일부만 들어온 것을 포함한 체결 부품 네 개가 보인다. 얼굴은 중앙부에 있으나 양어깨와 가슴 장갑도 아래쪽 화면을 크게 차지한다. 작업대의 흠집과 차갑고 어두운 연구실 배경은 이전 장면과 부합한다. 지도 화면은 프레임 밖이라 확인할 수 없다. 설비 중복이나 광학적으로 불가능한 반사는 없다.",
        "entities": "찰리 한 개체만 등장한다. 흰 마스크, 닫힌 기계식 원형 눈, 선형 입, 이마의 직사각형 부품, 원형 귀 관절과 샌드 베이지 장갑이 참조의 정체성을 유지한다. 푸른 가슴 원형 부품도 보인다. 얼굴과 장갑에는 물리적인 긁힘과 마모가 있으며 인간 피부나 얼굴 근육은 없다. 주변 장갑판과 체결 부품은 이전 장면의 재조립 상황에 해당한다. 추가 인물이나 삽입 문자는 없고, 전신 비례는 크롭 때문에 확인할 수 없다.",
        "hard_violations": [],
        "physics": "몸통과 어깨는 조립대 위에 누워 있고 머리는 목의 기계 연결부를 통해 몸통에 연결되어 있다. 뒤통수와 상판의 정확한 접촉선은 가려져 있지만, 머리가 연결부 없이 떠 있지는 않다. 왼쪽 장갑판과 체결 부품은 모두 상판에 놓여 있다. 노출된 신체 부분에서 무지지 부유나 움직임을 요구하는 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 1889,
   "B": 1875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "요청된 클로즈업 구도와 카메라 앵글을 정확히 구현하였으며, 이전 샷의 수술대 디테일과 찰리의 낡은 금속 질감을 매우 훌륭하게 살려냈습니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "구도와 찰리의 디테일이 훌륭하나, 주변 부품 등의 묘사에서 A에 비해 디테일이 약간 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S90sh2_sel.png",
    "asset_id": "6214d07f-85e5-43d8-be91-e9cc8b2fd588",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1299876>",
    "asset_id": "99dc4ad5-1251-4608-89ee-6005d5fc811c",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0eea-aeeb-780e-aaf4-a7ba2edc193a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S90sh2"
  }
 },
 "S91sh4::signage": {
  "fp": "02d14bfd02603a18",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::abf5ca27b890b10a": {
  "subjects": [],
  "subject_text": "섬바위 주변 얕은 바다·수중\n투명하게 맑은 에메랄드빛 해안의 바위섬 일대다.",
  "identity": "canonical",
  "scope_id": "L156",
  "scope_role": "location_exterior",
  "scope_sha": "3963eed2a60258a2"
 },
 "groupbg::island_rock_shallows": {
  "input_fingerprint": "5916d364447520b0",
  "meta": {
   "model": "gpt-image-2.5-sunburst",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "island_rock_shallows",
    "tags": [
     "S91sh13",
     "S91sh4",
     "S91sh7"
    ]
   },
   "context_sig": "71d46ba77f68d4e3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n섬바위 주변 얕은 바다·수중: 투명하게 맑은 에메랄드빛 해안의 바위섬 일대다. (특징: 바닷속으로 헤엄쳐 다니는 물고기들의 실루엣; 수면으로 강하게 다이빙하는 찰리의 둔탁한 파열음과 물보라; 그물로 고기를 잡는 앰버와 라울의 모습)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 91. 에필로그 / 어느 바다 섬바위 – D 자막: 한 달 후\n- 앰버와 라울도 함께 고기를 잡는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities: In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n섬바위 주변 얕은 바다·수중: 투명하게 맑은 에메랄드빛 해안의 바위섬 일대다. (특징: 바닷속으로 헤엄쳐 다니는 물고기들의 실루엣; 수면으로 강하게 다이빙하는 찰리의 둔탁한 파열음과 물보라; 그물로 고기를 잡는 앰버와 라울의 모습)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 91. 에필로그 / 어느 바다 섬바위 – D 자막: 한 달 후\n- 앰버와 라울도 함께 고기를 잡는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Writing the attached references already show may be readable — a sign, a plate, a building marking — each in the script that object itself carries; never substitute or translate a script the references fix. Wording that only one shot would need does not belong on a shared plate: leave it out rather than fixing it into the place. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_island_rock_shallows_e00e29.png",
  "asset_id": "0ea26009-f738-4493-90fd-bd836df1381d",
  "input_asset_ids": [
   "f2e25c03-b38c-4eb2-9ec4-3f02992dd74a"
  ],
  "origin_tag": "S91sh4",
  "place_text": "In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.",
  "origin_inputs": {
   "place_text": "In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.",
   "time_of_day_en": "day",
   "conti_asset_id": "f2e25c03-b38c-4eb2-9ec4-3f02992dd74a"
  }
 },
 "S91sh4::bgfirst_bg": {
  "input_fingerprint": "d2d4ba0dcde5aa8c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리 곁의 바닷물 속에서 앰버와 라울이 해맑게 웃으며 허리를 굽힌 채 수면 아래로 손을 뻗은 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at 앰버 and 라울's waist height, looking obliquely downward across their bent bodies toward their submerged hands, with 찰리 farther along the same viewing direction. Place 앰버 at left and 라울 at center-right, their smiles turned toward the fish below rather than the camera, and retain 찰리 in the upper-center background within the shared water space. Show different phases of the same fishing action—앰버 reaching farther down, 라울 bending into his reach, and 찰리 pitched toward the fish—while holding camera distance and lighting steady so lateral subject placement supplies the changing emphasis.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clear seawater (Transparent enough to reveal fish and the reaching hands below the surface) — Viewed obliquely downward through the surface toward submerged hands and fish; used as Connects the three fishing figures through a shared, visible target space; Fish beneath the surface (Visible in the seawater); used as Provide the concrete focus of the downward gazes and reaching gestures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime illumination preserves visibility through the clear seawater and readable smiles, balancing subdued contrast with the warmth of shared play.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 찰리 곁의 바닷물 속에서 앰버와 라울이 해맑게 웃으며 허리를 굽힌 채 수면 아래로 손을 뻗은 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at 앰버 and 라울's waist height, looking obliquely downward across their bent bodies toward their submerged hands, with 찰리 farther along the same viewing direction. Place 앰버 at left and 라울 at center-right, their smiles turned toward the fish below rather than the camera, and retain 찰리 in the upper-center background within the shared water space. Show different phases of the same fishing action—앰버 reaching farther down, 라울 bending into his reach, and 찰리 pitched toward the fish—while holding camera distance and lighting steady so lateral subject placement supplies the changing emphasis.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clear seawater (Transparent enough to reveal fish and the reaching hands below the surface) — Viewed obliquely downward through the surface toward submerged hands and fish; used as Connects the three fishing figures through a shared, visible target space; Fish beneath the surface (Visible in the seawater); used as Provide the concrete focus of the downward gazes and reaching gestures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime illumination preserves visibility through the clear seawater and readable smiles, balancing subdued contrast with the warmth of shared play.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Writing the attached references already show may be readable — a sign, a plate, a building or vehicle marking — each in the script that object itself carries; never substitute or translate a script the references fix. Invent no other wording: do not derive writing from what a place like this would have. Nothing is laid on top of the photograph: no caption, watermark or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S91sh4__bgfirst_bg.png",
  "asset_id": "dbc2b9b7-db16-49f5-b40e-13e7ac802e82",
  "input_asset_ids": [
   "f2e25c03-b38c-4eb2-9ec4-3f02992dd74a",
   "0ea26009-f738-4493-90fd-bd836df1381d"
  ]
 },
 "S91sh4": {
  "input_fingerprint": "693b12cb1b3d5365",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리 곁의 바닷물 속에서 앰버와 라울이 해맑게 웃으며 허리를 굽힌 채 수면 아래로 손을 뻗은 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at 앰버 and 라울's waist height, looking obliquely downward across their bent bodies toward their submerged hands, with 찰리 farther along the same viewing direction. Place 앰버 at left and 라울 at center-right, their smiles turned toward the fish below rather than the camera, and retain 찰리 in the upper-center background within the shared water space. Show different phases of the same fishing action—앰버 reaching farther down, 라울 bending into his reach, and 찰리 pitched toward the fish—while holding camera distance and lighting steady so lateral subject placement supplies the changing emphasis.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clear seawater (Transparent enough to reveal fish and the reaching hands below the surface) — Viewed obliquely downward through the surface toward submerged hands and fish; used as Connects the three fishing figures through a shared, visible target space; Fish beneath the surface (Visible in the seawater); used as Provide the concrete focus of the downward gazes and reaching gestures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime illumination preserves visibility through the clear seawater and readable smiles, balancing subdued contrast with the warmth of shared play.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A month later, the daylight sea is clear enough to reveal fish beneath the surface, and a table is being set with food. Charlie is reassembled and operational, attempting to catch fish. 앰버: She is fishing in the water. 라울: He is fishing in the water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리 곁의 바닷물 속에서 앰버와 라울이 해맑게 웃으며 허리를 굽힌 채 수면 아래로 손을 뻗은 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at 앰버 and 라울's waist height, looking obliquely downward across their bent bodies toward their submerged hands, with 찰리 farther along the same viewing direction. Place 앰버 at left and 라울 at center-right, their smiles turned toward the fish below rather than the camera, and retain 찰리 in the upper-center background within the shared water space. Show different phases of the same fishing action—앰버 reaching farther down, 라울 bending into his reach, and 찰리 pitched toward the fish—while holding camera distance and lighting steady so lateral subject placement supplies the changing emphasis.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clear seawater (Transparent enough to reveal fish and the reaching hands below the surface) — Viewed obliquely downward through the surface toward submerged hands and fish; used as Connects the three fishing figures through a shared, visible target space; Fish beneath the surface (Visible in the seawater); used as Provide the concrete focus of the downward gazes and reaching gestures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime illumination preserves visibility through the clear seawater and readable smiles, balancing subdued contrast with the warmth of shared play.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A month later, the daylight sea is clear enough to reveal fish beneath the surface, and a table is being set with food. Charlie is reassembled and operational, attempting to catch fish. 앰버: She is fishing in the water. 라울: He is fishing in the water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리 곁의 바닷물 속에서 앰버와 라울이 해맑게 웃으며 허리를 굽힌 채 수면 아래로 손을 뻗은 역동적인 구도.\n\nLOCATION (lock): In clear shallow seawater beside an island rock outcrop, close enough to reach below the surface while standing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Continue the lateral track at 앰버 and 라울's waist height, looking obliquely downward across their bent bodies toward their submerged hands, with 찰리 farther along the same viewing direction. Place 앰버 at left and 라울 at center-right, their smiles turned toward the fish below rather than the camera, and retain 찰리 in the upper-center background within the shared water space. Show different phases of the same fishing action—앰버 reaching farther down, 라울 bending into his reach, and 찰리 pitched toward the fish—while holding camera distance and lighting steady so lateral subject placement supplies the changing emphasis.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clear seawater (Transparent enough to reveal fish and the reaching hands below the surface) — Viewed obliquely downward through the surface toward submerged hands and fish; used as Connects the three fishing figures through a shared, visible target space; Fish beneath the surface (Visible in the seawater); used as Provide the concrete focus of the downward gazes and reaching gestures.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime illumination preserves visibility through the clear seawater and readable smiles, balancing subdued contrast with the warmth of shared play.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A month later, the daylight sea is clear enough to reveal fish beneath the surface, and a table is being set with food. Charlie is reassembled and operational, attempting to catch fish. 앰버: She is fishing in the water. 라울: He is fishing in the water.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 앰버 (한국계 백인 혼혈 여자아이, 10세의 어린 얼굴, 금발 머리, 커다란 눈) — wearing: 금발 머리와 창백한 피부에 대비되는 정교한 미래형 하이테크 방진 마스크, 기름때가 묻은 카키색 오버롤 작업복과 허리에 찬 공구 가죽 벨트.; 라울 (라틴계 흑인 혼혈 남자아이, 10세의 어린 얼굴, 뒤로 묶은 꽁지머리) — wearing: 난민촌 환경에 맞게 색이 바래고 얼룩진 헐렁한 티셔츠와 반바지. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S91sh4__bgfirst_bg.png",
     "asset_id": "dbc2b9b7-db16-49f5-b40e-13e7ac802e82",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/conti/conti_S91sh4.png",
     "asset_id": "f2e25c03-b38c-4eb2-9ec4-3f02992dd74a",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_island_rock_shallows_e00e29.png",
     "asset_id": "0ea26009-f738-4493-90fd-bd836df1381d",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1075117>",
     "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:876078>",
     "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:958670>",
     "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "B",
    "direction": "앰버와 라울 모두 웃으며 전경의 수중 물고기 무리를 내려다보고, 각각 한 손을 그 무리 쪽으로 뻗는다. 카메라를 응시하지 않는다. 중앙 뒤의 찰리도 얼굴과 몸을 아래로 기울이고 기계 손을 앞쪽 수면으로 내민다. 다만 두 아이의 뻗기 깊이와 동작 단계가 상당히 비슷하다.",
    "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽 소나무와 암벽 해안, 중앙 뒤 암초, 오른쪽 먼 섬, 돌이 비치는 얕은 바다가 장소 참고와 대응한다. 앰버는 왼쪽, 라울은 오른쪽, 찰리는 두 아이 사이 뒤쪽의 같은 물에 있다. 아이들의 상체가 화면을 크게 차지하며, 수평선과 얼굴을 정면에 가깝게 보는 낮은 시점이라 지정된 와이드·사선 하향 구도는 약하다.",
    "entities": "금발과 밝은 피부의 어린 여자아이, 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이, 기계 찰리만 보인다. 아이들의 외형은 각각 앰버와 라울의 참고에 대체로 부합하나 혼혈 배경 자체는 영상만으로 확정할 수 없다. 앰버의 때 묻은 카키 작업복, 가죽 공구 벨트와 목에 걸린 방진 마스크, 라울의 낡은 녹색 티셔츠와 갈색 반바지가 보인다. 찰리는 베이지 장갑판, 긴 팔, 흰 마스크 얼굴, 두 발광 눈과 선형 입, 푸른 가슴 원자로를 갖췄다. 수중 물고기와 손이 선명하며 추가 인물이나 문자는 없다.",
    "hard_violations": [],
    "physics": "두 아이는 물에 잠긴 다리를 굽혀 낮춘 자세이며 얕은 바닥이 체중을 받는 것으로 읽힌다. 라울은 다른 손을 자기 무릎에 얹고 있다. 찰리도 잠긴 하체와 낮춘 팔로 몸을 지지하는 자세다. 손과 팔은 연결되어 있고 수면의 굴절과 물결이 겹친다. 물고기는 물속을 헤엄치며, 지지 없이 공중에 뜬 몸이나 물체는 없다."
   },
   {
    "label": "A",
    "direction": "앰버는 웃으며 자기 손끝 앞의 전경 물고기를 내려다보고 팔을 멀리 아래로 뻗는다. 라울은 중앙 오른쪽에서 허리를 숙여 자기 손 아래의 물고기로 손을 내민다. 찰리는 상단 중앙에서 고개와 몸통을 앞으로 기울이고 중앙 수중 물고기 방향으로 긴 팔을 뻗는다. 세 인물의 시선과 동작이 공유된 물고기 공간으로 모이며 카메라를 보지 않는다.",
    "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 소나무 암벽 해안, 뒤쪽 암초와 오른쪽 먼 섬, 투명한 돌바닥 바다가 장소 참고의 주요 요소를 유지한다. 앰버는 왼쪽 전경, 라울은 중앙 오른쪽, 찰리는 상단 중앙 배경에 있다. 전경 수중 손과 물고기까지 넓게 포함하고 몸 위에서 물 쪽으로 내려다보는 경사가 A보다 명확해 지정된 와이드 구도에 가깝다. 다만 배경 찰리가 아이들에 비해 상당히 크게 보인다.",
    "entities": "앰버는 밝은 피부와 금발의 어린 여자아이로, 카키 작업복·공구 가죽 벨트·목에 걸린 기계식 마스크가 참고와 대응한다. 라울은 짙은 피부와 묶은 곱슬머리의 어린 남자아이이며, 얼룩진 헐렁한 티셔츠와 반바지를 입었다. 혼혈 정체성은 외형만으로 단정할 수 없지만 두 아이의 연령대와 주요 식별 특징은 맞는다. 찰리는 베이지 기계 장갑, 흰 마스크, 두 발광 눈과 입 선, 푸른 원자로, 긴 팔을 갖췄으나 짧고 뚱뚱한 성인 정도라는 크기보다 육중하게 읽힌다. 수중에는 여러 물고기가 있으며 추가 인물이나 글자는 없다.",
    "hard_violations": [],
    "physics": "앰버와 라울은 물속의 굽힌 다리로 바닥을 딛는 낮은 자세다. 앰버가 반대 팔을 옆으로 벌린 것은 앞으로 뻗는 몸의 균형 동작으로 가능하다. 라울의 다른 손은 수면 가까이에 내려와 있다. 찰리의 하체와 반대쪽 팔 끝은 물에 잠겨 지지 자세를 이루며, 앞으로 내민 기계 손은 팔에 정상적으로 연결되어 있다. 수중 손에는 물결과 굴절이 겹치고 물고기는 물 안에 있다. 무지지 부유나 명백히 불가능한 자세는 보이지 않는다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": null,
     "normalized": null,
     "ok": false
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "물고기를 향한 시선과 수중 손, 찰리의 크기는 잘 맞지만, 아이들을 크게 잡은 낮고 정면적인 구도가 지정된 와이드 숏과 비스듬한 하향 시점에 덜 충실하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 앰버·중앙 오른쪽 라울·상단 중앙 찰리의 와이드 배치와 서로 다른 뻗기 동작이 더 정확하지만, 뒤쪽 찰리의 몸집은 지정된 작은 성인 체격보다 크게 읽힌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앰버와 라울 모두 웃으며 전경의 수중 물고기 무리를 내려다보고, 각각 한 손을 그 무리 쪽으로 뻗는다. 카메라를 응시하지 않는다. 중앙 뒤의 찰리도 얼굴과 몸을 아래로 기울이고 기계 손을 앞쪽 수면으로 내민다. 다만 두 아이의 뻗기 깊이와 동작 단계가 상당히 비슷하다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽 소나무와 암벽 해안, 중앙 뒤 암초, 오른쪽 먼 섬, 돌이 비치는 얕은 바다가 장소 참고와 대응한다. 앰버는 왼쪽, 라울은 오른쪽, 찰리는 두 아이 사이 뒤쪽의 같은 물에 있다. 아이들의 상체가 화면을 크게 차지하며, 수평선과 얼굴을 정면에 가깝게 보는 낮은 시점이라 지정된 와이드·사선 하향 구도는 약하다.",
        "entities": "금발과 밝은 피부의 어린 여자아이, 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이, 기계 찰리만 보인다. 아이들의 외형은 각각 앰버와 라울의 참고에 대체로 부합하나 혼혈 배경 자체는 영상만으로 확정할 수 없다. 앰버의 때 묻은 카키 작업복, 가죽 공구 벨트와 목에 걸린 방진 마스크, 라울의 낡은 녹색 티셔츠와 갈색 반바지가 보인다. 찰리는 베이지 장갑판, 긴 팔, 흰 마스크 얼굴, 두 발광 눈과 선형 입, 푸른 가슴 원자로를 갖췄다. 수중 물고기와 손이 선명하며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 아이는 물에 잠긴 다리를 굽혀 낮춘 자세이며 얕은 바닥이 체중을 받는 것으로 읽힌다. 라울은 다른 손을 자기 무릎에 얹고 있다. 찰리도 잠긴 하체와 낮춘 팔로 몸을 지지하는 자세다. 손과 팔은 연결되어 있고 수면의 굴절과 물결이 겹친다. 물고기는 물속을 헤엄치며, 지지 없이 공중에 뜬 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "앰버는 웃으며 자기 손끝 앞의 전경 물고기를 내려다보고 팔을 멀리 아래로 뻗는다. 라울은 중앙 오른쪽에서 허리를 숙여 자기 손 아래의 물고기로 손을 내민다. 찰리는 상단 중앙에서 고개와 몸통을 앞으로 기울이고 중앙 수중 물고기 방향으로 긴 팔을 뻗는다. 세 인물의 시선과 동작이 공유된 물고기 공간으로 모이며 카메라를 보지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 소나무 암벽 해안, 뒤쪽 암초와 오른쪽 먼 섬, 투명한 돌바닥 바다가 장소 참고의 주요 요소를 유지한다. 앰버는 왼쪽 전경, 라울은 중앙 오른쪽, 찰리는 상단 중앙 배경에 있다. 전경 수중 손과 물고기까지 넓게 포함하고 몸 위에서 물 쪽으로 내려다보는 경사가 A보다 명확해 지정된 와이드 구도에 가깝다. 다만 배경 찰리가 아이들에 비해 상당히 크게 보인다.",
        "entities": "앰버는 밝은 피부와 금발의 어린 여자아이로, 카키 작업복·공구 가죽 벨트·목에 걸린 기계식 마스크가 참고와 대응한다. 라울은 짙은 피부와 묶은 곱슬머리의 어린 남자아이이며, 얼룩진 헐렁한 티셔츠와 반바지를 입었다. 혼혈 정체성은 외형만으로 단정할 수 없지만 두 아이의 연령대와 주요 식별 특징은 맞는다. 찰리는 베이지 기계 장갑, 흰 마스크, 두 발광 눈과 입 선, 푸른 원자로, 긴 팔을 갖췄으나 짧고 뚱뚱한 성인 정도라는 크기보다 육중하게 읽힌다. 수중에는 여러 물고기가 있으며 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버와 라울은 물속의 굽힌 다리로 바닥을 딛는 낮은 자세다. 앰버가 반대 팔을 옆으로 벌린 것은 앞으로 뻗는 몸의 균형 동작으로 가능하다. 라울의 다른 손은 수면 가까이에 내려와 있다. 찰리의 하체와 반대쪽 팔 끝은 물에 잠겨 지지 자세를 이루며, 앞으로 내민 기계 손은 팔에 정상적으로 연결되어 있다. 수중 손에는 물결과 굴절이 겹치고 물고기는 물 안에 있다. 무지지 부유나 명백히 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "물고기를 향한 시선과 수중 손, 찰리의 크기는 잘 맞지만, 아이들을 크게 잡은 낮고 정면적인 구도가 지정된 와이드 숏과 비스듬한 하향 시점에 덜 충실하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 앰버·중앙 오른쪽 라울·상단 중앙 찰리의 와이드 배치와 서로 다른 뻗기 동작이 더 정확하지만, 뒤쪽 찰리의 몸집은 지정된 작은 성인 체격보다 크게 읽힌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앰버와 라울 모두 웃으며 전경의 수중 물고기 무리를 내려다보고, 각각 한 손을 그 무리 쪽으로 뻗는다. 카메라를 응시하지 않는다. 중앙 뒤의 찰리도 얼굴과 몸을 아래로 기울이고 기계 손을 앞쪽 수면으로 내민다. 다만 두 아이의 뻗기 깊이와 동작 단계가 상당히 비슷하다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽 소나무와 암벽 해안, 중앙 뒤 암초, 오른쪽 먼 섬, 돌이 비치는 얕은 바다가 장소 참고와 대응한다. 앰버는 왼쪽, 라울은 오른쪽, 찰리는 두 아이 사이 뒤쪽의 같은 물에 있다. 아이들의 상체가 화면을 크게 차지하며, 수평선과 얼굴을 정면에 가깝게 보는 낮은 시점이라 지정된 와이드·사선 하향 구도는 약하다.",
        "entities": "금발과 밝은 피부의 어린 여자아이, 짙은 피부와 뒤로 묶은 곱슬머리의 어린 남자아이, 기계 찰리만 보인다. 아이들의 외형은 각각 앰버와 라울의 참고에 대체로 부합하나 혼혈 배경 자체는 영상만으로 확정할 수 없다. 앰버의 때 묻은 카키 작업복, 가죽 공구 벨트와 목에 걸린 방진 마스크, 라울의 낡은 녹색 티셔츠와 갈색 반바지가 보인다. 찰리는 베이지 장갑판, 긴 팔, 흰 마스크 얼굴, 두 발광 눈과 선형 입, 푸른 가슴 원자로를 갖췄다. 수중 물고기와 손이 선명하며 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "두 아이는 물에 잠긴 다리를 굽혀 낮춘 자세이며 얕은 바닥이 체중을 받는 것으로 읽힌다. 라울은 다른 손을 자기 무릎에 얹고 있다. 찰리도 잠긴 하체와 낮춘 팔로 몸을 지지하는 자세다. 손과 팔은 연결되어 있고 수면의 굴절과 물결이 겹친다. 물고기는 물속을 헤엄치며, 지지 없이 공중에 뜬 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "앰버는 웃으며 자기 손끝 앞의 전경 물고기를 내려다보고 팔을 멀리 아래로 뻗는다. 라울은 중앙 오른쪽에서 허리를 숙여 자기 손 아래의 물고기로 손을 내민다. 찰리는 상단 중앙에서 고개와 몸통을 앞으로 기울이고 중앙 수중 물고기 방향으로 긴 팔을 뻗는다. 세 인물의 시선과 동작이 공유된 물고기 공간으로 모이며 카메라를 보지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 소나무 암벽 해안, 뒤쪽 암초와 오른쪽 먼 섬, 투명한 돌바닥 바다가 장소 참고의 주요 요소를 유지한다. 앰버는 왼쪽 전경, 라울은 중앙 오른쪽, 찰리는 상단 중앙 배경에 있다. 전경 수중 손과 물고기까지 넓게 포함하고 몸 위에서 물 쪽으로 내려다보는 경사가 A보다 명확해 지정된 와이드 구도에 가깝다. 다만 배경 찰리가 아이들에 비해 상당히 크게 보인다.",
        "entities": "앰버는 밝은 피부와 금발의 어린 여자아이로, 카키 작업복·공구 가죽 벨트·목에 걸린 기계식 마스크가 참고와 대응한다. 라울은 짙은 피부와 묶은 곱슬머리의 어린 남자아이이며, 얼룩진 헐렁한 티셔츠와 반바지를 입었다. 혼혈 정체성은 외형만으로 단정할 수 없지만 두 아이의 연령대와 주요 식별 특징은 맞는다. 찰리는 베이지 기계 장갑, 흰 마스크, 두 발광 눈과 입 선, 푸른 원자로, 긴 팔을 갖췄으나 짧고 뚱뚱한 성인 정도라는 크기보다 육중하게 읽힌다. 수중에는 여러 물고기가 있으며 추가 인물이나 글자는 없다.",
        "hard_violations": [],
        "physics": "앰버와 라울은 물속의 굽힌 다리로 바닥을 딛는 낮은 자세다. 앰버가 반대 팔을 옆으로 벌린 것은 앞으로 뻗는 몸의 균형 동작으로 가능하다. 라울의 다른 손은 수면 가까이에 내려와 있다. 찰리의 하체와 반대쪽 팔 끝은 물에 잠겨 지지 자세를 이루며, 앞으로 내민 기계 손은 팔에 정상적으로 연결되어 있다. 수중 손에는 물결과 굴절이 겹치고 물고기는 물 안에 있다. 무지지 부유나 명백히 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "failed": [
    "gemini-pro"
   ],
   "route": "single_reverse"
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "B": 7,
   "A": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "물고기를 향한 시선과 수중 손, 찰리의 크기는 잘 맞지만, 아이들을 크게 잡은 낮고 정면적인 구도가 지정된 와이드 숏과 비스듬한 하향 시점에 덜 충실하다."
   },
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "왼쪽 앰버·중앙 오른쪽 라울·상단 중앙 찰리의 와이드 배치와 서로 다른 뻗기 동작이 더 정확하지만, 뒤쪽 찰리의 몸집은 지정된 작은 성인 체격보다 크게 읽힌다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/groupbg_island_rock_shallows_e00e29.png",
    "asset_id": "0ea26009-f738-4493-90fd-bd836df1381d",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 앰버: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:876078>",
    "asset_id": "f99d3a60-0ac2-4ed8-8626-d50120839208",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 라울: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:958670>",
    "asset_id": "7ec394fd-3470-4cbf-9b3d-64634ac5c720",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0ef0-75cb-7a62-94bc-cf0aea9a642c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S91sh4__bgfirst_bg.png",
   "bg_asset_id": "dbc2b9b7-db16-49f5-b40e-13e7ac802e82",
   "bg_record_key": "S91sh4::bgfirst_bg",
   "chain_winner": true,
   "winner_origin": "A",
   "authority": "groupbg",
   "group_key": "island_rock_shallows",
   "groupbg_asset_id": "0ea26009-f738-4493-90fd-bd836df1381d"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S91sh7::signage": {
  "fp": "e3785937db9f87be",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S91sh7": {
  "input_fingerprint": "112100f5978ce170",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 'B-200이 돌보던 아기 새'가 물속에 선 찰리의 낡은 어깨 위로 사뿐히 내려앉아 발을 막 딛은 근접 찰나.\n\nLOCATION (lock): In the sunlit shallows beside the island rocks, above the waterline at the standing robot's shoulder height. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane almost to a stop beside 찰리's shoulder, slightly above it and looking obliquely downward from the established front-side angle. Frame 아기 새 in the upper-left with its feet just touching the shoulder near center, while 찰리's partial three-quarter face occupies the right edge and remains inclined toward the fish below the crop. Observe the landing directly, keeping the feet-to-shoulder contact sharply readable and the bird's body angled toward its landing point; its settling weight, rather than a camera reframe or 찰리 turning to acknowledge it, supplies the change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater below 찰리 (Visible beyond his shoulder); used as Provides a restrained background field that preserves the fishing location without competing with the landing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding restrained daylight and controlled contrast, preserving the small feet and worn shoulder detail without introducing a special highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight sea remains clear, with fish visible below the surface and food being arranged on the nearby table. Charlie is operational in the water, and a baby bird has settled on his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 'B-200이 돌보던 아기 새'가 물속에 선 찰리의 낡은 어깨 위로 사뿐히 내려앉아 발을 막 딛은 근접 찰나.\n\nLOCATION (lock): In the sunlit shallows beside the island rocks, above the waterline at the standing robot's shoulder height. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane almost to a stop beside 찰리's shoulder, slightly above it and looking obliquely downward from the established front-side angle. Frame 아기 새 in the upper-left with its feet just touching the shoulder near center, while 찰리's partial three-quarter face occupies the right edge and remains inclined toward the fish below the crop. Observe the landing directly, keeping the feet-to-shoulder contact sharply readable and the bird's body angled toward its landing point; its settling weight, rather than a camera reframe or 찰리 turning to acknowledge it, supplies the change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater below 찰리 (Visible beyond his shoulder); used as Provides a restrained background field that preserves the fishing location without competing with the landing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding restrained daylight and controlled contrast, preserving the small feet and worn shoulder detail without introducing a special highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight sea remains clear, with fish visible below the surface and food being arranged on the nearby table. Charlie is operational in the water, and a baby bird has settled on his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에서 날아온 'B-200이 돌보던 아기 새'가 물속에 선 찰리의 낡은 어깨 위로 사뿐히 내려앉아 발을 막 딛은 근접 찰나.\n\nLOCATION (lock): In the sunlit shallows beside the island rocks, above the waterline at the standing robot's shoulder height. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Ease the crane almost to a stop beside 찰리's shoulder, slightly above it and looking obliquely downward from the established front-side angle. Frame 아기 새 in the upper-left with its feet just touching the shoulder near center, while 찰리's partial three-quarter face occupies the right edge and remains inclined toward the fish below the crop. Observe the landing directly, keeping the feet-to-shoulder contact sharply readable and the bird's body angled toward its landing point; its settling weight, rather than a camera reframe or 찰리 turning to acknowledge it, supplies the change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater below 찰리 (Visible beyond his shoulder); used as Provides a restrained background field that preserves the fishing location without competing with the landing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding restrained daylight and controlled contrast, preserving the small feet and worn shoulder detail without introducing a special highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight sea remains clear, with fish visible below the surface and food being arranged on the nearby table. Charlie is operational in the water, and a baby bird has settled on his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형.; 아기 새 (어린 새, 짧은 부리, 작은 날개, 가는 발가락, 둥근 깃털 몸체) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "찰리의 시선은 화면 아래(물속)를 향하고 있으며, 아기 새는 어깨의 착지점을 향해 몸을 기울이고 있음.",
    "built_space": "얕은 바닷물과 바위 바닥이 배경으로 깔려 있으며, 카메라는 찰리의 어깨 위에서 비스듬히 내려다보는 정확한 앵글을 취함.",
    "entities": "찰리(흰색 마스크, 모래색 장갑, 가슴의 푸른 원자로)와 아기 새 모두 레퍼런스의 특징을 정확히 반영함.",
    "hard_violations": [],
    "physics": "아기 새의 발이 어깨 표면에 정확히 닿아 체중을 지탱하고 있으며, 펼친 날개가 막 착지한 동작을 물리적으로 뒷받침함."
   },
   {
    "label": "B",
    "direction": "찰리의 시선은 화면 아래를 향하고, 아기 새 역시 어깨 쪽을 바라봄.",
    "built_space": "바닷물이 있는 야외 배경이며, 카메라는 요구된 어깨 근접 앵글을 유지함.",
    "entities": "찰리와 아기 새가 등장하나, 찰리의 흰색 마스크 이마에 레퍼런스에 없는 패널 선이 추가됨.",
    "hard_violations": [],
    "physics": "아기 새의 발이 어깨에 닿아 몸을 지탱하고 있으며, 날개는 펼쳐진 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 클로즈업 구도와 피사체의 배치(왼쪽 상단의 아기 새, 오른쪽 가장자리의 찰리 얼굴)를 매우 정확하게 구현했으며, 어깨에 막 내려앉는 발의 접촉면이 선명합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 피사체 배치는 양호하나, 찰리의 마스크 표면에 레퍼런스에 없는 불필요한 선이 들어갔고 로봇 관절의 구조적 자연스러움이 A에 비해 다소 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선은 화면 아래(물속)를 향하고 있으며, 아기 새는 어깨의 착지점을 향해 몸을 기울이고 있음.",
        "built_space": "얕은 바닷물과 바위 바닥이 배경으로 깔려 있으며, 카메라는 찰리의 어깨 위에서 비스듬히 내려다보는 정확한 앵글을 취함.",
        "entities": "찰리(흰색 마스크, 모래색 장갑, 가슴의 푸른 원자로)와 아기 새 모두 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "아기 새의 발이 어깨 표면에 정확히 닿아 체중을 지탱하고 있으며, 펼친 날개가 막 착지한 동작을 물리적으로 뒷받침함."
       },
       {
        "label": "B",
        "direction": "찰리의 시선은 화면 아래를 향하고, 아기 새 역시 어깨 쪽을 바라봄.",
        "built_space": "바닷물이 있는 야외 배경이며, 카메라는 요구된 어깨 근접 앵글을 유지함.",
        "entities": "찰리와 아기 새가 등장하나, 찰리의 흰색 마스크 이마에 레퍼런스에 없는 패널 선이 추가됨.",
        "hard_violations": [],
        "physics": "아기 새의 발이 어깨에 닿아 몸을 지탱하고 있으며, 날개는 펼쳐진 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구된 클로즈업 구도와 피사체의 배치(왼쪽 상단의 아기 새, 오른쪽 가장자리의 찰리 얼굴)를 매우 정확하게 구현했으며, 어깨에 막 내려앉는 발의 접촉면이 선명합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "구도와 피사체 배치는 양호하나, 찰리의 마스크 표면에 레퍼런스에 없는 불필요한 선이 들어갔고 로봇 관절의 구조적 자연스러움이 A에 비해 다소 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "찰리의 시선은 화면 아래(물속)를 향하고 있으며, 아기 새는 어깨의 착지점을 향해 몸을 기울이고 있음.",
        "built_space": "얕은 바닷물과 바위 바닥이 배경으로 깔려 있으며, 카메라는 찰리의 어깨 위에서 비스듬히 내려다보는 정확한 앵글을 취함.",
        "entities": "찰리(흰색 마스크, 모래색 장갑, 가슴의 푸른 원자로)와 아기 새 모두 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "아기 새의 발이 어깨 표면에 정확히 닿아 체중을 지탱하고 있으며, 펼친 날개가 막 착지한 동작을 물리적으로 뒷받침함."
       },
       {
        "label": "B",
        "direction": "찰리의 시선은 화면 아래를 향하고, 아기 새 역시 어깨 쪽을 바라봄.",
        "built_space": "바닷물이 있는 야외 배경이며, 카메라는 요구된 어깨 근접 앵글을 유지함.",
        "entities": "찰리와 아기 새가 등장하나, 찰리의 흰색 마스크 이마에 레퍼런스에 없는 패널 선이 추가됨.",
        "hard_violations": [],
        "physics": "아기 새의 발이 어깨에 닿아 몸을 지탱하고 있으며, 날개는 펼쳐진 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "새의 착지와 발 접촉은 명확하지만, 찰리의 얼굴이 카메라 쪽으로 더 열려 있어 화면 아래 물고기를 향한 고개 숙임이 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "좌상단 새와 중앙 부근의 발 접촉, 오른쪽 가장자리에서 아래로 숙인 찰리의 얼굴을 근접 구도로 구현해 지정된 착지 순간과 시선 유지에 더 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "새의 부리와 머리는 오른쪽 아래의 어깨 착지면을 향하고, 몸 아래로 내린 두 발이 그 면에 닿는다. 찰리는 고개가 비스듬하지만 흰 얼굴 전면이 카메라 쪽으로 많이 열려 있다. 새를 돌아보지는 않으나, 화면 아래 물고기를 계속 내려다보는 방향은 약하게 읽힌다. 물고기 자체는 이 구도에서 확인되지 않는다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않는다. 화면 왼쪽과 어깨 너머에 맑은 바닷물과 수중 돌이 보이며 이전 장면의 얕은 바다와 부합한다. 전경에는 착지 대상인 어깨 장갑 하나와 연결 관절, 오른쪽에는 잘린 머리와 상체가 있다. 어깨보다 조금 높은 전측면 근접 시점이며, 얼굴은 오른쪽 가장자리에 있지만 마스크 대부분이 드러난다.",
        "entities": "아기 새 한 마리와 찰리 한 대만 보인다. 새는 갈색 솜털, 둥근 몸, 짧은 부리, 검은 눈과 가는 발가락으로 참조의 어린 새와 일치한다. 찰리는 마모된 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개와 선형 입을 갖춰 참조 정체성을 유지한다. 하단에는 푸른 원자로 일부가 보인다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "새의 두 발과 발가락이 어깨 장갑 표면에 접촉하고 있어 몸을 지지한다. 펼쳐 올린 양 날개와 앞으로 기운 몸은 내려앉으며 감속하는 동작으로 가능하다. 어깨 장갑은 관절을 통해 찰리의 몸통에 연결된다. 찰리의 하체와 해저 접촉은 근접 프레임 밖이므로 직접 확인할 수 없지만, 보이는 부분에 지지 없는 부유나 불가능한 관절은 없다."
       },
       {
        "label": "B",
        "direction": "새는 오른쪽 아래의 어깨 착지면을 향해 머리와 몸을 기울이고 두 발을 내딛는다. 찰리는 새 쪽으로 고개를 돌리지 않고 얼굴을 화면 아래쪽으로 더 숙이고 있어, 아래 프레임 밖 물고기에 주의를 유지한다는 지시에 더 가깝다. 실제 물고기는 잘려 있어 정확한 시선 종착점까지 확인되지는 않는다.",
        "built_space": "건축물과 고정 설비는 없고, 어깨 뒤와 왼쪽을 맑은 얕은 바닷물 및 수중 돌이 채운다. 착지하는 어깨 장갑 하나가 중앙 왼쪽, 연결 관절과 몸통이 중앙 오른쪽, 부분적으로 잘린 얼굴이 오른쪽 가장자리에 있다. 어깨 위에서 비스듬히 내려다보는 전측면 근접 구도가 유지되며, 새는 좌상단에 놓인다. 섬 바위와 식탁은 프레임 밖이라 확인 대상이 아니다.",
        "entities": "참조와 같은 갈색 어린 새 한 마리와 기계 찰리 한 대가 보인다. 새의 둥근 솜털 몸, 짧은 부리, 작은 검은 눈, 가는 다리와 발가락이 일치한다. 찰리의 베이지색 마모 장갑, 흰 마스크, 주황색 눈 두 개, 검은 입 선 및 하단의 푸른 원자로가 참조와 부합한다. 추가 인물이나 새, 삽입 문자는 없다.",
        "hard_violations": [],
        "physics": "새의 양발이 곡면 어깨 장갑에 닿고 발가락이 표면을 짚어 체중을 받는다. 위로 펼친 날개와 약간 굽힌 다리는 비행 직후 착지 충격을 받아들이는 자세로 자연스럽다. 어깨와 몸통은 기계 관절로 연결되어 있으며 머리도 목 구조에 지지된다. 물속 하체는 화면 밖이지만 보이는 몸이나 물체 중 근거 없이 떠 있는 것은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "새의 착지와 발 접촉은 명확하지만, 찰리의 얼굴이 카메라 쪽으로 더 열려 있어 화면 아래 물고기를 향한 고개 숙임이 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "좌상단 새와 중앙 부근의 발 접촉, 오른쪽 가장자리에서 아래로 숙인 찰리의 얼굴을 근접 구도로 구현해 지정된 착지 순간과 시선 유지에 더 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "새의 부리와 머리는 오른쪽 아래의 어깨 착지면을 향하고, 몸 아래로 내린 두 발이 그 면에 닿는다. 찰리는 고개가 비스듬하지만 흰 얼굴 전면이 카메라 쪽으로 많이 열려 있다. 새를 돌아보지는 않으나, 화면 아래 물고기를 계속 내려다보는 방향은 약하게 읽힌다. 물고기 자체는 이 구도에서 확인되지 않는다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않는다. 화면 왼쪽과 어깨 너머에 맑은 바닷물과 수중 돌이 보이며 이전 장면의 얕은 바다와 부합한다. 전경에는 착지 대상인 어깨 장갑 하나와 연결 관절, 오른쪽에는 잘린 머리와 상체가 있다. 어깨보다 조금 높은 전측면 근접 시점이며, 얼굴은 오른쪽 가장자리에 있지만 마스크 대부분이 드러난다.",
        "entities": "아기 새 한 마리와 찰리 한 대만 보인다. 새는 갈색 솜털, 둥근 몸, 짧은 부리, 검은 눈과 가는 발가락으로 참조의 어린 새와 일치한다. 찰리는 마모된 샌드 베이지 장갑, 흰 마스크, 주황색 원형 눈 두 개와 선형 입을 갖춰 참조 정체성을 유지한다. 하단에는 푸른 원자로 일부가 보인다. 추가 인물이나 문자 표시는 없다.",
        "hard_violations": [],
        "physics": "새의 두 발과 발가락이 어깨 장갑 표면에 접촉하고 있어 몸을 지지한다. 펼쳐 올린 양 날개와 앞으로 기운 몸은 내려앉으며 감속하는 동작으로 가능하다. 어깨 장갑은 관절을 통해 찰리의 몸통에 연결된다. 찰리의 하체와 해저 접촉은 근접 프레임 밖이므로 직접 확인할 수 없지만, 보이는 부분에 지지 없는 부유나 불가능한 관절은 없다."
       },
       {
        "label": "A",
        "direction": "새는 오른쪽 아래의 어깨 착지면을 향해 머리와 몸을 기울이고 두 발을 내딛는다. 찰리는 새 쪽으로 고개를 돌리지 않고 얼굴을 화면 아래쪽으로 더 숙이고 있어, 아래 프레임 밖 물고기에 주의를 유지한다는 지시에 더 가깝다. 실제 물고기는 잘려 있어 정확한 시선 종착점까지 확인되지는 않는다.",
        "built_space": "건축물과 고정 설비는 없고, 어깨 뒤와 왼쪽을 맑은 얕은 바닷물 및 수중 돌이 채운다. 착지하는 어깨 장갑 하나가 중앙 왼쪽, 연결 관절과 몸통이 중앙 오른쪽, 부분적으로 잘린 얼굴이 오른쪽 가장자리에 있다. 어깨 위에서 비스듬히 내려다보는 전측면 근접 구도가 유지되며, 새는 좌상단에 놓인다. 섬 바위와 식탁은 프레임 밖이라 확인 대상이 아니다.",
        "entities": "참조와 같은 갈색 어린 새 한 마리와 기계 찰리 한 대가 보인다. 새의 둥근 솜털 몸, 짧은 부리, 작은 검은 눈, 가는 다리와 발가락이 일치한다. 찰리의 베이지색 마모 장갑, 흰 마스크, 주황색 눈 두 개, 검은 입 선 및 하단의 푸른 원자로가 참조와 부합한다. 추가 인물이나 새, 삽입 문자는 없다.",
        "hard_violations": [],
        "physics": "새의 양발이 곡면 어깨 장갑에 닿고 발가락이 표면을 짚어 체중을 받는다. 위로 펼친 날개와 약간 굽힌 다리는 비행 직후 착지 충격을 받아들이는 자세로 자연스럽다. 어깨와 몸통은 기계 관절로 연결되어 있으며 머리도 목 구조에 지지된다. 물속 하체는 화면 밖이지만 보이는 몸이나 물체 중 근거 없이 떠 있는 것은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.603
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.603
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1603
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 클로즈업 구도와 피사체의 배치(왼쪽 상단의 아기 새, 오른쪽 가장자리의 찰리 얼굴)를 매우 정확하게 구현했으며, 어깨에 막 내려앉는 발의 접촉면이 선명합니다."
   },
   {
    "label": "B",
    "score": 1603,
    "verdict_ko": "구도와 피사체 배치는 양호하나, 찰리의 마스크 표면에 레퍼런스에 없는 불필요한 선이 들어갔고 로봇 관절의 구조적 자연스러움이 A에 비해 다소 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S91sh4_sel.png",
    "asset_id": "b9bcc068-9f93-417e-8e82-c825ef7403f0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 아기 새: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761765>",
    "asset_id": "4f173cfe-d96c-4d9d-9779-afb9d9a8dd4b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0efb-4233-7c92-b378-e69917159f28",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S91sh4"
  }
 },
 "S91sh13::signage": {
  "fp": "a178d77dad65c6c2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S91sh13": {
  "input_fingerprint": "b99a266041118637",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링이 강렬한 진동과 함께 맹렬한 속도로 빙글빙글 회전하며 잔상을 띠는 근접 찰나.\n\nLOCATION (lock): In the same clear shallows beside the island rock outcrop, with the robot's chest above the water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly inward at 찰리's chest height, retaining the same side of his frontal axis and an oblique view across the ring rather than squaring the lens to his torso. Keep the spinning ring near center at less than a third of the frame, surrounded by chest structure, the edges of his forward-reaching arms, and a narrow view of the water; his head and the perched bird remain above the crop as his attention stays on fishing below. Observe the mechanism directly with rotational motion trails confined to the ring and sharply rendered surrounding structure, emphasizing the final decrease in camera distance without changing the lighting or adding an energy discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater (Visible in narrow areas beyond the torso and arms); used as Maintains the fishing context and prevents the ring from becoming an isolated, oversized mechanical object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained daylight keeps the chest structure precise and the rotating ring legible, with no added emission or anticipatory beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's chest ring is beginning to spin faster as he concentrates on catching fish; the baby bird remains perched on his shoulder. The clear daylight sea and the food table remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링이 강렬한 진동과 함께 맹렬한 속도로 빙글빙글 회전하며 잔상을 띠는 근접 찰나.\n\nLOCATION (lock): In the same clear shallows beside the island rock outcrop, with the robot's chest above the water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly inward at 찰리's chest height, retaining the same side of his frontal axis and an oblique view across the ring rather than squaring the lens to his torso. Keep the spinning ring near center at less than a third of the frame, surrounded by chest structure, the edges of his forward-reaching arms, and a narrow view of the water; his head and the perched bird remain above the crop as his attention stays on fishing below. Observe the mechanism directly with rotational motion trails confined to the ring and sharply rendered surrounding structure, emphasizing the final decrease in camera distance without changing the lighting or adding an energy discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater (Visible in narrow areas beyond the torso and arms); used as Maintains the fishing context and prevents the ring from becoming an isolated, oversized mechanical object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained daylight keeps the chest structure precise and the rotating ring legible, with no added emission or anticipatory beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's chest ring is beginning to spin faster as he concentrates on catching fish; the baby bird remains perched on his shoulder. The clear daylight sea and the food table remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — primarily South Korea, 2069; Korean-speaking characters unless otherwise specified, with multiethnic, multinational refugee communities. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 찰리의 가슴 링이 강렬한 진동과 함께 맹렬한 속도로 빙글빙글 회전하며 잔상을 띠는 근접 찰나.\n\nLOCATION (lock): In the same clear shallows beside the island rock outcrop, with the robot's chest above the water surface. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly inward at 찰리's chest height, retaining the same side of his frontal axis and an oblique view across the ring rather than squaring the lens to his torso. Keep the spinning ring near center at less than a third of the frame, surrounded by chest structure, the edges of his forward-reaching arms, and a narrow view of the water; his head and the perched bird remain above the crop as his attention stays on fishing below. Observe the mechanism directly with rotational motion trails confined to the ring and sharply rendered surrounding structure, emphasizing the final decrease in camera distance without changing the lighting or adding an energy discharge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seawater (Visible in narrow areas beyond the torso and arms); used as Maintains the fishing context and prevents the ring from becoming an isolated, oversized mechanical object.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged restrained daylight keeps the chest structure precise and the rotating ring legible, with no added emission or anticipatory beam.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nNOT HUMAN — 찰리: each of these is not a human being but whatever its own reference sheet shows (a machine body, a constructed body, or a body that only looks human). Read each one's face, eyes and expression from that sheet exactly as it describes them — its anatomy and its materials are whatever the sheet gives it, and nothing here overrides that. Do not add human eyes, skin, teeth or facial muscles that the sheet does not give it. Any human-form or acted-performance rule in this brief governs only the people other than these.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Charlie's chest ring is beginning to spin faster as he concentrates on catching fish; the baby bird remains perched on his shoulder. The clear daylight sea and the food table remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 찰리 (고릴라형 기계 몸체, 긴 팔과 짧은 다리, 각진 샌드 베이지 장갑판, 흰 마스크형 얼굴, 점과 선으로 표현되는 얼굴, 키 작고 뚱뚱한 성인 남성과(와) 견주어 외형을 가리면 키 작고 뚱뚱한 성인 남성처럼 보이는 크기다.) — wearing: 거대한 고릴라 신체 비율(육중한 팔과 짧은 다리)의 샌드 베이지 장갑판 파츠 마감, 눈 두 개와 입 선 하나가 표출되는 귀여운 화이트 마스크 얼굴, 가슴 중앙에 푸른 에너지가 회전하는 거대한 원자로 심볼이 박힌 기계 외형. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nWriting that the shot text and the attached references already fix may be readable — a sign they show, a plate or label, a logo stamped on a machine, markings on a vehicle. Render each one in the script that object itself carries: equipment and its maker's mark in the lettering that maker uses, and the signage of a street in the script of that street. Never substitute one script for another, and never translate lettering the references fix.\n\nInvent no other wording: do not derive writing from what a place like this would have, and never print the shot description itself into the image. Nothing is laid on top of the photograph: no caption, subtitle, speech balloon, watermark, sticker or overlay.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Whatever this brief and its reference images already fix — the framing, the layout, where each thing stands — stays exactly as fixed; make only the choices they leave open, and make those the more cinematic ones.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "로봇의 가슴이 프레임 우측을 향하고, 카메라는 로봇의 우측에서 비스듬하게 가슴 링을 바라봄.",
    "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
    "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
    "hard_violations": [],
    "physics": "가슴 링이 제자리에서 맹렬하게 회전하며 방출 효과 없이 링 내부에만 잔상을 남김."
   },
   {
    "label": "B",
    "direction": "로봇의 가슴이 카메라를 향해 거의 정면으로 배치됨.",
    "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
    "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
    "hard_violations": [],
    "physics": "가슴 링이 제자리에서 회전하며 잔상을 남기는 움직임이 안정적으로 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 대로 로봇의 우측에서 비스듬한 카메라 각도를 유지하며 가슴 링의 회전 잔상과 프레이밍을 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라가 몸통과 정면으로 직각을 이루지 않도록 하라는 명시적 지시를 어기고 거의 정면 구도를 취했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 가슴이 프레임 우측을 향하고, 카메라는 로봇의 우측에서 비스듬하게 가슴 링을 바라봄.",
        "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
        "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
        "hard_violations": [],
        "physics": "가슴 링이 제자리에서 맹렬하게 회전하며 방출 효과 없이 링 내부에만 잔상을 남김."
       },
       {
        "label": "B",
        "direction": "로봇의 가슴이 카메라를 향해 거의 정면으로 배치됨.",
        "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
        "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
        "hard_violations": [],
        "physics": "가슴 링이 제자리에서 회전하며 잔상을 남기는 움직임이 안정적으로 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 대로 로봇의 우측에서 비스듬한 카메라 각도를 유지하며 가슴 링의 회전 잔상과 프레이밍을 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "카메라가 몸통과 정면으로 직각을 이루지 않도록 하라는 명시적 지시를 어기고 거의 정면 구도를 취했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "로봇의 가슴이 프레임 우측을 향하고, 카메라는 로봇의 우측에서 비스듬하게 가슴 링을 바라봄.",
        "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
        "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
        "hard_violations": [],
        "physics": "가슴 링이 제자리에서 맹렬하게 회전하며 방출 효과 없이 링 내부에만 잔상을 남김."
       },
       {
        "label": "B",
        "direction": "로봇의 가슴이 카메라를 향해 거의 정면으로 배치됨.",
        "built_space": "야외의 얕은 바닷물이 배경에 위치하며, 자연스러운 빛 반사가 보임.",
        "entities": "찰리의 샌드 베이지색 가슴 장갑과 회전하는 푸른 에너지 링이 중앙에 있으며, 지시대로 머리와 새는 프레임 위로 잘려 보이지 않음.",
        "hard_violations": [],
        "physics": "가슴 링이 제자리에서 회전하며 잔상을 남기는 움직임이 안정적으로 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "링의 중앙 배치와 국소적인 회전 잔상은 좋지만, 이전 숏과 반대편 사선에서 가슴을 보여 카메라 축 유지 지시를 어기며 얼굴 하단도 프레임에 들어온다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 숏과 같은 쪽의 사선 시점과 젖은 장갑판의 연속성이 우세하지만, 링이 오른쪽으로 치우치고 얼굴 일부가 보여 지정된 중앙 배치와 머리 제외 크롭은 미달이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "링은 카메라에 비스듬히 노출되어 있으며 화면 오른쪽 흉갑 면이 더 넓게 보인다. 왼쪽 어깨가 크게 보였던 이전 숏과 비교하면 가슴 정면축의 반대편으로 이동한 시점이다. 양팔은 화면 아래쪽으로 이어지지만 손과 물고기는 보이지 않아 낚시 대상에 대한 정확한 손 방향은 확인할 수 없다. 눈은 잘려 시선도 확인되지 않는다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않는다. 가슴 링 한 개와 양쪽 어깨·팔 일부가 있으며, 몸통 바깥과 팔 사이로 맑은 얕은 바다가 보인다. 가슴은 수면 위에 있고 수중 돌과 낮의 반짝임은 이전 장소와 부합한다. 링은 화면 중앙에 있지만 머리를 완전히 제외하라는 지시와 달리 입과 턱이 상단에 남아 있다.",
        "entities": "등장하는 몸체는 찰리 한 대이며, 샌드 베이지 장갑판, 검은 기계 관절, 마모와 물방울, 흰 마스크 하단이 참조와 부합한다. 가슴에는 푸른 원형 장치가 한 개 있다. 다만 소품 참조의 열린 금속 링보다는 이전 숏의 발광 원자로 원판에 가깝고, 금속 체결부 세부도 다르다. 새와 음식상은 크롭 밖이므로 유지 여부를 판단할 수 없다. 다른 사람이나 문자는 없다.",
        "hard_violations": [],
        "physics": "링은 흉갑의 원형 하우징에 장착되어 있으며 팔도 어깨와 팔꿈치 관절에 연결되어 있다. 지지 없이 떠 있는 물체는 없다. 회전 잔상은 원형 장치 내부에 집중되고 주변 흉갑과 물방울은 선명해 회전 운동 자체는 설득력 있다. 다만 금속 링의 강한 진동보다 밝은 푸른 원판의 회전이 더 두드러진다. 하체 지지는 크롭 밖이다."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽 어깨가 크게 전경을 차지하고 가슴을 같은 쪽에서 비스듬히 보는 방향이어서 이전 숏의 카메라 측면과 이어진다. 머리는 아래로 숙여져 있으나 눈이 대부분 잘려 물속의 특정 대상을 응시하는지는 확인되지 않는다. 팔은 아래쪽 전경으로 이어지며 손과 물고기는 프레임 밖이다.",
        "built_space": "건축물이나 고정 설비는 없으며 가슴 링 한 개, 양쪽 어깨와 팔 일부가 보인다. 몸통과 팔 주변의 좁은 틈에는 맑은 바닷물과 수중 돌이 나타나 이전 장소의 재질과 낮 조명을 유지한다. 가슴 높이의 근접 사선 구도이지만 링 중심이 화면 오른쪽으로 치우쳤고, 상단에 입과 눈 일부까지 들어와 머리 제외 지시를 지키지 못했다.",
        "entities": "찰리의 육중한 기계 몸체, 샌드 베이지 장갑판, 흰 마스크, 주황색 눈 일부, 젖고 벗겨진 표면이 참조와 일치한다. 가슴 장치는 하나이며 금속 테두리와 체결부가 보이지만, 소품 참조의 열린 링과 달리 내부가 푸른 회전 원판으로 채워져 있다. 새와 음식상은 보이지 않는 크롭 영역이므로 누락으로 단정하지 않는다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "회전 장치는 가슴의 금속 하우징에 고정되어 있고 어깨와 팔은 관절로 연결되어 있어 지지 관계가 자연스럽다. 근거 없이 부유하는 부품은 없다. 장치 내부의 원주 방향 잔상과 선명한 주변 장갑판은 빠른 회전 표현에 부합한다. 외부로 뻗는 에너지 방출은 없지만 밝은 중심부 때문에 기계 링의 진동보다 발광 원판의 회전처럼 읽힌다. 발과 지면 접촉은 크롭 밖이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "링의 중앙 배치와 국소적인 회전 잔상은 좋지만, 이전 숏과 반대편 사선에서 가슴을 보여 카메라 축 유지 지시를 어기며 얼굴 하단도 프레임에 들어온다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "이전 숏과 같은 쪽의 사선 시점과 젖은 장갑판의 연속성이 우세하지만, 링이 오른쪽으로 치우치고 얼굴 일부가 보여 지정된 중앙 배치와 머리 제외 크롭은 미달이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "링은 카메라에 비스듬히 노출되어 있으며 화면 오른쪽 흉갑 면이 더 넓게 보인다. 왼쪽 어깨가 크게 보였던 이전 숏과 비교하면 가슴 정면축의 반대편으로 이동한 시점이다. 양팔은 화면 아래쪽으로 이어지지만 손과 물고기는 보이지 않아 낚시 대상에 대한 정확한 손 방향은 확인할 수 없다. 눈은 잘려 시선도 확인되지 않는다.",
        "built_space": "인공 건축물이나 고정 설비는 보이지 않는다. 가슴 링 한 개와 양쪽 어깨·팔 일부가 있으며, 몸통 바깥과 팔 사이로 맑은 얕은 바다가 보인다. 가슴은 수면 위에 있고 수중 돌과 낮의 반짝임은 이전 장소와 부합한다. 링은 화면 중앙에 있지만 머리를 완전히 제외하라는 지시와 달리 입과 턱이 상단에 남아 있다.",
        "entities": "등장하는 몸체는 찰리 한 대이며, 샌드 베이지 장갑판, 검은 기계 관절, 마모와 물방울, 흰 마스크 하단이 참조와 부합한다. 가슴에는 푸른 원형 장치가 한 개 있다. 다만 소품 참조의 열린 금속 링보다는 이전 숏의 발광 원자로 원판에 가깝고, 금속 체결부 세부도 다르다. 새와 음식상은 크롭 밖이므로 유지 여부를 판단할 수 없다. 다른 사람이나 문자는 없다.",
        "hard_violations": [],
        "physics": "링은 흉갑의 원형 하우징에 장착되어 있으며 팔도 어깨와 팔꿈치 관절에 연결되어 있다. 지지 없이 떠 있는 물체는 없다. 회전 잔상은 원형 장치 내부에 집중되고 주변 흉갑과 물방울은 선명해 회전 운동 자체는 설득력 있다. 다만 금속 링의 강한 진동보다 밝은 푸른 원판의 회전이 더 두드러진다. 하체 지지는 크롭 밖이다."
       },
       {
        "label": "A",
        "direction": "화면 왼쪽 어깨가 크게 전경을 차지하고 가슴을 같은 쪽에서 비스듬히 보는 방향이어서 이전 숏의 카메라 측면과 이어진다. 머리는 아래로 숙여져 있으나 눈이 대부분 잘려 물속의 특정 대상을 응시하는지는 확인되지 않는다. 팔은 아래쪽 전경으로 이어지며 손과 물고기는 프레임 밖이다.",
        "built_space": "건축물이나 고정 설비는 없으며 가슴 링 한 개, 양쪽 어깨와 팔 일부가 보인다. 몸통과 팔 주변의 좁은 틈에는 맑은 바닷물과 수중 돌이 나타나 이전 장소의 재질과 낮 조명을 유지한다. 가슴 높이의 근접 사선 구도이지만 링 중심이 화면 오른쪽으로 치우쳤고, 상단에 입과 눈 일부까지 들어와 머리 제외 지시를 지키지 못했다.",
        "entities": "찰리의 육중한 기계 몸체, 샌드 베이지 장갑판, 흰 마스크, 주황색 눈 일부, 젖고 벗겨진 표면이 참조와 일치한다. 가슴 장치는 하나이며 금속 테두리와 체결부가 보이지만, 소품 참조의 열린 링과 달리 내부가 푸른 회전 원판으로 채워져 있다. 새와 음식상은 보이지 않는 크롭 영역이므로 누락으로 단정하지 않는다. 추가 인물이나 문자는 없다.",
        "hard_violations": [],
        "physics": "회전 장치는 가슴의 금속 하우징에 고정되어 있고 어깨와 팔은 관절로 연결되어 있어 지지 관계가 자연스럽다. 근거 없이 부유하는 부품은 없다. 장치 내부의 원주 방향 잔상과 선명한 주변 장갑판은 빠른 회전 표현에 부합한다. 외부로 뻗는 에너지 방출은 없지만 밝은 중심부 때문에 기계 링의 진동보다 발광 원판의 회전처럼 읽힌다. 발과 지면 접촉은 크롭 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.429
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.429
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "gate": {
   "outcome": "not_applicable",
   "policy": "winner_violation_gate_v1_admissible_first",
   "applicable": false
  },
  "totals": {
   "A": 2000,
   "B": 1429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 대로 로봇의 우측에서 비스듬한 카메라 각도를 유지하며 가슴 링의 회전 잔상과 프레이밍을 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "카메라가 몸통과 정면으로 직각을 이루지 않도록 하라는 명시적 지시를 어기고 거의 정면 구도를 취했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 찰리 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/91b9626e-83ee-4a85-9208-0c8e2d153a63/images/e565764e-23fa-4991-a743-975fd8d061df/scene/recipe/S91sh7_sel.png",
    "asset_id": "576d1723-dd3b-4689-bb69-134c701ba121",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 찰리: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1075117>",
    "asset_id": "da31c121-c39d-46d3-9070-57d1757f7d39",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 찰리의 가슴 링 장치: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:871281>",
    "asset_id": "f8976a4b-5c57-41bb-b0f3-8c1869fc65c1",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": true,
  "shot_run_uid": "06ab0f00-16ab-7777-b708-a98704861263",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S91sh7"
  }
 }
}